From 9a7ff2834b92f33fd7d424410bf05dfd0f5db252 Mon Sep 17 00:00:00 2001 From: Ed_ Date: Mon, 22 Jun 2026 01:29:28 -0400 Subject: [PATCH] conductor(creikey_dl_cv): Phase 1 Acquire - transcript (2082 clean segments, 74KB) + 815MB mp4 --- .../artifacts/phase1.log | 11 + .../artifacts/transcript.json | 29268 ++++++++++++++++ .../artifacts/transcript_clean.txt | 2082 ++ .../artifacts/video.log | 16 + 4 files changed, 31377 insertions(+) create mode 100644 conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/phase1.log create mode 100644 conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript.json create mode 100644 conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript_clean.txt create mode 100644 conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/video.log diff --git a/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/phase1.log b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/phase1.log new file mode 100644 index 00000000..6a6826d7 --- /dev/null +++ b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/phase1.log @@ -0,0 +1,11 @@ +Phase 1 Acquire for creikey_dl_cv: https://youtu.be/yxkUvXs-hoQ +Artifacts: C:\projects\manual_slop\conductor\tracks\video_analysis_creikey_dl_cv_20260621\artifacts +Step 1: extract_transcript (yt-dlp VTT directly) + OK: wrote C:\projects\manual_slop\conductor\tracks\video_analysis_creikey_dl_cv_20260621\artifacts\transcript.json (4186 segments) +Step 2: download_video + OK: wrote C:\projects\manual_slop\conductor\tracks\video_analysis_creikey_dl_cv_20260621\artifacts\video.mp4 (855142416 bytes) +{ + "status": "ok", + "video_path": "C:\\projects\\manual_slop\\conductor\\tracks\\video_analysis_creikey_dl_cv_20260621\\artifacts\\video.mp4", + "transcript_path": "C:\\projects\\manual_slop\\conductor\\tracks\\video_analysis_creikey_dl_cv_20260621\\artifacts\\transcript.json" +} diff --git a/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript.json b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript.json new file mode 100644 index 00000000..f68f23ce --- /dev/null +++ b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript.json @@ -0,0 +1,29268 @@ +{ + "video_id": "yxkUvXs-hoQ", + "segments": [ + { + "start": 0.52, + "duration": 0.0, + "text": "Okay,<00:00:00.800> so<00:00:00.960> this<00:00:01.200> is<00:00:01.320> very<00:00:01.600> difficult<00:00:02.200> because" + }, + { + "start": 2.63, + "duration": 0.0, + "text": "Okay, so this is very difficult because" + }, + { + "start": 2.64, + "duration": 0.0, + "text": "Okay, so this is very difficult because Cameron<00:00:02.960> is<00:00:03.080> such<00:00:03.280> a<00:00:03.520> is<00:00:03.600> such<00:00:03.800> an<00:00:03.920> like" + }, + { + "start": 4.15, + "duration": 0.0, + "text": "Cameron is such a is such an like" + }, + { + "start": 4.16, + "duration": 0.0, + "text": "Cameron is such a is such an like indescribable<00:00:05.280> guy." + }, + { + "start": 6.51, + "duration": 0.0, + "text": "indescribable guy." + }, + { + "start": 6.52, + "duration": 0.0, + "text": "indescribable guy. Um" + }, + { + "start": 7.91, + "duration": 0.0, + "text": "Um" + }, + { + "start": 7.92, + "duration": 0.0, + "text": "Um you<00:00:08.040> know," + }, + { + "start": 9.67, + "duration": 0.0, + "text": "you know," + }, + { + "start": 9.68, + "duration": 0.0, + "text": "you know, yeah,<00:00:09.880> this<00:00:10.080> is<00:00:10.360> very<00:00:10.720> very<00:00:11.040> very<00:00:11.280> difficult." + }, + { + "start": 11.83, + "duration": 0.0, + "text": "yeah, this is very very very difficult." + }, + { + "start": 11.84, + "duration": 0.0, + "text": "yeah, this is very very very difficult. I<00:00:11.960> know<00:00:12.200> Cameron<00:00:12.720> just<00:00:12.960> as<00:00:13.120> like<00:00:13.360> the<00:00:13.480> guy<00:00:13.880> who" + }, + { + "start": 15.15, + "duration": 0.0, + "text": "I know Cameron just as like the guy who" + }, + { + "start": 15.16, + "duration": 0.0, + "text": "I know Cameron just as like the guy who uh<00:00:15.640> you<00:00:15.760> know," + }, + { + "start": 17.83, + "duration": 0.0, + "text": "uh you know," + }, + { + "start": 17.84, + "duration": 0.0, + "text": "uh you know, he" + }, + { + "start": 18.87, + "duration": 0.0, + "text": "he" + }, + { + "start": 18.88, + "duration": 0.0, + "text": "he he's<00:00:19.080> like<00:00:19.200> the<00:00:19.320> guy<00:00:19.560> crazy<00:00:20.000> enough<00:00:20.360> to<00:00:20.520> just" + }, + { + "start": 20.75, + "duration": 0.0, + "text": "he's like the guy crazy enough to just" + }, + { + "start": 20.76, + "duration": 0.0, + "text": "he's like the guy crazy enough to just like<00:00:21.320> dig<00:00:21.600> through<00:00:21.920> the<00:00:22.040> mountain<00:00:22.640> and<00:00:22.800> do" + }, + { + "start": 23.15, + "duration": 0.0, + "text": "like dig through the mountain and do" + }, + { + "start": 23.16, + "duration": 0.0, + "text": "like dig through the mountain and do just<00:00:23.360> like<00:00:23.840> as<00:00:24.560> sudden" + }, + { + "start": 26.83, + "duration": 0.0, + "text": "just like as sudden" + }, + { + "start": 26.84, + "duration": 0.0, + "text": "just like as sudden sudden<00:00:28.040> huge<00:00:28.640> amounts<00:00:29.080> of<00:00:29.240> work" + }, + { + "start": 30.55, + "duration": 0.0, + "text": "sudden huge amounts of work" + }, + { + "start": 30.56, + "duration": 0.0, + "text": "sudden huge amounts of work uh<00:00:30.920> of<00:00:31.480> like,<00:00:31.640> \"Oh,<00:00:31.680> I<00:00:31.760> cloned<00:00:32.240> an<00:00:32.360> application" + }, + { + "start": 32.83, + "duration": 0.0, + "text": "uh of like, \"Oh, I cloned an application" + }, + { + "start": 32.84, + "duration": 0.0, + "text": "uh of like, \"Oh, I cloned an application or<00:00:32.920> I<00:00:32.960> did<00:00:33.160> this.\"<00:00:33.440> And<00:00:33.600> then<00:00:33.800> at<00:00:33.920> other<00:00:34.120> times" + }, + { + "start": 34.55, + "duration": 0.0, + "text": "or I did this.\" And then at other times" + }, + { + "start": 34.56, + "duration": 0.0, + "text": "or I did this.\" And then at other times he,<00:00:35.320> you<00:00:35.400> know,<00:00:35.480> goes<00:00:35.760> and<00:00:35.840> becomes<00:00:36.120> a<00:00:36.160> web" + }, + { + "start": 36.39, + "duration": 0.0, + "text": "he, you know, goes and becomes a web" + }, + { + "start": 36.4, + "duration": 0.0, + "text": "he, you know, goes and becomes a web developer<00:00:36.960> and<00:00:37.080> then<00:00:37.240> complains<00:00:37.640> about<00:00:38.000> like" + }, + { + "start": 38.71, + "duration": 0.0, + "text": "developer and then complains about like" + }, + { + "start": 38.72, + "duration": 0.0, + "text": "developer and then complains about like not<00:00:38.960> being<00:00:39.160> of<00:00:39.280> the<00:00:39.400> program<00:00:39.760> anymore." + }, + { + "start": 42.11, + "duration": 0.0, + "text": "not being of the program anymore." + }, + { + "start": 42.12, + "duration": 0.0, + "text": "not being of the program anymore. So,<00:00:42.560> anyway,<00:00:43.320> um" + }, + { + "start": 44.35, + "duration": 0.0, + "text": "So, anyway, um" + }, + { + "start": 44.36, + "duration": 0.0, + "text": "So, anyway, um can<00:00:44.440> we<00:00:44.520> get<00:00:44.640> a<00:00:44.680> round<00:00:44.880> of<00:00:45.000> applause<00:00:45.400> for<00:00:45.480> the" + }, + { + "start": 45.59, + "duration": 0.0, + "text": "can we get a round of applause for the" + }, + { + "start": 45.6, + "duration": 0.0, + "text": "can we get a round of applause for the completely<00:00:46.200> undescribable" + }, + { + "start": 47.75, + "duration": 0.0, + "text": "completely undescribable" + }, + { + "start": 47.76, + "duration": 0.0, + "text": "completely undescribable Cameron<00:00:48.120> Wrights?" + }, + { + "start": 55.08, + "duration": 0.0, + "text": "What's<00:00:55.280> up,<00:00:55.480> Craigsters?<00:00:56.360> LET'S<00:00:56.520> MAKE<00:00:56.640> SOME" + }, + { + "start": 56.75, + "duration": 0.0, + "text": "What's up, Craigsters? LET'S MAKE SOME" + }, + { + "start": 56.76, + "duration": 0.0, + "text": "What's up, Craigsters? LET'S MAKE SOME NOISE." + }, + { + "start": 62.04, + "duration": 0.0, + "text": "LET'S<00:01:02.280> SAY<00:01:02.400> YOU<00:01:02.520> WANTED<00:01:02.760> TO<00:01:02.800> WRITE<00:01:02.960> A<00:01:03.000> program" + }, + { + "start": 63.95, + "duration": 0.0, + "text": "LET'S SAY YOU WANTED TO WRITE A program" + }, + { + "start": 63.96, + "duration": 0.0, + "text": "LET'S SAY YOU WANTED TO WRITE A program that<00:01:04.080> could<00:01:04.239> tell<00:01:04.400> the<00:01:04.519> difference<00:01:04.960> between<00:01:05.320> a" + }, + { + "start": 65.35, + "duration": 0.0, + "text": "that could tell the difference between a" + }, + { + "start": 65.36, + "duration": 0.0, + "text": "that could tell the difference between a cat<00:01:05.920> and<00:01:06.040> a<00:01:06.120> dog." + }, + { + "start": 67.39, + "duration": 0.0, + "text": "cat and a dog." + }, + { + "start": 67.4, + "duration": 0.0, + "text": "cat and a dog. I<00:01:07.480> want<00:01:07.640> you<00:01:07.720> to<00:01:07.840> actually<00:01:08.200> think<00:01:08.400> about<00:01:08.680> how" + }, + { + "start": 68.79, + "duration": 0.0, + "text": "I want you to actually think about how" + }, + { + "start": 68.8, + "duration": 0.0, + "text": "I want you to actually think about how you<00:01:08.880> would<00:01:09.000> do<00:01:09.160> this<00:01:09.360> for<00:01:09.480> a<00:01:09.520> second,<00:01:10.000> okay?" + }, + { + "start": 70.23, + "duration": 0.0, + "text": "you would do this for a second, okay?" + }, + { + "start": 70.24, + "duration": 0.0, + "text": "you would do this for a second, okay? Imagine<00:01:10.600> you<00:01:10.680> were<00:01:10.800> given<00:01:11.040> the<00:01:11.160> pixels<00:01:11.880> and" + }, + { + "start": 71.99, + "duration": 0.0, + "text": "Imagine you were given the pixels and" + }, + { + "start": 72.0, + "duration": 0.0, + "text": "Imagine you were given the pixels and you<00:01:12.080> had<00:01:12.200> to<00:01:12.280> make<00:01:12.440> the<00:01:12.520> program<00:01:13.480> that<00:01:13.640> said," + }, + { + "start": 73.87, + "duration": 0.0, + "text": "you had to make the program that said," + }, + { + "start": 73.88, + "duration": 0.0, + "text": "you had to make the program that said, you<00:01:13.960> know,<00:01:14.080> cat<00:01:14.360> or<00:01:14.440> dog." + }, + { + "start": 75.55, + "duration": 0.0, + "text": "you know, cat or dog." + }, + { + "start": 75.56, + "duration": 0.0, + "text": "you know, cat or dog. You<00:01:15.640> might<00:01:15.800> try<00:01:15.960> something<00:01:16.320> like,<00:01:16.480> you<00:01:16.560> know," + }, + { + "start": 76.99, + "duration": 0.0, + "text": "You might try something like, you know," + }, + { + "start": 77.0, + "duration": 0.0, + "text": "You might try something like, you know, look<00:01:17.160> for<00:01:17.320> ear<00:01:17.520> shapes<00:01:17.800> with<00:01:18.000> edge<00:01:18.160> detection," + }, + { + "start": 78.83, + "duration": 0.0, + "text": "look for ear shapes with edge detection," + }, + { + "start": 78.84, + "duration": 0.0, + "text": "look for ear shapes with edge detection, maybe<00:01:19.120> like<00:01:19.320> detect<00:01:19.680> like<00:01:20.320> the<00:01:20.400> little<00:01:20.640> angle" + }, + { + "start": 80.91, + "duration": 0.0, + "text": "maybe like detect like the little angle" + }, + { + "start": 80.92, + "duration": 0.0, + "text": "maybe like detect like the little angle there,<00:01:21.120> you<00:01:21.200> know,<00:01:21.240> if<00:01:21.320> it's<00:01:21.440> a<00:01:21.480> pointy<00:01:21.840> ear," + }, + { + "start": 81.95, + "duration": 0.0, + "text": "there, you know, if it's a pointy ear," + }, + { + "start": 81.96, + "duration": 0.0, + "text": "there, you know, if it's a pointy ear, it's<00:01:22.080> a<00:01:22.120> cat,<00:01:22.440> if<00:01:22.480> it's<00:01:22.600> a<00:01:22.680> floppy<00:01:23.040> ear,<00:01:23.200> it's<00:01:23.320> a" + }, + { + "start": 83.35, + "duration": 0.0, + "text": "it's a cat, if it's a floppy ear, it's a" + }, + { + "start": 83.36, + "duration": 0.0, + "text": "it's a cat, if it's a floppy ear, it's a dog." + }, + { + "start": 84.87, + "duration": 0.0, + "text": "dog." + }, + { + "start": 84.88, + "duration": 0.0, + "text": "dog. Problem<00:01:25.240> solved,<00:01:25.640> right?<00:01:25.800> You<00:01:25.880> don't<00:01:26.000> need" + }, + { + "start": 86.15, + "duration": 0.0, + "text": "Problem solved, right? You don't need" + }, + { + "start": 86.16, + "duration": 0.0, + "text": "Problem solved, right? You don't need any<00:01:26.280> more<00:01:26.440> tech,<00:01:26.680> you<00:01:26.720> don't<00:01:26.840> need<00:01:27.000> any<00:01:27.160> fancy" + }, + { + "start": 87.71, + "duration": 0.0, + "text": "any more tech, you don't need any fancy" + }, + { + "start": 87.72, + "duration": 0.0, + "text": "any more tech, you don't need any fancy whatever." + }, + { + "start": 88.79, + "duration": 0.0, + "text": "whatever." + }, + { + "start": 88.8, + "duration": 0.0, + "text": "whatever. Uh-oh." + }, + { + "start": 92.04, + "duration": 0.0, + "text": "You<00:01:32.120> know,<00:01:32.200> if<00:01:32.280> we<00:01:32.400> look<00:01:32.560> at<00:01:32.640> our<00:01:32.720> program" + }, + { + "start": 93.07, + "duration": 0.0, + "text": "You know, if we look at our program" + }, + { + "start": 93.08, + "duration": 0.0, + "text": "You know, if we look at our program here,<00:01:33.280> it's<00:01:33.520> just<00:01:33.720> a<00:01:34.440> system<00:01:34.840> with<00:01:34.960> an<00:01:35.080> input" + }, + { + "start": 95.51, + "duration": 0.0, + "text": "here, it's just a system with an input" + }, + { + "start": 95.52, + "duration": 0.0, + "text": "here, it's just a system with an input and<00:01:35.640> an<00:01:35.760> output,<00:01:36.120> right?<00:01:37.080> Um" + }, + { + "start": 98.03, + "duration": 0.0, + "text": "and an output, right? Um" + }, + { + "start": 98.04, + "duration": 0.0, + "text": "and an output, right? Um take<00:01:38.280> in<00:01:38.400> the<00:01:38.520> image" + }, + { + "start": 100.55, + "duration": 0.0, + "text": "take in the image" + }, + { + "start": 100.56, + "duration": 0.0, + "text": "take in the image and<00:01:40.680> then,<00:01:40.800> you<00:01:40.880> know,<00:01:41.000> look<00:01:41.240> for<00:01:41.760> whatever" + }, + { + "start": 102.03, + "duration": 0.0, + "text": "and then, you know, look for whatever" + }, + { + "start": 102.04, + "duration": 0.0, + "text": "and then, you know, look for whatever heuristic<00:01:42.480> we've<00:01:42.600> decided<00:01:43.000> we're<00:01:43.080> looking" + }, + { + "start": 103.31, + "duration": 0.0, + "text": "heuristic we've decided we're looking" + }, + { + "start": 103.32, + "duration": 0.0, + "text": "heuristic we've decided we're looking for,<00:01:43.960> and<00:01:44.080> then<00:01:44.200> we<00:01:44.600> we,<00:01:45.200> you<00:01:45.280> know,<00:01:45.400> make<00:01:45.560> our" + }, + { + "start": 105.67, + "duration": 0.0, + "text": "for, and then we we, you know, make our" + }, + { + "start": 105.68, + "duration": 0.0, + "text": "for, and then we we, you know, make our output,<00:01:46.000> right?<00:01:46.320> But,<00:01:46.520> um<00:01:46.960> if<00:01:47.120> we<00:01:47.200> keep<00:01:47.400> trying" + }, + { + "start": 107.55, + "duration": 0.0, + "text": "output, right? But, um if we keep trying" + }, + { + "start": 107.56, + "duration": 0.0, + "text": "output, right? But, um if we keep trying to<00:01:47.640> iterate<00:01:47.920> on<00:01:48.000> this,<00:01:48.120> it's<00:01:48.240> like,<00:01:48.360> okay," + }, + { + "start": 108.55, + "duration": 0.0, + "text": "to iterate on this, it's like, okay," + }, + { + "start": 108.56, + "duration": 0.0, + "text": "to iterate on this, it's like, okay, maybe<00:01:48.840> some<00:01:49.000> of<00:01:49.080> them<00:01:49.200> have<00:01:49.280> pointy<00:01:49.520> ears,<00:01:49.760> but" + }, + { + "start": 109.83, + "duration": 0.0, + "text": "maybe some of them have pointy ears, but" + }, + { + "start": 109.84, + "duration": 0.0, + "text": "maybe some of them have pointy ears, but only<00:01:50.000> if<00:01:50.120> the<00:01:50.200> fur<00:01:50.280> is<00:01:50.440> orange,<00:01:51.240> and<00:01:51.320> we<00:01:51.400> try<00:01:51.520> to" + }, + { + "start": 111.63, + "duration": 0.0, + "text": "only if the fur is orange, and we try to" + }, + { + "start": 111.64, + "duration": 0.0, + "text": "only if the fur is orange, and we try to keep,<00:01:51.880> you<00:01:51.920> know,<00:01:52.080> coming<00:01:52.320> up<00:01:52.400> with<00:01:52.520> patterns" + }, + { + "start": 112.91, + "duration": 0.0, + "text": "keep, you know, coming up with patterns" + }, + { + "start": 112.92, + "duration": 0.0, + "text": "keep, you know, coming up with patterns that<00:01:53.040> try<00:01:53.240> to<00:01:53.440> get<00:01:53.680> at<00:01:53.880> what<00:01:54.040> we<00:01:54.160> mean<00:01:54.480> by<00:01:54.640> a<00:01:54.720> dog" + }, + { + "start": 115.31, + "duration": 0.0, + "text": "that try to get at what we mean by a dog" + }, + { + "start": 115.32, + "duration": 0.0, + "text": "that try to get at what we mean by a dog or<00:01:55.440> cat," + }, + { + "start": 116.59, + "duration": 0.0, + "text": "or cat," + }, + { + "start": 116.6, + "duration": 0.0, + "text": "or cat, um<00:01:57.000> you'll<00:01:57.200> realize<00:01:57.640> that<00:01:57.800> it<00:01:57.960> quickly" + }, + { + "start": 118.27, + "duration": 0.0, + "text": "um you'll realize that it quickly" + }, + { + "start": 118.28, + "duration": 0.0, + "text": "um you'll realize that it quickly becomes<00:01:58.640> untenable,<00:01:59.360> right?" + }, + { + "start": 120.35, + "duration": 0.0, + "text": "becomes untenable, right?" + }, + { + "start": 120.36, + "duration": 0.0, + "text": "becomes untenable, right? So,<00:02:00.440> the<00:02:00.560> whole<00:02:00.840> idea<00:02:01.240> of<00:02:01.440> machine<00:02:02.080> learning" + }, + { + "start": 122.99, + "duration": 0.0, + "text": "So, the whole idea of machine learning" + }, + { + "start": 123.0, + "duration": 0.0, + "text": "So, the whole idea of machine learning is<00:02:03.200> you<00:02:03.280> essentially<00:02:03.840> have<00:02:04.000> a<00:02:04.080> program<00:02:04.720> that" + }, + { + "start": 124.91, + "duration": 0.0, + "text": "is you essentially have a program that" + }, + { + "start": 124.92, + "duration": 0.0, + "text": "is you essentially have a program that creates<00:02:05.360> that<00:02:05.520> program.<00:02:06.120> It's<00:02:06.280> basically" + }, + { + "start": 126.71, + "duration": 0.0, + "text": "creates that program. It's basically" + }, + { + "start": 126.72, + "duration": 0.0, + "text": "creates that program. It's basically meta<00:02:06.960> programming,<00:02:07.479> right?<00:02:07.800> That's<00:02:08.280> sort<00:02:08.479> of" + }, + { + "start": 128.55, + "duration": 0.0, + "text": "meta programming, right? That's sort of" + }, + { + "start": 128.56, + "duration": 0.0, + "text": "meta programming, right? That's sort of a<00:02:08.600> metaphor<00:02:08.960> for<00:02:09.080> what<00:02:09.200> it<00:02:09.320> is." + }, + { + "start": 130.23, + "duration": 0.0, + "text": "a metaphor for what it is." + }, + { + "start": 130.24, + "duration": 0.0, + "text": "a metaphor for what it is. Um,<00:02:10.720> you<00:02:10.840> set<00:02:11.000> up<00:02:11.120> some<00:02:11.320> architecture<00:02:11.880> and" + }, + { + "start": 131.99, + "duration": 0.0, + "text": "Um, you set up some architecture and" + }, + { + "start": 132.0, + "duration": 0.0, + "text": "Um, you set up some architecture and repeatedly<00:02:12.640> optimize<00:02:13.240> with<00:02:13.360> respect<00:02:13.760> to,<00:02:14.080> you" + }, + { + "start": 134.15, + "duration": 0.0, + "text": "repeatedly optimize with respect to, you" + }, + { + "start": 134.16, + "duration": 0.0, + "text": "repeatedly optimize with respect to, you know,<00:02:14.240> your<00:02:14.360> examples,<00:02:15.520> um,<00:02:16.000> a<00:02:16.320> machine" + }, + { + "start": 136.67, + "duration": 0.0, + "text": "know, your examples, um, a machine" + }, + { + "start": 136.68, + "duration": 0.0, + "text": "know, your examples, um, a machine learning<00:02:16.959> model,<00:02:17.240> a<00:02:17.320> black<00:02:17.640> box<00:02:17.959> of<00:02:18.320> many" + }, + { + "start": 138.51, + "duration": 0.0, + "text": "learning model, a black box of many" + }, + { + "start": 138.52, + "duration": 0.0, + "text": "learning model, a black box of many parameters.<00:02:19.200> It's<00:02:19.360> a<00:02:19.440> learned<00:02:19.959> program," + }, + { + "start": 140.63, + "duration": 0.0, + "text": "parameters. It's a learned program," + }, + { + "start": 140.64, + "duration": 0.0, + "text": "parameters. It's a learned program, right?<00:02:21.400> It's<00:02:21.760> automatic<00:02:22.400> programming." + }, + { + "start": 144.19, + "duration": 0.0, + "text": "right? It's automatic programming." + }, + { + "start": 144.2, + "duration": 0.0, + "text": "right? It's automatic programming. As<00:02:24.320> a<00:02:24.360> quick<00:02:24.680> aside,<00:02:25.040> why<00:02:25.160> is<00:02:25.240> it<00:02:25.320> all<00:02:25.400> Python?" + }, + { + "start": 145.87, + "duration": 0.0, + "text": "As a quick aside, why is it all Python?" + }, + { + "start": 145.88, + "duration": 0.0, + "text": "As a quick aside, why is it all Python? It's<00:02:26.040> because<00:02:26.560> it's<00:02:26.720> basically<00:02:27.160> meta" + }, + { + "start": 147.35, + "duration": 0.0, + "text": "It's because it's basically meta" + }, + { + "start": 147.36, + "duration": 0.0, + "text": "It's because it's basically meta programming,<00:02:27.840> right?<00:02:28.040> It's<00:02:28.160> a<00:02:28.200> complicated" + }, + { + "start": 148.71, + "duration": 0.0, + "text": "programming, right? It's a complicated" + }, + { + "start": 148.72, + "duration": 0.0, + "text": "programming, right? It's a complicated build<00:02:28.959> system.<00:02:29.800> Uh,<00:02:29.920> these<00:02:30.120> training<00:02:30.520> scripts" + }, + { + "start": 150.87, + "duration": 0.0, + "text": "build system. Uh, these training scripts" + }, + { + "start": 150.88, + "duration": 0.0, + "text": "build system. Uh, these training scripts basically<00:02:31.360> are<00:02:31.480> just<00:02:31.720> there<00:02:32.080> to<00:02:32.200> eventually," + }, + { + "start": 152.75, + "duration": 0.0, + "text": "basically are just there to eventually," + }, + { + "start": 152.76, + "duration": 0.0, + "text": "basically are just there to eventually, you<00:02:32.840> know,<00:02:33.400> automatically<00:02:33.920> create<00:02:34.200> the" + }, + { + "start": 154.31, + "duration": 0.0, + "text": "you know, automatically create the" + }, + { + "start": 154.32, + "duration": 0.0, + "text": "you know, automatically create the program<00:02:34.800> that<00:02:34.959> does<00:02:35.160> what<00:02:35.280> you<00:02:35.400> want<00:02:35.560> it<00:02:35.640> to" + }, + { + "start": 155.71, + "duration": 0.0, + "text": "program that does what you want it to" + }, + { + "start": 155.72, + "duration": 0.0, + "text": "program that does what you want it to do.<00:02:36.520> Um,<00:02:37.440> and<00:02:37.560> most<00:02:37.760> of<00:02:37.840> the<00:02:38.200> heavy<00:02:38.400> lifting<00:02:38.720> is" + }, + { + "start": 158.83, + "duration": 0.0, + "text": "do. Um, and most of the heavy lifting is" + }, + { + "start": 158.84, + "duration": 0.0, + "text": "do. Um, and most of the heavy lifting is not<00:02:39.040> necessarily<00:02:39.560> done<00:02:39.959> by<00:02:40.160> the<00:02:40.240> actual<00:02:40.520> code" + }, + { + "start": 160.71, + "duration": 0.0, + "text": "not necessarily done by the actual code" + }, + { + "start": 160.72, + "duration": 0.0, + "text": "not necessarily done by the actual code you<00:02:40.800> write<00:02:40.959> anyways,<00:02:41.480> it's<00:02:41.680> done" + }, + { + "start": 162.59, + "duration": 0.0, + "text": "you write anyways, it's done" + }, + { + "start": 162.6, + "duration": 0.0, + "text": "you write anyways, it's done in<00:02:42.720> the<00:02:42.920> in<00:02:43.040> the<00:02:43.120> dot<00:02:43.360> products<00:02:43.760> on<00:02:43.959> the<00:02:44.040> GPU" + }, + { + "start": 164.59, + "duration": 0.0, + "text": "in the in the dot products on the GPU" + }, + { + "start": 164.6, + "duration": 0.0, + "text": "in the in the dot products on the GPU or,<00:02:44.959> you<00:02:45.040> know," + }, + { + "start": 166.39, + "duration": 0.0, + "text": "or, you know," + }, + { + "start": 166.4, + "duration": 0.0, + "text": "or, you know, piece<00:02:46.600> of<00:02:46.680> compute<00:02:47.000> that<00:02:47.120> you're<00:02:47.240> using<00:02:47.600> to" + }, + { + "start": 168.43, + "duration": 0.0, + "text": "piece of compute that you're using to" + }, + { + "start": 168.44, + "duration": 0.0, + "text": "piece of compute that you're using to actually<00:02:48.760> do<00:02:49.040> the<00:02:49.200> machine<00:02:49.480> learning.<00:02:50.360> Um," + }, + { + "start": 170.87, + "duration": 0.0, + "text": "actually do the machine learning. Um," + }, + { + "start": 170.88, + "duration": 0.0, + "text": "actually do the machine learning. Um, and<00:02:51.000> John<00:02:51.240> Carmack<00:02:51.560> for<00:02:51.680> his<00:02:51.880> current<00:02:52.400> AI" + }, + { + "start": 172.63, + "duration": 0.0, + "text": "and John Carmack for his current AI" + }, + { + "start": 172.64, + "duration": 0.0, + "text": "and John Carmack for his current AI efforts,<00:02:52.959> he's<00:02:53.080> in<00:02:53.160> the<00:02:53.240> race<00:02:53.520> to<00:02:53.640> win<00:02:53.760> the" + }, + { + "start": 173.83, + "duration": 0.0, + "text": "efforts, he's in the race to win the" + }, + { + "start": 173.84, + "duration": 0.0, + "text": "efforts, he's in the race to win the world." + }, + { + "start": 174.949, + "duration": 0.0, + "text": "world." + }, + { + "start": 174.959, + "duration": 0.0, + "text": "world. Um," + }, + { + "start": 176.19, + "duration": 0.0, + "text": "Um," + }, + { + "start": 176.2, + "duration": 0.0, + "text": "Um, he<00:02:56.360> has<00:02:56.760> switched<00:02:57.040> to<00:02:57.120> Python.<00:02:57.880> I<00:02:57.959> thought" + }, + { + "start": 178.07, + "duration": 0.0, + "text": "he has switched to Python. I thought" + }, + { + "start": 178.08, + "duration": 0.0, + "text": "he has switched to Python. I thought that<00:02:58.160> was<00:02:58.280> notable." + }, + { + "start": 179.47, + "duration": 0.0, + "text": "that was notable." + }, + { + "start": 179.48, + "duration": 0.0, + "text": "that was notable. So,<00:02:59.800> a<00:02:59.880> little<00:03:00.040> bit<00:03:00.160> about<00:03:00.440> me,<00:03:01.200> right?<00:03:01.480> I'm" + }, + { + "start": 181.67, + "duration": 0.0, + "text": "So, a little bit about me, right? I'm" + }, + { + "start": 181.68, + "duration": 0.0, + "text": "So, a little bit about me, right? I'm mostly<00:03:02.040> a<00:03:02.080> game<00:03:02.320> developer.<00:03:02.720> If<00:03:02.800> we<00:03:02.880> go<00:03:03.040> back" + }, + { + "start": 183.39, + "duration": 0.0, + "text": "mostly a game developer. If we go back" + }, + { + "start": 183.4, + "duration": 0.0, + "text": "mostly a game developer. If we go back to<00:03:03.480> 2017,<00:03:04.440> one<00:03:04.519> of<00:03:04.600> my<00:03:04.680> first<00:03:04.920> projects<00:03:05.280> was<00:03:05.400> a" + }, + { + "start": 185.47, + "duration": 0.0, + "text": "to 2017, one of my first projects was a" + }, + { + "start": 185.48, + "duration": 0.0, + "text": "to 2017, one of my first projects was a game<00:03:05.959> a<00:03:06.000> game<00:03:06.280> engine." + }, + { + "start": 187.27, + "duration": 0.0, + "text": "game a game engine." + }, + { + "start": 187.28, + "duration": 0.0, + "text": "game a game engine. Um,<00:03:07.640> I<00:03:07.680> think<00:03:07.880> this<00:03:08.000> is<00:03:08.080> probably<00:03:08.280> pretty" + }, + { + "start": 188.47, + "duration": 0.0, + "text": "Um, I think this is probably pretty" + }, + { + "start": 188.48, + "duration": 0.0, + "text": "Um, I think this is probably pretty common<00:03:08.800> for<00:03:08.920> people<00:03:09.239> who<00:03:10.120> I<00:03:10.160> don't<00:03:10.280> know<00:03:10.320> like" + }, + { + "start": 190.43, + "duration": 0.0, + "text": "common for people who I don't know like" + }, + { + "start": 190.44, + "duration": 0.0, + "text": "common for people who I don't know like to<00:03:10.519> program<00:03:10.920> is<00:03:11.040> the<00:03:11.120> first<00:03:11.360> thing<00:03:11.480> they<00:03:11.560> try" + }, + { + "start": 191.67, + "duration": 0.0, + "text": "to program is the first thing they try" + }, + { + "start": 191.68, + "duration": 0.0, + "text": "to program is the first thing they try to<00:03:11.760> make<00:03:11.920> is<00:03:12.000> a<00:03:12.040> game<00:03:12.320> engine.<00:03:12.560> It's<00:03:12.680> kind<00:03:12.840> of" + }, + { + "start": 192.91, + "duration": 0.0, + "text": "to make is a game engine. It's kind of" + }, + { + "start": 192.92, + "duration": 0.0, + "text": "to make is a game engine. It's kind of funny." + }, + { + "start": 193.79, + "duration": 0.0, + "text": "funny." + }, + { + "start": 193.8, + "duration": 0.0, + "text": "funny. Um,<00:03:14.440> you<00:03:14.480> don't<00:03:14.600> know<00:03:14.680> anything<00:03:14.959> about<00:03:15.120> what" + }, + { + "start": 195.229, + "duration": 0.0, + "text": "Um, you don't know anything about what" + }, + { + "start": 195.239, + "duration": 0.0, + "text": "Um, you don't know anything about what you're<00:03:15.320> doing.<00:03:15.560> You're<00:03:15.640> like,<00:03:15.760> I'm<00:03:15.840> going<00:03:15.959> to," + }, + { + "start": 196.229, + "duration": 0.0, + "text": "you're doing. You're like, I'm going to," + }, + { + "start": 196.239, + "duration": 0.0, + "text": "you're doing. You're like, I'm going to, you<00:03:16.320> know,<00:03:16.440> do<00:03:16.560> the<00:03:16.720> thing<00:03:16.959> that's<00:03:17.200> the" + }, + { + "start": 197.27, + "duration": 0.0, + "text": "you know, do the thing that's the" + }, + { + "start": 197.28, + "duration": 0.0, + "text": "you know, do the thing that's the ultimate<00:03:17.640> thing." + }, + { + "start": 198.55, + "duration": 0.0, + "text": "ultimate thing." + }, + { + "start": 198.56, + "duration": 0.0, + "text": "ultimate thing. Um,<00:03:19.000> and<00:03:19.080> then<00:03:19.200> much<00:03:19.400> later,<00:03:19.600> you<00:03:19.680> know,<00:03:19.800> 3" + }, + { + "start": 199.99, + "duration": 0.0, + "text": "Um, and then much later, you know, 3" + }, + { + "start": 200.0, + "duration": 0.0, + "text": "Um, and then much later, you know, 3 years<00:03:20.280> later,<00:03:20.519> you<00:03:20.640> know,<00:03:20.720> it's<00:03:20.800> like<00:03:20.920> I'm" + }, + { + "start": 200.99, + "duration": 0.0, + "text": "years later, you know, it's like I'm" + }, + { + "start": 201.0, + "duration": 0.0, + "text": "years later, you know, it's like I'm making<00:03:21.239> a<00:03:21.280> lot<00:03:21.400> of<00:03:21.480> individual<00:03:21.880> games,<00:03:22.280> I'm" + }, + { + "start": 202.35, + "duration": 0.0, + "text": "making a lot of individual games, I'm" + }, + { + "start": 202.36, + "duration": 0.0, + "text": "making a lot of individual games, I'm working<00:03:22.600> on<00:03:22.680> the<00:03:22.760> Godot<00:03:23.040> engine." + }, + { + "start": 204.07, + "duration": 0.0, + "text": "working on the Godot engine." + }, + { + "start": 204.08, + "duration": 0.0, + "text": "working on the Godot engine. Um,<00:03:25.000> you<00:03:25.080> know,<00:03:25.160> these<00:03:25.320> are<00:03:25.400> a<00:03:25.440> couple<00:03:25.640> of<00:03:25.720> my" + }, + { + "start": 205.83, + "duration": 0.0, + "text": "Um, you know, these are a couple of my" + }, + { + "start": 205.84, + "duration": 0.0, + "text": "Um, you know, these are a couple of my game<00:03:26.120> projects." + }, + { + "start": 207.31, + "duration": 0.0, + "text": "game projects." + }, + { + "start": 207.32, + "duration": 0.0, + "text": "game projects. Um,<00:03:27.840> that<00:03:28.040> there<00:03:28.200> on<00:03:28.320> the<00:03:28.400> upper<00:03:28.600> left<00:03:28.840> is<00:03:28.920> a" + }, + { + "start": 208.949, + "duration": 0.0, + "text": "Um, that there on the upper left is a" + }, + { + "start": 208.959, + "duration": 0.0, + "text": "Um, that there on the upper left is a space<00:03:29.239> game<00:03:29.440> called<00:03:29.720> Asteris.<00:03:30.239> It's" + }, + { + "start": 210.35, + "duration": 0.0, + "text": "space game called Asteris. It's" + }, + { + "start": 210.36, + "duration": 0.0, + "text": "space game called Asteris. It's releasing<00:03:30.720> in<00:03:30.800> 2035.<00:03:32.280> Um,<00:03:32.519> it's<00:03:32.640> got" + }, + { + "start": 212.83, + "duration": 0.0, + "text": "releasing in 2035. Um, it's got" + }, + { + "start": 212.84, + "duration": 0.0, + "text": "releasing in 2035. Um, it's got something<00:03:33.160> like<00:03:33.360> a<00:03:33.800> Overwatch<00:03:34.280> net<00:03:34.480> code<00:03:34.680> in" + }, + { + "start": 214.79, + "duration": 0.0, + "text": "something like a Overwatch net code in" + }, + { + "start": 214.8, + "duration": 0.0, + "text": "something like a Overwatch net code in it,<00:03:34.920> right?<00:03:35.239> It's<00:03:35.400> like<00:03:35.560> a" + }, + { + "start": 216.35, + "duration": 0.0, + "text": "it, right? It's like a" + }, + { + "start": 216.36, + "duration": 0.0, + "text": "it, right? It's like a big<00:03:37.000> solar<00:03:37.280> system<00:03:37.640> with<00:03:37.760> many<00:03:37.959> people" + }, + { + "start": 218.229, + "duration": 0.0, + "text": "big solar system with many people" + }, + { + "start": 218.239, + "duration": 0.0, + "text": "big solar system with many people playing<00:03:38.560> at<00:03:38.640> the<00:03:38.720> same<00:03:38.920> time.<00:03:39.720> Um," + }, + { + "start": 221.07, + "duration": 0.0, + "text": "playing at the same time. Um," + }, + { + "start": 221.08, + "duration": 0.0, + "text": "playing at the same time. Um, yeah,<00:03:41.480> I<00:03:41.760> I<00:03:41.800> was<00:03:41.959> pretty<00:03:42.160> proud<00:03:42.440> of<00:03:42.600> how<00:03:43.000> stable" + }, + { + "start": 223.35, + "duration": 0.0, + "text": "yeah, I I was pretty proud of how stable" + }, + { + "start": 223.36, + "duration": 0.0, + "text": "yeah, I I was pretty proud of how stable the<00:03:43.440> net<00:03:43.560> code<00:03:43.760> was<00:03:43.920> considering<00:03:44.600> how<00:03:44.720> chaotic" + }, + { + "start": 225.07, + "duration": 0.0, + "text": "the net code was considering how chaotic" + }, + { + "start": 225.08, + "duration": 0.0, + "text": "the net code was considering how chaotic that<00:03:45.239> game<00:03:45.400> was,<00:03:45.720> right?<00:03:46.560> Um," + }, + { + "start": 227.509, + "duration": 0.0, + "text": "that game was, right? Um," + }, + { + "start": 227.519, + "duration": 0.0, + "text": "that game was, right? Um, and<00:03:47.560> then<00:03:47.680> Dante's<00:03:48.040> Cowboy.<00:03:49.120> Um,<00:03:49.320> this<00:03:49.480> was<00:03:49.600> no" + }, + { + "start": 229.71, + "duration": 0.0, + "text": "and then Dante's Cowboy. Um, this was no" + }, + { + "start": 229.72, + "duration": 0.0, + "text": "and then Dante's Cowboy. Um, this was no engine<00:03:49.959> in<00:03:50.080> C,<00:03:50.320> 3D<00:03:50.600> rendering,<00:03:51.000> animated" + }, + { + "start": 231.31, + "duration": 0.0, + "text": "engine in C, 3D rendering, animated" + }, + { + "start": 231.32, + "duration": 0.0, + "text": "engine in C, 3D rendering, animated armatures,<00:03:52.440> all<00:03:52.560> from<00:03:52.720> scratch,<00:03:53.240> and<00:03:53.440> all<00:03:53.560> the" + }, + { + "start": 233.67, + "duration": 0.0, + "text": "armatures, all from scratch, and all the" + }, + { + "start": 233.68, + "duration": 0.0, + "text": "armatures, all from scratch, and all the NPCs<00:03:54.120> in<00:03:54.160> this<00:03:54.280> game<00:03:54.560> were<00:03:55.040> uh" + }, + { + "start": 235.71, + "duration": 0.0, + "text": "NPCs in this game were uh" + }, + { + "start": 235.72, + "duration": 0.0, + "text": "NPCs in this game were uh you<00:03:55.760> know,<00:03:55.880> they<00:03:55.960> were<00:03:56.040> controlled<00:03:56.440> by<00:03:56.560> an" + }, + { + "start": 236.67, + "duration": 0.0, + "text": "you know, they were controlled by an" + }, + { + "start": 236.68, + "duration": 0.0, + "text": "you know, they were controlled by an LLM,<00:03:57.400> and<00:03:57.560> this<00:03:57.720> was<00:03:57.840> never<00:03:58.040> released<00:03:58.360> because" + }, + { + "start": 238.63, + "duration": 0.0, + "text": "LLM, and this was never released because" + }, + { + "start": 238.64, + "duration": 0.0, + "text": "LLM, and this was never released because it's<00:03:58.760> fundamentally<00:03:59.280> flawed.<00:04:00.240> Um" + }, + { + "start": 241.35, + "duration": 0.0, + "text": "it's fundamentally flawed. Um" + }, + { + "start": 241.36, + "duration": 0.0, + "text": "it's fundamentally flawed. Um I'm<00:04:01.600> a<00:04:01.640> deep<00:04:01.960> learning<00:04:02.240> hobbyist." + }, + { + "start": 243.51, + "duration": 0.0, + "text": "I'm a deep learning hobbyist." + }, + { + "start": 243.52, + "duration": 0.0, + "text": "I'm a deep learning hobbyist. The<00:04:03.560> majority<00:04:04.040> of<00:04:04.160> my<00:04:04.320> work<00:04:04.560> is<00:04:04.680> not<00:04:05.080> in<00:04:05.200> deep" + }, + { + "start": 245.39, + "duration": 0.0, + "text": "The majority of my work is not in deep" + }, + { + "start": 245.4, + "duration": 0.0, + "text": "The majority of my work is not in deep learning,<00:04:05.800> so<00:04:05.920> there's<00:04:06.080> going<00:04:06.200> to<00:04:06.280> be<00:04:06.400> some" + }, + { + "start": 246.67, + "duration": 0.0, + "text": "learning, so there's going to be some" + }, + { + "start": 246.68, + "duration": 0.0, + "text": "learning, so there's going to be some inaccuracies<00:04:07.480> in<00:04:07.600> what<00:04:07.720> I<00:04:07.800> say.<00:04:08.560> It's<00:04:08.760> just" + }, + { + "start": 248.949, + "duration": 0.0, + "text": "inaccuracies in what I say. It's just" + }, + { + "start": 248.959, + "duration": 0.0, + "text": "inaccuracies in what I say. It's just always<00:04:09.240> been<00:04:09.440> interesting<00:04:09.840> to<00:04:09.959> me<00:04:10.120> ever<00:04:10.360> since" + }, + { + "start": 250.59, + "duration": 0.0, + "text": "always been interesting to me ever since" + }, + { + "start": 250.6, + "duration": 0.0, + "text": "always been interesting to me ever since back<00:04:10.800> when<00:04:10.920> I<00:04:10.959> started,<00:04:11.320> you<00:04:11.360> know,<00:04:11.480> first" + }, + { + "start": 251.67, + "duration": 0.0, + "text": "back when I started, you know, first" + }, + { + "start": 251.68, + "duration": 0.0, + "text": "back when I started, you know, first doing<00:04:11.840> games,<00:04:12.280> I've<00:04:12.400> also<00:04:12.600> been<00:04:12.720> kind<00:04:12.920> of" + }, + { + "start": 253.03, + "duration": 0.0, + "text": "doing games, I've also been kind of" + }, + { + "start": 253.04, + "duration": 0.0, + "text": "doing games, I've also been kind of looking<00:04:13.320> at<00:04:13.920> deep<00:04:14.160> learning,<00:04:14.480> which<00:04:14.640> is<00:04:14.720> a" + }, + { + "start": 254.79, + "duration": 0.0, + "text": "looking at deep learning, which is a" + }, + { + "start": 254.8, + "duration": 0.0, + "text": "looking at deep learning, which is a subset<00:04:15.160> of<00:04:15.240> machine<00:04:15.520> learning,<00:04:16.239> and<00:04:16.400> it's" + }, + { + "start": 256.47, + "duration": 0.0, + "text": "subset of machine learning, and it's" + }, + { + "start": 256.48, + "duration": 0.0, + "text": "subset of machine learning, and it's just<00:04:16.600> always<00:04:16.799> kind<00:04:16.959> of<00:04:17.079> interested<00:04:17.440> me.<00:04:17.560> Like" + }, + { + "start": 257.71, + "duration": 0.0, + "text": "just always kind of interested me. Like" + }, + { + "start": 257.72, + "duration": 0.0, + "text": "just always kind of interested me. Like I've<00:04:17.799> been<00:04:17.920> following<00:04:18.320> it,<00:04:18.440> right?" + }, + { + "start": 259.39, + "duration": 0.0, + "text": "I've been following it, right?" + }, + { + "start": 259.4, + "duration": 0.0, + "text": "I've been following it, right? Um" + }, + { + "start": 260.349, + "duration": 0.0, + "text": "Um" + }, + { + "start": 260.359, + "duration": 0.0, + "text": "Um Ironically<00:04:20.799> enough,<00:04:21.040> actually,<00:04:21.320> my<00:04:21.440> first" + }, + { + "start": 261.75, + "duration": 0.0, + "text": "Ironically enough, actually, my first" + }, + { + "start": 261.76, + "duration": 0.0, + "text": "Ironically enough, actually, my first ever<00:04:21.959> professional<00:04:22.440> programming<00:04:22.880> experience" + }, + { + "start": 264.07, + "duration": 0.0, + "text": "ever professional programming experience" + }, + { + "start": 264.08, + "duration": 0.0, + "text": "ever professional programming experience was<00:04:24.440> to<00:04:24.919> architect<00:04:25.360> a<00:04:25.400> deep<00:04:25.640> learning<00:04:25.840> model" + }, + { + "start": 266.03, + "duration": 0.0, + "text": "was to architect a deep learning model" + }, + { + "start": 266.04, + "duration": 0.0, + "text": "was to architect a deep learning model from<00:04:26.200> scratch,<00:04:27.200> um<00:04:27.480> make<00:04:27.680> it<00:04:27.840> work,<00:04:28.120> and" + }, + { + "start": 268.23, + "duration": 0.0, + "text": "from scratch, um make it work, and" + }, + { + "start": 268.24, + "duration": 0.0, + "text": "from scratch, um make it work, and deploy<00:04:28.600> it<00:04:28.720> in<00:04:28.800> production.<00:04:29.840> Um<00:04:30.080> this<00:04:30.240> was<00:04:30.400> for" + }, + { + "start": 270.67, + "duration": 0.0, + "text": "deploy it in production. Um this was for" + }, + { + "start": 270.68, + "duration": 0.0, + "text": "deploy it in production. Um this was for my<00:04:30.880> college<00:04:31.280> roommate's<00:04:31.600> company<00:04:31.960> called" + }, + { + "start": 272.19, + "duration": 0.0, + "text": "my college roommate's company called" + }, + { + "start": 272.2, + "duration": 0.0, + "text": "my college roommate's company called MacroHard.<00:04:33.520> Um<00:04:34.400> he<00:04:34.560> basically<00:04:35.040> made<00:04:35.280> software" + }, + { + "start": 275.83, + "duration": 0.0, + "text": "MacroHard. Um he basically made software" + }, + { + "start": 275.84, + "duration": 0.0, + "text": "MacroHard. Um he basically made software that<00:04:36.040> like<00:04:36.600> was<00:04:36.760> a<00:04:36.800> custom<00:04:37.160> broadcast<00:04:37.720> overlay" + }, + { + "start": 278.11, + "duration": 0.0, + "text": "that like was a custom broadcast overlay" + }, + { + "start": 278.12, + "duration": 0.0, + "text": "that like was a custom broadcast overlay for<00:04:38.360> for<00:04:38.560> League<00:04:38.760> of<00:04:38.840> Legends<00:04:39.240> teams," + }, + { + "start": 280.31, + "duration": 0.0, + "text": "for for League of Legends teams," + }, + { + "start": 280.32, + "duration": 0.0, + "text": "for for League of Legends teams, and<00:04:40.840> I<00:04:41.000> made<00:04:41.280> a<00:04:41.400> deep<00:04:41.760> learning<00:04:42.000> model<00:04:42.240> that" + }, + { + "start": 282.35, + "duration": 0.0, + "text": "and I made a deep learning model that" + }, + { + "start": 282.36, + "duration": 0.0, + "text": "and I made a deep learning model that predicted<00:04:42.880> which<00:04:43.040> team<00:04:43.360> was<00:04:43.680> uh<00:04:43.760> going<00:04:43.960> to<00:04:44.040> win" + }, + { + "start": 284.15, + "duration": 0.0, + "text": "predicted which team was uh going to win" + }, + { + "start": 284.16, + "duration": 0.0, + "text": "predicted which team was uh going to win the<00:04:44.240> game,<00:04:44.919> right?<00:04:45.520> So,<00:04:45.600> that<00:04:45.720> little" + }, + { + "start": 285.87, + "duration": 0.0, + "text": "the game, right? So, that little" + }, + { + "start": 285.88, + "duration": 0.0, + "text": "the game, right? So, that little visualization<00:04:46.440> there<00:04:46.600> is<00:04:46.720> what<00:04:46.880> I<00:04:47.240> created." + }, + { + "start": 288.15, + "duration": 0.0, + "text": "visualization there is what I created." + }, + { + "start": 288.16, + "duration": 0.0, + "text": "visualization there is what I created. Uh<00:04:48.240> this<00:04:48.360> was<00:04:48.480> trained<00:04:48.919> and,<00:04:49.280> you<00:04:49.360> know,<00:04:49.520> on<00:04:49.880> a" + }, + { + "start": 290.03, + "duration": 0.0, + "text": "Uh this was trained and, you know, on a" + }, + { + "start": 290.04, + "duration": 0.0, + "text": "Uh this was trained and, you know, on a and<00:04:50.120> optimized<00:04:50.560> on<00:04:50.680> many,<00:04:50.880> many<00:04:51.400> professional" + }, + { + "start": 291.67, + "duration": 0.0, + "text": "and optimized on many, many professional" + }, + { + "start": 291.68, + "duration": 0.0, + "text": "and optimized on many, many professional games." + }, + { + "start": 293.07, + "duration": 0.0, + "text": "games." + }, + { + "start": 293.08, + "duration": 0.0, + "text": "games. Um" + }, + { + "start": 294.43, + "duration": 0.0, + "text": "Um" + }, + { + "start": 294.44, + "duration": 0.0, + "text": "Um and<00:04:54.560> you<00:04:54.600> know,<00:04:54.720> it's<00:04:54.840> it's<00:04:55.000> just<00:04:55.120> a<00:04:55.160> pretty" + }, + { + "start": 295.39, + "duration": 0.0, + "text": "and you know, it's it's just a pretty" + }, + { + "start": 295.4, + "duration": 0.0, + "text": "and you know, it's it's just a pretty interesting<00:04:55.720> problem,<00:04:56.080> right?<00:04:56.840> And<00:04:57.000> at<00:04:57.120> that" + }, + { + "start": 297.27, + "duration": 0.0, + "text": "interesting problem, right? And at that" + }, + { + "start": 297.28, + "duration": 0.0, + "text": "interesting problem, right? And at that time,<00:04:57.880> right,<00:04:58.120> I<00:04:58.240> implemented<00:04:59.040> uh<00:04:59.200> League<00:04:59.400> of" + }, + { + "start": 299.47, + "duration": 0.0, + "text": "time, right, I implemented uh League of" + }, + { + "start": 299.48, + "duration": 0.0, + "text": "time, right, I implemented uh League of Legends<00:04:59.919> game<00:05:00.160> predicting<00:05:00.480> paper," + }, + { + "start": 301.51, + "duration": 0.0, + "text": "Legends game predicting paper," + }, + { + "start": 301.52, + "duration": 0.0, + "text": "Legends game predicting paper, and<00:05:01.720> I<00:05:01.760> couldn't<00:05:02.080> reproduce<00:05:02.400> the<00:05:02.520> results," + }, + { + "start": 302.95, + "duration": 0.0, + "text": "and I couldn't reproduce the results," + }, + { + "start": 302.96, + "duration": 0.0, + "text": "and I couldn't reproduce the results, you<00:05:03.040> know,<00:05:03.200> I<00:05:03.280> was<00:05:03.440> like,<00:05:04.000> \"Well," + }, + { + "start": 304.909, + "duration": 0.0, + "text": "you know, I was like, \"Well," + }, + { + "start": 304.919, + "duration": 0.0, + "text": "you know, I was like, \"Well, I<00:05:05.000> copied<00:05:05.360> the<00:05:05.440> thing<00:05:05.640> that<00:05:05.760> the<00:05:05.840> paper<00:05:06.080> said" + }, + { + "start": 306.27, + "duration": 0.0, + "text": "I copied the thing that the paper said" + }, + { + "start": 306.28, + "duration": 0.0, + "text": "I copied the thing that the paper said to<00:05:06.360> do,<00:05:06.600> and<00:05:06.720> then<00:05:06.880> it's,<00:05:07.320> you<00:05:07.400> know,<00:05:07.480> it" + }, + { + "start": 307.55, + "duration": 0.0, + "text": "to do, and then it's, you know, it" + }, + { + "start": 307.56, + "duration": 0.0, + "text": "to do, and then it's, you know, it doesn't<00:05:07.800> work.\"" + }, + { + "start": 308.75, + "duration": 0.0, + "text": "doesn't work.\"" + }, + { + "start": 308.76, + "duration": 0.0, + "text": "doesn't work.\" And<00:05:09.720> basically," + }, + { + "start": 311.19, + "duration": 0.0, + "text": "And basically," + }, + { + "start": 311.2, + "duration": 0.0, + "text": "And basically, there's<00:05:11.360> a<00:05:11.400> there<00:05:11.520> was<00:05:11.640> a<00:05:11.720> bug<00:05:12.040> in<00:05:12.160> the<00:05:12.240> paper," + }, + { + "start": 312.55, + "duration": 0.0, + "text": "there's a there was a bug in the paper," + }, + { + "start": 312.56, + "duration": 0.0, + "text": "there's a there was a bug in the paper, right?<00:05:12.720> This<00:05:12.800> is<00:05:12.880> a<00:05:12.919> very<00:05:13.080> important<00:05:13.440> lesson" + }, + { + "start": 313.71, + "duration": 0.0, + "text": "right? This is a very important lesson" + }, + { + "start": 313.72, + "duration": 0.0, + "text": "right? This is a very important lesson of<00:05:13.880> not<00:05:14.120> just<00:05:14.320> how<00:05:14.440> hard<00:05:14.760> machine<00:05:15.000> learning" + }, + { + "start": 315.27, + "duration": 0.0, + "text": "of not just how hard machine learning" + }, + { + "start": 315.28, + "duration": 0.0, + "text": "of not just how hard machine learning is,<00:05:15.480> but<00:05:15.960> how<00:05:16.120> often<00:05:16.400> people<00:05:16.680> mess<00:05:16.880> up,<00:05:17.160> right?" + }, + { + "start": 317.31, + "duration": 0.0, + "text": "is, but how often people mess up, right?" + }, + { + "start": 317.32, + "duration": 0.0, + "text": "is, but how often people mess up, right? And<00:05:17.400> how<00:05:17.520> easy<00:05:17.760> it<00:05:17.840> is<00:05:17.960> to<00:05:18.040> mess<00:05:18.280> up<00:05:18.840> and<00:05:18.960> be<00:05:19.040> so" + }, + { + "start": 319.19, + "duration": 0.0, + "text": "And how easy it is to mess up and be so" + }, + { + "start": 319.2, + "duration": 0.0, + "text": "And how easy it is to mess up and be so confident<00:05:19.680> in<00:05:19.760> what<00:05:19.880> you<00:05:19.960> did<00:05:20.600> that<00:05:20.720> you" + }, + { + "start": 320.83, + "duration": 0.0, + "text": "confident in what you did that you" + }, + { + "start": 320.84, + "duration": 0.0, + "text": "confident in what you did that you publish<00:05:21.200> it<00:05:21.280> to<00:05:21.360> the<00:05:21.480> world<00:05:21.880> and<00:05:22.040> waste<00:05:22.400> my," + }, + { + "start": 322.87, + "duration": 0.0, + "text": "publish it to the world and waste my," + }, + { + "start": 322.88, + "duration": 0.0, + "text": "publish it to the world and waste my, you<00:05:22.960> know,<00:05:23.080> 2<00:05:23.280> weeks,<00:05:23.720> like" + }, + { + "start": 325.23, + "duration": 0.0, + "text": "you know, 2 weeks, like" + }, + { + "start": 325.24, + "duration": 0.0, + "text": "you know, 2 weeks, like it's<00:05:25.440> just<00:05:25.720> very,<00:05:25.960> very<00:05:26.200> easy<00:05:26.400> to<00:05:26.480> mess<00:05:26.720> up." + }, + { + "start": 326.909, + "duration": 0.0, + "text": "it's just very, very easy to mess up." + }, + { + "start": 326.919, + "duration": 0.0, + "text": "it's just very, very easy to mess up. And<00:05:27.040> just<00:05:27.520> to<00:05:27.600> give<00:05:27.720> a<00:05:27.800> quick<00:05:28.000> overview<00:05:28.400> of<00:05:28.520> the" + }, + { + "start": 328.63, + "duration": 0.0, + "text": "And just to give a quick overview of the" + }, + { + "start": 328.64, + "duration": 0.0, + "text": "And just to give a quick overview of the bug<00:05:28.919> they<00:05:29.040> had,<00:05:29.480> right?<00:05:30.280> When<00:05:30.400> you<00:05:30.480> want<00:05:30.600> to" + }, + { + "start": 330.63, + "duration": 0.0, + "text": "bug they had, right? When you want to" + }, + { + "start": 330.64, + "duration": 0.0, + "text": "bug they had, right? When you want to make<00:05:30.800> sure<00:05:31.040> a<00:05:31.120> machine<00:05:31.640> learning<00:05:31.919> model" + }, + { + "start": 332.23, + "duration": 0.0, + "text": "make sure a machine learning model" + }, + { + "start": 332.24, + "duration": 0.0, + "text": "make sure a machine learning model works,<00:05:32.960> you<00:05:33.080> usually<00:05:33.400> have<00:05:33.560> a<00:05:33.640> you<00:05:33.720> give<00:05:33.840> it<00:05:33.919> a" + }, + { + "start": 333.95, + "duration": 0.0, + "text": "works, you usually have a you give it a" + }, + { + "start": 333.96, + "duration": 0.0, + "text": "works, you usually have a you give it a test,<00:05:34.320> just<00:05:34.440> like<00:05:34.560> a<00:05:34.600> person,<00:05:35.000> right?<00:05:35.800> And<00:05:36.320> if" + }, + { + "start": 336.59, + "duration": 0.0, + "text": "test, just like a person, right? And if" + }, + { + "start": 336.6, + "duration": 0.0, + "text": "test, just like a person, right? And if that<00:05:36.800> machine<00:05:37.040> learning<00:05:37.240> model<00:05:37.480> were<00:05:37.600> to<00:05:37.720> see" + }, + { + "start": 337.83, + "duration": 0.0, + "text": "that machine learning model were to see" + }, + { + "start": 337.84, + "duration": 0.0, + "text": "that machine learning model were to see the<00:05:38.040> answers<00:05:38.480> of<00:05:38.560> the<00:05:38.680> test<00:05:39.120> before<00:05:39.400> it<00:05:39.520> took" + }, + { + "start": 339.67, + "duration": 0.0, + "text": "the answers of the test before it took" + }, + { + "start": 339.68, + "duration": 0.0, + "text": "the answers of the test before it took it,<00:05:40.400> uh<00:05:40.800> you<00:05:40.880> know,<00:05:41.080> its<00:05:41.280> success<00:05:41.640> in<00:05:41.720> that" + }, + { + "start": 341.83, + "duration": 0.0, + "text": "it, uh you know, its success in that" + }, + { + "start": 341.84, + "duration": 0.0, + "text": "it, uh you know, its success in that test<00:05:42.120> would<00:05:42.280> obviously<00:05:42.640> not<00:05:42.800> be" + }, + { + "start": 342.87, + "duration": 0.0, + "text": "test would obviously not be" + }, + { + "start": 342.88, + "duration": 0.0, + "text": "test would obviously not be representative<00:05:43.520> of<00:05:43.640> it<00:05:43.840> of<00:05:44.080> if<00:05:44.200> it<00:05:44.280> could<00:05:44.400> do" + }, + { + "start": 344.51, + "duration": 0.0, + "text": "representative of it of if it could do" + }, + { + "start": 344.52, + "duration": 0.0, + "text": "representative of it of if it could do the,<00:05:44.840> you<00:05:44.919> know,<00:05:45.000> a<00:05:45.240> problem,<00:05:45.640> right?" + }, + { + "start": 346.11, + "duration": 0.0, + "text": "the, you know, a problem, right?" + }, + { + "start": 346.12, + "duration": 0.0, + "text": "the, you know, a problem, right? Arbitrary<00:05:46.600> new<00:05:46.760> examples,<00:05:47.400> it's<00:05:47.600> not" + }, + { + "start": 347.79, + "duration": 0.0, + "text": "Arbitrary new examples, it's not" + }, + { + "start": 347.8, + "duration": 0.0, + "text": "Arbitrary new examples, it's not evidence<00:05:48.080> that<00:05:48.200> it<00:05:48.280> generalizes.<00:05:49.400> And<00:05:49.520> in" + }, + { + "start": 349.55, + "duration": 0.0, + "text": "evidence that it generalizes. And in" + }, + { + "start": 349.56, + "duration": 0.0, + "text": "evidence that it generalizes. And in this<00:05:49.720> specific<00:05:50.120> case,<00:05:50.480> they<00:05:50.640> had<00:05:50.800> a<00:05:50.840> metric," + }, + { + "start": 351.43, + "duration": 0.0, + "text": "this specific case, they had a metric," + }, + { + "start": 351.44, + "duration": 0.0, + "text": "this specific case, they had a metric, the<00:05:51.600> champion<00:05:52.040> win<00:05:52.200> rate,<00:05:52.880> that<00:05:53.080> was<00:05:53.240> trained" + }, + { + "start": 353.51, + "duration": 0.0, + "text": "the champion win rate, that was trained" + }, + { + "start": 353.52, + "duration": 0.0, + "text": "the champion win rate, that was trained on<00:05:53.600> their<00:05:53.720> entire<00:05:54.760> um" + }, + { + "start": 355.63, + "duration": 0.0, + "text": "on their entire um" + }, + { + "start": 355.64, + "duration": 0.0, + "text": "on their entire um that<00:05:55.760> champion<00:05:56.040> win<00:05:56.200> rate<00:05:56.320> was<00:05:56.440> based<00:05:56.640> on" + }, + { + "start": 356.71, + "duration": 0.0, + "text": "that champion win rate was based on" + }, + { + "start": 356.72, + "duration": 0.0, + "text": "that champion win rate was based on their<00:05:56.800> entire<00:05:57.200> data<00:05:57.440> set,<00:05:57.800> right?<00:05:58.400> And<00:05:58.520> so" + }, + { + "start": 358.59, + "duration": 0.0, + "text": "their entire data set, right? And so" + }, + { + "start": 358.6, + "duration": 0.0, + "text": "their entire data set, right? And so then<00:05:58.720> when<00:05:58.880> it<00:05:59.000> went<00:05:59.200> to<00:05:59.920> test<00:06:00.240> the<00:06:00.360> model,<00:06:00.720> it" + }, + { + "start": 360.83, + "duration": 0.0, + "text": "then when it went to test the model, it" + }, + { + "start": 360.84, + "duration": 0.0, + "text": "then when it went to test the model, it had<00:06:00.960> the<00:06:01.080> win<00:06:01.320> rate<00:06:01.480> for<00:06:01.600> the<00:06:01.680> champions<00:06:02.080> that" + }, + { + "start": 362.19, + "duration": 0.0, + "text": "had the win rate for the champions that" + }, + { + "start": 362.2, + "duration": 0.0, + "text": "had the win rate for the champions that were<00:06:02.320> in<00:06:02.480> the<00:06:02.560> test<00:06:02.880> set.<00:06:03.560> And<00:06:03.760> so<00:06:04.000> it" + }, + { + "start": 364.11, + "duration": 0.0, + "text": "were in the test set. And so it" + }, + { + "start": 364.12, + "duration": 0.0, + "text": "were in the test set. And so it purported<00:06:04.560> to<00:06:04.640> have<00:06:04.840> amazing<00:06:05.240> accuracy<00:06:05.640> at" + }, + { + "start": 365.75, + "duration": 0.0, + "text": "purported to have amazing accuracy at" + }, + { + "start": 365.76, + "duration": 0.0, + "text": "purported to have amazing accuracy at predicting<00:06:06.160> who<00:06:06.280> would<00:06:06.400> win<00:06:07.080> on<00:06:07.200> the<00:06:07.280> test" + }, + { + "start": 367.55, + "duration": 0.0, + "text": "predicting who would win on the test" + }, + { + "start": 367.56, + "duration": 0.0, + "text": "predicting who would win on the test set,<00:06:07.760> but<00:06:07.880> there<00:06:08.000> was<00:06:08.160> data<00:06:08.560> information" + }, + { + "start": 369.07, + "duration": 0.0, + "text": "set, but there was data information" + }, + { + "start": 369.08, + "duration": 0.0, + "text": "set, but there was data information pollution,<00:06:09.760> right?<00:06:10.600> These<00:06:10.760> are<00:06:10.800> the<00:06:10.880> kinds<00:06:11.080> of" + }, + { + "start": 371.15, + "duration": 0.0, + "text": "pollution, right? These are the kinds of" + }, + { + "start": 371.16, + "duration": 0.0, + "text": "pollution, right? These are the kinds of bugs<00:06:11.360> you<00:06:11.480> find<00:06:11.720> in<00:06:11.960> deep<00:06:12.200> learning.<00:06:12.480> They're" + }, + { + "start": 372.59, + "duration": 0.0, + "text": "bugs you find in deep learning. They're" + }, + { + "start": 372.6, + "duration": 0.0, + "text": "bugs you find in deep learning. They're not<00:06:12.760> something<00:06:13.080> you<00:06:13.160> could<00:06:13.320> really" + }, + { + "start": 373.51, + "duration": 0.0, + "text": "not something you could really" + }, + { + "start": 373.52, + "duration": 0.0, + "text": "not something you could really breakpoint<00:06:13.960> in<00:06:14.040> the<00:06:14.120> debugger,<00:06:14.960> but<00:06:15.040> things" + }, + { + "start": 375.23, + "duration": 0.0, + "text": "breakpoint in the debugger, but things" + }, + { + "start": 375.24, + "duration": 0.0, + "text": "breakpoint in the debugger, but things you<00:06:15.280> have<00:06:15.400> to<00:06:15.480> find<00:06:15.760> like<00:06:15.880> a<00:06:15.920> scientist," + }, + { + "start": 376.55, + "duration": 0.0, + "text": "you have to find like a scientist," + }, + { + "start": 376.56, + "duration": 0.0, + "text": "you have to find like a scientist, right?<00:06:17.360> Um<00:06:17.520> you<00:06:17.600> almost<00:06:17.800> have<00:06:17.960> to<00:06:18.040> like<00:06:18.280> uh" + }, + { + "start": 378.63, + "duration": 0.0, + "text": "right? Um you almost have to like uh" + }, + { + "start": 378.64, + "duration": 0.0, + "text": "right? Um you almost have to like uh binary<00:06:19.000> search<00:06:19.240> over<00:06:19.400> the<00:06:19.760> uh<00:06:20.120> problem<00:06:20.480> space" + }, + { + "start": 380.71, + "duration": 0.0, + "text": "binary search over the uh problem space" + }, + { + "start": 380.72, + "duration": 0.0, + "text": "binary search over the uh problem space sometimes<00:06:21.160> to<00:06:21.280> see<00:06:21.400> what's<00:06:21.520> going<00:06:21.760> wrong.<00:06:22.320> It" + }, + { + "start": 382.39, + "duration": 0.0, + "text": "sometimes to see what's going wrong. It" + }, + { + "start": 382.4, + "duration": 0.0, + "text": "sometimes to see what's going wrong. It could<00:06:22.480> be<00:06:22.600> anything,<00:06:23.360> right?" + }, + { + "start": 384.31, + "duration": 0.0, + "text": "could be anything, right?" + }, + { + "start": 384.32, + "duration": 0.0, + "text": "could be anything, right? And<00:06:24.400> I<00:06:24.440> want<00:06:24.560> to<00:06:24.640> stress<00:06:24.800> again<00:06:24.960> that<00:06:25.040> machine" + }, + { + "start": 385.31, + "duration": 0.0, + "text": "And I want to stress again that machine" + }, + { + "start": 385.32, + "duration": 0.0, + "text": "And I want to stress again that machine learning<00:06:25.560> is<00:06:25.720> automatic<00:06:26.160> programming," + }, + { + "start": 387.23, + "duration": 0.0, + "text": "learning is automatic programming," + }, + { + "start": 387.24, + "duration": 0.0, + "text": "learning is automatic programming, right?<00:06:27.480> In<00:06:27.560> this<00:06:27.720> paper,<00:06:28.120> this<00:06:28.280> is<00:06:28.360> a<00:06:28.400> machine" + }, + { + "start": 388.67, + "duration": 0.0, + "text": "right? In this paper, this is a machine" + }, + { + "start": 388.68, + "duration": 0.0, + "text": "right? In this paper, this is a machine learning<00:06:28.919> technique,<00:06:29.400> not<00:06:29.560> a<00:06:29.600> deep<00:06:29.760> learning" + }, + { + "start": 389.99, + "duration": 0.0, + "text": "learning technique, not a deep learning" + }, + { + "start": 390.0, + "duration": 0.0, + "text": "learning technique, not a deep learning technique,<00:06:30.280> called<00:06:30.480> random<00:06:30.760> forest,<00:06:31.640> where" + }, + { + "start": 391.75, + "duration": 0.0, + "text": "technique, called random forest, where" + }, + { + "start": 391.76, + "duration": 0.0, + "text": "technique, called random forest, where it's<00:06:31.880> basically<00:06:32.440> a<00:06:32.520> series<00:06:32.840> of<00:06:33.000> learned<00:06:33.320> if" + }, + { + "start": 393.43, + "duration": 0.0, + "text": "it's basically a series of learned if" + }, + { + "start": 393.44, + "duration": 0.0, + "text": "it's basically a series of learned if statements<00:06:33.840> directly,<00:06:34.480> right?" + }, + { + "start": 395.51, + "duration": 0.0, + "text": "statements directly, right?" + }, + { + "start": 395.52, + "duration": 0.0, + "text": "statements directly, right? And<00:06:35.640> you<00:06:35.680> know,<00:06:35.760> there<00:06:35.840> might<00:06:36.000> be<00:06:36.080> some" + }, + { + "start": 396.23, + "duration": 0.0, + "text": "And you know, there might be some" + }, + { + "start": 396.24, + "duration": 0.0, + "text": "And you know, there might be some majority<00:06:36.560> voting<00:06:36.880> scheme,<00:06:37.200> and<00:06:37.280> they<00:06:37.360> have" + }, + { + "start": 397.51, + "duration": 0.0, + "text": "majority voting scheme, and they have" + }, + { + "start": 397.52, + "duration": 0.0, + "text": "majority voting scheme, and they have some<00:06:37.720> algorithm<00:06:38.080> to<00:06:38.200> optimize<00:06:38.720> like<00:06:39.000> based<00:06:39.240> on" + }, + { + "start": 399.31, + "duration": 0.0, + "text": "some algorithm to optimize like based on" + }, + { + "start": 399.32, + "duration": 0.0, + "text": "some algorithm to optimize like based on the<00:06:39.400> training<00:06:39.680> data,<00:06:39.960> oh,<00:06:40.520> this<00:06:40.720> if<00:06:40.840> statement" + }, + { + "start": 401.07, + "duration": 0.0, + "text": "the training data, oh, this if statement" + }, + { + "start": 401.08, + "duration": 0.0, + "text": "the training data, oh, this if statement should<00:06:41.200> be<00:06:41.320> this,<00:06:41.560> and<00:06:41.680> this<00:06:41.800> if<00:06:41.880> statement" + }, + { + "start": 402.15, + "duration": 0.0, + "text": "should be this, and this if statement" + }, + { + "start": 402.16, + "duration": 0.0, + "text": "should be this, and this if statement should<00:06:42.240> be<00:06:42.360> this.<00:06:43.080> They<00:06:43.200> basically<00:06:43.640> are" + }, + { + "start": 403.79, + "duration": 0.0, + "text": "should be this. They basically are" + }, + { + "start": 403.8, + "duration": 0.0, + "text": "should be this. They basically are attempting<00:06:44.200> to<00:06:44.320> tweak<00:06:44.640> the<00:06:44.760> if<00:06:44.960> statements" + }, + { + "start": 405.31, + "duration": 0.0, + "text": "attempting to tweak the if statements" + }, + { + "start": 405.32, + "duration": 0.0, + "text": "attempting to tweak the if statements such<00:06:45.520> that<00:06:45.680> it,<00:06:45.880> you<00:06:45.960> know,<00:06:46.040> is<00:06:46.480> as<00:06:46.680> accurate" + }, + { + "start": 406.99, + "duration": 0.0, + "text": "such that it, you know, is as accurate" + }, + { + "start": 407.0, + "duration": 0.0, + "text": "such that it, you know, is as accurate as<00:06:47.080> possible<00:06:47.480> on<00:06:47.560> the<00:06:47.640> training<00:06:47.919> data." + }, + { + "start": 408.87, + "duration": 0.0, + "text": "as possible on the training data." + }, + { + "start": 408.88, + "duration": 0.0, + "text": "as possible on the training data. Um" + }, + { + "start": 410.23, + "duration": 0.0, + "text": "Um" + }, + { + "start": 410.24, + "duration": 0.0, + "text": "Um and<00:06:50.440> deep<00:06:50.680> learning<00:06:50.919> is<00:06:51.040> separate<00:06:51.360> from<00:06:51.480> that." + }, + { + "start": 411.59, + "duration": 0.0, + "text": "and deep learning is separate from that." + }, + { + "start": 411.6, + "duration": 0.0, + "text": "and deep learning is separate from that. It's<00:06:51.720> a<00:06:51.760> subset<00:06:52.320> of<00:06:52.440> machine<00:06:52.760> learning." + }, + { + "start": 413.67, + "duration": 0.0, + "text": "It's a subset of machine learning." + }, + { + "start": 413.68, + "duration": 0.0, + "text": "It's a subset of machine learning. And<00:06:53.760> the<00:06:53.840> story<00:06:54.080> kind<00:06:54.200> of<00:06:54.280> goes<00:06:54.440> like<00:06:54.600> this," + }, + { + "start": 414.909, + "duration": 0.0, + "text": "And the story kind of goes like this," + }, + { + "start": 414.919, + "duration": 0.0, + "text": "And the story kind of goes like this, right?<00:06:55.800> Some<00:06:55.960> guy<00:06:56.080> in<00:06:56.160> the<00:06:56.240> '60s<00:06:56.600> was<00:06:56.720> looking" + }, + { + "start": 416.95, + "duration": 0.0, + "text": "right? Some guy in the '60s was looking" + }, + { + "start": 416.96, + "duration": 0.0, + "text": "right? Some guy in the '60s was looking at<00:06:57.040> this,<00:06:57.240> and<00:06:57.360> he<00:06:57.440> was<00:06:57.600> like,<00:06:58.200> I<00:06:58.320> think<00:06:58.800> I<00:06:58.880> know" + }, + { + "start": 418.99, + "duration": 0.0, + "text": "at this, and he was like, I think I know" + }, + { + "start": 419.0, + "duration": 0.0, + "text": "at this, and he was like, I think I know what<00:06:59.080> these<00:06:59.280> are.<00:06:59.919> He<00:07:00.000> was<00:07:00.120> very<00:07:00.280> foolish." + }, + { + "start": 421.43, + "duration": 0.0, + "text": "what these are. He was very foolish." + }, + { + "start": 421.44, + "duration": 0.0, + "text": "what these are. He was very foolish. Um" + }, + { + "start": 422.75, + "duration": 0.0, + "text": "Um" + }, + { + "start": 422.76, + "duration": 0.0, + "text": "Um you<00:07:02.800> know,<00:07:02.880> he<00:07:03.000> he<00:07:03.160> he<00:07:03.360> basically,<00:07:04.280> on<00:07:04.400> a<00:07:04.440> very" + }, + { + "start": 424.75, + "duration": 0.0, + "text": "you know, he he he basically, on a very" + }, + { + "start": 424.76, + "duration": 0.0, + "text": "you know, he he he basically, on a very abstract<00:07:05.160> level,<00:07:05.400> said,<00:07:05.680> \"Okay,<00:07:05.919> you<00:07:06.000> know" + }, + { + "start": 426.07, + "duration": 0.0, + "text": "abstract level, said, \"Okay, you know" + }, + { + "start": 426.08, + "duration": 0.0, + "text": "abstract level, said, \"Okay, you know what<00:07:06.200> neurons<00:07:06.520> are?<00:07:06.760> You<00:07:06.840> know<00:07:06.919> what's<00:07:07.040> going" + }, + { + "start": 427.23, + "duration": 0.0, + "text": "what neurons are? You know what's going" + }, + { + "start": 427.24, + "duration": 0.0, + "text": "what neurons are? You know what's going on<00:07:07.320> in<00:07:07.400> the<00:07:07.440> brain?<00:07:07.760> You<00:07:07.800> know<00:07:07.880> what's" + }, + { + "start": 427.99, + "duration": 0.0, + "text": "on in the brain? You know what's" + }, + { + "start": 428.0, + "duration": 0.0, + "text": "on in the brain? You know what's enabling<00:07:08.360> us<00:07:08.480> to<00:07:08.560> do<00:07:08.680> all<00:07:08.760> these<00:07:08.880> amazing" + }, + { + "start": 429.19, + "duration": 0.0, + "text": "enabling us to do all these amazing" + }, + { + "start": 429.2, + "duration": 0.0, + "text": "enabling us to do all these amazing things?<00:07:10.040> Is<00:07:10.480> little,<00:07:10.919> you<00:07:11.000> know,<00:07:11.640> boxes.<00:07:12.080> They" + }, + { + "start": 432.19, + "duration": 0.0, + "text": "things? Is little, you know, boxes. They" + }, + { + "start": 432.2, + "duration": 0.0, + "text": "things? Is little, you know, boxes. They take<00:07:12.440> in<00:07:12.520> a<00:07:12.560> bunch<00:07:12.720> of<00:07:12.760> numbers,<00:07:13.680> and<00:07:13.919> then" + }, + { + "start": 434.15, + "duration": 0.0, + "text": "take in a bunch of numbers, and then" + }, + { + "start": 434.16, + "duration": 0.0, + "text": "take in a bunch of numbers, and then they<00:07:14.280> mutate<00:07:14.600> those<00:07:14.760> numbers<00:07:15.000> somehow.<00:07:15.640> Maybe" + }, + { + "start": 436.03, + "duration": 0.0, + "text": "they mutate those numbers somehow. Maybe" + }, + { + "start": 436.04, + "duration": 0.0, + "text": "they mutate those numbers somehow. Maybe when<00:07:16.200> they<00:07:16.320> have<00:07:16.480> enough<00:07:16.680> of<00:07:16.760> that<00:07:16.960> number," + }, + { + "start": 437.19, + "duration": 0.0, + "text": "when they have enough of that number," + }, + { + "start": 437.2, + "duration": 0.0, + "text": "when they have enough of that number, they<00:07:17.320> they<00:07:17.520> activate<00:07:17.880> and<00:07:18.000> propagate<00:07:18.400> that" + }, + { + "start": 438.51, + "duration": 0.0, + "text": "they they activate and propagate that" + }, + { + "start": 438.52, + "duration": 0.0, + "text": "they they activate and propagate that information<00:07:18.960> to<00:07:19.040> the<00:07:19.160> other<00:07:19.360> neurons.\"" + }, + { + "start": 440.47, + "duration": 0.0, + "text": "information to the other neurons.\"" + }, + { + "start": 440.48, + "duration": 0.0, + "text": "information to the other neurons.\" Um<00:07:20.960> if<00:07:21.080> you<00:07:21.120> were<00:07:21.200> to<00:07:21.280> zoom<00:07:21.520> in<00:07:21.720> really,<00:07:21.919> really" + }, + { + "start": 442.11, + "duration": 0.0, + "text": "Um if you were to zoom in really, really" + }, + { + "start": 442.12, + "duration": 0.0, + "text": "Um if you were to zoom in really, really quickly<00:07:22.919> on<00:07:23.160> what<00:07:23.360> one<00:07:23.480> of<00:07:23.560> those,<00:07:24.000> you<00:07:24.040> know," + }, + { + "start": 444.15, + "duration": 0.0, + "text": "quickly on what one of those, you know," + }, + { + "start": 444.16, + "duration": 0.0, + "text": "quickly on what one of those, you know, very<00:07:24.400> abstract<00:07:24.840> human<00:07:25.080> neurons<00:07:25.400> that<00:07:25.520> they" + }, + { + "start": 445.59, + "duration": 0.0, + "text": "very abstract human neurons that they" + }, + { + "start": 445.6, + "duration": 0.0, + "text": "very abstract human neurons that they designed<00:07:26.440> is<00:07:26.560> actually<00:07:26.840> doing,<00:07:27.520> right?<00:07:28.200> Um<00:07:29.040> a" + }, + { + "start": 449.11, + "duration": 0.0, + "text": "designed is actually doing, right? Um a" + }, + { + "start": 449.12, + "duration": 0.0, + "text": "designed is actually doing, right? Um a single<00:07:29.440> neuron<00:07:29.800> in<00:07:29.919> a<00:07:29.960> neural<00:07:30.240> network," + }, + { + "start": 451.03, + "duration": 0.0, + "text": "single neuron in a neural network," + }, + { + "start": 451.04, + "duration": 0.0, + "text": "single neuron in a neural network, right?<00:07:31.760> The<00:07:31.880> first<00:07:32.160> step<00:07:32.400> is<00:07:32.480> it<00:07:32.640> adds<00:07:32.960> all<00:07:33.080> the" + }, + { + "start": 453.35, + "duration": 0.0, + "text": "right? The first step is it adds all the" + }, + { + "start": 453.36, + "duration": 0.0, + "text": "right? The first step is it adds all the weights<00:07:33.600> multiplied<00:07:34.000> by<00:07:34.080> the<00:07:34.200> incoming" + }, + { + "start": 454.63, + "duration": 0.0, + "text": "weights multiplied by the incoming" + }, + { + "start": 454.64, + "duration": 0.0, + "text": "weights multiplied by the incoming activations,<00:07:35.480> right?<00:07:35.960> So<00:07:36.040> you<00:07:36.120> see<00:07:36.200> that" + }, + { + "start": 456.39, + "duration": 0.0, + "text": "activations, right? So you see that" + }, + { + "start": 456.4, + "duration": 0.0, + "text": "activations, right? So you see that bigger<00:07:36.720> arrow<00:07:37.000> is<00:07:37.160> a<00:07:37.200> much<00:07:37.440> stronger<00:07:37.760> weight." + }, + { + "start": 458.03, + "duration": 0.0, + "text": "bigger arrow is a much stronger weight." + }, + { + "start": 458.04, + "duration": 0.0, + "text": "bigger arrow is a much stronger weight. That<00:07:38.160> tiny<00:07:38.440> arrow,<00:07:39.080> you<00:07:39.160> know,<00:07:39.240> you<00:07:39.280> can<00:07:39.400> even" + }, + { + "start": 459.51, + "duration": 0.0, + "text": "That tiny arrow, you know, you can even" + }, + { + "start": 459.52, + "duration": 0.0, + "text": "That tiny arrow, you know, you can even have<00:07:39.600> a<00:07:39.640> negative<00:07:40.000> weight,<00:07:40.320> right?<00:07:40.800> There's" + }, + { + "start": 460.909, + "duration": 0.0, + "text": "have a negative weight, right? There's" + }, + { + "start": 460.919, + "duration": 0.0, + "text": "have a negative weight, right? There's basically<00:07:41.320> a<00:07:41.360> set<00:07:41.680> of<00:07:41.840> weights<00:07:42.440> of<00:07:42.800> coming" + }, + { + "start": 463.07, + "duration": 0.0, + "text": "basically a set of weights of coming" + }, + { + "start": 463.08, + "duration": 0.0, + "text": "basically a set of weights of coming from<00:07:43.240> the<00:07:43.560> uh<00:07:43.920> neurons<00:07:44.360> on<00:07:44.480> the<00:07:44.560> previous" + }, + { + "start": 464.91, + "duration": 0.0, + "text": "from the uh neurons on the previous" + }, + { + "start": 464.92, + "duration": 0.0, + "text": "from the uh neurons on the previous layer,<00:07:45.640> right?<00:07:45.800> And<00:07:45.880> then<00:07:45.960> it<00:07:46.000> just<00:07:46.160> adds<00:07:46.360> them" + }, + { + "start": 466.63, + "duration": 0.0, + "text": "layer, right? And then it just adds them" + }, + { + "start": 466.64, + "duration": 0.0, + "text": "layer, right? And then it just adds them like<00:07:46.800> numbers.<00:07:47.080> They're<00:07:47.160> all<00:07:47.240> floats,<00:07:47.880> right?" + }, + { + "start": 468.87, + "duration": 0.0, + "text": "like numbers. They're all floats, right?" + }, + { + "start": 468.88, + "duration": 0.0, + "text": "like numbers. They're all floats, right? Next,<00:07:49.080> there's<00:07:49.240> a<00:07:49.320> bias<00:07:49.720> that<00:07:49.800> adds<00:07:49.960> a<00:07:50.040> bias<00:07:50.480> to" + }, + { + "start": 470.59, + "duration": 0.0, + "text": "Next, there's a bias that adds a bias to" + }, + { + "start": 470.6, + "duration": 0.0, + "text": "Next, there's a bias that adds a bias to that<00:07:50.840> weighted<00:07:51.200> sum." + }, + { + "start": 472.11, + "duration": 0.0, + "text": "that weighted sum." + }, + { + "start": 472.12, + "duration": 0.0, + "text": "that weighted sum. Um<00:07:52.240> so,<00:07:52.360> this<00:07:52.520> could<00:07:52.640> be<00:07:52.760> considered<00:07:53.120> like<00:07:53.240> the" + }, + { + "start": 473.35, + "duration": 0.0, + "text": "Um so, this could be considered like the" + }, + { + "start": 473.36, + "duration": 0.0, + "text": "Um so, this could be considered like the activation<00:07:53.840> threshold<00:07:54.240> for<00:07:54.360> that<00:07:54.480> neuron," + }, + { + "start": 474.83, + "duration": 0.0, + "text": "activation threshold for that neuron," + }, + { + "start": 474.84, + "duration": 0.0, + "text": "activation threshold for that neuron, right?<00:07:55.040> So,<00:07:55.160> like<00:07:56.000> if<00:07:56.120> the<00:07:56.200> bias<00:07:56.560> is<00:07:56.680> -6,<00:07:57.280> then" + }, + { + "start": 477.39, + "duration": 0.0, + "text": "right? So, like if the bias is -6, then" + }, + { + "start": 477.4, + "duration": 0.0, + "text": "right? So, like if the bias is -6, then you<00:07:57.480> would<00:07:57.600> require,<00:07:58.280> you<00:07:58.360> know,<00:07:58.600> six<00:07:58.960> to<00:07:59.120> be" + }, + { + "start": 479.39, + "duration": 0.0, + "text": "you would require, you know, six to be" + }, + { + "start": 479.4, + "duration": 0.0, + "text": "you would require, you know, six to be above<00:07:59.760> zero<00:08:00.080> and<00:08:00.240> to<00:08:00.320> continue<00:08:00.640> to<00:08:00.760> propagate" + }, + { + "start": 481.19, + "duration": 0.0, + "text": "above zero and to continue to propagate" + }, + { + "start": 481.2, + "duration": 0.0, + "text": "above zero and to continue to propagate that<00:08:01.320> information.<00:08:02.240> Again,<00:08:02.440> this<00:08:02.560> is<00:08:02.640> a<00:08:02.680> very" + }, + { + "start": 482.95, + "duration": 0.0, + "text": "that information. Again, this is a very" + }, + { + "start": 482.96, + "duration": 0.0, + "text": "that information. Again, this is a very rough,<00:08:03.320> bad<00:08:03.680> model<00:08:04.040> of<00:08:04.120> biological<00:08:04.720> neurons," + }, + { + "start": 485.51, + "duration": 0.0, + "text": "rough, bad model of biological neurons," + }, + { + "start": 485.52, + "duration": 0.0, + "text": "rough, bad model of biological neurons, and<00:08:05.640> they<00:08:05.720> stumbled<00:08:06.040> upon<00:08:06.280> something<00:08:06.560> that's" + }, + { + "start": 486.71, + "duration": 0.0, + "text": "and they stumbled upon something that's" + }, + { + "start": 486.72, + "duration": 0.0, + "text": "and they stumbled upon something that's actually<00:08:06.960> pretty<00:08:07.160> crazy<00:08:08.040> um<00:08:08.200> just<00:08:08.360> by<00:08:08.440> doing" + }, + { + "start": 488.71, + "duration": 0.0, + "text": "actually pretty crazy um just by doing" + }, + { + "start": 488.72, + "duration": 0.0, + "text": "actually pretty crazy um just by doing that,<00:08:08.880> right?<00:08:09.080> But,<00:08:09.520> the<00:08:09.640> final<00:08:09.920> step<00:08:10.160> is<00:08:10.280> to" + }, + { + "start": 490.39, + "duration": 0.0, + "text": "that, right? But, the final step is to" + }, + { + "start": 490.4, + "duration": 0.0, + "text": "that, right? But, the final step is to pass<00:08:10.680> it<00:08:10.800> through<00:08:11.000> any<00:08:11.240> nonlinear<00:08:12.120> like<00:08:12.320> sine" + }, + { + "start": 492.63, + "duration": 0.0, + "text": "pass it through any nonlinear like sine" + }, + { + "start": 492.64, + "duration": 0.0, + "text": "pass it through any nonlinear like sine or<00:08:13.000> or<00:08:13.120> x<00:08:13.240> squared.<00:08:13.600> I'll<00:08:13.680> only<00:08:13.800> get<00:08:13.919> to<00:08:13.960> why" + }, + { + "start": 494.03, + "duration": 0.0, + "text": "or or x squared. I'll only get to why" + }, + { + "start": 494.04, + "duration": 0.0, + "text": "or or x squared. I'll only get to why they<00:08:14.160> do<00:08:14.320> that<00:08:14.520> in<00:08:14.640> a<00:08:14.680> second,<00:08:15.120> right?<00:08:15.680> That's" + }, + { + "start": 495.87, + "duration": 0.0, + "text": "they do that in a second, right? That's" + }, + { + "start": 495.88, + "duration": 0.0, + "text": "they do that in a second, right? That's what<00:08:16.000> happens<00:08:16.280> in<00:08:16.360> a<00:08:16.400> neuron.<00:08:16.919> Here's<00:08:17.080> some" + }, + { + "start": 497.19, + "duration": 0.0, + "text": "what happens in a neuron. Here's some" + }, + { + "start": 497.2, + "duration": 0.0, + "text": "what happens in a neuron. Here's some like<00:08:17.320> pseudocode<00:08:17.919> for<00:08:18.120> how<00:08:18.240> you<00:08:18.360> would<00:08:18.480> find" + }, + { + "start": 498.75, + "duration": 0.0, + "text": "like pseudocode for how you would find" + }, + { + "start": 498.76, + "duration": 0.0, + "text": "like pseudocode for how you would find the<00:08:18.919> float<00:08:19.240> value<00:08:19.520> for<00:08:19.640> a<00:08:19.720> neuron," + }, + { + "start": 500.83, + "duration": 0.0, + "text": "the float value for a neuron," + }, + { + "start": 500.84, + "duration": 0.0, + "text": "the float value for a neuron, right?<00:08:21.760> Um" + }, + { + "start": 502.43, + "duration": 0.0, + "text": "right? Um" + }, + { + "start": 502.44, + "duration": 0.0, + "text": "right? Um and<00:08:22.560> you<00:08:22.640> can<00:08:22.760> take<00:08:22.919> a<00:08:22.960> look<00:08:23.080> at<00:08:23.120> this<00:08:23.240> and" + }, + { + "start": 503.35, + "duration": 0.0, + "text": "and you can take a look at this and" + }, + { + "start": 503.36, + "duration": 0.0, + "text": "and you can take a look at this and ponder<00:08:23.600> it<00:08:23.680> for<00:08:23.800> a<00:08:23.840> second,<00:08:24.200> but" + }, + { + "start": 505.39, + "duration": 0.0, + "text": "ponder it for a second, but" + }, + { + "start": 505.4, + "duration": 0.0, + "text": "ponder it for a second, but you'll<00:08:25.640> you'll<00:08:26.000> notice<00:08:26.400> that<00:08:26.680> well,<00:08:27.080> if<00:08:27.200> we're" + }, + { + "start": 507.31, + "duration": 0.0, + "text": "you'll you'll notice that well, if we're" + }, + { + "start": 507.32, + "duration": 0.0, + "text": "you'll you'll notice that well, if we're constantly<00:08:27.720> just<00:08:27.880> multiplying<00:08:28.280> and<00:08:28.400> adding" + }, + { + "start": 508.59, + "duration": 0.0, + "text": "constantly just multiplying and adding" + }, + { + "start": 508.6, + "duration": 0.0, + "text": "constantly just multiplying and adding numbers,<00:08:28.880> that's<00:08:29.040> kind<00:08:29.240> of<00:08:29.320> what<00:08:29.440> linear" + }, + { + "start": 509.71, + "duration": 0.0, + "text": "numbers, that's kind of what linear" + }, + { + "start": 509.72, + "duration": 0.0, + "text": "numbers, that's kind of what linear algebra<00:08:30.040> is," + }, + { + "start": 510.99, + "duration": 0.0, + "text": "algebra is," + }, + { + "start": 511.0, + "duration": 0.0, + "text": "algebra is, and<00:08:31.320> that<00:08:31.480> is<00:08:31.680> how<00:08:31.800> they<00:08:31.919> accelerate<00:08:32.360> this<00:08:32.520> and" + }, + { + "start": 512.589, + "duration": 0.0, + "text": "and that is how they accelerate this and" + }, + { + "start": 512.599, + "duration": 0.0, + "text": "and that is how they accelerate this and talk<00:08:32.800> about<00:08:33.039> this<00:08:33.280> in<00:08:33.400> the<00:08:33.479> actual,<00:08:34.000> you<00:08:34.080> know," + }, + { + "start": 514.19, + "duration": 0.0, + "text": "talk about this in the actual, you know," + }, + { + "start": 514.2, + "duration": 0.0, + "text": "talk about this in the actual, you know, field,<00:08:34.640> right?<00:08:35.400> They<00:08:35.479> don't<00:08:35.640> express<00:08:36.039> each" + }, + { + "start": 516.19, + "duration": 0.0, + "text": "field, right? They don't express each" + }, + { + "start": 516.2, + "duration": 0.0, + "text": "field, right? They don't express each individual<00:08:36.640> weight<00:08:36.840> in<00:08:36.960> some<00:08:37.120> kind<00:08:37.240> of<00:08:37.320> array." + }, + { + "start": 517.709, + "duration": 0.0, + "text": "individual weight in some kind of array." + }, + { + "start": 517.719, + "duration": 0.0, + "text": "individual weight in some kind of array. They<00:08:37.840> express<00:08:38.159> it<00:08:38.280> as<00:08:38.440> a<00:08:38.520> matrix,<00:08:39.159> right?" + }, + { + "start": 520.07, + "duration": 0.0, + "text": "They express it as a matrix, right?" + }, + { + "start": 520.08, + "duration": 0.0, + "text": "They express it as a matrix, right? Um" + }, + { + "start": 521.23, + "duration": 0.0, + "text": "Um" + }, + { + "start": 521.24, + "duration": 0.0, + "text": "Um and<00:08:41.360> then<00:08:41.440> there's<00:08:41.760> a<00:08:41.840> lot<00:08:42.039> of<00:08:42.120> math<00:08:42.280> that" + }, + { + "start": 522.39, + "duration": 0.0, + "text": "and then there's a lot of math that" + }, + { + "start": 522.4, + "duration": 0.0, + "text": "and then there's a lot of math that applies<00:08:42.680> when<00:08:42.800> you<00:08:42.880> do<00:08:43.039> that,<00:08:43.280> right?<00:08:44.160> Um<00:08:44.520> you" + }, + { + "start": 524.67, + "duration": 0.0, + "text": "applies when you do that, right? Um you" + }, + { + "start": 524.68, + "duration": 0.0, + "text": "applies when you do that, right? Um you optimize<00:08:45.280> those<00:08:45.560> weights<00:08:45.880> and<00:08:45.960> the<00:08:46.080> bias<00:08:46.880> with" + }, + { + "start": 526.99, + "duration": 0.0, + "text": "optimize those weights and the bias with" + }, + { + "start": 527.0, + "duration": 0.0, + "text": "optimize those weights and the bias with a<00:08:47.040> process<00:08:47.440> called<00:08:47.839> gradient<00:08:48.200> descent," + }, + { + "start": 528.79, + "duration": 0.0, + "text": "a process called gradient descent," + }, + { + "start": 528.8, + "duration": 0.0, + "text": "a process called gradient descent, right?<00:08:49.280> Stochastic<00:08:49.760> gradient<00:08:50.080> descent.<00:08:51.040> Um" + }, + { + "start": 531.87, + "duration": 0.0, + "text": "right? Stochastic gradient descent. Um" + }, + { + "start": 531.88, + "duration": 0.0, + "text": "right? Stochastic gradient descent. Um I'll<00:08:52.000> also<00:08:52.240> get<00:08:52.360> to<00:08:52.400> that<00:08:52.600> in<00:08:52.680> a<00:08:52.720> second," + }, + { + "start": 533.07, + "duration": 0.0, + "text": "I'll also get to that in a second," + }, + { + "start": 533.08, + "duration": 0.0, + "text": "I'll also get to that in a second, right?<00:08:53.720> But,<00:08:53.839> fundamentally,<00:08:54.360> this<00:08:54.520> is<00:08:54.600> just" + }, + { + "start": 534.71, + "duration": 0.0, + "text": "right? But, fundamentally, this is just" + }, + { + "start": 534.72, + "duration": 0.0, + "text": "right? But, fundamentally, this is just a<00:08:54.800> diagram<00:08:55.280> for<00:08:55.400> this<00:08:55.600> expression<00:08:56.080> or<00:08:56.160> that" + }, + { + "start": 536.31, + "duration": 0.0, + "text": "a diagram for this expression or that" + }, + { + "start": 536.32, + "duration": 0.0, + "text": "a diagram for this expression or that code,<00:08:56.720> right?<00:08:57.200> When<00:08:57.360> you<00:08:57.480> see<00:08:57.680> this<00:08:57.880> diagram," + }, + { + "start": 538.47, + "duration": 0.0, + "text": "code, right? When you see this diagram," + }, + { + "start": 538.48, + "duration": 0.0, + "text": "code, right? When you see this diagram, it<00:08:58.680> is<00:08:58.839> simply<00:08:59.560> that<00:08:59.800> code,<00:09:00.360> right?<00:09:01.160> Um" + }, + { + "start": 542.15, + "duration": 0.0, + "text": "it is simply that code, right? Um" + }, + { + "start": 542.16, + "duration": 0.0, + "text": "it is simply that code, right? Um that's<00:09:02.320> what<00:09:02.400> it's<00:09:02.520> intending<00:09:02.880> to<00:09:02.960> describe," + }, + { + "start": 543.43, + "duration": 0.0, + "text": "that's what it's intending to describe," + }, + { + "start": 543.44, + "duration": 0.0, + "text": "that's what it's intending to describe, right?<00:09:04.240> And<00:09:04.360> so,<00:09:04.440> just<00:09:04.600> to<00:09:04.680> kind<00:09:04.800> of<00:09:04.880> get<00:09:05.000> at" + }, + { + "start": 545.15, + "duration": 0.0, + "text": "right? And so, just to kind of get at" + }, + { + "start": 545.16, + "duration": 0.0, + "text": "right? And so, just to kind of get at like<00:09:05.360> what<00:09:05.520> the<00:09:05.640> act<00:09:05.839> what's<00:09:06.000> going<00:09:06.160> on<00:09:06.240> with" + }, + { + "start": 546.31, + "duration": 0.0, + "text": "like what the act what's going on with" + }, + { + "start": 546.32, + "duration": 0.0, + "text": "like what the act what's going on with the<00:09:06.400> activation<00:09:06.839> function,<00:09:07.160> you<00:09:07.240> know,<00:09:07.360> why" + }, + { + "start": 547.55, + "duration": 0.0, + "text": "the activation function, you know, why" + }, + { + "start": 547.56, + "duration": 0.0, + "text": "the activation function, you know, why do<00:09:07.680> they<00:09:08.160> pass<00:09:08.400> it<00:09:08.480> through<00:09:08.720> sine,<00:09:09.240> right?" + }, + { + "start": 550.11, + "duration": 0.0, + "text": "do they pass it through sine, right?" + }, + { + "start": 550.12, + "duration": 0.0, + "text": "do they pass it through sine, right? Each<00:09:10.400> layer<00:09:10.680> of<00:09:10.760> a<00:09:10.800> neural<00:09:11.000> network<00:09:11.280> is" + }, + { + "start": 551.35, + "duration": 0.0, + "text": "Each layer of a neural network is" + }, + { + "start": 551.36, + "duration": 0.0, + "text": "Each layer of a neural network is basically<00:09:12.000> a<00:09:12.080> node<00:09:12.440> in<00:09:12.520> a<00:09:12.600> transform" + }, + { + "start": 552.95, + "duration": 0.0, + "text": "basically a node in a transform" + }, + { + "start": 552.96, + "duration": 0.0, + "text": "basically a node in a transform hierarchy<00:09:13.480> of<00:09:13.600> like<00:09:13.800> n<00:09:13.920> dimensions,<00:09:14.520> right?" + }, + { + "start": 555.71, + "duration": 0.0, + "text": "hierarchy of like n dimensions, right?" + }, + { + "start": 555.72, + "duration": 0.0, + "text": "hierarchy of like n dimensions, right? Um<00:09:15.880> so,<00:09:16.000> if<00:09:16.080> you're<00:09:16.200> a<00:09:16.240> game<00:09:16.440> developer,<00:09:16.960> which" + }, + { + "start": 557.99, + "duration": 0.0, + "text": "Um so, if you're a game developer, which" + }, + { + "start": 558.0, + "duration": 0.0, + "text": "Um so, if you're a game developer, which you're<00:09:18.080> the<00:09:18.200> target<00:09:18.480> audience<00:09:18.760> of<00:09:18.800> this<00:09:18.920> talk," + }, + { + "start": 559.15, + "duration": 0.0, + "text": "you're the target audience of this talk," + }, + { + "start": 559.16, + "duration": 0.0, + "text": "you're the target audience of this talk, so<00:09:19.240> you<00:09:19.320> should<00:09:19.480> be<00:09:19.560> a<00:09:19.600> game<00:09:19.760> developer." + }, + { + "start": 560.91, + "duration": 0.0, + "text": "so you should be a game developer." + }, + { + "start": 560.92, + "duration": 0.0, + "text": "so you should be a game developer. Um" + }, + { + "start": 561.59, + "duration": 0.0, + "text": "Um" + }, + { + "start": 561.6, + "duration": 0.0, + "text": "Um You're<00:09:21.680> very<00:09:21.800> familiar<00:09:22.160> with<00:09:22.320> like<00:09:22.440> trees<00:09:22.800> of" + }, + { + "start": 562.87, + "duration": 0.0, + "text": "You're very familiar with like trees of" + }, + { + "start": 562.88, + "duration": 0.0, + "text": "You're very familiar with like trees of transforms,<00:09:23.520> right?" + }, + { + "start": 564.79, + "duration": 0.0, + "text": "transforms, right?" + }, + { + "start": 564.8, + "duration": 0.0, + "text": "transforms, right? Um<00:09:25.120> and<00:09:25.280> you<00:09:25.400> know<00:09:25.560> that<00:09:25.760> if<00:09:25.920> you<00:09:26.040> have<00:09:26.560> a" + }, + { + "start": 566.63, + "duration": 0.0, + "text": "Um and you know that if you have a" + }, + { + "start": 566.64, + "duration": 0.0, + "text": "Um and you know that if you have a series<00:09:26.960> of<00:09:27.040> transforms<00:09:27.560> one<00:09:27.760> after<00:09:27.960> the" + }, + { + "start": 568.07, + "duration": 0.0, + "text": "series of transforms one after the" + }, + { + "start": 568.08, + "duration": 0.0, + "text": "series of transforms one after the other,<00:09:28.640> you<00:09:28.760> can<00:09:29.000> multiply<00:09:29.400> them<00:09:29.560> all<00:09:29.800> and<00:09:29.880> it" + }, + { + "start": 569.95, + "duration": 0.0, + "text": "other, you can multiply them all and it" + }, + { + "start": 569.96, + "duration": 0.0, + "text": "other, you can multiply them all and it becomes<00:09:30.280> one<00:09:30.640> linear<00:09:30.839> transform,<00:09:31.280> right?" + }, + { + "start": 571.39, + "duration": 0.0, + "text": "becomes one linear transform, right?" + }, + { + "start": 571.4, + "duration": 0.0, + "text": "becomes one linear transform, right? That's<00:09:31.560> like<00:09:31.720> what<00:09:31.839> it<00:09:31.920> means<00:09:32.120> for<00:09:32.200> something" + }, + { + "start": 572.43, + "duration": 0.0, + "text": "That's like what it means for something" + }, + { + "start": 572.44, + "duration": 0.0, + "text": "That's like what it means for something to<00:09:32.520> be<00:09:32.800> be<00:09:32.920> linear.<00:09:33.880> And<00:09:34.120> when<00:09:34.240> we<00:09:34.320> want" + }, + { + "start": 574.47, + "duration": 0.0, + "text": "to be be linear. And when we want" + }, + { + "start": 574.48, + "duration": 0.0, + "text": "to be be linear. And when we want complicated<00:09:34.920> behavior<00:09:35.440> from<00:09:35.600> these<00:09:35.720> deep" + }, + { + "start": 575.91, + "duration": 0.0, + "text": "complicated behavior from these deep" + }, + { + "start": 575.92, + "duration": 0.0, + "text": "complicated behavior from these deep learning<00:09:36.120> models,<00:09:37.080> uh<00:09:37.160> we<00:09:37.240> don't<00:09:37.400> want<00:09:37.600> it<00:09:37.680> to" + }, + { + "start": 577.71, + "duration": 0.0, + "text": "learning models, uh we don't want it to" + }, + { + "start": 577.72, + "duration": 0.0, + "text": "learning models, uh we don't want it to be<00:09:37.840> reducible<00:09:38.320> like<00:09:38.480> that<00:09:38.680> to<00:09:38.800> one<00:09:39.080> linear<00:09:39.880> uh" + }, + { + "start": 580.07, + "duration": 0.0, + "text": "be reducible like that to one linear uh" + }, + { + "start": 580.08, + "duration": 0.0, + "text": "be reducible like that to one linear uh translation<00:09:40.680> and<00:09:40.880> and<00:09:41.000> rotation<00:09:41.440> whatever," + }, + { + "start": 581.75, + "duration": 0.0, + "text": "translation and and rotation whatever," + }, + { + "start": 581.76, + "duration": 0.0, + "text": "translation and and rotation whatever, right?<00:09:42.360> So,<00:09:42.480> you<00:09:42.560> basically<00:09:43.360> think<00:09:43.520> about<00:09:43.680> it" + }, + { + "start": 583.75, + "duration": 0.0, + "text": "right? So, you basically think about it" + }, + { + "start": 583.76, + "duration": 0.0, + "text": "right? So, you basically think about it as<00:09:43.840> like<00:09:44.080> warping<00:09:44.520> the<00:09:44.600> space" + }, + { + "start": 585.79, + "duration": 0.0, + "text": "as like warping the space" + }, + { + "start": 585.8, + "duration": 0.0, + "text": "as like warping the space such<00:09:46.000> that<00:09:46.520> the<00:09:46.920> translation<00:09:47.480> and<00:09:47.640> offset<00:09:47.960> you" + }, + { + "start": 588.03, + "duration": 0.0, + "text": "such that the translation and offset you" + }, + { + "start": 588.04, + "duration": 0.0, + "text": "such that the translation and offset you do<00:09:48.800> kind<00:09:49.000> of<00:09:49.120> like<00:09:49.360> it<00:09:49.480> it<00:09:49.840> enables<00:09:50.240> you<00:09:50.360> to" + }, + { + "start": 590.47, + "duration": 0.0, + "text": "do kind of like it it enables you to" + }, + { + "start": 590.48, + "duration": 0.0, + "text": "do kind of like it it enables you to basically<00:09:51.040> draw<00:09:51.320> a<00:09:51.360> much<00:09:51.680> more<00:09:52.480> richer" + }, + { + "start": 592.75, + "duration": 0.0, + "text": "basically draw a much more richer" + }, + { + "start": 592.76, + "duration": 0.0, + "text": "basically draw a much more richer contour<00:09:53.240> around<00:09:53.600> the<00:09:53.680> data,<00:09:54.280> right?<00:09:55.080> Um" + }, + { + "start": 596.35, + "duration": 0.0, + "text": "contour around the data, right? Um" + }, + { + "start": 596.36, + "duration": 0.0, + "text": "contour around the data, right? Um And<00:09:56.560> so,<00:09:56.880> just<00:09:57.120> as<00:09:57.240> a<00:09:57.280> as<00:09:57.520> as<00:09:57.680> another<00:09:57.920> metaphor" + }, + { + "start": 598.23, + "duration": 0.0, + "text": "And so, just as a as as another metaphor" + }, + { + "start": 598.24, + "duration": 0.0, + "text": "And so, just as a as as another metaphor to<00:09:58.320> help<00:09:58.680> you<00:09:58.920> as<00:09:59.040> a<00:09:59.080> game<00:09:59.280> developer" + }, + { + "start": 599.59, + "duration": 0.0, + "text": "to help you as a game developer" + }, + { + "start": 599.6, + "duration": 0.0, + "text": "to help you as a game developer understand<00:09:59.920> what's<00:10:00.040> going<00:10:00.240> on<00:10:00.360> here,<00:10:00.760> right?" + }, + { + "start": 601.47, + "duration": 0.0, + "text": "understand what's going on here, right?" + }, + { + "start": 601.48, + "duration": 0.0, + "text": "understand what's going on here, right? Both<00:10:01.880> that<00:10:02.160> neural<00:10:02.360> network<00:10:02.760> and<00:10:02.880> that<00:10:03.040> shader" + }, + { + "start": 603.27, + "duration": 0.0, + "text": "Both that neural network and that shader" + }, + { + "start": 603.28, + "duration": 0.0, + "text": "Both that neural network and that shader graph,<00:10:03.520> these<00:10:03.680> are<00:10:03.760> both<00:10:03.920> compute<00:10:04.320> graphs," + }, + { + "start": 604.87, + "duration": 0.0, + "text": "graph, these are both compute graphs," + }, + { + "start": 604.88, + "duration": 0.0, + "text": "graph, these are both compute graphs, right?<00:10:05.600> There's<00:10:05.800> no<00:10:05.920> branching.<00:10:06.560> Information" + }, + { + "start": 607.03, + "duration": 0.0, + "text": "right? There's no branching. Information" + }, + { + "start": 607.04, + "duration": 0.0, + "text": "right? There's no branching. Information is<00:10:07.160> flowing<00:10:07.440> in<00:10:07.600> from<00:10:07.760> the<00:10:07.840> left<00:10:08.200> and<00:10:08.360> it's" + }, + { + "start": 608.55, + "duration": 0.0, + "text": "is flowing in from the left and it's" + }, + { + "start": 608.56, + "duration": 0.0, + "text": "is flowing in from the left and it's going<00:10:08.840> out<00:10:09.040> the<00:10:09.160> right.<00:10:09.920> Um" + }, + { + "start": 611.27, + "duration": 0.0, + "text": "going out the right. Um" + }, + { + "start": 611.28, + "duration": 0.0, + "text": "going out the right. Um A<00:10:11.360> shader<00:10:11.680> operates<00:10:12.000> on<00:10:12.120> vertices<00:10:12.600> while<00:10:12.720> a" + }, + { + "start": 612.75, + "duration": 0.0, + "text": "A shader operates on vertices while a" + }, + { + "start": 612.76, + "duration": 0.0, + "text": "A shader operates on vertices while a neural<00:10:12.920> network<00:10:13.240> operates<00:10:13.720> on<00:10:13.800> an<00:10:13.920> image<00:10:14.280> or" + }, + { + "start": 614.35, + "duration": 0.0, + "text": "neural network operates on an image or" + }, + { + "start": 614.36, + "duration": 0.0, + "text": "neural network operates on an image or some<00:10:14.520> state<00:10:14.840> or<00:10:14.920> anything<00:10:15.240> you<00:10:15.320> want,<00:10:15.720> right?" + }, + { + "start": 616.63, + "duration": 0.0, + "text": "some state or anything you want, right?" + }, + { + "start": 616.64, + "duration": 0.0, + "text": "some state or anything you want, right? And<00:10:16.920> in<00:10:17.040> the<00:10:17.120> case<00:10:17.360> of<00:10:17.440> game<00:10:17.600> development,<00:10:18.120> an" + }, + { + "start": 618.23, + "duration": 0.0, + "text": "And in the case of game development, an" + }, + { + "start": 618.24, + "duration": 0.0, + "text": "And in the case of game development, an artist<00:10:18.520> might<00:10:18.640> be<00:10:18.760> manually<00:10:19.080> tuning<00:10:19.480> the" + }, + { + "start": 619.59, + "duration": 0.0, + "text": "artist might be manually tuning the" + }, + { + "start": 619.6, + "duration": 0.0, + "text": "artist might be manually tuning the inputs<00:10:19.960> to<00:10:20.040> that<00:10:20.160> shader<00:10:20.560> and<00:10:20.880> making<00:10:21.160> it<00:10:21.280> look" + }, + { + "start": 621.43, + "duration": 0.0, + "text": "inputs to that shader and making it look" + }, + { + "start": 621.44, + "duration": 0.0, + "text": "inputs to that shader and making it look good,<00:10:21.720> right?<00:10:22.240> They<00:10:22.360> might<00:10:22.480> be<00:10:22.600> nudging,<00:10:23.040> you" + }, + { + "start": 623.11, + "duration": 0.0, + "text": "good, right? They might be nudging, you" + }, + { + "start": 623.12, + "duration": 0.0, + "text": "good, right? They might be nudging, you know,<00:10:23.200> the<00:10:23.320> color<00:10:23.600> and<00:10:23.680> trying<00:10:23.840> to<00:10:23.920> make<00:10:24.040> the" + }, + { + "start": 624.11, + "duration": 0.0, + "text": "know, the color and trying to make the" + }, + { + "start": 624.12, + "duration": 0.0, + "text": "know, the color and trying to make the output<00:10:24.400> look<00:10:24.520> good<00:10:24.720> or<00:10:24.800> some<00:10:24.960> other" + }, + { + "start": 625.23, + "duration": 0.0, + "text": "output look good or some other" + }, + { + "start": 625.24, + "duration": 0.0, + "text": "output look good or some other parameter.<00:10:26.120> In<00:10:26.240> the<00:10:26.320> case<00:10:26.600> of<00:10:26.800> deep<00:10:27.000> learning," + }, + { + "start": 627.47, + "duration": 0.0, + "text": "parameter. In the case of deep learning," + }, + { + "start": 627.48, + "duration": 0.0, + "text": "parameter. In the case of deep learning, that's<00:10:27.720> actually<00:10:28.160> replaced<00:10:28.560> by<00:10:28.680> a<00:10:28.720> machine," + }, + { + "start": 629.27, + "duration": 0.0, + "text": "that's actually replaced by a machine," + }, + { + "start": 629.28, + "duration": 0.0, + "text": "that's actually replaced by a machine, right?<00:10:29.880> Which<00:10:30.080> is<00:10:30.400> which<00:10:30.600> direction" + }, + { + "start": 630.95, + "duration": 0.0, + "text": "right? Which is which direction" + }, + { + "start": 630.96, + "duration": 0.0, + "text": "right? Which is which direction minimizes<00:10:31.480> error<00:10:31.720> on<00:10:31.840> training<00:10:32.200> for<00:10:32.360> each<00:10:32.520> one" + }, + { + "start": 632.63, + "duration": 0.0, + "text": "minimizes error on training for each one" + }, + { + "start": 632.64, + "duration": 0.0, + "text": "minimizes error on training for each one of<00:10:32.720> these<00:10:32.840> little<00:10:33.240> uh<00:10:33.360> parameters,<00:10:33.840> right?" + }, + { + "start": 633.99, + "duration": 0.0, + "text": "of these little uh parameters, right?" + }, + { + "start": 634.0, + "duration": 0.0, + "text": "of these little uh parameters, right? Each<00:10:34.160> one<00:10:34.240> of<00:10:34.320> those<00:10:34.440> weights<00:10:34.680> and<00:10:34.760> all<00:10:34.840> those" + }, + { + "start": 635.03, + "duration": 0.0, + "text": "Each one of those weights and all those" + }, + { + "start": 635.04, + "duration": 0.0, + "text": "Each one of those weights and all those biases,<00:10:36.000> it<00:10:36.160> knows<00:10:36.520> which<00:10:36.720> direction<00:10:37.120> it" + }, + { + "start": 637.23, + "duration": 0.0, + "text": "biases, it knows which direction it" + }, + { + "start": 637.24, + "duration": 0.0, + "text": "biases, it knows which direction it should<00:10:37.400> go<00:10:37.600> in<00:10:37.960> assuming<00:10:38.320> all<00:10:38.400> the<00:10:38.520> other" + }, + { + "start": 638.71, + "duration": 0.0, + "text": "should go in assuming all the other" + }, + { + "start": 638.72, + "duration": 0.0, + "text": "should go in assuming all the other parameters<00:10:39.160> were,<00:10:39.360> you<00:10:39.440> know,<00:10:40.000> not<00:10:40.200> moving." + }, + { + "start": 640.71, + "duration": 0.0, + "text": "parameters were, you know, not moving." + }, + { + "start": 640.72, + "duration": 0.0, + "text": "parameters were, you know, not moving. It<00:10:40.800> knows<00:10:40.920> approximately<00:10:41.360> which<00:10:41.480> direction" + }, + { + "start": 641.75, + "duration": 0.0, + "text": "It knows approximately which direction" + }, + { + "start": 641.76, + "duration": 0.0, + "text": "It knows approximately which direction to<00:10:41.839> go<00:10:42.000> in<00:10:42.120> to<00:10:42.520> get<00:10:42.720> better<00:10:42.920> at<00:10:42.960> the<00:10:43.080> task.<00:10:43.839> And" + }, + { + "start": 643.91, + "duration": 0.0, + "text": "to go in to get better at the task. And" + }, + { + "start": 643.92, + "duration": 0.0, + "text": "to go in to get better at the task. And then<00:10:44.040> it<00:10:44.080> just<00:10:44.200> kind<00:10:44.360> of<00:10:44.400> repeatedly<00:10:44.760> does" + }, + { + "start": 644.95, + "duration": 0.0, + "text": "then it just kind of repeatedly does" + }, + { + "start": 644.96, + "duration": 0.0, + "text": "then it just kind of repeatedly does that<00:10:45.120> thousands<00:10:45.480> and<00:10:45.600> millions<00:10:45.880> of<00:10:45.960> times." + }, + { + "start": 647.19, + "duration": 0.0, + "text": "that thousands and millions of times." + }, + { + "start": 647.2, + "duration": 0.0, + "text": "that thousands and millions of times. Um<00:10:47.920> so,<00:10:48.040> there<00:10:48.200> is<00:10:48.360> a<00:10:48.400> code<00:10:48.640> base<00:10:48.920> for<00:10:49.000> this" + }, + { + "start": 649.15, + "duration": 0.0, + "text": "Um so, there is a code base for this" + }, + { + "start": 649.16, + "duration": 0.0, + "text": "Um so, there is a code base for this talk.<00:10:49.480> It's<00:10:49.640> here.<00:10:50.320> github.com/krickey/ai" + }, + { + "start": 652.829, + "duration": 0.0, + "text": "talk. It's here. github.com/krickey/ai" + }, + { + "start": 652.839, + "duration": 0.0, + "text": "talk. It's here. github.com/krickey/ai Um<00:10:53.240> and<00:10:53.360> to<00:10:53.440> just<00:10:53.600> get<00:10:53.720> at<00:10:53.800> what<00:10:53.920> this<00:10:54.040> code" + }, + { + "start": 654.19, + "duration": 0.0, + "text": "Um and to just get at what this code" + }, + { + "start": 654.2, + "duration": 0.0, + "text": "Um and to just get at what this code base<00:10:54.360> is<00:10:54.480> supposed<00:10:54.720> to<00:10:54.800> be<00:10:55.240> for<00:10:55.360> a<00:10:55.400> second," + }, + { + "start": 655.75, + "duration": 0.0, + "text": "base is supposed to be for a second," + }, + { + "start": 655.76, + "duration": 0.0, + "text": "base is supposed to be for a second, right?<00:10:56.360> Deep<00:10:56.600> learning<00:10:56.920> libraries<00:10:57.480> or,<00:10:57.600> you" + }, + { + "start": 657.67, + "duration": 0.0, + "text": "right? Deep learning libraries or, you" + }, + { + "start": 657.68, + "duration": 0.0, + "text": "right? Deep learning libraries or, you know,<00:10:57.839> autograd,<00:10:58.800> they<00:10:58.880> let<00:10:59.080> you<00:10:59.160> perform" + }, + { + "start": 659.63, + "duration": 0.0, + "text": "know, autograd, they let you perform" + }, + { + "start": 659.64, + "duration": 0.0, + "text": "know, autograd, they let you perform some<00:10:59.880> arbitrary,<00:11:00.600> you<00:11:00.680> know,<00:11:00.839> computations," + }, + { + "start": 661.59, + "duration": 0.0, + "text": "some arbitrary, you know, computations," + }, + { + "start": 661.6, + "duration": 0.0, + "text": "some arbitrary, you know, computations, right?<00:11:02.080> Might<00:11:02.240> be<00:11:02.320> a<00:11:02.400> matrix<00:11:02.760> multiplication," + }, + { + "start": 663.39, + "duration": 0.0, + "text": "right? Might be a matrix multiplication," + }, + { + "start": 663.4, + "duration": 0.0, + "text": "right? Might be a matrix multiplication, you<00:11:03.480> know,<00:11:03.560> it<00:11:03.640> might<00:11:03.800> be<00:11:04.120> x<00:11:04.320> squared<00:11:04.600> or<00:11:04.680> one" + }, + { + "start": 664.79, + "duration": 0.0, + "text": "you know, it might be x squared or one" + }, + { + "start": 664.8, + "duration": 0.0, + "text": "you know, it might be x squared or one of<00:11:04.839> those<00:11:05.120> activation<00:11:05.600> functions<00:11:05.839> we<00:11:05.920> talked" + }, + { + "start": 666.15, + "duration": 0.0, + "text": "of those activation functions we talked" + }, + { + "start": 666.16, + "duration": 0.0, + "text": "of those activation functions we talked about.<00:11:07.120> Um<00:11:07.360> and<00:11:07.480> then<00:11:07.720> find<00:11:08.120> out<00:11:08.680> how<00:11:08.920> you<00:11:09.040> need" + }, + { + "start": 669.19, + "duration": 0.0, + "text": "about. Um and then find out how you need" + }, + { + "start": 669.2, + "duration": 0.0, + "text": "about. Um and then find out how you need to<00:11:09.640> turn<00:11:10.080> or<00:11:10.280> or<00:11:10.480> or<00:11:10.680> adjust<00:11:11.200> those" + }, + { + "start": 671.39, + "duration": 0.0, + "text": "to turn or or or adjust those" + }, + { + "start": 671.4, + "duration": 0.0, + "text": "to turn or or or adjust those parameters,<00:11:12.000> right?<00:11:12.160> The<00:11:12.240> weights<00:11:12.480> and<00:11:12.560> the" + }, + { + "start": 672.59, + "duration": 0.0, + "text": "parameters, right? The weights and the" + }, + { + "start": 672.6, + "duration": 0.0, + "text": "parameters, right? The weights and the biases.<00:11:13.720> It<00:11:14.320> automatically<00:11:15.000> figures<00:11:15.440> out" + }, + { + "start": 675.87, + "duration": 0.0, + "text": "biases. It automatically figures out" + }, + { + "start": 675.88, + "duration": 0.0, + "text": "biases. It automatically figures out what<00:11:16.040> the<00:11:16.120> expression<00:11:16.640> is<00:11:17.080> to<00:11:17.960> make<00:11:18.240> those" + }, + { + "start": 678.63, + "duration": 0.0, + "text": "what the expression is to make those" + }, + { + "start": 678.64, + "duration": 0.0, + "text": "what the expression is to make those move<00:11:18.880> in<00:11:18.960> a<00:11:19.000> direction<00:11:19.400> that<00:11:19.600> makes<00:11:19.839> the<00:11:19.920> model" + }, + { + "start": 680.11, + "duration": 0.0, + "text": "move in a direction that makes the model" + }, + { + "start": 680.12, + "duration": 0.0, + "text": "move in a direction that makes the model get<00:11:20.280> better<00:11:20.480> at<00:11:20.560> the<00:11:20.640> task,<00:11:21.200> right?" + }, + { + "start": 682.35, + "duration": 0.0, + "text": "get better at the task, right?" + }, + { + "start": 682.36, + "duration": 0.0, + "text": "get better at the task, right? And<00:11:23.080> that's<00:11:23.280> basically<00:11:23.600> that" + }, + { + "start": 684.39, + "duration": 0.0, + "text": "And that's basically that" + }, + { + "start": 684.4, + "duration": 0.0, + "text": "And that's basically that the<00:11:24.520> what<00:11:24.800> these<00:11:25.040> deep<00:11:25.200> learning<00:11:25.400> libraries" + }, + { + "start": 685.75, + "duration": 0.0, + "text": "the what these deep learning libraries" + }, + { + "start": 685.76, + "duration": 0.0, + "text": "the what these deep learning libraries are<00:11:25.880> supposed<00:11:26.120> to<00:11:26.160> be<00:11:26.240> in<00:11:26.280> a<00:11:26.320> nutshell,<00:11:26.680> you" + }, + { + "start": 686.71, + "duration": 0.0, + "text": "are supposed to be in a nutshell, you" + }, + { + "start": 686.72, + "duration": 0.0, + "text": "are supposed to be in a nutshell, you know,<00:11:26.839> TensorFlow,<00:11:27.320> PyTorch.<00:11:28.040> Um" + }, + { + "start": 689.51, + "duration": 0.0, + "text": "know, TensorFlow, PyTorch. Um" + }, + { + "start": 689.52, + "duration": 0.0, + "text": "know, TensorFlow, PyTorch. Um And<00:11:30.120> you<00:11:30.200> know," + }, + { + "start": 691.15, + "duration": 0.0, + "text": "And you know," + }, + { + "start": 691.16, + "duration": 0.0, + "text": "And you know, just<00:11:31.360> as<00:11:31.480> a<00:11:31.600> little<00:11:31.839> preliminary<00:11:32.200> for<00:11:32.320> this" + }, + { + "start": 692.43, + "duration": 0.0, + "text": "just as a little preliminary for this" + }, + { + "start": 692.44, + "duration": 0.0, + "text": "just as a little preliminary for this code<00:11:32.680> base,<00:11:32.920> there's<00:11:33.040> a<00:11:33.080> lot<00:11:33.200> of<00:11:33.280> like<00:11:33.440> tensors" + }, + { + "start": 693.87, + "duration": 0.0, + "text": "code base, there's a lot of like tensors" + }, + { + "start": 693.88, + "duration": 0.0, + "text": "code base, there's a lot of like tensors in<00:11:34.040> it.<00:11:34.440> Um" + }, + { + "start": 695.11, + "duration": 0.0, + "text": "in it. Um" + }, + { + "start": 695.12, + "duration": 0.0, + "text": "in it. Um but<00:11:35.240> really<00:11:35.440> tensors<00:11:35.800> are<00:11:35.840> just<00:11:36.040> arrays<00:11:36.360> of" + }, + { + "start": 696.47, + "duration": 0.0, + "text": "but really tensors are just arrays of" + }, + { + "start": 696.48, + "duration": 0.0, + "text": "but really tensors are just arrays of matrices.<00:11:37.360> Um" + }, + { + "start": 698.07, + "duration": 0.0, + "text": "matrices. Um" + }, + { + "start": 698.08, + "duration": 0.0, + "text": "matrices. Um like<00:11:38.240> you<00:11:38.320> might<00:11:38.480> want<00:11:38.600> to<00:11:38.760> talk<00:11:39.000> about<00:11:39.240> like" + }, + { + "start": 699.47, + "duration": 0.0, + "text": "like you might want to talk about like" + }, + { + "start": 699.48, + "duration": 0.0, + "text": "like you might want to talk about like oh<00:11:40.320> uh" + }, + { + "start": 701.39, + "duration": 0.0, + "text": "oh uh" + }, + { + "start": 701.4, + "duration": 0.0, + "text": "oh uh We're<00:11:42.120> we're<00:11:42.240> feeding<00:11:42.560> in<00:11:42.680> many<00:11:42.960> examples" + }, + { + "start": 703.59, + "duration": 0.0, + "text": "We're we're feeding in many examples" + }, + { + "start": 703.6, + "duration": 0.0, + "text": "We're we're feeding in many examples into<00:11:44.080> this<00:11:44.360> deep<00:11:44.640> learning<00:11:44.840> model.<00:11:45.480> And<00:11:45.680> so<00:11:45.760> we" + }, + { + "start": 705.87, + "duration": 0.0, + "text": "into this deep learning model. And so we" + }, + { + "start": 705.88, + "duration": 0.0, + "text": "into this deep learning model. And so we want<00:11:46.080> like<00:11:46.200> a<00:11:46.240> third<00:11:46.480> dimension<00:11:46.840> to<00:11:47.000> it<00:11:47.160> where" + }, + { + "start": 707.27, + "duration": 0.0, + "text": "want like a third dimension to it where" + }, + { + "start": 707.28, + "duration": 0.0, + "text": "want like a third dimension to it where it's<00:11:47.360> just<00:11:47.560> an<00:11:47.680> array<00:11:48.040> of<00:11:48.280> the<00:11:48.360> matrices<00:11:48.800> that" + }, + { + "start": 708.91, + "duration": 0.0, + "text": "it's just an array of the matrices that" + }, + { + "start": 708.92, + "duration": 0.0, + "text": "it's just an array of the matrices that we're<00:11:49.000> feeding<00:11:49.240> into<00:11:49.440> the<00:11:49.480> model,<00:11:50.280> right?<00:11:51.120> Um" + }, + { + "start": 711.95, + "duration": 0.0, + "text": "we're feeding into the model, right? Um" + }, + { + "start": 711.96, + "duration": 0.0, + "text": "we're feeding into the model, right? Um And<00:11:52.160> so<00:11:52.280> my<00:11:52.880> idea<00:11:53.280> with<00:11:53.400> this<00:11:53.520> code<00:11:53.720> base<00:11:54.520> is<00:11:54.600> it" + }, + { + "start": 714.67, + "duration": 0.0, + "text": "And so my idea with this code base is it" + }, + { + "start": 714.68, + "duration": 0.0, + "text": "And so my idea with this code base is it would<00:11:54.760> be<00:11:54.880> really<00:11:55.120> cool<00:11:55.480> if<00:11:55.600> instead<00:11:55.920> of<00:11:56.000> me" + }, + { + "start": 716.11, + "duration": 0.0, + "text": "would be really cool if instead of me" + }, + { + "start": 716.12, + "duration": 0.0, + "text": "would be really cool if instead of me having<00:11:56.360> a<00:11:56.720> you<00:11:56.800> know,<00:11:56.880> have<00:11:57.080> a<00:11:57.160> Python" + }, + { + "start": 717.51, + "duration": 0.0, + "text": "having a you know, have a Python" + }, + { + "start": 717.52, + "duration": 0.0, + "text": "having a you know, have a Python interpreter,<00:11:58.360> there<00:11:58.440> was<00:11:58.600> one<00:11:58.800> C<00:11:59.000> file<00:11:59.240> called" + }, + { + "start": 719.51, + "duration": 0.0, + "text": "interpreter, there was one C file called" + }, + { + "start": 719.52, + "duration": 0.0, + "text": "interpreter, there was one C file called ai.c,<00:12:00.560> right?<00:12:01.160> We<00:12:01.240> had<00:12:01.400> CPU<00:12:01.880> and<00:12:02.000> GPU<00:12:02.480> deep" + }, + { + "start": 722.71, + "duration": 0.0, + "text": "ai.c, right? We had CPU and GPU deep" + }, + { + "start": 722.72, + "duration": 0.0, + "text": "ai.c, right? We had CPU and GPU deep learning<00:12:03.160> computer<00:12:03.480> vision<00:12:03.720> algorithms<00:12:04.080> and" + }, + { + "start": 724.15, + "duration": 0.0, + "text": "learning computer vision algorithms and" + }, + { + "start": 724.16, + "duration": 0.0, + "text": "learning computer vision algorithms and implementations,<00:12:04.960> right?<00:12:05.600> I<00:12:05.680> would<00:12:05.840> want<00:12:05.960> it" + }, + { + "start": 725.99, + "duration": 0.0, + "text": "implementations, right? I would want it" + }, + { + "start": 726.0, + "duration": 0.0, + "text": "implementations, right? I would want it to<00:12:06.080> be<00:12:06.120> able<00:12:06.320> to<00:12:06.640> execute<00:12:07.720> LLM<00:12:08.480> local<00:12:08.800> LLMs," + }, + { + "start": 729.31, + "duration": 0.0, + "text": "to be able to execute LLM local LLMs," + }, + { + "start": 729.32, + "duration": 0.0, + "text": "to be able to execute LLM local LLMs, right?<00:12:09.920> I<00:12:09.960> want<00:12:10.120> you<00:12:10.160> to<00:12:10.240> be<00:12:10.280> able<00:12:10.400> to" + }, + { + "start": 730.47, + "duration": 0.0, + "text": "right? I want you to be able to" + }, + { + "start": 730.48, + "duration": 0.0, + "text": "right? I want you to be able to fine-tune<00:12:10.880> those<00:12:11.040> LLMs,<00:12:11.520> not<00:12:11.720> just<00:12:12.000> the" + }, + { + "start": 732.07, + "duration": 0.0, + "text": "fine-tune those LLMs, not just the" + }, + { + "start": 732.08, + "duration": 0.0, + "text": "fine-tune those LLMs, not just the inference.<00:12:13.080> Um<00:12:13.600> I<00:12:13.640> would<00:12:13.760> want<00:12:13.880> to<00:12:13.960> hardware" + }, + { + "start": 734.23, + "duration": 0.0, + "text": "inference. Um I would want to hardware" + }, + { + "start": 734.24, + "duration": 0.0, + "text": "inference. Um I would want to hardware accelerate<00:12:14.560> with<00:12:14.680> something<00:12:14.920> portable<00:12:15.320> like" + }, + { + "start": 735.43, + "duration": 0.0, + "text": "accelerate with something portable like" + }, + { + "start": 735.44, + "duration": 0.0, + "text": "accelerate with something portable like Sokol,<00:12:16.120> right?<00:12:16.680> I<00:12:16.720> have<00:12:16.840> some<00:12:16.960> computer" + }, + { + "start": 737.27, + "duration": 0.0, + "text": "Sokol, right? I have some computer" + }, + { + "start": 737.28, + "duration": 0.0, + "text": "Sokol, right? I have some computer vision<00:12:17.480> algorithms<00:12:17.800> as<00:12:17.920> well<00:12:18.080> cuz<00:12:18.200> those<00:12:18.320> are" + }, + { + "start": 738.43, + "duration": 0.0, + "text": "vision algorithms as well cuz those are" + }, + { + "start": 738.44, + "duration": 0.0, + "text": "vision algorithms as well cuz those are fun." + }, + { + "start": 739.43, + "duration": 0.0, + "text": "fun." + }, + { + "start": 739.44, + "duration": 0.0, + "text": "fun. Um<00:12:20.480> and<00:12:20.560> unfortunately,<00:12:21.000> this<00:12:21.200> this<00:12:21.480> this" + }, + { + "start": 741.71, + "duration": 0.0, + "text": "Um and unfortunately, this this this" + }, + { + "start": 741.72, + "duration": 0.0, + "text": "Um and unfortunately, this this this talk<00:12:21.960> code<00:12:22.120> base<00:12:22.320> right<00:12:22.440> now<00:12:22.680> is<00:12:22.800> like<00:12:23.360> super" + }, + { + "start": 743.63, + "duration": 0.0, + "text": "talk code base right now is like super" + }, + { + "start": 743.64, + "duration": 0.0, + "text": "talk code base right now is like super super<00:12:23.920> not<00:12:24.200> that,<00:12:24.480> right?<00:12:24.720> It<00:12:24.800> has<00:12:24.920> a" + }, + { + "start": 744.95, + "duration": 0.0, + "text": "super not that, right? It has a" + }, + { + "start": 744.96, + "duration": 0.0, + "text": "super not that, right? It has a hand-coded<00:12:25.520> gradient<00:12:25.880> calculation<00:12:26.400> for<00:12:26.520> a" + }, + { + "start": 746.55, + "duration": 0.0, + "text": "hand-coded gradient calculation for a" + }, + { + "start": 746.56, + "duration": 0.0, + "text": "hand-coded gradient calculation for a two-layer<00:12:27.040> perceptron<00:12:27.560> that<00:12:27.680> classifies" + }, + { + "start": 748.19, + "duration": 0.0, + "text": "two-layer perceptron that classifies" + }, + { + "start": 748.2, + "duration": 0.0, + "text": "two-layer perceptron that classifies digits.<00:12:28.920> It's<00:12:29.120> super<00:12:29.360> primitive<00:12:29.760> and<00:12:29.880> it's" + }, + { + "start": 749.99, + "duration": 0.0, + "text": "digits. It's super primitive and it's" + }, + { + "start": 750.0, + "duration": 0.0, + "text": "digits. It's super primitive and it's like<00:12:30.200> 1970s<00:12:30.880> stuff.<00:12:31.640> There's<00:12:31.800> no<00:12:31.920> working" + }, + { + "start": 752.27, + "duration": 0.0, + "text": "like 1970s stuff. There's no working" + }, + { + "start": 752.28, + "duration": 0.0, + "text": "like 1970s stuff. There's no working auto<00:12:32.480> grad<00:12:32.760> yet,<00:12:33.000> but<00:12:33.280> I<00:12:33.400> almost<00:12:33.720> have<00:12:33.840> it" + }, + { + "start": 753.95, + "duration": 0.0, + "text": "auto grad yet, but I almost have it" + }, + { + "start": 753.96, + "duration": 0.0, + "text": "auto grad yet, but I almost have it working." + }, + { + "start": 754.91, + "duration": 0.0, + "text": "working." + }, + { + "start": 754.92, + "duration": 0.0, + "text": "working. Um<00:12:35.040> and<00:12:35.120> I<00:12:35.400> I<00:12:35.440> intend<00:12:35.800> to<00:12:35.880> continue<00:12:36.600> working<00:12:36.880> on" + }, + { + "start": 757.03, + "duration": 0.0, + "text": "Um and I I intend to continue working on" + }, + { + "start": 757.04, + "duration": 0.0, + "text": "Um and I I intend to continue working on ai.c<00:12:37.680> cuz<00:12:37.800> it's<00:12:37.920> kind<00:12:38.080> of<00:12:38.160> fun.<00:12:38.640> I'll<00:12:38.720> be" + }, + { + "start": 758.87, + "duration": 0.0, + "text": "ai.c cuz it's kind of fun. I'll be" + }, + { + "start": 758.88, + "duration": 0.0, + "text": "ai.c cuz it's kind of fun. I'll be sending<00:12:39.160> updates<00:12:39.480> in<00:12:39.600> the" + }, + { + "start": 760.43, + "duration": 0.0, + "text": "sending updates in the" + }, + { + "start": 760.44, + "duration": 0.0, + "text": "sending updates in the in<00:12:40.560> the<00:12:40.680> BSC<00:12:41.080> Discord<00:12:41.400> about<00:12:41.600> that." + }, + { + "start": 764.4, + "duration": 0.0, + "text": "Um" + }, + { + "start": 765.829, + "duration": 0.0, + "text": "Um" + }, + { + "start": 765.839, + "duration": 0.0, + "text": "Um but<00:12:46.360> just<00:12:46.600> as<00:12:46.720> an<00:12:46.760> example<00:12:47.160> of<00:12:47.280> what<00:12:47.360> I<00:12:47.400> was" + }, + { + "start": 767.51, + "duration": 0.0, + "text": "but just as an example of what I was" + }, + { + "start": 767.52, + "duration": 0.0, + "text": "but just as an example of what I was trying<00:12:47.760> to<00:12:47.839> do<00:12:48.360> with<00:12:49.040> this<00:12:49.520> little<00:12:49.760> tiny,<00:12:50.120> you" + }, + { + "start": 770.19, + "duration": 0.0, + "text": "trying to do with this little tiny, you" + }, + { + "start": 770.2, + "duration": 0.0, + "text": "trying to do with this little tiny, you know,<00:12:50.440> deep<00:12:50.640> learning<00:12:50.880> library<00:12:51.320> is" + }, + { + "start": 772.95, + "duration": 0.0, + "text": "know, deep learning library is" + }, + { + "start": 772.96, + "duration": 0.0, + "text": "know, deep learning library is each<00:12:53.280> of<00:12:53.360> these<00:12:53.520> videos<00:12:53.960> are<00:12:54.120> videos<00:12:54.440> I<00:12:54.480> got" + }, + { + "start": 774.63, + "duration": 0.0, + "text": "each of these videos are videos I got" + }, + { + "start": 774.64, + "duration": 0.0, + "text": "each of these videos are videos I got from<00:12:54.800> YouTube<00:12:55.200> of<00:12:55.320> some<00:12:55.480> kind<00:12:55.680> of<00:12:55.920> ideally" + }, + { + "start": 776.51, + "duration": 0.0, + "text": "from YouTube of some kind of ideally" + }, + { + "start": 776.52, + "duration": 0.0, + "text": "from YouTube of some kind of ideally animating<00:12:57.120> a<00:12:57.240> scene,<00:12:57.600> right?<00:12:57.800> Where<00:12:57.880> the" + }, + { + "start": 778.23, + "duration": 0.0, + "text": "animating a scene, right? Where the" + }, + { + "start": 778.24, + "duration": 0.0, + "text": "animating a scene, right? Where the camera<00:12:58.520> is<00:12:58.600> stationary.<00:12:59.680> Um<00:13:00.160> the<00:13:00.240> theory<00:13:00.600> is" + }, + { + "start": 780.829, + "duration": 0.0, + "text": "camera is stationary. Um the theory is" + }, + { + "start": 780.839, + "duration": 0.0, + "text": "camera is stationary. Um the theory is you<00:13:00.920> should<00:13:01.080> be<00:13:01.160> able<00:13:01.320> to<00:13:01.400> train<00:13:01.560> a<00:13:01.600> deep" + }, + { + "start": 781.75, + "duration": 0.0, + "text": "you should be able to train a deep" + }, + { + "start": 781.76, + "duration": 0.0, + "text": "you should be able to train a deep learning<00:13:01.960> model<00:13:02.360> to<00:13:02.560> generate<00:13:02.880> these<00:13:03.000> videos" + }, + { + "start": 783.51, + "duration": 0.0, + "text": "learning model to generate these videos" + }, + { + "start": 783.52, + "duration": 0.0, + "text": "learning model to generate these videos from<00:13:03.800> the<00:13:03.920> image.<00:13:04.960> Um" + }, + { + "start": 785.67, + "duration": 0.0, + "text": "from the image. Um" + }, + { + "start": 785.68, + "duration": 0.0, + "text": "from the image. Um you<00:13:05.760> know,<00:13:05.839> the<00:13:05.960> assumption<00:13:06.400> is<00:13:06.520> that<00:13:06.640> the" + }, + { + "start": 786.71, + "duration": 0.0, + "text": "you know, the assumption is that the" + }, + { + "start": 786.72, + "duration": 0.0, + "text": "you know, the assumption is that the behavior<00:13:07.080> of<00:13:07.160> the<00:13:07.240> pixels<00:13:07.640> is<00:13:07.720> relatively" + }, + { + "start": 788.15, + "duration": 0.0, + "text": "behavior of the pixels is relatively" + }, + { + "start": 788.16, + "duration": 0.0, + "text": "behavior of the pixels is relatively simple.<00:13:09.120> Um<00:13:09.440> and<00:13:09.520> so<00:13:09.640> I<00:13:09.680> actually,<00:13:10.440> even" + }, + { + "start": 790.63, + "duration": 0.0, + "text": "simple. Um and so I actually, even" + }, + { + "start": 790.64, + "duration": 0.0, + "text": "simple. Um and so I actually, even though<00:13:10.720> I<00:13:10.800> wasn't<00:13:11.080> able<00:13:11.240> to<00:13:11.400> get<00:13:12.120> this<00:13:12.440> working" + }, + { + "start": 792.75, + "duration": 0.0, + "text": "though I wasn't able to get this working" + }, + { + "start": 792.76, + "duration": 0.0, + "text": "though I wasn't able to get this working in<00:13:12.839> time<00:13:13.040> for<00:13:13.120> the<00:13:13.200> talk<00:13:13.520> with<00:13:14.200> ai.c,<00:13:15.080> I" + }, + { + "start": 797.43, + "duration": 0.0, + "text": "in time for the talk with ai.c, I" + }, + { + "start": 797.44, + "duration": 0.0, + "text": "in time for the talk with ai.c, I Hold<00:13:17.800> on.<00:13:17.920> Sorry,<00:13:18.120> one<00:13:18.280> sec." + }, + { + "start": 800.76, + "duration": 0.0, + "text": "Oh,<00:13:20.920> here<00:13:21.080> we<00:13:21.160> go.<00:13:21.520> Okay.<00:13:22.160> So,<00:13:22.480> I<00:13:23.320> basically" + }, + { + "start": 803.829, + "duration": 0.0, + "text": "Oh, here we go. Okay. So, I basically" + }, + { + "start": 803.839, + "duration": 0.0, + "text": "Oh, here we go. Okay. So, I basically just,<00:13:24.120> you<00:13:24.160> know,<00:13:24.280> let's<00:13:24.480> just<00:13:24.600> try<00:13:24.760> it<00:13:24.839> in" + }, + { + "start": 804.91, + "duration": 0.0, + "text": "just, you know, let's just try it in" + }, + { + "start": 804.92, + "duration": 0.0, + "text": "just, you know, let's just try it in Python<00:13:25.240> really<00:13:25.400> quick<00:13:25.600> to<00:13:25.680> see<00:13:25.800> if<00:13:25.920> this<00:13:26.040> will" + }, + { + "start": 806.11, + "duration": 0.0, + "text": "Python really quick to see if this will" + }, + { + "start": 806.12, + "duration": 0.0, + "text": "Python really quick to see if this will even<00:13:26.280> work<00:13:26.520> at<00:13:26.640> all.<00:13:27.120> This<00:13:27.280> is<00:13:27.360> an<00:13:27.600> 80<00:13:27.840> kilobyte" + }, + { + "start": 808.67, + "duration": 0.0, + "text": "even work at all. This is an 80 kilobyte" + }, + { + "start": 808.68, + "duration": 0.0, + "text": "even work at all. This is an 80 kilobyte of<00:13:28.800> parameters<00:13:29.320> neural<00:13:29.480> network<00:13:30.320> and<00:13:30.560> it's" + }, + { + "start": 810.63, + "duration": 0.0, + "text": "of parameters neural network and it's" + }, + { + "start": 810.64, + "duration": 0.0, + "text": "of parameters neural network and it's garbage,<00:13:31.280> right?<00:13:31.880> This<00:13:32.120> kind<00:13:32.280> of<00:13:32.320> happens<00:13:32.560> a" + }, + { + "start": 812.59, + "duration": 0.0, + "text": "garbage, right? This kind of happens a" + }, + { + "start": 812.6, + "duration": 0.0, + "text": "garbage, right? This kind of happens a lot<00:13:32.760> in<00:13:32.800> deep<00:13:32.960> learning<00:13:33.200> when<00:13:33.280> you<00:13:33.360> just<00:13:33.480> have" + }, + { + "start": 813.55, + "duration": 0.0, + "text": "lot in deep learning when you just have" + }, + { + "start": 813.56, + "duration": 0.0, + "text": "lot in deep learning when you just have a<00:13:33.600> random<00:13:33.880> idea<00:13:34.560> and<00:13:34.680> you<00:13:34.720> just<00:13:34.920> try<00:13:35.200> it.<00:13:35.360> It" + }, + { + "start": 815.43, + "duration": 0.0, + "text": "a random idea and you just try it. It" + }, + { + "start": 815.44, + "duration": 0.0, + "text": "a random idea and you just try it. It makes<00:13:35.600> something<00:13:35.880> really<00:13:36.640> sometimes<00:13:37.040> hellish" + }, + { + "start": 817.47, + "duration": 0.0, + "text": "makes something really sometimes hellish" + }, + { + "start": 817.48, + "duration": 0.0, + "text": "makes something really sometimes hellish like<00:13:38.080> I<00:13:38.160> didn't<00:13:38.280> even<00:13:38.440> know<00:13:38.520> what<00:13:38.600> that<00:13:38.760> is," + }, + { + "start": 818.91, + "duration": 0.0, + "text": "like I didn't even know what that is," + }, + { + "start": 818.92, + "duration": 0.0, + "text": "like I didn't even know what that is, honestly." + }, + { + "start": 820.11, + "duration": 0.0, + "text": "honestly." + }, + { + "start": 820.12, + "duration": 0.0, + "text": "honestly. I<00:13:40.200> think<00:13:40.360> that<00:13:40.560> 80<00:13:40.680> kilobyte<00:13:41.040> is<00:13:41.160> way<00:13:41.280> too" + }, + { + "start": 821.43, + "duration": 0.0, + "text": "I think that 80 kilobyte is way too" + }, + { + "start": 821.44, + "duration": 0.0, + "text": "I think that 80 kilobyte is way too small<00:13:41.760> to<00:13:41.920> encode<00:13:42.280> the<00:13:42.320> dynamics<00:13:42.800> of<00:13:42.880> a<00:13:42.920> waving" + }, + { + "start": 823.19, + "duration": 0.0, + "text": "small to encode the dynamics of a waving" + }, + { + "start": 823.2, + "duration": 0.0, + "text": "small to encode the dynamics of a waving tree,<00:13:43.440> probably.<00:13:43.920> So<00:13:44.040> I<00:13:44.080> would<00:13:44.200> just,<00:13:44.400> you" + }, + { + "start": 824.47, + "duration": 0.0, + "text": "tree, probably. So I would just, you" + }, + { + "start": 824.48, + "duration": 0.0, + "text": "tree, probably. So I would just, you know,<00:13:45.000> 10x<00:13:45.360> it<00:13:45.520> like" + }, + { + "start": 826.91, + "duration": 0.0, + "text": "know, 10x it like" + }, + { + "start": 826.92, + "duration": 0.0, + "text": "know, 10x it like Deep<00:13:47.360> learning<00:13:47.560> is<00:13:47.640> just<00:13:47.760> very<00:13:47.960> experimental" + }, + { + "start": 828.39, + "duration": 0.0, + "text": "Deep learning is just very experimental" + }, + { + "start": 828.4, + "duration": 0.0, + "text": "Deep learning is just very experimental like<00:13:48.520> this.<00:13:48.720> It's<00:13:48.840> not<00:13:49.000> exactly<00:13:49.320> known,<00:13:49.960> you" + }, + { + "start": 830.03, + "duration": 0.0, + "text": "like this. It's not exactly known, you" + }, + { + "start": 830.04, + "duration": 0.0, + "text": "like this. It's not exactly known, you know,<00:13:50.160> what<00:13:50.280> would<00:13:50.400> cause<00:13:50.920> the<00:13:51.040> model<00:13:51.320> to" + }, + { + "start": 831.43, + "duration": 0.0, + "text": "know, what would cause the model to" + }, + { + "start": 831.44, + "duration": 0.0, + "text": "know, what would cause the model to correctly<00:13:51.760> learn<00:13:52.000> how<00:13:52.080> to<00:13:52.200> ideally<00:13:52.520> animate" + }, + { + "start": 832.79, + "duration": 0.0, + "text": "correctly learn how to ideally animate" + }, + { + "start": 832.8, + "duration": 0.0, + "text": "correctly learn how to ideally animate that<00:13:52.960> scene<00:13:53.200> or<00:13:53.280> make<00:13:53.400> the<00:13:53.480> clouds<00:13:53.800> move," + }, + { + "start": 833.99, + "duration": 0.0, + "text": "that scene or make the clouds move," + }, + { + "start": 834.0, + "duration": 0.0, + "text": "that scene or make the clouds move, which<00:13:54.160> is<00:13:54.280> what<00:13:54.360> I'm<00:13:54.480> hoping<00:13:54.720> for,<00:13:55.000> right?<00:13:55.560> You" + }, + { + "start": 835.67, + "duration": 0.0, + "text": "which is what I'm hoping for, right? You" + }, + { + "start": 835.68, + "duration": 0.0, + "text": "which is what I'm hoping for, right? You can<00:13:55.800> imagine<00:13:56.080> with<00:13:56.200> a<00:13:56.240> model<00:13:56.480> like<00:13:56.640> this<00:13:56.800> you" + }, + { + "start": 836.87, + "duration": 0.0, + "text": "can imagine with a model like this you" + }, + { + "start": 836.88, + "duration": 0.0, + "text": "can imagine with a model like this you could<00:13:57.000> take<00:13:57.200> any<00:13:57.360> desktop<00:13:57.720> wallpaper<00:13:58.600> and" + }, + { + "start": 838.67, + "duration": 0.0, + "text": "could take any desktop wallpaper and" + }, + { + "start": 838.68, + "duration": 0.0, + "text": "could take any desktop wallpaper and just<00:13:58.800> make<00:13:58.960> it<00:13:59.080> ideally<00:13:59.400> animate<00:13:59.680> in<00:13:59.760> the" + }, + { + "start": 839.79, + "duration": 0.0, + "text": "just make it ideally animate in the" + }, + { + "start": 839.8, + "duration": 0.0, + "text": "just make it ideally animate in the background,<00:14:00.280> right?<00:14:00.440> Especially<00:14:00.760> you<00:14:00.840> got<00:14:00.960> it" + }, + { + "start": 841.03, + "duration": 0.0, + "text": "background, right? Especially you got it" + }, + { + "start": 841.04, + "duration": 0.0, + "text": "background, right? Especially you got it like" + }, + { + "start": 841.95, + "duration": 0.0, + "text": "like" + }, + { + "start": 841.96, + "duration": 0.0, + "text": "like you<00:14:02.040> know,<00:14:02.280> good<00:14:02.440> enough<00:14:02.600> on<00:14:02.680> the<00:14:02.720> battery.<00:14:03.000> It" + }, + { + "start": 843.07, + "duration": 0.0, + "text": "you know, good enough on the battery. It" + }, + { + "start": 843.08, + "duration": 0.0, + "text": "you know, good enough on the battery. It just<00:14:03.320> might<00:14:03.440> be<00:14:03.520> kind<00:14:03.720> of<00:14:03.800> fun<00:14:03.960> to,<00:14:04.080> you<00:14:04.120> know," + }, + { + "start": 844.63, + "duration": 0.0, + "text": "just might be kind of fun to, you know," + }, + { + "start": 844.64, + "duration": 0.0, + "text": "just might be kind of fun to, you know, minimize<00:14:05.000> it<00:14:05.080> and<00:14:05.160> then<00:14:05.240> see<00:14:05.400> stuff<00:14:05.640> always" + }, + { + "start": 845.83, + "duration": 0.0, + "text": "minimize it and then see stuff always" + }, + { + "start": 845.84, + "duration": 0.0, + "text": "minimize it and then see stuff always moving." + }, + { + "start": 846.87, + "duration": 0.0, + "text": "moving." + }, + { + "start": 846.88, + "duration": 0.0, + "text": "moving. Um" + }, + { + "start": 848.07, + "duration": 0.0, + "text": "Um" + }, + { + "start": 848.08, + "duration": 0.0, + "text": "Um So<00:14:08.160> just<00:14:08.320> as<00:14:08.400> a<00:14:08.440> little<00:14:08.640> bit<00:14:08.720> of<00:14:08.800> a<00:14:09.080> historical" + }, + { + "start": 849.71, + "duration": 0.0, + "text": "So just as a little bit of a historical" + }, + { + "start": 849.72, + "duration": 0.0, + "text": "So just as a little bit of a historical detour,<00:14:10.400> right?<00:14:10.960> That<00:14:11.240> what<00:14:11.400> I<00:14:11.440> just" + }, + { + "start": 851.63, + "duration": 0.0, + "text": "detour, right? That what I just" + }, + { + "start": 851.64, + "duration": 0.0, + "text": "detour, right? That what I just explained<00:14:11.960> to<00:14:12.000> you,<00:14:12.080> that<00:14:12.240> deep<00:14:12.440> learning" + }, + { + "start": 852.75, + "duration": 0.0, + "text": "explained to you, that deep learning" + }, + { + "start": 852.76, + "duration": 0.0, + "text": "explained to you, that deep learning model,<00:14:13.640> it<00:14:13.800> the<00:14:13.920> perceptron,<00:14:14.720> it<00:14:14.800> was<00:14:14.920> this<00:14:15.120> is" + }, + { + "start": 855.15, + "duration": 0.0, + "text": "model, it the perceptron, it was this is" + }, + { + "start": 855.16, + "duration": 0.0, + "text": "model, it the perceptron, it was this is an<00:14:15.280> article<00:14:15.680> about<00:14:15.920> it<00:14:16.000> in<00:14:16.080> 1958,<00:14:17.160> right?" + }, + { + "start": 858.23, + "duration": 0.0, + "text": "an article about it in 1958, right?" + }, + { + "start": 858.24, + "duration": 0.0, + "text": "an article about it in 1958, right? They<00:14:18.320> were<00:14:18.400> saying<00:14:18.640> it<00:14:18.720> can<00:14:18.840> tell<00:14:19.040> the" + }, + { + "start": 859.11, + "duration": 0.0, + "text": "They were saying it can tell the" + }, + { + "start": 859.12, + "duration": 0.0, + "text": "They were saying it can tell the difference<00:14:19.480> between<00:14:19.720> a<00:14:19.760> cat<00:14:20.120> and<00:14:20.240> a<00:14:20.320> dog.<00:14:20.760> It" + }, + { + "start": 860.83, + "duration": 0.0, + "text": "difference between a cat and a dog. It" + }, + { + "start": 860.84, + "duration": 0.0, + "text": "difference between a cat and a dog. It is<00:14:20.960> of<00:14:21.080> no<00:14:21.240> practical<00:14:21.840> use.<00:14:22.720> One<00:14:22.920> day<00:14:23.040> it<00:14:23.120> might" + }, + { + "start": 863.23, + "duration": 0.0, + "text": "is of no practical use. One day it might" + }, + { + "start": 863.24, + "duration": 0.0, + "text": "is of no practical use. One day it might be<00:14:23.320> useful<00:14:23.600> to<00:14:23.680> send<00:14:23.960> one<00:14:24.080> into<00:14:24.280> outer<00:14:24.480> space" + }, + { + "start": 864.79, + "duration": 0.0, + "text": "be useful to send one into outer space" + }, + { + "start": 864.8, + "duration": 0.0, + "text": "be useful to send one into outer space to<00:14:24.920> take<00:14:25.200> impressions<00:14:25.720> for<00:14:25.880> us." + }, + { + "start": 866.83, + "duration": 0.0, + "text": "to take impressions for us." + }, + { + "start": 866.84, + "duration": 0.0, + "text": "to take impressions for us. So<00:14:27.360> AI<00:14:27.680> has<00:14:27.880> always<00:14:28.160> been<00:14:28.320> a<00:14:28.360> grift," + }, + { + "start": 868.67, + "duration": 0.0, + "text": "So AI has always been a grift," + }, + { + "start": 868.68, + "duration": 0.0, + "text": "So AI has always been a grift, basically,<00:14:29.360> right?<00:14:29.680> And<00:14:29.760> it's<00:14:29.960> followed" + }, + { + "start": 870.39, + "duration": 0.0, + "text": "basically, right? And it's followed" + }, + { + "start": 870.4, + "duration": 0.0, + "text": "basically, right? And it's followed basically<00:14:30.839> this<00:14:31.080> this<00:14:31.320> phase<00:14:31.760> where" + }, + { + "start": 872.71, + "duration": 0.0, + "text": "basically this this phase where" + }, + { + "start": 872.72, + "duration": 0.0, + "text": "basically this this phase where all<00:14:32.800> the<00:14:32.880> investors<00:14:33.280> pour<00:14:33.440> their<00:14:33.560> money<00:14:33.800> in" + }, + { + "start": 873.91, + "duration": 0.0, + "text": "all the investors pour their money in" + }, + { + "start": 873.92, + "duration": 0.0, + "text": "all the investors pour their money in and<00:14:34.000> they're<00:14:34.080> like,<00:14:34.240> \"This<00:14:34.440> is<00:14:34.520> going<00:14:34.720> to<00:14:34.760> be<00:14:34.880> a" + }, + { + "start": 874.91, + "duration": 0.0, + "text": "and they're like, \"This is going to be a" + }, + { + "start": 874.92, + "duration": 0.0, + "text": "and they're like, \"This is going to be a new<00:14:35.080> form<00:14:35.320> of<00:14:35.480> life<00:14:35.720> that<00:14:35.880> will<00:14:36.000> like<00:14:36.200> do<00:14:36.320> our" + }, + { + "start": 876.43, + "duration": 0.0, + "text": "new form of life that will like do our" + }, + { + "start": 876.44, + "duration": 0.0, + "text": "new form of life that will like do our bidding,<00:14:36.760> right?\"<00:14:37.600> And<00:14:37.720> then<00:14:37.839> it<00:14:37.920> doesn't" + }, + { + "start": 878.15, + "duration": 0.0, + "text": "bidding, right?\" And then it doesn't" + }, + { + "start": 878.16, + "duration": 0.0, + "text": "bidding, right?\" And then it doesn't work<00:14:38.560> and<00:14:38.680> then<00:14:38.800> it<00:14:38.880> happens<00:14:39.160> again<00:14:39.440> and<00:14:39.520> then" + }, + { + "start": 879.63, + "duration": 0.0, + "text": "work and then it happens again and then" + }, + { + "start": 879.64, + "duration": 0.0, + "text": "work and then it happens again and then it<00:14:39.720> doesn't<00:14:39.920> work.<00:14:40.839> And<00:14:40.880> now<00:14:41.000> it's<00:14:41.120> happening" + }, + { + "start": 881.43, + "duration": 0.0, + "text": "it doesn't work. And now it's happening" + }, + { + "start": 881.44, + "duration": 0.0, + "text": "it doesn't work. And now it's happening again<00:14:41.800> and<00:14:41.920> it's,<00:14:42.120> you<00:14:42.240> know," + }, + { + "start": 883.03, + "duration": 0.0, + "text": "again and it's, you know," + }, + { + "start": 883.04, + "duration": 0.0, + "text": "again and it's, you know, who<00:14:43.120> knows<00:14:43.280> what's<00:14:43.400> going<00:14:43.520> to<00:14:43.560> happen." + }, + { + "start": 884.55, + "duration": 0.0, + "text": "who knows what's going to happen." + }, + { + "start": 884.56, + "duration": 0.0, + "text": "who knows what's going to happen. Um" + }, + { + "start": 889.36, + "duration": 0.0, + "text": "He<00:14:49.480> just<00:14:49.600> says,<00:14:49.760> \"What<00:14:49.880> did<00:14:49.960> we<00:14:50.080> learn," + }, + { + "start": 890.31, + "duration": 0.0, + "text": "He just says, \"What did we learn," + }, + { + "start": 890.32, + "duration": 0.0, + "text": "He just says, \"What did we learn, Palmer?\"<00:14:51.160> Um" + }, + { + "start": 892.59, + "duration": 0.0, + "text": "Palmer?\" Um" + }, + { + "start": 892.6, + "duration": 0.0, + "text": "Palmer?\" Um And<00:14:52.800> so<00:14:52.960> me<00:14:53.160> personally,<00:14:53.680> you<00:14:53.760> know,<00:14:53.960> I<00:14:54.440> my" + }, + { + "start": 894.71, + "duration": 0.0, + "text": "And so me personally, you know, I my" + }, + { + "start": 894.72, + "duration": 0.0, + "text": "And so me personally, you know, I my kind<00:14:54.920> of<00:14:55.000> like<00:14:55.440> AI<00:14:55.720> boom<00:14:55.960> bust<00:14:56.280> was<00:14:56.600> I<00:14:56.720> got" + }, + { + "start": 896.91, + "duration": 0.0, + "text": "kind of like AI boom bust was I got" + }, + { + "start": 896.92, + "duration": 0.0, + "text": "kind of like AI boom bust was I got really<00:14:57.160> into<00:14:57.520> deep<00:14:57.760> learning.<00:14:58.680> Um<00:14:59.200> you<00:14:59.280> know," + }, + { + "start": 899.39, + "duration": 0.0, + "text": "really into deep learning. Um you know," + }, + { + "start": 899.4, + "duration": 0.0, + "text": "really into deep learning. Um you know, and<00:14:59.520> then<00:14:59.680> I<00:14:59.960> right<00:15:00.160> when<00:15:00.320> GPT-2<00:15:00.839> started<00:15:01.120> to" + }, + { + "start": 901.19, + "duration": 0.0, + "text": "and then I right when GPT-2 started to" + }, + { + "start": 901.2, + "duration": 0.0, + "text": "and then I right when GPT-2 started to happen,<00:15:01.520> I<00:15:01.560> was<00:15:01.640> like,<00:15:01.760> \"This<00:15:01.880> thing's<00:15:02.120> kind" + }, + { + "start": 902.27, + "duration": 0.0, + "text": "happen, I was like, \"This thing's kind" + }, + { + "start": 902.28, + "duration": 0.0, + "text": "happen, I was like, \"This thing's kind of<00:15:02.320> nonsense<00:15:02.720> and<00:15:02.800> it's<00:15:02.920> not<00:15:03.040> going<00:15:03.160> to" + }, + { + "start": 903.23, + "duration": 0.0, + "text": "of nonsense and it's not going to" + }, + { + "start": 903.24, + "duration": 0.0, + "text": "of nonsense and it's not going to scale.\"<00:15:03.920> So<00:15:04.080> I<00:15:04.160> basically<00:15:04.480> was<00:15:04.560> like,<00:15:04.720> \"All" + }, + { + "start": 904.75, + "duration": 0.0, + "text": "scale.\" So I basically was like, \"All" + }, + { + "start": 904.76, + "duration": 0.0, + "text": "scale.\" So I basically was like, \"All right,<00:15:04.880> I'm<00:15:04.960> kind<00:15:05.120> of<00:15:05.200> done<00:15:05.320> with<00:15:05.400> deep" + }, + { + "start": 905.51, + "duration": 0.0, + "text": "right, I'm kind of done with deep" + }, + { + "start": 905.52, + "duration": 0.0, + "text": "right, I'm kind of done with deep learning.<00:15:05.760> I<00:15:05.800> think<00:15:05.920> it's<00:15:06.040> useless.\"<00:15:07.080> Um<00:15:07.560> and" + }, + { + "start": 907.829, + "duration": 0.0, + "text": "learning. I think it's useless.\" Um and" + }, + { + "start": 907.839, + "duration": 0.0, + "text": "learning. I think it's useless.\" Um and so<00:15:07.920> I<00:15:08.000> basically<00:15:08.720> it's<00:15:08.800> kind<00:15:08.920> of<00:15:09.000> like<00:15:09.120> selling" + }, + { + "start": 909.39, + "duration": 0.0, + "text": "so I basically it's kind of like selling" + }, + { + "start": 909.4, + "duration": 0.0, + "text": "so I basically it's kind of like selling Bitcoin<00:15:09.839> at<00:15:10.040> $100.<00:15:11.280> I<00:15:11.400> never<00:15:11.640> would<00:15:11.800> want<00:15:11.920> to" + }, + { + "start": 911.99, + "duration": 0.0, + "text": "Bitcoin at $100. I never would want to" + }, + { + "start": 912.0, + "duration": 0.0, + "text": "Bitcoin at $100. I never would want to buy<00:15:12.200> back<00:15:12.440> in<00:15:12.640> cuz<00:15:12.800> that's<00:15:13.000> realizing<00:15:13.440> my" + }, + { + "start": 913.51, + "duration": 0.0, + "text": "buy back in cuz that's realizing my" + }, + { + "start": 913.52, + "duration": 0.0, + "text": "buy back in cuz that's realizing my losses.<00:15:14.000> So<00:15:14.040> I'm<00:15:14.120> just<00:15:14.200> kind<00:15:14.320> of<00:15:14.400> waiting<00:15:14.600> for" + }, + { + "start": 914.71, + "duration": 0.0, + "text": "losses. So I'm just kind of waiting for" + }, + { + "start": 914.72, + "duration": 0.0, + "text": "losses. So I'm just kind of waiting for it<00:15:14.760> to<00:15:15.200> collapse<00:15:15.520> again,<00:15:15.680> hopefully,<00:15:15.960> maybe," + }, + { + "start": 916.23, + "duration": 0.0, + "text": "it to collapse again, hopefully, maybe," + }, + { + "start": 916.24, + "duration": 0.0, + "text": "it to collapse again, hopefully, maybe, and<00:15:16.320> then<00:15:16.400> I<00:15:16.440> can<00:15:16.560> buy." + }, + { + "start": 917.51, + "duration": 0.0, + "text": "and then I can buy." + }, + { + "start": 917.52, + "duration": 0.0, + "text": "and then I can buy. Um<00:15:17.640> and<00:15:17.720> if<00:15:17.800> you<00:15:17.880> think<00:15:18.000> that's<00:15:18.160> a<00:15:18.200> funny<00:15:18.440> joke" + }, + { + "start": 919.43, + "duration": 0.0, + "text": "Um and if you think that's a funny joke" + }, + { + "start": 919.44, + "duration": 0.0, + "text": "Um and if you think that's a funny joke in<00:15:19.520> the<00:15:19.600> audience<00:15:19.960> right<00:15:20.120> now,<00:15:20.480> how<00:15:20.600> many<00:15:20.760> of" + }, + { + "start": 920.829, + "duration": 0.0, + "text": "in the audience right now, how many of" + }, + { + "start": 920.839, + "duration": 0.0, + "text": "in the audience right now, how many of you<00:15:20.880> have<00:15:21.040> any<00:15:21.200> amount<00:15:21.400> of<00:15:21.480> retirement<00:15:21.880> or" + }, + { + "start": 921.95, + "duration": 0.0, + "text": "you have any amount of retirement or" + }, + { + "start": 921.96, + "duration": 0.0, + "text": "you have any amount of retirement or capital<00:15:22.400> in<00:15:22.640> ETFs,<00:15:23.520> right?<00:15:23.760> So,<00:15:23.840> I<00:15:23.880> need<00:15:23.960> you" + }, + { + "start": 924.03, + "duration": 0.0, + "text": "capital in ETFs, right? So, I need you" + }, + { + "start": 924.04, + "duration": 0.0, + "text": "capital in ETFs, right? So, I need you to<00:15:24.080> raise<00:15:24.240> your<00:15:24.360> hand." + }, + { + "start": 926.51, + "duration": 0.0, + "text": "to raise your hand." + }, + { + "start": 926.52, + "duration": 0.0, + "text": "to raise your hand. Okay." + }, + { + "start": 927.95, + "duration": 0.0, + "text": "Okay." + }, + { + "start": 927.96, + "duration": 0.0, + "text": "Okay. Your<00:15:28.080> money's<00:15:28.280> being<00:15:28.480> gambled<00:15:28.920> on<00:15:29.000> the" + }, + { + "start": 929.07, + "duration": 0.0, + "text": "Your money's being gambled on the" + }, + { + "start": 929.08, + "duration": 0.0, + "text": "Your money's being gambled on the success<00:15:29.440> of<00:15:29.600> AGI<00:15:30.160> right<00:15:30.480> now.<00:15:31.080> Roughly<00:15:31.360> 30%<00:15:31.920> of" + }, + { + "start": 931.99, + "duration": 0.0, + "text": "success of AGI right now. Roughly 30% of" + }, + { + "start": 932.0, + "duration": 0.0, + "text": "success of AGI right now. Roughly 30% of the<00:15:32.080> S&P<00:15:32.440> 500<00:15:32.920> is<00:15:33.080> heavily<00:15:33.440> weighted<00:15:33.680> on<00:15:33.840> AI" + }, + { + "start": 933.99, + "duration": 0.0, + "text": "the S&P 500 is heavily weighted on AI" + }, + { + "start": 934.0, + "duration": 0.0, + "text": "the S&P 500 is heavily weighted on AI succeeding.<00:15:34.840> AI<00:15:35.000> growth<00:15:35.280> stories,<00:15:35.760> you<00:15:35.840> know," + }, + { + "start": 936.71, + "duration": 0.0, + "text": "succeeding. AI growth stories, you know," + }, + { + "start": 936.72, + "duration": 0.0, + "text": "succeeding. AI growth stories, you know, high<00:15:36.840> price<00:15:37.080> to<00:15:37.200> earnings." + }, + { + "start": 938.23, + "duration": 0.0, + "text": "high price to earnings." + }, + { + "start": 938.24, + "duration": 0.0, + "text": "high price to earnings. Um<00:15:38.880> so,<00:15:39.120> I<00:15:39.240> propose<00:15:39.800> a<00:15:39.880> model<00:15:40.120> of<00:15:40.240> history" + }, + { + "start": 940.59, + "duration": 0.0, + "text": "Um so, I propose a model of history" + }, + { + "start": 940.6, + "duration": 0.0, + "text": "Um so, I propose a model of history here.<00:15:41.040> There's<00:15:41.200> basically<00:15:41.560> two<00:15:41.760> grifts" + }, + { + "start": 942.03, + "duration": 0.0, + "text": "here. There's basically two grifts" + }, + { + "start": 942.04, + "duration": 0.0, + "text": "here. There's basically two grifts orbiting<00:15:42.360> each<00:15:42.520> other,<00:15:42.720> right?<00:15:43.640> Fake<00:15:43.840> money" + }, + { + "start": 944.07, + "duration": 0.0, + "text": "orbiting each other, right? Fake money" + }, + { + "start": 944.08, + "duration": 0.0, + "text": "orbiting each other, right? Fake money pumps<00:15:44.320> the<00:15:44.480> AI.<00:15:44.960> Capital<00:15:45.360> flees<00:15:45.640> to<00:15:45.760> Bitcoin" + }, + { + "start": 946.19, + "duration": 0.0, + "text": "pumps the AI. Capital flees to Bitcoin" + }, + { + "start": 946.2, + "duration": 0.0, + "text": "pumps the AI. Capital flees to Bitcoin instead<00:15:46.440> of<00:15:46.560> US-related,<00:15:47.320> you<00:15:47.360> know,<00:15:47.680> assets." + }, + { + "start": 948.43, + "duration": 0.0, + "text": "instead of US-related, you know, assets." + }, + { + "start": 948.44, + "duration": 0.0, + "text": "instead of US-related, you know, assets. No<00:15:48.560> more<00:15:48.680> easy<00:15:48.880> capital,<00:15:49.280> no<00:15:49.400> more<00:15:49.640> AI." + }, + { + "start": 950.31, + "duration": 0.0, + "text": "No more easy capital, no more AI." + }, + { + "start": 950.32, + "duration": 0.0, + "text": "No more easy capital, no more AI. Deflationary<00:15:50.880> doom<00:15:51.120> spiral<00:15:51.400> where<00:15:51.480> GDP" + }, + { + "start": 951.75, + "duration": 0.0, + "text": "Deflationary doom spiral where GDP" + }, + { + "start": 951.76, + "duration": 0.0, + "text": "Deflationary doom spiral where GDP becomes<00:15:52.040> 1/10<00:15:52.400> of<00:15:52.480> what<00:15:52.600> it<00:15:52.680> is<00:15:52.800> now.<00:15:53.400> This<00:15:53.600> is" + }, + { + "start": 953.71, + "duration": 0.0, + "text": "becomes 1/10 of what it is now. This is" + }, + { + "start": 953.72, + "duration": 0.0, + "text": "becomes 1/10 of what it is now. This is an<00:15:53.840> idea<00:15:54.280> that<00:15:54.440> is<00:15:54.560> basically<00:15:54.960> a" + }, + { + "start": 954.99, + "duration": 0.0, + "text": "an idea that is basically a" + }, + { + "start": 955.0, + "duration": 0.0, + "text": "an idea that is basically a bubble-generating<00:15:55.680> phenomenon,<00:15:56.360> right?" + }, + { + "start": 956.95, + "duration": 0.0, + "text": "bubble-generating phenomenon, right?" + }, + { + "start": 956.96, + "duration": 0.0, + "text": "bubble-generating phenomenon, right? Both<00:15:57.200> of<00:15:57.280> these,<00:15:57.680> you<00:15:57.720> know,<00:15:57.840> one<00:15:58.080> paper" + }, + { + "start": 958.43, + "duration": 0.0, + "text": "Both of these, you know, one paper" + }, + { + "start": 958.44, + "duration": 0.0, + "text": "Both of these, you know, one paper basically<00:15:58.960> started<00:15:59.280> the<00:15:59.360> whole" + }, + { + "start": 960.55, + "duration": 0.0, + "text": "basically started the whole" + }, + { + "start": 960.56, + "duration": 0.0, + "text": "basically started the whole There's<00:16:00.680> like<00:16:00.800> the<00:16:00.880> transformer<00:16:01.400> paper<00:16:01.640> and" + }, + { + "start": 961.71, + "duration": 0.0, + "text": "There's like the transformer paper and" + }, + { + "start": 961.72, + "duration": 0.0, + "text": "There's like the transformer paper and there's<00:16:01.800> the<00:16:01.880> Bitcoin<00:16:02.240> papers,<00:16:02.760> parallels," + }, + { + "start": 963.23, + "duration": 0.0, + "text": "there's the Bitcoin papers, parallels," + }, + { + "start": 963.24, + "duration": 0.0, + "text": "there's the Bitcoin papers, parallels, right?<00:16:04.160> And<00:16:04.200> if<00:16:04.320> you<00:16:04.360> look<00:16:04.520> at<00:16:04.640> this<00:16:04.800> like,<00:16:04.960> you" + }, + { + "start": 965.03, + "duration": 0.0, + "text": "right? And if you look at this like, you" + }, + { + "start": 965.04, + "duration": 0.0, + "text": "right? And if you look at this like, you know,<00:16:05.160> innovators-laggards<00:16:06.240> curve,<00:16:06.600> right?" + }, + { + "start": 967.43, + "duration": 0.0, + "text": "know, innovators-laggards curve, right?" + }, + { + "start": 967.44, + "duration": 0.0, + "text": "know, innovators-laggards curve, right? I<00:16:07.480> think<00:16:07.680> we're<00:16:07.840> right<00:16:08.080> about<00:16:08.320> here<00:16:08.480> if<00:16:08.600> you" + }, + { + "start": 968.67, + "duration": 0.0, + "text": "I think we're right about here if you" + }, + { + "start": 968.68, + "duration": 0.0, + "text": "I think we're right about here if you look<00:16:08.840> at<00:16:08.920> the<00:16:09.080> Nvidia<00:16:09.400> stock<00:16:09.680> price.<00:16:10.320> It<00:16:10.440> fits" + }, + { + "start": 970.59, + "duration": 0.0, + "text": "look at the Nvidia stock price. It fits" + }, + { + "start": 970.6, + "duration": 0.0, + "text": "look at the Nvidia stock price. It fits the<00:16:10.720> curve." + }, + { + "start": 972.23, + "duration": 0.0, + "text": "the curve." + }, + { + "start": 972.24, + "duration": 0.0, + "text": "the curve. So,<00:16:12.320> we're<00:16:12.440> at<00:16:12.560> peak<00:16:12.800> of<00:16:12.880> grift<00:16:13.120> at<00:16:13.240> the" + }, + { + "start": 973.27, + "duration": 0.0, + "text": "So, we're at peak of grift at the" + }, + { + "start": 973.28, + "duration": 0.0, + "text": "So, we're at peak of grift at the moment.<00:16:13.720> It'll<00:16:13.880> fall<00:16:14.120> off<00:16:14.280> soon,<00:16:14.520> I<00:16:14.600> think." + }, + { + "start": 974.83, + "duration": 0.0, + "text": "moment. It'll fall off soon, I think." + }, + { + "start": 974.84, + "duration": 0.0, + "text": "moment. It'll fall off soon, I think. And<00:16:14.960> so,<00:16:15.080> I<00:16:15.120> think<00:16:15.320> now's<00:16:15.480> a<00:16:15.520> good<00:16:15.680> time<00:16:15.840> to" + }, + { + "start": 975.91, + "duration": 0.0, + "text": "And so, I think now's a good time to" + }, + { + "start": 975.92, + "duration": 0.0, + "text": "And so, I think now's a good time to short,<00:16:16.280> but" + }, + { + "start": 979.8, + "duration": 0.0, + "text": "that<00:16:20.040> may<00:16:20.160> or<00:16:20.240> may<00:16:20.320> not<00:16:20.480> be<00:16:20.560> financial<00:16:21.000> advice" + }, + { + "start": 981.27, + "duration": 0.0, + "text": "that may or may not be financial advice" + }, + { + "start": 981.28, + "duration": 0.0, + "text": "that may or may not be financial advice if<00:16:21.400> it<00:16:21.520> works<00:16:21.720> out<00:16:21.880> for<00:16:22.000> you." + }, + { + "start": 984.39, + "duration": 0.0, + "text": "if it works out for you." + }, + { + "start": 984.4, + "duration": 0.0, + "text": "if it works out for you. I<00:16:24.480> do<00:16:24.640> just<00:16:24.839> want<00:16:24.960> to<00:16:25.040> say<00:16:25.280> that<00:16:25.520> shorting,<00:16:26.360> I" + }, + { + "start": 986.47, + "duration": 0.0, + "text": "I do just want to say that shorting, I" + }, + { + "start": 986.48, + "duration": 0.0, + "text": "I do just want to say that shorting, I believe,<00:16:26.920> has<00:16:27.160> unlimited<00:16:28.000> uh<00:16:28.400> risk<00:16:28.800> and<00:16:28.960> you" + }, + { + "start": 989.19, + "duration": 0.0, + "text": "believe, has unlimited uh risk and you" + }, + { + "start": 989.2, + "duration": 0.0, + "text": "believe, has unlimited uh risk and you at<00:16:29.400> best<00:16:29.640> would<00:16:29.720> like<00:16:29.880> double<00:16:30.120> your<00:16:30.240> money,<00:16:30.480> so" + }, + { + "start": 990.59, + "duration": 0.0, + "text": "at best would like double your money, so" + }, + { + "start": 990.6, + "duration": 0.0, + "text": "at best would like double your money, so it's<00:16:30.720> probably<00:16:30.960> not<00:16:31.120> like<00:16:31.280> a<00:16:31.320> good<00:16:31.480> idea." + }, + { + "start": 992.51, + "duration": 0.0, + "text": "it's probably not like a good idea." + }, + { + "start": 992.52, + "duration": 0.0, + "text": "it's probably not like a good idea. Um" + }, + { + "start": 993.79, + "duration": 0.0, + "text": "Um" + }, + { + "start": 993.8, + "duration": 0.0, + "text": "Um Here's<00:16:33.920> some<00:16:34.040> examples<00:16:34.400> of<00:16:34.720> uh<00:16:35.120> just<00:16:35.640> machine" + }, + { + "start": 995.99, + "duration": 0.0, + "text": "Here's some examples of uh just machine" + }, + { + "start": 996.0, + "duration": 0.0, + "text": "Here's some examples of uh just machine learning<00:16:36.240> in<00:16:36.320> general,<00:16:36.600> right?<00:16:36.720> Let's<00:16:36.839> go" + }, + { + "start": 996.95, + "duration": 0.0, + "text": "learning in general, right? Let's go" + }, + { + "start": 996.96, + "duration": 0.0, + "text": "learning in general, right? Let's go back<00:16:37.160> to<00:16:37.280> not,<00:16:38.200> you<00:16:38.240> know,<00:16:38.360> being<00:16:38.560> kind<00:16:38.680> of" + }, + { + "start": 998.79, + "duration": 0.0, + "text": "back to not, you know, being kind of" + }, + { + "start": 998.8, + "duration": 0.0, + "text": "back to not, you know, being kind of schizo.<00:16:39.800> Um" + }, + { + "start": 1001.03, + "duration": 0.0, + "text": "schizo. Um" + }, + { + "start": 1001.04, + "duration": 0.0, + "text": "schizo. Um So,<00:16:41.160> this<00:16:41.480> is<00:16:41.680> a<00:16:41.760> little<00:16:41.960> piece<00:16:42.160> of<00:16:42.200> history" + }, + { + "start": 1002.51, + "duration": 0.0, + "text": "So, this is a little piece of history" + }, + { + "start": 1002.52, + "duration": 0.0, + "text": "So, this is a little piece of history again,<00:16:42.800> right?<00:16:43.440> This<00:16:43.560> is<00:16:43.640> called<00:16:43.920> Cyc.<00:16:44.240> In" + }, + { + "start": 1004.31, + "duration": 0.0, + "text": "again, right? This is called Cyc. In" + }, + { + "start": 1004.32, + "duration": 0.0, + "text": "again, right? This is called Cyc. In 1984,<00:16:45.400> some<00:16:45.560> people<00:16:45.839> were<00:16:45.920> very<00:16:46.160> upset<00:16:46.440> at<00:16:46.520> the" + }, + { + "start": 1006.67, + "duration": 0.0, + "text": "1984, some people were very upset at the" + }, + { + "start": 1006.68, + "duration": 0.0, + "text": "1984, some people were very upset at the deep<00:16:46.920> learning<00:16:47.160> people.<00:16:47.760> They<00:16:47.880> were<00:16:48.000> like," + }, + { + "start": 1008.15, + "duration": 0.0, + "text": "deep learning people. They were like," + }, + { + "start": 1008.16, + "duration": 0.0, + "text": "deep learning people. They were like, \"Look<00:16:48.320> at<00:16:48.400> this<00:16:48.560> arbitrary<00:16:49.040> floats<00:16:49.360> and" + }, + { + "start": 1009.43, + "duration": 0.0, + "text": "\"Look at this arbitrary floats and" + }, + { + "start": 1009.44, + "duration": 0.0, + "text": "\"Look at this arbitrary floats and numbers.<00:16:49.720> I<00:16:49.760> want<00:16:49.920> like<00:16:50.080> an<00:16:50.160> interpretable" + }, + { + "start": 1011.15, + "duration": 0.0, + "text": "numbers. I want like an interpretable" + }, + { + "start": 1011.16, + "duration": 0.0, + "text": "numbers. I want like an interpretable program<00:16:51.640> that's<00:16:51.839> intelligent,<00:16:52.360> right?\"<00:16:52.880> So," + }, + { + "start": 1012.99, + "duration": 0.0, + "text": "program that's intelligent, right?\" So," + }, + { + "start": 1013.0, + "duration": 0.0, + "text": "program that's intelligent, right?\" So, they<00:16:53.160> basically<00:16:53.720> set<00:16:53.960> out<00:16:54.120> to<00:16:54.760> manually" + }, + { + "start": 1015.19, + "duration": 0.0, + "text": "they basically set out to manually" + }, + { + "start": 1015.2, + "duration": 0.0, + "text": "they basically set out to manually program<00:16:55.640> all<00:16:55.760> of<00:16:55.839> human<00:16:56.080> intelligence.<00:16:57.040> And" + }, + { + "start": 1017.15, + "duration": 0.0, + "text": "program all of human intelligence. And" + }, + { + "start": 1017.16, + "duration": 0.0, + "text": "program all of human intelligence. And the<00:16:57.240> theory<00:16:57.520> was<00:16:57.720> if<00:16:57.800> you<00:16:57.880> just<00:16:58.040> throw<00:16:58.160> enough" + }, + { + "start": 1018.35, + "duration": 0.0, + "text": "the theory was if you just throw enough" + }, + { + "start": 1018.36, + "duration": 0.0, + "text": "the theory was if you just throw enough bodies<00:16:58.680> and<00:16:58.800> people<00:16:59.080> at<00:16:59.200> it,<00:16:59.280> it'll" + }, + { + "start": 1019.43, + "duration": 0.0, + "text": "bodies and people at it, it'll" + }, + { + "start": 1019.44, + "duration": 0.0, + "text": "bodies and people at it, it'll eventually<00:16:59.800> be<00:16:59.920> smart,<00:17:00.320> right?<00:17:00.880> They've<00:17:01.040> been" + }, + { + "start": 1021.15, + "duration": 0.0, + "text": "eventually be smart, right? They've been" + }, + { + "start": 1021.16, + "duration": 0.0, + "text": "eventually be smart, right? They've been doing<00:17:01.360> this<00:17:01.480> since<00:17:01.640> 1984.<00:17:02.920> Okay,<00:17:03.120> if<00:17:03.240> you<00:17:03.360> ask" + }, + { + "start": 1023.59, + "duration": 0.0, + "text": "doing this since 1984. Okay, if you ask" + }, + { + "start": 1023.6, + "duration": 0.0, + "text": "doing this since 1984. Okay, if you ask this<00:17:03.760> system<00:17:04.040> questions<00:17:04.439> like,<00:17:04.880> \"If<00:17:04.959> I<00:17:05.040> put<00:17:05.199> a" + }, + { + "start": 1025.51, + "duration": 0.0, + "text": "this system questions like, \"If I put a" + }, + { + "start": 1025.52, + "duration": 0.0, + "text": "this system questions like, \"If I put a a<00:17:05.560> cup<00:17:05.760> of<00:17:05.839> coffee<00:17:06.120> on<00:17:06.240> a<00:17:06.280> table<00:17:06.600> and<00:17:06.680> then<00:17:06.800> I" + }, + { + "start": 1026.91, + "duration": 0.0, + "text": "a cup of coffee on a table and then I" + }, + { + "start": 1026.92, + "duration": 0.0, + "text": "a cup of coffee on a table and then I push<00:17:07.199> on<00:17:07.319> it,<00:17:07.400> what<00:17:07.520> happens?\"<00:17:07.880> It'll<00:17:08.439> say" + }, + { + "start": 1028.55, + "duration": 0.0, + "text": "push on it, what happens?\" It'll say" + }, + { + "start": 1028.56, + "duration": 0.0, + "text": "push on it, what happens?\" It'll say like,<00:17:08.680> \"Oh,<00:17:08.839> well,<00:17:09.000> you<00:17:09.120> know,<00:17:09.199> it<00:17:09.280> explodes.\"" + }, + { + "start": 1029.79, + "duration": 0.0, + "text": "like, \"Oh, well, you know, it explodes.\"" + }, + { + "start": 1029.8, + "duration": 0.0, + "text": "like, \"Oh, well, you know, it explodes.\" It's<00:17:10.160> like<00:17:10.360> still<00:17:10.560> not<00:17:10.720> working<00:17:11.000> very<00:17:11.199> well." + }, + { + "start": 1031.55, + "duration": 0.0, + "text": "It's like still not working very well." + }, + { + "start": 1031.56, + "duration": 0.0, + "text": "It's like still not working very well. the<00:17:11.680> reasoning<00:17:12.040> traces<00:17:12.439> are<00:17:12.520> like<00:17:12.640> a<00:17:12.720> thousand" + }, + { + "start": 1033.63, + "duration": 0.0, + "text": "the reasoning traces are like a thousand" + }, + { + "start": 1033.64, + "duration": 0.0, + "text": "the reasoning traces are like a thousand deep,<00:17:14.079> you<00:17:14.160> know,<00:17:14.280> it's<00:17:14.360> kind<00:17:14.520> of<00:17:14.600> nuts.<00:17:14.800> It" + }, + { + "start": 1034.91, + "duration": 0.0, + "text": "deep, you know, it's kind of nuts. It" + }, + { + "start": 1034.92, + "duration": 0.0, + "text": "deep, you know, it's kind of nuts. It didn't<00:17:15.120> really<00:17:15.319> work,<00:17:15.600> right?<00:17:16.240> The<00:17:16.400> First<00:17:16.680> of" + }, + { + "start": 1036.71, + "duration": 0.0, + "text": "didn't really work, right? The First of" + }, + { + "start": 1036.72, + "duration": 0.0, + "text": "didn't really work, right? The First of all,<00:17:16.839> the<00:17:16.920> reasoning<00:17:17.240> isn't<00:17:17.400> even" + }, + { + "start": 1037.55, + "duration": 0.0, + "text": "all, the reasoning isn't even" + }, + { + "start": 1037.56, + "duration": 0.0, + "text": "all, the reasoning isn't even understandable." + }, + { + "start": 1038.79, + "duration": 0.0, + "text": "understandable." + }, + { + "start": 1038.8, + "duration": 0.0, + "text": "understandable. Um<00:17:19.240> this<00:17:19.400> approach<00:17:19.760> just<00:17:19.959> doesn't<00:17:20.240> It's" + }, + { + "start": 1040.35, + "duration": 0.0, + "text": "Um this approach just doesn't It's" + }, + { + "start": 1040.36, + "duration": 0.0, + "text": "Um this approach just doesn't It's called<00:17:20.640> symbolic<00:17:21.160> AI,<00:17:21.360> right?<00:17:21.720> And<00:17:21.880> I<00:17:21.920> think" + }, + { + "start": 1042.11, + "duration": 0.0, + "text": "called symbolic AI, right? And I think" + }, + { + "start": 1042.12, + "duration": 0.0, + "text": "called symbolic AI, right? And I think that<00:17:22.240> there's<00:17:22.400> actually<00:17:22.720> a<00:17:22.760> modern<00:17:23.079> parallel," + }, + { + "start": 1043.63, + "duration": 0.0, + "text": "that there's actually a modern parallel," + }, + { + "start": 1043.64, + "duration": 0.0, + "text": "that there's actually a modern parallel, right?<00:17:24.240> Waymo<00:17:24.560> is<00:17:24.640> like<00:17:24.839> psych,<00:17:25.160> that<00:17:25.360> project" + }, + { + "start": 1045.67, + "duration": 0.0, + "text": "right? Waymo is like psych, that project" + }, + { + "start": 1045.68, + "duration": 0.0, + "text": "right? Waymo is like psych, that project where<00:17:25.760> they<00:17:25.880> were<00:17:25.959> trying<00:17:26.120> to<00:17:26.240> manually" + }, + { + "start": 1046.67, + "duration": 0.0, + "text": "where they were trying to manually" + }, + { + "start": 1046.68, + "duration": 0.0, + "text": "where they were trying to manually encode<00:17:27.040> intelligence<00:17:27.520> instead<00:17:27.760> of<00:17:28.160> automatic" + }, + { + "start": 1048.67, + "duration": 0.0, + "text": "encode intelligence instead of automatic" + }, + { + "start": 1048.68, + "duration": 0.0, + "text": "encode intelligence instead of automatic programming,<00:17:29.280> right?<00:17:29.480> This<00:17:29.640> is<00:17:29.760> like<00:17:30.280> deep" + }, + { + "start": 1050.47, + "duration": 0.0, + "text": "programming, right? This is like deep" + }, + { + "start": 1050.48, + "duration": 0.0, + "text": "programming, right? This is like deep learning<00:17:30.720> is<00:17:30.840> really<00:17:31.040> a<00:17:31.080> new<00:17:31.280> paradigm<00:17:31.640> of" + }, + { + "start": 1051.71, + "duration": 0.0, + "text": "learning is really a new paradigm of" + }, + { + "start": 1051.72, + "duration": 0.0, + "text": "learning is really a new paradigm of programming<00:17:32.280> where<00:17:32.480> it's<00:17:32.680> automatic.<00:17:33.160> You're" + }, + { + "start": 1053.27, + "duration": 0.0, + "text": "programming where it's automatic. You're" + }, + { + "start": 1053.28, + "duration": 0.0, + "text": "programming where it's automatic. You're automatically<00:17:33.800> creating<00:17:34.200> the<00:17:34.640> the" + }, + { + "start": 1054.75, + "duration": 0.0, + "text": "automatically creating the the" + }, + { + "start": 1054.76, + "duration": 0.0, + "text": "automatically creating the the nonlinearities<00:17:35.440> that<00:17:35.600> do<00:17:35.760> the<00:17:35.840> thing<00:17:36.000> that" + }, + { + "start": 1056.11, + "duration": 0.0, + "text": "nonlinearities that do the thing that" + }, + { + "start": 1056.12, + "duration": 0.0, + "text": "nonlinearities that do the thing that you<00:17:36.240> want<00:17:36.440> the<00:17:36.560> computer<00:17:36.960> to<00:17:37.040> do,<00:17:37.200> right?<00:17:37.920> Um" + }, + { + "start": 1058.19, + "duration": 0.0, + "text": "you want the computer to do, right? Um" + }, + { + "start": 1058.2, + "duration": 0.0, + "text": "you want the computer to do, right? Um Waymo,<00:17:38.680> they<00:17:38.840> basically<00:17:39.280> scan<00:17:39.560> the<00:17:39.680> city<00:17:40.000> and" + }, + { + "start": 1060.11, + "duration": 0.0, + "text": "Waymo, they basically scan the city and" + }, + { + "start": 1060.12, + "duration": 0.0, + "text": "Waymo, they basically scan the city and they<00:17:40.200> put<00:17:40.320> a<00:17:40.360> bunch<00:17:40.480> of<00:17:40.560> if<00:17:40.680> statements<00:17:40.960> where" + }, + { + "start": 1061.07, + "duration": 0.0, + "text": "they put a bunch of if statements where" + }, + { + "start": 1061.08, + "duration": 0.0, + "text": "they put a bunch of if statements where the<00:17:41.160> car<00:17:41.400> should<00:17:41.560> go.<00:17:42.160> So,<00:17:42.280> they're<00:17:42.360> very<00:17:42.760> slow" + }, + { + "start": 1062.99, + "duration": 0.0, + "text": "the car should go. So, they're very slow" + }, + { + "start": 1063.0, + "duration": 0.0, + "text": "the car should go. So, they're very slow to<00:17:43.120> roll<00:17:43.360> out.<00:17:43.640> So,<00:17:43.760> Tesla<00:17:44.520> has<00:17:44.720> basically" + }, + { + "start": 1065.15, + "duration": 0.0, + "text": "to roll out. So, Tesla has basically" + }, + { + "start": 1065.16, + "duration": 0.0, + "text": "to roll out. So, Tesla has basically just<00:17:45.360> recorded<00:17:46.120> millions<00:17:46.440> and<00:17:46.520> millions<00:17:46.760> of" + }, + { + "start": 1066.87, + "duration": 0.0, + "text": "just recorded millions and millions of" + }, + { + "start": 1066.88, + "duration": 0.0, + "text": "just recorded millions and millions of hours<00:17:47.080> of<00:17:47.160> just<00:17:47.320> people<00:17:47.560> driving<00:17:48.440> and<00:17:48.680> they're" + }, + { + "start": 1068.83, + "duration": 0.0, + "text": "hours of just people driving and they're" + }, + { + "start": 1068.84, + "duration": 0.0, + "text": "hours of just people driving and they're training<00:17:49.240> massive<00:17:49.640> massive<00:17:50.040> transformers<00:17:50.840> on" + }, + { + "start": 1071.15, + "duration": 0.0, + "text": "training massive massive transformers on" + }, + { + "start": 1071.16, + "duration": 0.0, + "text": "training massive massive transformers on human<00:17:51.400> behavior<00:17:51.840> and<00:17:51.920> they're<00:17:52.000> just<00:17:52.120> trying" + }, + { + "start": 1072.23, + "duration": 0.0, + "text": "human behavior and they're just trying" + }, + { + "start": 1072.24, + "duration": 0.0, + "text": "human behavior and they're just trying to<00:17:52.320> model<00:17:52.640> it,<00:17:52.800> right?" + }, + { + "start": 1074.11, + "duration": 0.0, + "text": "to model it, right?" + }, + { + "start": 1074.12, + "duration": 0.0, + "text": "to model it, right? It's<00:17:54.280> also<00:17:54.640> There's<00:17:54.800> another<00:17:55.040> parallel<00:17:55.400> with" + }, + { + "start": 1075.47, + "duration": 0.0, + "text": "It's also There's another parallel with" + }, + { + "start": 1075.48, + "duration": 0.0, + "text": "It's also There's another parallel with like<00:17:55.680> Siri<00:17:56.040> and<00:17:56.160> Echo<00:17:56.360> Dot,<00:17:56.680> right?<00:17:56.800> These<00:17:56.960> are" + }, + { + "start": 1077.03, + "duration": 0.0, + "text": "like Siri and Echo Dot, right? These are" + }, + { + "start": 1077.04, + "duration": 0.0, + "text": "like Siri and Echo Dot, right? These are symbolic<00:17:57.480> systems.<00:17:57.920> They're<00:17:58.080> often" + }, + { + "start": 1078.31, + "duration": 0.0, + "text": "symbolic systems. They're often" + }, + { + "start": 1078.32, + "duration": 0.0, + "text": "symbolic systems. They're often frustratingly<00:17:58.880> unintelligent.<00:17:59.680> Um" + }, + { + "start": 1080.47, + "duration": 0.0, + "text": "frustratingly unintelligent. Um" + }, + { + "start": 1080.48, + "duration": 0.0, + "text": "frustratingly unintelligent. Um to<00:18:00.600> add<00:18:00.760> new<00:18:00.880> intelligent<00:18:01.240> behavior,<00:18:01.760> an" + }, + { + "start": 1081.91, + "duration": 0.0, + "text": "to add new intelligent behavior, an" + }, + { + "start": 1081.92, + "duration": 0.0, + "text": "to add new intelligent behavior, an engineer<00:18:02.320> has<00:18:02.480> to<00:18:02.560> basically,<00:18:03.520> you<00:18:03.600> know," + }, + { + "start": 1083.67, + "duration": 0.0, + "text": "engineer has to basically, you know," + }, + { + "start": 1083.68, + "duration": 0.0, + "text": "engineer has to basically, you know, design<00:18:04.040> the<00:18:04.120> algorithm<00:18:04.480> and<00:18:04.560> all<00:18:04.640> that<00:18:04.800> stuff." + }, + { + "start": 1085.03, + "duration": 0.0, + "text": "design the algorithm and all that stuff." + }, + { + "start": 1085.04, + "duration": 0.0, + "text": "design the algorithm and all that stuff. And<00:18:05.200> ChatGPT<00:18:05.800> is<00:18:05.920> statistical.<00:18:06.520> It's<00:18:06.680> deep" + }, + { + "start": 1086.87, + "duration": 0.0, + "text": "And ChatGPT is statistical. It's deep" + }, + { + "start": 1086.88, + "duration": 0.0, + "text": "And ChatGPT is statistical. It's deep learning.<00:18:07.560> It's<00:18:07.800> automatic<00:18:08.360> programming," + }, + { + "start": 1088.83, + "duration": 0.0, + "text": "learning. It's automatic programming," + }, + { + "start": 1088.84, + "duration": 0.0, + "text": "learning. It's automatic programming, right?<00:18:09.560> Um" + }, + { + "start": 1090.71, + "duration": 0.0, + "text": "right? Um" + }, + { + "start": 1090.72, + "duration": 0.0, + "text": "right? Um Deep<00:18:10.920> learning<00:18:11.080> models<00:18:11.320> don't<00:18:11.600> learn.<00:18:11.960> That" + }, + { + "start": 1092.11, + "duration": 0.0, + "text": "Deep learning models don't learn. That" + }, + { + "start": 1092.12, + "duration": 0.0, + "text": "Deep learning models don't learn. That always<00:18:12.320> bugs<00:18:12.520> me.<00:18:12.600> I<00:18:12.680> think<00:18:12.840> it's<00:18:12.960> a<00:18:13.000> misnomer." + }, + { + "start": 1094.19, + "duration": 0.0, + "text": "always bugs me. I think it's a misnomer." + }, + { + "start": 1094.2, + "duration": 0.0, + "text": "always bugs me. I think it's a misnomer. Um<00:18:14.280> that's<00:18:14.400> like<00:18:14.520> saying<00:18:14.680> when<00:18:14.800> you're<00:18:15.000> game" + }, + { + "start": 1095.15, + "duration": 0.0, + "text": "Um that's like saying when you're game" + }, + { + "start": 1095.16, + "duration": 0.0, + "text": "Um that's like saying when you're game developer,<00:18:15.400> you're<00:18:15.520> creating<00:18:16.000> a<00:18:16.040> universe." + }, + { + "start": 1096.51, + "duration": 0.0, + "text": "developer, you're creating a universe." + }, + { + "start": 1096.52, + "duration": 0.0, + "text": "developer, you're creating a universe. It's<00:18:16.640> like<00:18:16.760> a<00:18:16.800> very<00:18:17.080> like<00:18:17.680> hand-wavy<00:18:18.400> way<00:18:18.560> to" + }, + { + "start": 1098.67, + "duration": 0.0, + "text": "It's like a very like hand-wavy way to" + }, + { + "start": 1098.68, + "duration": 0.0, + "text": "It's like a very like hand-wavy way to say<00:18:18.920> something<00:18:19.160> that<00:18:19.320> I<00:18:19.360> think<00:18:19.560> is<00:18:19.680> really" + }, + { + "start": 1099.87, + "duration": 0.0, + "text": "say something that I think is really" + }, + { + "start": 1099.88, + "duration": 0.0, + "text": "say something that I think is really mechanical<00:18:20.400> and<00:18:20.480> mundane,<00:18:20.880> right?<00:18:21.360> I<00:18:21.480> think" + }, + { + "start": 1101.67, + "duration": 0.0, + "text": "mechanical and mundane, right? I think" + }, + { + "start": 1101.68, + "duration": 0.0, + "text": "mechanical and mundane, right? I think it's<00:18:22.000> more<00:18:22.280> accurate<00:18:22.560> way<00:18:22.680> to<00:18:22.960> uh<00:18:23.480> describe" + }, + { + "start": 1103.87, + "duration": 0.0, + "text": "it's more accurate way to uh describe" + }, + { + "start": 1103.88, + "duration": 0.0, + "text": "it's more accurate way to uh describe what<00:18:24.000> happens<00:18:24.240> with<00:18:24.320> deep<00:18:24.480> learning<00:18:24.720> is" + }, + { + "start": 1104.87, + "duration": 0.0, + "text": "what happens with deep learning is" + }, + { + "start": 1104.88, + "duration": 0.0, + "text": "what happens with deep learning is optimizing,<00:18:25.440> right?<00:18:25.600> And<00:18:25.680> it's<00:18:25.800> an<00:18:25.960> automatic" + }, + { + "start": 1106.43, + "duration": 0.0, + "text": "optimizing, right? And it's an automatic" + }, + { + "start": 1106.44, + "duration": 0.0, + "text": "optimizing, right? And it's an automatic optimization.<00:18:27.560> Um" + }, + { + "start": 1108.83, + "duration": 0.0, + "text": "optimization. Um" + }, + { + "start": 1108.84, + "duration": 0.0, + "text": "optimization. Um Here's<00:18:29.000> just<00:18:29.120> another<00:18:29.360> side<00:18:29.640> note,<00:18:29.800> right?" + }, + { + "start": 1109.95, + "duration": 0.0, + "text": "Here's just another side note, right?" + }, + { + "start": 1109.96, + "duration": 0.0, + "text": "Here's just another side note, right? So,<00:18:30.040> there's<00:18:30.200> big<00:18:30.360> copyright<00:18:30.760> disputes<00:18:31.080> with" + }, + { + "start": 1111.15, + "duration": 0.0, + "text": "So, there's big copyright disputes with" + }, + { + "start": 1111.16, + "duration": 0.0, + "text": "So, there's big copyright disputes with these<00:18:31.320> deep<00:18:31.520> learning<00:18:31.720> models,<00:18:32.080> right?<00:18:32.200> And<00:18:32.320> I" + }, + { + "start": 1112.35, + "duration": 0.0, + "text": "these deep learning models, right? And I" + }, + { + "start": 1112.36, + "duration": 0.0, + "text": "these deep learning models, right? And I just<00:18:32.760> had<00:18:32.920> a<00:18:32.960> funny<00:18:33.200> idea,<00:18:33.560> right?<00:18:33.760> Like" + }, + { + "start": 1115.03, + "duration": 0.0, + "text": "just had a funny idea, right? Like" + }, + { + "start": 1115.04, + "duration": 0.0, + "text": "just had a funny idea, right? Like you<00:18:35.160> could<00:18:35.280> almost<00:18:35.520> make<00:18:35.640> the<00:18:35.760> ar-<00:18:36.120> if<00:18:36.280> you" + }, + { + "start": 1116.39, + "duration": 0.0, + "text": "you could almost make the ar- if you" + }, + { + "start": 1116.4, + "duration": 0.0, + "text": "you could almost make the ar- if you take<00:18:36.560> the<00:18:36.720> argument<00:18:37.120> that<00:18:37.400> uh<00:18:37.480> the<00:18:37.560> deep" + }, + { + "start": 1117.75, + "duration": 0.0, + "text": "take the argument that uh the deep" + }, + { + "start": 1117.76, + "duration": 0.0, + "text": "take the argument that uh the deep learning<00:18:37.960> people<00:18:38.200> are<00:18:38.240> trying<00:18:38.440> to<00:18:38.480> make," + }, + { + "start": 1118.75, + "duration": 0.0, + "text": "learning people are trying to make," + }, + { + "start": 1118.76, + "duration": 0.0, + "text": "learning people are trying to make, right?<00:18:39.080> They're<00:18:39.160> trying<00:18:39.320> to<00:18:39.360> secure<00:18:39.640> their" + }, + { + "start": 1119.75, + "duration": 0.0, + "text": "right? They're trying to secure their" + }, + { + "start": 1119.76, + "duration": 0.0, + "text": "right? They're trying to secure their bag.<00:18:40.000> They're<00:18:40.080> trying<00:18:40.240> to<00:18:40.320> say,<00:18:40.480> \"Well,<00:18:41.120> we" + }, + { + "start": 1121.31, + "duration": 0.0, + "text": "bag. They're trying to say, \"Well, we" + }, + { + "start": 1121.32, + "duration": 0.0, + "text": "bag. They're trying to say, \"Well, we learned<00:18:41.640> the<00:18:41.720> patterns<00:18:42.240> from<00:18:42.440> what<00:18:42.560> you<00:18:42.640> gave" + }, + { + "start": 1122.87, + "duration": 0.0, + "text": "learned the patterns from what you gave" + }, + { + "start": 1122.88, + "duration": 0.0, + "text": "learned the patterns from what you gave us.<00:18:43.040> We<00:18:43.120> didn't,<00:18:43.480> you<00:18:43.560> know,<00:18:43.720> copy<00:18:44.040> the<00:18:44.120> pixels" + }, + { + "start": 1124.47, + "duration": 0.0, + "text": "us. We didn't, you know, copy the pixels" + }, + { + "start": 1124.48, + "duration": 0.0, + "text": "us. We didn't, you know, copy the pixels or<00:18:44.520> whatever.\"<00:18:45.200> That's<00:18:45.400> like<00:18:45.520> what<00:18:45.640> JPEG" + }, + { + "start": 1125.99, + "duration": 0.0, + "text": "or whatever.\" That's like what JPEG" + }, + { + "start": 1126.0, + "duration": 0.0, + "text": "or whatever.\" That's like what JPEG does,<00:18:46.520> like<00:18:46.720> learns<00:18:46.960> a<00:18:47.000> model<00:18:47.360> of<00:18:47.480> the<00:18:47.560> pixels." + }, + { + "start": 1128.51, + "duration": 0.0, + "text": "does, like learns a model of the pixels." + }, + { + "start": 1128.52, + "duration": 0.0, + "text": "does, like learns a model of the pixels. And<00:18:48.640> so,<00:18:48.720> it's<00:18:48.840> almost<00:18:49.080> like<00:18:49.200> arguing<00:18:49.560> that" + }, + { + "start": 1129.71, + "duration": 0.0, + "text": "And so, it's almost like arguing that" + }, + { + "start": 1129.72, + "duration": 0.0, + "text": "And so, it's almost like arguing that because,<00:18:50.240> you<00:18:50.360> know,<00:18:50.440> it's<00:18:50.600> JPEG<00:18:51.000> and<00:18:51.080> not" + }, + { + "start": 1131.19, + "duration": 0.0, + "text": "because, you know, it's JPEG and not" + }, + { + "start": 1131.2, + "duration": 0.0, + "text": "because, you know, it's JPEG and not PNG,<00:18:52.200> uh<00:18:52.280> because<00:18:52.520> it's<00:18:52.640> a<00:18:52.680> model<00:18:52.960> of<00:18:53.040> the" + }, + { + "start": 1133.11, + "duration": 0.0, + "text": "PNG, uh because it's a model of the" + }, + { + "start": 1133.12, + "duration": 0.0, + "text": "PNG, uh because it's a model of the pixels,<00:18:53.560> you<00:18:53.600> know,<00:18:53.800> I<00:18:53.880> didn't<00:18:54.120> steal<00:18:54.280> your" + }, + { + "start": 1134.39, + "duration": 0.0, + "text": "pixels, you know, I didn't steal your" + }, + { + "start": 1134.4, + "duration": 0.0, + "text": "pixels, you know, I didn't steal your thing." + }, + { + "start": 1135.39, + "duration": 0.0, + "text": "thing." + }, + { + "start": 1135.4, + "duration": 0.0, + "text": "thing. Um<00:18:55.560> so,<00:18:55.680> it's<00:18:55.760> a<00:18:55.800> little<00:18:56.040> strange,<00:18:56.760> right?<00:18:57.040> Um" + }, + { + "start": 1137.19, + "duration": 0.0, + "text": "Um so, it's a little strange, right? Um" + }, + { + "start": 1137.2, + "duration": 0.0, + "text": "Um so, it's a little strange, right? Um I<00:18:57.240> think<00:18:57.400> it's<00:18:57.560> kind<00:18:57.760> of<00:18:57.840> a<00:18:57.880> bad<00:18:58.160> faith" + }, + { + "start": 1138.35, + "duration": 0.0, + "text": "I think it's kind of a bad faith" + }, + { + "start": 1138.36, + "duration": 0.0, + "text": "I think it's kind of a bad faith argument<00:18:58.720> cuz<00:18:59.360> the<00:18:59.480> patterns<00:19:00.040> uh<00:19:00.320> that<00:19:00.480> are" + }, + { + "start": 1140.79, + "duration": 0.0, + "text": "argument cuz the patterns uh that are" + }, + { + "start": 1140.8, + "duration": 0.0, + "text": "argument cuz the patterns uh that are found<00:19:01.120> in<00:19:01.240> deep<00:19:01.440> learning<00:19:02.320> um<00:19:02.960> are<00:19:03.120> very" + }, + { + "start": 1143.35, + "duration": 0.0, + "text": "found in deep learning um are very" + }, + { + "start": 1143.36, + "duration": 0.0, + "text": "found in deep learning um are very generic<00:19:03.800> and<00:19:03.880> they're<00:19:03.960> not<00:19:04.120> so<00:19:04.240> specific<00:19:04.720> to" + }, + { + "start": 1144.79, + "duration": 0.0, + "text": "generic and they're not so specific to" + }, + { + "start": 1144.8, + "duration": 0.0, + "text": "generic and they're not so specific to the<00:19:04.920> image.<00:19:05.280> And<00:19:05.400> so," + }, + { + "start": 1146.47, + "duration": 0.0, + "text": "the image. And so," + }, + { + "start": 1146.48, + "duration": 0.0, + "text": "the image. And so, it's<00:19:06.520> very<00:19:06.680> debatable,<00:19:07.120> I<00:19:07.200> think.<00:19:07.360> I'm<00:19:07.560> not" + }, + { + "start": 1147.67, + "duration": 0.0, + "text": "it's very debatable, I think. I'm not" + }, + { + "start": 1147.68, + "duration": 0.0, + "text": "it's very debatable, I think. I'm not sure<00:19:07.760> what's<00:19:07.880> going<00:19:08.000> to<00:19:08.040> happen.<00:19:08.960> But<00:19:09.040> just<00:19:09.200> to" + }, + { + "start": 1149.27, + "duration": 0.0, + "text": "sure what's going to happen. But just to" + }, + { + "start": 1149.28, + "duration": 0.0, + "text": "sure what's going to happen. But just to go<00:19:09.360> back<00:19:09.520> to<00:19:09.600> my<00:19:09.720> League<00:19:09.880> of<00:19:10.000> Legends<00:19:10.360> task," + }, + { + "start": 1150.71, + "duration": 0.0, + "text": "go back to my League of Legends task," + }, + { + "start": 1150.72, + "duration": 0.0, + "text": "go back to my League of Legends task, right?<00:19:10.920> If<00:19:11.000> we<00:19:11.120> if<00:19:11.200> we<00:19:11.280> just<00:19:11.400> want<00:19:11.520> to<00:19:11.560> think" + }, + { + "start": 1151.71, + "duration": 0.0, + "text": "right? If we if we just want to think" + }, + { + "start": 1151.72, + "duration": 0.0, + "text": "right? If we if we just want to think about<00:19:11.920> that<00:19:12.040> little<00:19:12.200> bit<00:19:12.320> more<00:19:12.440> as<00:19:12.520> an" + }, + { + "start": 1152.59, + "duration": 0.0, + "text": "about that little bit more as an" + }, + { + "start": 1152.6, + "duration": 0.0, + "text": "about that little bit more as an engineer," + }, + { + "start": 1153.71, + "duration": 0.0, + "text": "engineer," + }, + { + "start": 1153.72, + "duration": 0.0, + "text": "engineer, um<00:19:13.840> you<00:19:13.960> might<00:19:14.120> naively<00:19:14.560> just<00:19:14.760> put<00:19:14.920> in<00:19:15.000> the" + }, + { + "start": 1155.07, + "duration": 0.0, + "text": "um you might naively just put in the" + }, + { + "start": 1155.08, + "duration": 0.0, + "text": "um you might naively just put in the current<00:19:15.360> state<00:19:15.640> of<00:19:15.760> where<00:19:15.960> all<00:19:16.080> the<00:19:16.160> people" + }, + { + "start": 1156.43, + "duration": 0.0, + "text": "current state of where all the people" + }, + { + "start": 1156.44, + "duration": 0.0, + "text": "current state of where all the people are<00:19:16.560> on<00:19:16.600> the<00:19:16.680> map,<00:19:17.000> how<00:19:17.120> much<00:19:17.280> gold<00:19:17.480> they<00:19:17.560> have," + }, + { + "start": 1157.71, + "duration": 0.0, + "text": "are on the map, how much gold they have," + }, + { + "start": 1157.72, + "duration": 0.0, + "text": "are on the map, how much gold they have, what<00:19:17.840> items<00:19:18.080> they<00:19:18.160> have,<00:19:18.840> throw<00:19:19.040> it<00:19:19.120> through" + }, + { + "start": 1159.23, + "duration": 0.0, + "text": "what items they have, throw it through" + }, + { + "start": 1159.24, + "duration": 0.0, + "text": "what items they have, throw it through that<00:19:19.360> perceptron<00:19:19.880> I<00:19:19.920> showed<00:19:20.200> you,<00:19:20.760> and<00:19:20.880> maybe" + }, + { + "start": 1161.11, + "duration": 0.0, + "text": "that perceptron I showed you, and maybe" + }, + { + "start": 1161.12, + "duration": 0.0, + "text": "that perceptron I showed you, and maybe it'll<00:19:21.320> automatically<00:19:21.840> find<00:19:22.080> ways<00:19:22.280> to<00:19:22.400> relate," + }, + { + "start": 1162.79, + "duration": 0.0, + "text": "it'll automatically find ways to relate," + }, + { + "start": 1162.8, + "duration": 0.0, + "text": "it'll automatically find ways to relate, oh,<00:19:23.240> like,<00:19:23.600> you<00:19:23.680> know,<00:19:23.800> when<00:19:23.920> the<00:19:24.000> gold<00:19:24.400> and" + }, + { + "start": 1164.47, + "duration": 0.0, + "text": "oh, like, you know, when the gold and" + }, + { + "start": 1164.48, + "duration": 0.0, + "text": "oh, like, you know, when the gold and this<00:19:24.720> item<00:19:24.920> is<00:19:25.040> here,<00:19:25.280> then<00:19:25.440> like<00:19:25.880> they're" + }, + { + "start": 1165.95, + "duration": 0.0, + "text": "this item is here, then like they're" + }, + { + "start": 1165.96, + "duration": 0.0, + "text": "this item is here, then like they're more<00:19:26.320> likely<00:19:26.600> to<00:19:26.680> win<00:19:26.840> than<00:19:26.960> you<00:19:27.000> would<00:19:27.120> think," + }, + { + "start": 1167.35, + "duration": 0.0, + "text": "more likely to win than you would think," + }, + { + "start": 1167.36, + "duration": 0.0, + "text": "more likely to win than you would think, and<00:19:27.440> stuff<00:19:27.600> like<00:19:27.760> that.<00:19:28.480> These<00:19:28.600> are<00:19:28.640> the<00:19:28.720> kind" + }, + { + "start": 1168.83, + "duration": 0.0, + "text": "and stuff like that. These are the kind" + }, + { + "start": 1168.84, + "duration": 0.0, + "text": "and stuff like that. These are the kind of<00:19:28.920> patterns<00:19:29.280> you<00:19:29.600> hope<00:19:29.840> a<00:19:29.920> deep<00:19:30.120> learning" + }, + { + "start": 1170.35, + "duration": 0.0, + "text": "of patterns you hope a deep learning" + }, + { + "start": 1170.36, + "duration": 0.0, + "text": "of patterns you hope a deep learning model<00:19:30.600> finds<00:19:31.040> as<00:19:31.200> it<00:19:31.360> optimizes<00:19:31.960> based<00:19:32.200> on" + }, + { + "start": 1172.47, + "duration": 0.0, + "text": "model finds as it optimizes based on" + }, + { + "start": 1172.48, + "duration": 0.0, + "text": "model finds as it optimizes based on massive<00:19:32.800> amount<00:19:32.960> of<00:19:33.040> data<00:19:33.200> you<00:19:33.320> have,<00:19:33.600> right?" + }, + { + "start": 1175.07, + "duration": 0.0, + "text": "massive amount of data you have, right?" + }, + { + "start": 1175.08, + "duration": 0.0, + "text": "massive amount of data you have, right? There's<00:19:35.200> a<00:19:35.240> problem<00:19:35.520> with<00:19:35.600> this,<00:19:35.760> though." + }, + { + "start": 1175.87, + "duration": 0.0, + "text": "There's a problem with this, though." + }, + { + "start": 1175.88, + "duration": 0.0, + "text": "There's a problem with this, though. What<00:19:36.000> about<00:19:36.200> past<00:19:36.680> states,<00:19:37.240> right?<00:19:37.440> Past" + }, + { + "start": 1177.67, + "duration": 0.0, + "text": "What about past states, right? Past" + }, + { + "start": 1177.68, + "duration": 0.0, + "text": "What about past states, right? Past state<00:19:38.160> informs<00:19:38.720> the<00:19:38.800> nature<00:19:39.080> of<00:19:39.200> present" + }, + { + "start": 1179.55, + "duration": 0.0, + "text": "state informs the nature of present" + }, + { + "start": 1179.56, + "duration": 0.0, + "text": "state informs the nature of present state.<00:19:40.280> If<00:19:40.400> somebody<00:19:40.600> was<00:19:40.720> performing<00:19:41.120> badly" + }, + { + "start": 1181.43, + "duration": 0.0, + "text": "state. If somebody was performing badly" + }, + { + "start": 1181.44, + "duration": 0.0, + "text": "state. If somebody was performing badly on<00:19:41.520> your<00:19:41.600> team<00:19:41.760> in<00:19:41.840> the<00:19:41.920> past,<00:19:42.400> right?<00:19:42.600> They" + }, + { + "start": 1182.67, + "duration": 0.0, + "text": "on your team in the past, right? They" + }, + { + "start": 1182.68, + "duration": 0.0, + "text": "on your team in the past, right? They were<00:19:42.840> like<00:19:43.080> running<00:19:43.240> it<00:19:43.320> down<00:19:43.480> mid,<00:19:44.000> this" + }, + { + "start": 1184.15, + "duration": 0.0, + "text": "were like running it down mid, this" + }, + { + "start": 1184.16, + "duration": 0.0, + "text": "were like running it down mid, this should<00:19:44.280> inform<00:19:44.640> your<00:19:44.760> current<00:19:45.000> prediction" + }, + { + "start": 1185.35, + "duration": 0.0, + "text": "should inform your current prediction" + }, + { + "start": 1185.36, + "duration": 0.0, + "text": "should inform your current prediction like<00:19:45.520> heavily,<00:19:45.840> right?<00:19:46.680> Um" + }, + { + "start": 1187.95, + "duration": 0.0, + "text": "like heavily, right? Um" + }, + { + "start": 1187.96, + "duration": 0.0, + "text": "like heavily, right? Um So,<00:19:48.080> this<00:19:48.560> uh<00:19:48.720> you<00:19:48.800> know" + }, + { + "start": 1189.99, + "duration": 0.0, + "text": "So, this uh you know" + }, + { + "start": 1190.0, + "duration": 0.0, + "text": "So, this uh you know You<00:19:50.040> need<00:19:50.200> a<00:19:50.240> recurrent<00:19:50.640> neural<00:19:50.760> network<00:19:51.040> for" + }, + { + "start": 1191.11, + "duration": 0.0, + "text": "You need a recurrent neural network for" + }, + { + "start": 1191.12, + "duration": 0.0, + "text": "You need a recurrent neural network for this,<00:19:51.360> right?<00:19:51.880> A<00:19:51.960> recurrent<00:19:52.320> neural<00:19:52.440> network" + }, + { + "start": 1192.67, + "duration": 0.0, + "text": "this, right? A recurrent neural network" + }, + { + "start": 1192.68, + "duration": 0.0, + "text": "this, right? A recurrent neural network is<00:19:52.800> basically<00:19:53.200> two<00:19:53.480> of<00:19:53.560> those<00:19:53.760> perceptrons" + }, + { + "start": 1194.43, + "duration": 0.0, + "text": "is basically two of those perceptrons" + }, + { + "start": 1194.44, + "duration": 0.0, + "text": "is basically two of those perceptrons that<00:19:54.560> I<00:19:54.600> showed<00:19:54.880> you.<00:19:55.520> One<00:19:55.760> is<00:19:55.920> modifying<00:19:56.560> the" + }, + { + "start": 1196.63, + "duration": 0.0, + "text": "that I showed you. One is modifying the" + }, + { + "start": 1196.64, + "duration": 0.0, + "text": "that I showed you. One is modifying the state<00:19:56.960> almost,<00:19:57.360> and<00:19:57.440> then<00:19:57.560> another<00:19:57.800> one<00:19:57.960> is" + }, + { + "start": 1198.07, + "duration": 0.0, + "text": "state almost, and then another one is" + }, + { + "start": 1198.08, + "duration": 0.0, + "text": "state almost, and then another one is visualizing<00:19:58.680> the<00:19:58.760> state<00:19:59.080> towards<00:19:59.400> whatever" + }, + { + "start": 1199.79, + "duration": 0.0, + "text": "visualizing the state towards whatever" + }, + { + "start": 1199.8, + "duration": 0.0, + "text": "visualizing the state towards whatever like<00:20:00.040> you<00:20:00.120> want<00:20:00.320> it<00:20:00.400> to<00:20:01.120> visualize,<00:20:01.560> right?<00:20:01.720> It" + }, + { + "start": 1201.79, + "duration": 0.0, + "text": "like you want it to visualize, right? It" + }, + { + "start": 1201.8, + "duration": 0.0, + "text": "like you want it to visualize, right? It might<00:20:01.920> be<00:20:02.040> probability<00:20:02.480> for<00:20:02.600> who's<00:20:02.760> going<00:20:02.880> to" + }, + { + "start": 1202.95, + "duration": 0.0, + "text": "might be probability for who's going to" + }, + { + "start": 1202.96, + "duration": 0.0, + "text": "might be probability for who's going to win,<00:20:03.280> or<00:20:03.360> it<00:20:03.440> might<00:20:03.600> be<00:20:03.680> like<00:20:04.400> you<00:20:04.520> could" + }, + { + "start": 1204.63, + "duration": 0.0, + "text": "win, or it might be like you could" + }, + { + "start": 1204.64, + "duration": 0.0, + "text": "win, or it might be like you could replace<00:20:04.920> it<00:20:05.000> with<00:20:05.120> another<00:20:05.640> deep<00:20:05.840> learning" + }, + { + "start": 1206.11, + "duration": 0.0, + "text": "replace it with another deep learning" + }, + { + "start": 1206.12, + "duration": 0.0, + "text": "replace it with another deep learning network<00:20:06.400> and<00:20:06.480> train<00:20:06.680> it<00:20:06.760> to<00:20:06.840> visualize<00:20:07.360> like" + }, + { + "start": 1208.63, + "duration": 0.0, + "text": "network and train it to visualize like" + }, + { + "start": 1208.64, + "duration": 0.0, + "text": "network and train it to visualize like I<00:20:08.680> don't<00:20:08.800> know,<00:20:08.880> which<00:20:09.200> which<00:20:09.400> champion<00:20:09.800> is" + }, + { + "start": 1209.91, + "duration": 0.0, + "text": "I don't know, which which champion is" + }, + { + "start": 1209.92, + "duration": 0.0, + "text": "I don't know, which which champion is like<00:20:10.280> has<00:20:10.480> the<00:20:10.520> most<00:20:10.680> synergy<00:20:10.920> with<00:20:11.040> the<00:20:11.080> other" + }, + { + "start": 1211.19, + "duration": 0.0, + "text": "like has the most synergy with the other" + }, + { + "start": 1211.2, + "duration": 0.0, + "text": "like has the most synergy with the other champion,<00:20:11.560> stuff<00:20:11.720> like<00:20:11.840> this,<00:20:12.160> right?<00:20:12.960> Um<00:20:13.400> it" + }, + { + "start": 1213.51, + "duration": 0.0, + "text": "champion, stuff like this, right? Um it" + }, + { + "start": 1213.52, + "duration": 0.0, + "text": "champion, stuff like this, right? Um it has<00:20:13.680> state.<00:20:14.040> It's<00:20:14.160> much<00:20:14.360> more<00:20:14.520> like<00:20:14.680> a" + }, + { + "start": 1214.83, + "duration": 0.0, + "text": "has state. It's much more like a" + }, + { + "start": 1214.84, + "duration": 0.0, + "text": "has state. It's much more like a automatic<00:20:15.240> program,<00:20:15.600> right?<00:20:15.840> Automatically" + }, + { + "start": 1216.23, + "duration": 0.0, + "text": "automatic program, right? Automatically" + }, + { + "start": 1216.24, + "duration": 0.0, + "text": "automatic program, right? Automatically learned<00:20:16.520> program.<00:20:17.480> Um<00:20:17.840> the<00:20:17.960> hidden<00:20:18.160> state<00:20:18.400> of" + }, + { + "start": 1218.47, + "duration": 0.0, + "text": "learned program. Um the hidden state of" + }, + { + "start": 1218.48, + "duration": 0.0, + "text": "learned program. Um the hidden state of a<00:20:18.520> recurrent<00:20:18.800> neural<00:20:18.920> network<00:20:19.200> is<00:20:19.320> like<00:20:19.560> RAM." + }, + { + "start": 1219.83, + "duration": 0.0, + "text": "a recurrent neural network is like RAM." + }, + { + "start": 1219.84, + "duration": 0.0, + "text": "a recurrent neural network is like RAM. It's<00:20:19.920> analogous<00:20:20.280> to<00:20:20.360> a<00:20:20.400> RAM." + }, + { + "start": 1221.35, + "duration": 0.0, + "text": "It's analogous to a RAM." + }, + { + "start": 1221.36, + "duration": 0.0, + "text": "It's analogous to a RAM. Um<00:20:21.920> and<00:20:22.040> that<00:20:22.160> and<00:20:22.240> that<00:20:22.360> is<00:20:22.440> why<00:20:22.560> I<00:20:22.640> think<00:20:22.840> that" + }, + { + "start": 1222.95, + "duration": 0.0, + "text": "Um and that and that is why I think that" + }, + { + "start": 1222.96, + "duration": 0.0, + "text": "Um and that and that is why I think that these<00:20:23.160> are" + }, + { + "start": 1224.07, + "duration": 0.0, + "text": "these are" + }, + { + "start": 1224.08, + "duration": 0.0, + "text": "these are best<00:20:24.320> described<00:20:24.640> as<00:20:24.760> automatic<00:20:25.120> programs," + }, + { + "start": 1225.51, + "duration": 0.0, + "text": "best described as automatic programs," + }, + { + "start": 1225.52, + "duration": 0.0, + "text": "best described as automatic programs, right?<00:20:26.280> Um<00:20:26.720> and<00:20:26.880> so,<00:20:27.000> what<00:20:27.160> I<00:20:27.240> did<00:20:27.760> at<00:20:27.880> the<00:20:28.000> time" + }, + { + "start": 1228.51, + "duration": 0.0, + "text": "right? Um and so, what I did at the time" + }, + { + "start": 1228.52, + "duration": 0.0, + "text": "right? Um and so, what I did at the time to<00:20:28.920> successfully<00:20:29.400> model<00:20:30.400> uh<00:20:30.520> who<00:20:30.640> would<00:20:30.760> win" + }, + { + "start": 1230.87, + "duration": 0.0, + "text": "to successfully model uh who would win" + }, + { + "start": 1230.88, + "duration": 0.0, + "text": "to successfully model uh who would win in<00:20:30.960> League<00:20:31.120> of<00:20:31.200> Legends<00:20:31.440> games<00:20:31.800> was<00:20:31.960> I<00:20:32.160> I<00:20:32.360> I<00:20:32.600> I" + }, + { + "start": 1232.71, + "duration": 0.0, + "text": "in League of Legends games was I I I I" + }, + { + "start": 1232.72, + "duration": 0.0, + "text": "in League of Legends games was I I I I selected<00:20:33.040> a<00:20:33.080> GRU<00:20:33.360> architecture,<00:20:34.280> and<00:20:34.400> I<00:20:34.440> spent" + }, + { + "start": 1234.71, + "duration": 0.0, + "text": "selected a GRU architecture, and I spent" + }, + { + "start": 1234.72, + "duration": 0.0, + "text": "selected a GRU architecture, and I spent weeks<00:20:34.960> tuning<00:20:35.240> it<00:20:35.400> and<00:20:35.640> and<00:20:36.400> messing<00:20:36.760> around," + }, + { + "start": 1237.19, + "duration": 0.0, + "text": "weeks tuning it and and messing around," + }, + { + "start": 1237.2, + "duration": 0.0, + "text": "weeks tuning it and and messing around, and<00:20:37.400> I<00:20:37.520> I<00:20:37.640> got<00:20:37.840> I<00:20:37.920> managed<00:20:38.200> to<00:20:38.280> make<00:20:38.440> it<00:20:38.560> so" + }, + { + "start": 1239.11, + "duration": 0.0, + "text": "and I I got I managed to make it so" + }, + { + "start": 1239.12, + "duration": 0.0, + "text": "and I I got I managed to make it so halfway<00:20:39.400> through<00:20:39.560> the<00:20:39.640> game<00:20:39.880> it<00:20:39.960> could" + }, + { + "start": 1240.07, + "duration": 0.0, + "text": "halfway through the game it could" + }, + { + "start": 1240.08, + "duration": 0.0, + "text": "halfway through the game it could predict<00:20:40.360> who<00:20:40.400> would<00:20:40.560> win<00:20:40.720> with<00:20:40.880> 86%<00:20:41.520> accuracy." + }, + { + "start": 1242.55, + "duration": 0.0, + "text": "predict who would win with 86% accuracy." + }, + { + "start": 1242.56, + "duration": 0.0, + "text": "predict who would win with 86% accuracy. Um<00:20:42.720> if<00:20:42.840> you<00:20:42.880> just<00:20:43.040> went<00:20:43.200> with<00:20:43.320> whoever<00:20:43.600> had" + }, + { + "start": 1243.75, + "duration": 0.0, + "text": "Um if you just went with whoever had" + }, + { + "start": 1243.76, + "duration": 0.0, + "text": "Um if you just went with whoever had more<00:20:43.920> gold,<00:20:44.200> it<00:20:44.280> was<00:20:44.360> like<00:20:44.520> 73%<00:20:45.240> accuracy.<00:20:46.240> Um" + }, + { + "start": 1246.43, + "duration": 0.0, + "text": "more gold, it was like 73% accuracy. Um" + }, + { + "start": 1246.44, + "duration": 0.0, + "text": "more gold, it was like 73% accuracy. Um and<00:20:46.560> then<00:20:46.680> this<00:20:46.880> continued<00:20:47.280> to<00:20:47.360> generalize" + }, + { + "start": 1247.99, + "duration": 0.0, + "text": "and then this continued to generalize" + }, + { + "start": 1248.0, + "duration": 0.0, + "text": "and then this continued to generalize even<00:20:48.200> on<00:20:48.320> arbitrarily<00:20:48.800> new<00:20:48.960> games.<00:20:49.920> Um<00:20:50.200> and" + }, + { + "start": 1250.31, + "duration": 0.0, + "text": "even on arbitrarily new games. Um and" + }, + { + "start": 1250.32, + "duration": 0.0, + "text": "even on arbitrarily new games. Um and just<00:20:50.520> as,<00:20:50.800> you<00:20:50.840> know,<00:20:51.040> as<00:20:51.240> a<00:20:51.280> side<00:20:51.520> note<00:20:51.640> for" + }, + { + "start": 1251.71, + "duration": 0.0, + "text": "just as, you know, as a side note for" + }, + { + "start": 1251.72, + "duration": 0.0, + "text": "just as, you know, as a side note for game<00:20:51.920> developers,<00:20:52.680> you<00:20:52.800> could<00:20:52.920> use<00:20:53.080> a" + }, + { + "start": 1253.11, + "duration": 0.0, + "text": "game developers, you could use a" + }, + { + "start": 1253.12, + "duration": 0.0, + "text": "game developers, you could use a technology<00:20:53.600> like<00:20:53.800> this<00:20:54.480> for<00:20:54.640> our<00:20:54.760> play<00:20:54.960> of<00:20:55.040> the" + }, + { + "start": 1255.11, + "duration": 0.0, + "text": "technology like this for our play of the" + }, + { + "start": 1255.12, + "duration": 0.0, + "text": "technology like this for our play of the game<00:20:55.400> kind<00:20:55.560> of<00:20:55.640> feature,<00:20:55.920> right?<00:20:56.120> To<00:20:56.200> see<00:20:56.360> big" + }, + { + "start": 1256.67, + "duration": 0.0, + "text": "game kind of feature, right? To see big" + }, + { + "start": 1256.68, + "duration": 0.0, + "text": "game kind of feature, right? To see big unpredictable<00:20:57.160> swings<00:20:57.440> in<00:20:57.520> who's<00:20:57.720> winning." + }, + { + "start": 1258.51, + "duration": 0.0, + "text": "unpredictable swings in who's winning." + }, + { + "start": 1258.52, + "duration": 0.0, + "text": "unpredictable swings in who's winning. Um<00:20:58.880> we<00:20:59.000> actually<00:20:59.320> did<00:20:59.560> do<00:20:59.720> that.<00:20:59.880> We<00:20:59.960> took<00:21:00.200> a" + }, + { + "start": 1260.43, + "duration": 0.0, + "text": "Um we actually did do that. We took a" + }, + { + "start": 1260.44, + "duration": 0.0, + "text": "Um we actually did do that. We took a our<00:21:00.560> giant<00:21:00.840> corpus<00:21:01.200> of<00:21:01.280> professional<00:21:01.680> games," + }, + { + "start": 1262.51, + "duration": 0.0, + "text": "our giant corpus of professional games," + }, + { + "start": 1262.52, + "duration": 0.0, + "text": "our giant corpus of professional games, and<00:21:02.760> I<00:21:02.800> ran<00:21:03.000> this<00:21:03.160> model<00:21:03.440> and<00:21:03.560> saw<00:21:03.720> what<00:21:03.840> were" + }, + { + "start": 1263.91, + "duration": 0.0, + "text": "and I ran this model and saw what were" + }, + { + "start": 1263.92, + "duration": 0.0, + "text": "and I ran this model and saw what were the<00:21:04.040> biggest<00:21:04.360> swings,<00:21:05.080> and<00:21:05.200> we<00:21:05.320> looked<00:21:05.520> at" + }, + { + "start": 1265.55, + "duration": 0.0, + "text": "the biggest swings, and we looked at" + }, + { + "start": 1265.56, + "duration": 0.0, + "text": "the biggest swings, and we looked at those<00:21:05.720> games,<00:21:06.040> and<00:21:06.120> they<00:21:06.200> were<00:21:06.280> the<00:21:06.360> most" + }, + { + "start": 1266.59, + "duration": 0.0, + "text": "those games, and they were the most" + }, + { + "start": 1266.6, + "duration": 0.0, + "text": "those games, and they were the most exciting<00:21:06.960> matches,<00:21:07.560> right?<00:21:08.320> So,<00:21:08.440> this<00:21:08.600> is" + }, + { + "start": 1268.71, + "duration": 0.0, + "text": "exciting matches, right? So, this is" + }, + { + "start": 1268.72, + "duration": 0.0, + "text": "exciting matches, right? So, this is kind<00:21:08.920> of<00:21:09.000> like<00:21:09.440> just<00:21:09.840> to<00:21:09.960> give<00:21:10.160> you<00:21:10.280> ideas<00:21:10.800> for" + }, + { + "start": 1270.87, + "duration": 0.0, + "text": "kind of like just to give you ideas for" + }, + { + "start": 1270.88, + "duration": 0.0, + "text": "kind of like just to give you ideas for like<00:21:11.000> what<00:21:11.120> this<00:21:11.240> technology<00:21:11.640> can<00:21:11.760> be<00:21:12.240> used" + }, + { + "start": 1272.51, + "duration": 0.0, + "text": "like what this technology can be used" + }, + { + "start": 1272.52, + "duration": 0.0, + "text": "like what this technology can be used for,<00:21:12.720> right?<00:21:13.600> Um<00:21:14.560> each<00:21:14.840> perceptron<00:21:15.320> in<00:21:15.400> a" + }, + { + "start": 1275.43, + "duration": 0.0, + "text": "for, right? Um each perceptron in a" + }, + { + "start": 1275.44, + "duration": 0.0, + "text": "for, right? Um each perceptron in a layer<00:21:15.680> is<00:21:15.800> like<00:21:16.000> roughly<00:21:16.360> analogous<00:21:16.880> to<00:21:17.000> an<00:21:17.120> if" + }, + { + "start": 1277.23, + "duration": 0.0, + "text": "layer is like roughly analogous to an if" + }, + { + "start": 1277.24, + "duration": 0.0, + "text": "layer is like roughly analogous to an if statement<00:21:17.560> in<00:21:17.600> the<00:21:17.680> program,<00:21:18.000> right?<00:21:18.120> Like" + }, + { + "start": 1278.23, + "duration": 0.0, + "text": "statement in the program, right? Like" + }, + { + "start": 1278.24, + "duration": 0.0, + "text": "statement in the program, right? Like another<00:21:18.440> branch.<00:21:19.240> It<00:21:19.400> lets<00:21:19.560> you<00:21:19.760> make<00:21:20.000> a<00:21:20.040> more" + }, + { + "start": 1280.23, + "duration": 0.0, + "text": "another branch. It lets you make a more" + }, + { + "start": 1280.24, + "duration": 0.0, + "text": "another branch. It lets you make a more specific<00:21:20.680> contour<00:21:21.160> around<00:21:21.960> what<00:21:22.080> you're" + }, + { + "start": 1282.19, + "duration": 0.0, + "text": "specific contour around what you're" + }, + { + "start": 1282.2, + "duration": 0.0, + "text": "specific contour around what you're trying<00:21:22.480> to<00:21:22.680> encapsulate,<00:21:23.360> right?<00:21:24.040> Um<00:21:24.800> you're" + }, + { + "start": 1284.91, + "duration": 0.0, + "text": "trying to encapsulate, right? Um you're" + }, + { + "start": 1284.92, + "duration": 0.0, + "text": "trying to encapsulate, right? Um you're adding<00:21:25.120> more<00:21:25.200> resolution<00:21:25.600> to<00:21:25.640> that<00:21:25.760> boundary." + }, + { + "start": 1286.71, + "duration": 0.0, + "text": "adding more resolution to that boundary." + }, + { + "start": 1286.72, + "duration": 0.0, + "text": "adding more resolution to that boundary. Um<00:21:26.800> this<00:21:26.920> is<00:21:27.000> why<00:21:27.040> it's<00:21:27.160> called<00:21:27.360> deep<00:21:27.640> learning" + }, + { + "start": 1287.95, + "duration": 0.0, + "text": "Um this is why it's called deep learning" + }, + { + "start": 1287.96, + "duration": 0.0, + "text": "Um this is why it's called deep learning cuz<00:21:28.080> there's<00:21:28.240> many<00:21:28.880> Once<00:21:29.040> you<00:21:29.160> add<00:21:29.320> many" + }, + { + "start": 1289.47, + "duration": 0.0, + "text": "cuz there's many Once you add many" + }, + { + "start": 1289.48, + "duration": 0.0, + "text": "cuz there's many Once you add many layers,<00:21:29.760> you<00:21:29.840> can<00:21:30.000> make<00:21:30.160> very<00:21:30.360> complicated" + }, + { + "start": 1290.83, + "duration": 0.0, + "text": "layers, you can make very complicated" + }, + { + "start": 1290.84, + "duration": 0.0, + "text": "layers, you can make very complicated behavior,<00:21:31.320> right?<00:21:32.160> Um<00:21:32.360> the<00:21:32.440> problem<00:21:32.800> if<00:21:32.880> you" + }, + { + "start": 1292.99, + "duration": 0.0, + "text": "behavior, right? Um the problem if you" + }, + { + "start": 1293.0, + "duration": 0.0, + "text": "behavior, right? Um the problem if you just<00:21:33.160> take<00:21:33.320> that<00:21:33.440> perceptron<00:21:34.000> and<00:21:34.080> you<00:21:34.200> add<00:21:34.360> a" + }, + { + "start": 1294.39, + "duration": 0.0, + "text": "just take that perceptron and you add a" + }, + { + "start": 1294.4, + "duration": 0.0, + "text": "just take that perceptron and you add a bunch<00:21:34.600> of<00:21:34.680> layers<00:21:35.480> is<00:21:35.600> it<00:21:35.720> adds<00:21:36.000> way,<00:21:36.320> way,<00:21:36.480> way" + }, + { + "start": 1296.63, + "duration": 0.0, + "text": "bunch of layers is it adds way, way, way" + }, + { + "start": 1296.64, + "duration": 0.0, + "text": "bunch of layers is it adds way, way, way too<00:21:36.760> many<00:21:36.920> parameters<00:21:37.400> and<00:21:37.480> the<00:21:37.520> gradients" + }, + { + "start": 1297.95, + "duration": 0.0, + "text": "too many parameters and the gradients" + }, + { + "start": 1297.96, + "duration": 0.0, + "text": "too many parameters and the gradients get<00:21:38.160> very<00:21:38.400> tiny,<00:21:38.720> right?<00:21:39.480> Um<00:21:39.720> it<00:21:39.840> doesn't" + }, + { + "start": 1300.11, + "duration": 0.0, + "text": "get very tiny, right? Um it doesn't" + }, + { + "start": 1300.12, + "duration": 0.0, + "text": "get very tiny, right? Um it doesn't optimize,<00:21:40.640> right?<00:21:40.760> The<00:21:40.840> optimizer<00:21:41.200> doesn't" + }, + { + "start": 1301.43, + "duration": 0.0, + "text": "optimize, right? The optimizer doesn't" + }, + { + "start": 1301.44, + "duration": 0.0, + "text": "optimize, right? The optimizer doesn't work.<00:21:41.600> I<00:21:41.640> mean,<00:21:41.760> you<00:21:41.880> saw<00:21:42.040> my<00:21:42.200> failure<00:21:42.680> of<00:21:42.800> deep" + }, + { + "start": 1302.99, + "duration": 0.0, + "text": "work. I mean, you saw my failure of deep" + }, + { + "start": 1303.0, + "duration": 0.0, + "text": "work. I mean, you saw my failure of deep learning.<00:21:43.280> It's<00:21:43.400> not<00:21:43.600> easy." + }, + { + "start": 1304.51, + "duration": 0.0, + "text": "learning. It's not easy." + }, + { + "start": 1304.52, + "duration": 0.0, + "text": "learning. It's not easy. Uh<00:21:44.720> you<00:21:44.880> almost<00:21:45.120> have<00:21:45.240> to<00:21:45.320> like<00:21:45.800> guess<00:21:46.000> and" + }, + { + "start": 1306.11, + "duration": 0.0, + "text": "Uh you almost have to like guess and" + }, + { + "start": 1306.12, + "duration": 0.0, + "text": "Uh you almost have to like guess and check<00:21:46.360> if<00:21:46.480> you<00:21:46.560> really<00:21:47.000> aren't,<00:21:47.560> you<00:21:47.640> know," + }, + { + "start": 1308.23, + "duration": 0.0, + "text": "check if you really aren't, you know," + }, + { + "start": 1308.24, + "duration": 0.0, + "text": "check if you really aren't, you know, branching<00:21:48.600> off<00:21:48.720> of<00:21:48.800> some<00:21:48.960> established<00:21:49.720> known" + }, + { + "start": 1309.99, + "duration": 0.0, + "text": "branching off of some established known" + }, + { + "start": 1310.0, + "duration": 0.0, + "text": "branching off of some established known working<00:21:50.440> uh<00:21:50.600> architecture,<00:21:51.120> right?<00:21:51.840> Um<00:21:52.320> and" + }, + { + "start": 1312.43, + "duration": 0.0, + "text": "working uh architecture, right? Um and" + }, + { + "start": 1312.44, + "duration": 0.0, + "text": "working uh architecture, right? Um and so<00:21:52.680> all<00:21:52.880> of<00:21:52.960> the<00:21:53.440> field<00:21:53.760> of<00:21:54.000> of<00:21:54.200> modern<00:21:54.480> deep" + }, + { + "start": 1314.67, + "duration": 0.0, + "text": "so all of the field of of modern deep" + }, + { + "start": 1314.68, + "duration": 0.0, + "text": "so all of the field of of modern deep learning<00:21:54.920> is<00:21:55.040> basically<00:21:55.520> about<00:21:56.200> how<00:21:56.320> do<00:21:56.400> we" + }, + { + "start": 1316.55, + "duration": 0.0, + "text": "learning is basically about how do we" + }, + { + "start": 1316.56, + "duration": 0.0, + "text": "learning is basically about how do we add<00:21:56.760> more<00:21:56.880> parameters<00:21:57.400> to<00:21:57.520> this<00:21:57.640> model<00:21:58.280> and" + }, + { + "start": 1318.39, + "duration": 0.0, + "text": "add more parameters to this model and" + }, + { + "start": 1318.4, + "duration": 0.0, + "text": "add more parameters to this model and still<00:21:58.560> get<00:21:58.680> those<00:21:58.840> parameters<00:21:59.280> to<00:21:59.400> be" + }, + { + "start": 1319.51, + "duration": 0.0, + "text": "still get those parameters to be" + }, + { + "start": 1319.52, + "duration": 0.0, + "text": "still get those parameters to be optimized<00:22:00.080> into<00:22:00.240> something<00:22:00.520> that" + }, + { + "start": 1320.63, + "duration": 0.0, + "text": "optimized into something that" + }, + { + "start": 1320.64, + "duration": 0.0, + "text": "optimized into something that generalizes<00:22:01.320> to<00:22:01.800> what<00:22:01.920> we<00:22:02.000> needed<00:22:02.200> to<00:22:02.320> do," + }, + { + "start": 1322.75, + "duration": 0.0, + "text": "generalizes to what we needed to do," + }, + { + "start": 1322.76, + "duration": 0.0, + "text": "generalizes to what we needed to do, right?<00:22:03.240> In<00:22:03.360> the<00:22:03.440> case<00:22:03.680> of<00:22:03.840> a<00:22:03.960> convolutional" + }, + { + "start": 1324.47, + "duration": 0.0, + "text": "right? In the case of a convolutional" + }, + { + "start": 1324.48, + "duration": 0.0, + "text": "right? In the case of a convolutional neural<00:22:04.640> network<00:22:04.920> over<00:22:05.080> here,<00:22:05.720> they<00:22:05.920> basically" + }, + { + "start": 1327.23, + "duration": 0.0, + "text": "neural network over here, they basically" + }, + { + "start": 1327.24, + "duration": 0.0, + "text": "neural network over here, they basically they<00:22:07.520> kind<00:22:07.680> of<00:22:07.760> start<00:22:07.880> with<00:22:07.960> the<00:22:08.040> inductive" + }, + { + "start": 1328.43, + "duration": 0.0, + "text": "they kind of start with the inductive" + }, + { + "start": 1328.44, + "duration": 0.0, + "text": "they kind of start with the inductive bias<00:22:08.800> that<00:22:08.960> like,<00:22:09.160> \"Okay," + }, + { + "start": 1330.19, + "duration": 0.0, + "text": "bias that like, \"Okay," + }, + { + "start": 1330.2, + "duration": 0.0, + "text": "bias that like, \"Okay, um<00:22:10.240> when<00:22:10.360> you're<00:22:10.520> looking<00:22:10.920> to<00:22:11.080> see<00:22:11.200> if<00:22:11.360> a<00:22:11.520> image" + }, + { + "start": 1331.79, + "duration": 0.0, + "text": "um when you're looking to see if a image" + }, + { + "start": 1331.8, + "duration": 0.0, + "text": "um when you're looking to see if a image is<00:22:11.920> of<00:22:12.000> a<00:22:12.040> cat<00:22:12.760> or<00:22:12.920> a<00:22:12.960> dog,<00:22:14.000> um<00:22:14.160> and<00:22:14.280> you" + }, + { + "start": 1334.43, + "duration": 0.0, + "text": "is of a cat or a dog, um and you" + }, + { + "start": 1334.44, + "duration": 0.0, + "text": "is of a cat or a dog, um and you basically<00:22:14.840> are<00:22:14.920> hoping<00:22:15.240> that<00:22:15.400> let's<00:22:15.600> say<00:22:15.720> this" + }, + { + "start": 1335.87, + "duration": 0.0, + "text": "basically are hoping that let's say this" + }, + { + "start": 1335.88, + "duration": 0.0, + "text": "basically are hoping that let's say this deep<00:22:16.040> learning<00:22:16.240> model<00:22:16.400> will<00:22:16.520> pick<00:22:16.760> up<00:22:16.880> on<00:22:16.960> the" + }, + { + "start": 1337.03, + "duration": 0.0, + "text": "deep learning model will pick up on the" + }, + { + "start": 1337.04, + "duration": 0.0, + "text": "deep learning model will pick up on the pointy<00:22:17.360> ear<00:22:17.520> thing.<00:22:18.240> It<00:22:18.360> doesn't<00:22:18.600> matter" + }, + { + "start": 1338.83, + "duration": 0.0, + "text": "pointy ear thing. It doesn't matter" + }, + { + "start": 1338.84, + "duration": 0.0, + "text": "pointy ear thing. It doesn't matter where<00:22:19.160> in<00:22:19.240> the<00:22:19.360> image<00:22:19.640> that<00:22:19.760> pointy<00:22:20.040> ear<00:22:20.280> is," + }, + { + "start": 1340.59, + "duration": 0.0, + "text": "where in the image that pointy ear is," + }, + { + "start": 1340.6, + "duration": 0.0, + "text": "where in the image that pointy ear is, right?\"<00:22:21.400> Um<00:22:21.640> they<00:22:21.800> basically<00:22:22.200> want<00:22:22.560> these" + }, + { + "start": 1342.95, + "duration": 0.0, + "text": "right?\" Um they basically want these" + }, + { + "start": 1342.96, + "duration": 0.0, + "text": "right?\" Um they basically want these deep<00:22:23.160> learning<00:22:23.360> models<00:22:23.600> to<00:22:23.720> learn,<00:22:24.440> you<00:22:24.520> know," + }, + { + "start": 1345.27, + "duration": 0.0, + "text": "deep learning models to learn, you know," + }, + { + "start": 1345.28, + "duration": 0.0, + "text": "deep learning models to learn, you know, features<00:22:25.960> that<00:22:26.320> it<00:22:26.400> doesn't<00:22:26.600> matter<00:22:26.800> where" + }, + { + "start": 1346.95, + "duration": 0.0, + "text": "features that it doesn't matter where" + }, + { + "start": 1346.96, + "duration": 0.0, + "text": "features that it doesn't matter where they<00:22:27.080> are<00:22:27.200> in<00:22:27.280> the<00:22:27.360> image.<00:22:28.160> Uh<00:22:28.240> if<00:22:28.360> you<00:22:28.440> just" + }, + { + "start": 1348.59, + "duration": 0.0, + "text": "they are in the image. Uh if you just" + }, + { + "start": 1348.6, + "duration": 0.0, + "text": "they are in the image. Uh if you just fed<00:22:28.800> every<00:22:29.000> single<00:22:29.200> pixel<00:22:29.640> and<00:22:29.760> all<00:22:29.800> the<00:22:29.880> RGB" + }, + { + "start": 1350.23, + "duration": 0.0, + "text": "fed every single pixel and all the RGB" + }, + { + "start": 1350.24, + "duration": 0.0, + "text": "fed every single pixel and all the RGB values<00:22:30.640> into<00:22:30.760> a<00:22:30.800> giant<00:22:31.040> perceptron,<00:22:32.040> then<00:22:32.200> it" + }, + { + "start": 1352.27, + "duration": 0.0, + "text": "values into a giant perceptron, then it" + }, + { + "start": 1352.28, + "duration": 0.0, + "text": "values into a giant perceptron, then it would<00:22:32.400> have<00:22:32.520> to<00:22:32.600> learn<00:22:33.000> that,<00:22:33.280> you<00:22:33.360> know,<00:22:33.480> cat" + }, + { + "start": 1353.71, + "duration": 0.0, + "text": "would have to learn that, you know, cat" + }, + { + "start": 1353.72, + "duration": 0.0, + "text": "would have to learn that, you know, cat ear<00:22:33.840> means<00:22:34.080> cat<00:22:34.560> in<00:22:34.680> every<00:22:34.880> single<00:22:35.280> spot," + }, + { + "start": 1355.71, + "duration": 0.0, + "text": "ear means cat in every single spot," + }, + { + "start": 1355.72, + "duration": 0.0, + "text": "ear means cat in every single spot, right?<00:22:36.160> That's<00:22:36.320> not<00:22:36.480> really<00:22:36.640> what<00:22:36.760> we<00:22:36.840> do<00:22:36.960> when" + }, + { + "start": 1357.03, + "duration": 0.0, + "text": "right? That's not really what we do when" + }, + { + "start": 1357.04, + "duration": 0.0, + "text": "right? That's not really what we do when we<00:22:37.080> recognize<00:22:37.480> images<00:22:37.840> and<00:22:37.960> so<00:22:38.560> they" + }, + { + "start": 1358.67, + "duration": 0.0, + "text": "we recognize images and so they" + }, + { + "start": 1358.68, + "duration": 0.0, + "text": "we recognize images and so they basically<00:22:39.120> That's<00:22:39.320> like<00:22:39.440> a<00:22:39.480> little<00:22:39.600> bit<00:22:39.720> of" + }, + { + "start": 1359.79, + "duration": 0.0, + "text": "basically That's like a little bit of" + }, + { + "start": 1359.8, + "duration": 0.0, + "text": "basically That's like a little bit of key<00:22:40.000> insight.<00:22:40.400> That's<00:22:40.560> what<00:22:40.880> deep<00:22:41.080> learning" + }, + { + "start": 1361.23, + "duration": 0.0, + "text": "key insight. That's what deep learning" + }, + { + "start": 1361.24, + "duration": 0.0, + "text": "key insight. That's what deep learning researchers<00:22:41.640> are<00:22:41.720> trying<00:22:41.920> to<00:22:42.040> do<00:22:42.280> is<00:22:42.400> like" + }, + { + "start": 1363.23, + "duration": 0.0, + "text": "researchers are trying to do is like" + }, + { + "start": 1363.24, + "duration": 0.0, + "text": "researchers are trying to do is like come<00:22:43.480> up<00:22:43.560> with<00:22:43.720> ways<00:22:44.000> to<00:22:44.280> to<00:22:44.480> formulate<00:22:44.920> and" + }, + { + "start": 1364.99, + "duration": 0.0, + "text": "come up with ways to to formulate and" + }, + { + "start": 1365.0, + "duration": 0.0, + "text": "come up with ways to to formulate and structure<00:22:45.360> the<00:22:45.480> computation,<00:22:46.440> the<00:22:46.560> graph<00:22:46.760> of" + }, + { + "start": 1366.87, + "duration": 0.0, + "text": "structure the computation, the graph of" + }, + { + "start": 1366.88, + "duration": 0.0, + "text": "structure the computation, the graph of computation,<00:22:47.960> such<00:22:48.200> that<00:22:48.560> when<00:22:48.680> you<00:22:48.840> optimize" + }, + { + "start": 1369.27, + "duration": 0.0, + "text": "computation, such that when you optimize" + }, + { + "start": 1369.28, + "duration": 0.0, + "text": "computation, such that when you optimize it,<00:22:49.480> it<00:22:49.600> gets<00:22:49.760> what<00:22:49.880> you<00:22:50.000> want,<00:22:50.320> right?<00:22:51.080> Um<00:22:51.200> and" + }, + { + "start": 1371.27, + "duration": 0.0, + "text": "it, it gets what you want, right? Um and" + }, + { + "start": 1371.28, + "duration": 0.0, + "text": "it, it gets what you want, right? Um and a<00:22:51.320> transformer<00:22:51.880> on<00:22:51.960> the<00:22:52.080> other<00:22:52.200> end<00:22:52.320> here<00:22:52.440> I'll" + }, + { + "start": 1372.51, + "duration": 0.0, + "text": "a transformer on the other end here I'll" + }, + { + "start": 1372.52, + "duration": 0.0, + "text": "a transformer on the other end here I'll talk<00:22:52.720> about<00:22:53.040> in<00:22:53.160> a<00:22:53.200> second.<00:22:53.800> Um<00:22:54.360> this<00:22:54.480> is" + }, + { + "start": 1374.55, + "duration": 0.0, + "text": "talk about in a second. Um this is" + }, + { + "start": 1374.56, + "duration": 0.0, + "text": "talk about in a second. Um this is another<00:22:54.800> example<00:22:55.120> of<00:22:55.200> an<00:22:55.280> architectural" + }, + { + "start": 1375.79, + "duration": 0.0, + "text": "another example of an architectural" + }, + { + "start": 1375.8, + "duration": 0.0, + "text": "another example of an architectural change.<00:22:56.480> You<00:22:56.560> can<00:22:56.720> add<00:22:57.000> way<00:22:57.120> more<00:22:57.320> parameters" + }, + { + "start": 1377.75, + "duration": 0.0, + "text": "change. You can add way more parameters" + }, + { + "start": 1377.76, + "duration": 0.0, + "text": "change. You can add way more parameters and<00:22:57.880> it's<00:22:58.520> still<00:22:58.720> stable<00:22:59.080> and<00:22:59.160> it's<00:22:59.280> still," + }, + { + "start": 1379.71, + "duration": 0.0, + "text": "and it's still stable and it's still," + }, + { + "start": 1379.72, + "duration": 0.0, + "text": "and it's still stable and it's still, you<00:22:59.760> know,<00:22:59.880> optimizing<00:23:00.400> correctly,<00:23:00.720> right?" + }, + { + "start": 1381.79, + "duration": 0.0, + "text": "you know, optimizing correctly, right?" + }, + { + "start": 1381.8, + "duration": 0.0, + "text": "you know, optimizing correctly, right? Just<00:23:02.040> as<00:23:02.120> another<00:23:02.320> little<00:23:02.480> side<00:23:02.720> note,<00:23:02.960> some" + }, + { + "start": 1383.11, + "duration": 0.0, + "text": "Just as another little side note, some" + }, + { + "start": 1383.12, + "duration": 0.0, + "text": "Just as another little side note, some deep<00:23:03.280> learning<00:23:03.480> techniques<00:23:03.840> are<00:23:03.920> a<00:23:03.960> little" + }, + { + "start": 1384.11, + "duration": 0.0, + "text": "deep learning techniques are a little" + }, + { + "start": 1384.12, + "duration": 0.0, + "text": "deep learning techniques are a little silly.<00:23:04.840> Uh<00:23:04.960> Uh<00:23:05.080> this<00:23:05.240> is<00:23:05.320> called<00:23:05.520> dropout.<00:23:06.560> You" + }, + { + "start": 1386.71, + "duration": 0.0, + "text": "silly. Uh Uh this is called dropout. You" + }, + { + "start": 1386.72, + "duration": 0.0, + "text": "silly. Uh Uh this is called dropout. You randomly<00:23:07.040> turn<00:23:07.280> off<00:23:07.440> neurons<00:23:07.880> during<00:23:08.120> the" + }, + { + "start": 1388.75, + "duration": 0.0, + "text": "randomly turn off neurons during the" + }, + { + "start": 1388.76, + "duration": 0.0, + "text": "randomly turn off neurons during the during<00:23:08.960> the<00:23:09.040> training.<00:23:09.520> It's<00:23:09.800> it's<00:23:10.080> supposed" + }, + { + "start": 1390.35, + "duration": 0.0, + "text": "during the training. It's it's supposed" + }, + { + "start": 1390.36, + "duration": 0.0, + "text": "during the training. It's it's supposed to<00:23:10.440> present<00:23:10.920> or<00:23:11.080> or<00:23:11.160> prevent<00:23:11.440> memorizing<00:23:11.840> the" + }, + { + "start": 1391.91, + "duration": 0.0, + "text": "to present or or prevent memorizing the" + }, + { + "start": 1391.92, + "duration": 0.0, + "text": "to present or or prevent memorizing the training<00:23:12.200> data<00:23:13.000> um<00:23:13.160> and<00:23:13.280> hopefully<00:23:13.600> make<00:23:13.800> it" + }, + { + "start": 1393.87, + "duration": 0.0, + "text": "training data um and hopefully make it" + }, + { + "start": 1393.88, + "duration": 0.0, + "text": "training data um and hopefully make it so<00:23:14.000> neurons<00:23:14.400> aren't<00:23:14.720> exactly<00:23:15.200> specialized" + }, + { + "start": 1395.79, + "duration": 0.0, + "text": "so neurons aren't exactly specialized" + }, + { + "start": 1395.8, + "duration": 0.0, + "text": "so neurons aren't exactly specialized like" + }, + { + "start": 1397.27, + "duration": 0.0, + "text": "like" + }, + { + "start": 1397.28, + "duration": 0.0, + "text": "like Who<00:23:17.400> knows<00:23:17.640> why<00:23:17.760> this<00:23:17.920> works?<00:23:18.160> Honestly,<00:23:18.560> I" + }, + { + "start": 1398.63, + "duration": 0.0, + "text": "Who knows why this works? Honestly, I" + }, + { + "start": 1398.64, + "duration": 0.0, + "text": "Who knows why this works? Honestly, I don't<00:23:18.880> I<00:23:18.920> haven't<00:23:19.160> seen<00:23:19.360> any<00:23:19.560> real<00:23:19.800> convincing" + }, + { + "start": 1400.27, + "duration": 0.0, + "text": "don't I haven't seen any real convincing" + }, + { + "start": 1400.28, + "duration": 0.0, + "text": "don't I haven't seen any real convincing argument<00:23:20.720> in<00:23:20.840> any<00:23:21.000> paper<00:23:21.360> for<00:23:21.800> what<00:23:21.960> exactly" + }, + { + "start": 1402.35, + "duration": 0.0, + "text": "argument in any paper for what exactly" + }, + { + "start": 1402.36, + "duration": 0.0, + "text": "argument in any paper for what exactly this<00:23:22.520> is<00:23:22.600> doing.<00:23:23.440> Like<00:23:23.960> basically<00:23:24.360> what's" + }, + { + "start": 1404.51, + "duration": 0.0, + "text": "this is doing. Like basically what's" + }, + { + "start": 1404.52, + "duration": 0.0, + "text": "this is doing. Like basically what's happening<00:23:24.840> in<00:23:24.880> deep<00:23:25.080> learning<00:23:25.280> is<00:23:25.360> people<00:23:25.600> try" + }, + { + "start": 1405.87, + "duration": 0.0, + "text": "happening in deep learning is people try" + }, + { + "start": 1405.88, + "duration": 0.0, + "text": "happening in deep learning is people try stuff<00:23:26.240> and<00:23:26.360> they<00:23:26.440> have<00:23:26.600> some<00:23:26.800> very<00:23:27.120> like<00:23:27.600> okay," + }, + { + "start": 1407.79, + "duration": 0.0, + "text": "stuff and they have some very like okay," + }, + { + "start": 1407.8, + "duration": 0.0, + "text": "stuff and they have some very like okay, yeah,<00:23:27.920> that<00:23:28.040> sounds<00:23:28.240> about<00:23:28.480> right.<00:23:28.600> I<00:23:28.640> think" + }, + { + "start": 1408.79, + "duration": 0.0, + "text": "yeah, that sounds about right. I think" + }, + { + "start": 1408.8, + "duration": 0.0, + "text": "yeah, that sounds about right. I think this<00:23:28.920> what's<00:23:29.080> happening<00:23:29.440> and<00:23:29.520> then<00:23:29.640> it's<00:23:29.760> like" + }, + { + "start": 1409.87, + "duration": 0.0, + "text": "this what's happening and then it's like" + }, + { + "start": 1409.88, + "duration": 0.0, + "text": "this what's happening and then it's like there's<00:23:30.000> no<00:23:30.120> rigor<00:23:30.400> in<00:23:30.480> this<00:23:30.600> field<00:23:30.840> basically" + }, + { + "start": 1411.27, + "duration": 0.0, + "text": "there's no rigor in this field basically" + }, + { + "start": 1411.28, + "duration": 0.0, + "text": "there's no rigor in this field basically like<00:23:32.080> um<00:23:32.800> it's<00:23:32.960> really<00:23:33.160> crazy.<00:23:33.520> This<00:23:33.680> is" + }, + { + "start": 1413.75, + "duration": 0.0, + "text": "like um it's really crazy. This is" + }, + { + "start": 1413.76, + "duration": 0.0, + "text": "like um it's really crazy. This is actually<00:23:34.000> used<00:23:34.200> in<00:23:34.280> production<00:23:34.680> for<00:23:34.800> chat<00:23:35.040> GPT" + }, + { + "start": 1415.43, + "duration": 0.0, + "text": "actually used in production for chat GPT" + }, + { + "start": 1415.44, + "duration": 0.0, + "text": "actually used in production for chat GPT without<00:23:35.760> it<00:23:35.840> like<00:23:36.000> it<00:23:36.120> wouldn't<00:23:36.320> work,<00:23:36.520> I" + }, + { + "start": 1416.55, + "duration": 0.0, + "text": "without it like it wouldn't work, I" + }, + { + "start": 1416.56, + "duration": 0.0, + "text": "without it like it wouldn't work, I think.<00:23:37.440> Um<00:23:37.560> and<00:23:37.640> it's<00:23:37.720> like<00:23:37.840> you<00:23:37.920> know,<00:23:38.000> we" + }, + { + "start": 1418.03, + "duration": 0.0, + "text": "think. Um and it's like you know, we" + }, + { + "start": 1418.04, + "duration": 0.0, + "text": "think. Um and it's like you know, we just<00:23:38.200> randomly<00:23:38.480> turn<00:23:38.640> off<00:23:38.720> the<00:23:38.800> neurons" + }, + { + "start": 1419.27, + "duration": 0.0, + "text": "just randomly turn off the neurons" + }, + { + "start": 1419.28, + "duration": 0.0, + "text": "just randomly turn off the neurons during<00:23:39.440> the<00:23:39.520> training.<00:23:39.920> Like<00:23:40.280> okay." + }, + { + "start": 1421.19, + "duration": 0.0, + "text": "during the training. Like okay." + }, + { + "start": 1421.2, + "duration": 0.0, + "text": "during the training. Like okay. Um" + }, + { + "start": 1422.55, + "duration": 0.0, + "text": "Um" + }, + { + "start": 1422.56, + "duration": 0.0, + "text": "Um So,<00:23:43.000> there's<00:23:43.240> a<00:23:43.280> zillion<00:23:43.960> videos<00:23:44.360> that" + }, + { + "start": 1424.47, + "duration": 0.0, + "text": "So, there's a zillion videos that" + }, + { + "start": 1424.48, + "duration": 0.0, + "text": "So, there's a zillion videos that explain<00:23:44.760> transformers<00:23:45.240> already.<00:23:45.600> I'm<00:23:45.680> not" + }, + { + "start": 1425.79, + "duration": 0.0, + "text": "explain transformers already. I'm not" + }, + { + "start": 1425.8, + "duration": 0.0, + "text": "explain transformers already. I'm not going<00:23:45.920> to<00:23:46.000> give<00:23:46.200> you<00:23:46.560> an<00:23:46.640> explanation<00:23:47.080> of<00:23:47.160> the" + }, + { + "start": 1427.23, + "duration": 0.0, + "text": "going to give you an explanation of the" + }, + { + "start": 1427.24, + "duration": 0.0, + "text": "going to give you an explanation of the transformer,<00:23:47.760> which<00:23:47.920> is<00:23:48.040> you<00:23:48.080> know,<00:23:48.160> the" + }, + { + "start": 1428.27, + "duration": 0.0, + "text": "transformer, which is you know, the" + }, + { + "start": 1428.28, + "duration": 0.0, + "text": "transformer, which is you know, the cause<00:23:48.520> of<00:23:48.600> the<00:23:48.680> current<00:23:48.920> rift,<00:23:49.240> right?<00:23:50.080> Um<00:23:50.360> and" + }, + { + "start": 1430.71, + "duration": 0.0, + "text": "cause of the current rift, right? Um and" + }, + { + "start": 1430.72, + "duration": 0.0, + "text": "cause of the current rift, right? Um and you<00:23:50.800> know,<00:23:50.880> you<00:23:50.960> can<00:23:51.120> watch<00:23:51.320> any<00:23:51.440> of<00:23:51.520> these<00:23:51.680> and" + }, + { + "start": 1431.75, + "duration": 0.0, + "text": "you know, you can watch any of these and" + }, + { + "start": 1431.76, + "duration": 0.0, + "text": "you know, you can watch any of these and they're<00:23:51.880> good<00:23:52.160> and<00:23:52.480> whatever.<00:23:53.360> But<00:23:53.560> I'm<00:23:53.640> just" + }, + { + "start": 1433.75, + "duration": 0.0, + "text": "they're good and whatever. But I'm just" + }, + { + "start": 1433.76, + "duration": 0.0, + "text": "they're good and whatever. But I'm just going<00:23:53.880> to<00:23:53.920> give<00:23:54.000> you<00:23:54.040> a<00:23:54.080> quick<00:23:54.320> overview.<00:23:54.600> This" + }, + { + "start": 1434.75, + "duration": 0.0, + "text": "going to give you a quick overview. This" + }, + { + "start": 1434.76, + "duration": 0.0, + "text": "going to give you a quick overview. This is<00:23:54.880> a<00:23:54.960> single<00:23:55.240> slide.<00:23:55.680> This<00:23:55.800> is<00:23:55.880> what<00:23:56.000> a" + }, + { + "start": 1436.03, + "duration": 0.0, + "text": "is a single slide. This is what a" + }, + { + "start": 1436.04, + "duration": 0.0, + "text": "is a single slide. This is what a transformer<00:23:56.480> is<00:23:56.560> doing,<00:23:56.880> right?<00:23:57.400> It's" + }, + { + "start": 1437.47, + "duration": 0.0, + "text": "transformer is doing, right? It's" + }, + { + "start": 1437.48, + "duration": 0.0, + "text": "transformer is doing, right? It's basically<00:23:57.800> going<00:23:58.040> from<00:23:58.200> a<00:23:58.240> sentence,<00:23:59.080> a" + }, + { + "start": 1439.15, + "duration": 0.0, + "text": "basically going from a sentence, a" + }, + { + "start": 1439.16, + "duration": 0.0, + "text": "basically going from a sentence, a lookup<00:23:59.440> table<00:23:59.680> per<00:23:59.840> token<00:24:00.280> into<00:24:00.480> a<00:24:00.560> point<00:24:00.920> in<00:24:01.000> a" + }, + { + "start": 1441.03, + "duration": 0.0, + "text": "lookup table per token into a point in a" + }, + { + "start": 1441.04, + "duration": 0.0, + "text": "lookup table per token into a point in a space<00:24:01.440> where<00:24:01.560> like<00:24:02.280> the<00:24:02.400> closer<00:24:02.720> two<00:24:02.840> points" + }, + { + "start": 1443.11, + "duration": 0.0, + "text": "space where like the closer two points" + }, + { + "start": 1443.12, + "duration": 0.0, + "text": "space where like the closer two points are<00:24:03.200> the<00:24:03.280> more<00:24:03.440> related<00:24:03.760> they<00:24:03.880> are,<00:24:04.040> right?<00:24:04.800> It" + }, + { + "start": 1444.95, + "duration": 0.0, + "text": "are the more related they are, right? It" + }, + { + "start": 1444.96, + "duration": 0.0, + "text": "are the more related they are, right? It adds<00:24:05.200> a<00:24:05.280> position<00:24:05.720> encoding<00:24:06.120> to<00:24:06.200> those<00:24:06.360> points" + }, + { + "start": 1446.83, + "duration": 0.0, + "text": "adds a position encoding to those points" + }, + { + "start": 1446.84, + "duration": 0.0, + "text": "adds a position encoding to those points so<00:24:06.960> that,<00:24:07.200> you<00:24:07.280> know,<00:24:07.920> if<00:24:08.080> we<00:24:08.200> have<00:24:08.400> an<00:24:08.520> array" + }, + { + "start": 1448.83, + "duration": 0.0, + "text": "so that, you know, if we have an array" + }, + { + "start": 1448.84, + "duration": 0.0, + "text": "so that, you know, if we have an array of<00:24:09.160> these<00:24:09.360> points,<00:24:09.680> then<00:24:09.960> the<00:24:10.040> points<00:24:10.320> are" + }, + { + "start": 1450.67, + "duration": 0.0, + "text": "of these points, then the points are" + }, + { + "start": 1450.68, + "duration": 0.0, + "text": "of these points, then the points are modulated<00:24:11.160> in<00:24:11.240> such<00:24:11.480> a<00:24:11.520> way<00:24:11.800> that<00:24:12.360> the<00:24:12.640> the" + }, + { + "start": 1452.75, + "duration": 0.0, + "text": "modulated in such a way that the the" + }, + { + "start": 1452.76, + "duration": 0.0, + "text": "modulated in such a way that the the deep<00:24:12.960> learning<00:24:13.160> models<00:24:13.640> they<00:24:13.760> can<00:24:13.880> they<00:24:14.000> can" + }, + { + "start": 1454.15, + "duration": 0.0, + "text": "deep learning models they can they can" + }, + { + "start": 1454.16, + "duration": 0.0, + "text": "deep learning models they can they can pick<00:24:14.480> up<00:24:14.800> where<00:24:14.960> they<00:24:15.120> are<00:24:15.240> in<00:24:15.320> the<00:24:15.400> sentence," + }, + { + "start": 1455.95, + "duration": 0.0, + "text": "pick up where they are in the sentence," + }, + { + "start": 1455.96, + "duration": 0.0, + "text": "pick up where they are in the sentence, right?<00:24:16.200> They're<00:24:16.320> almost<00:24:16.560> jiggle<00:24:16.880> in<00:24:17.000> a<00:24:17.040> way" + }, + { + "start": 1457.23, + "duration": 0.0, + "text": "right? They're almost jiggle in a way" + }, + { + "start": 1457.24, + "duration": 0.0, + "text": "right? They're almost jiggle in a way where<00:24:17.320> like<00:24:17.440> you<00:24:17.520> could<00:24:18.080> learn<00:24:18.320> that<00:24:18.440> pattern." + }, + { + "start": 1459.31, + "duration": 0.0, + "text": "where like you could learn that pattern." + }, + { + "start": 1459.32, + "duration": 0.0, + "text": "where like you could learn that pattern. Um" + }, + { + "start": 1460.47, + "duration": 0.0, + "text": "Um" + }, + { + "start": 1460.48, + "duration": 0.0, + "text": "Um I<00:24:20.560> keep<00:24:20.720> using<00:24:20.920> the<00:24:20.960> word<00:24:21.160> learn.<00:24:21.320> I'm<00:24:21.400> sorry." + }, + { + "start": 1461.63, + "duration": 0.0, + "text": "I keep using the word learn. I'm sorry." + }, + { + "start": 1461.64, + "duration": 0.0, + "text": "I keep using the word learn. I'm sorry. It's<00:24:21.760> pervasive<00:24:22.200> in<00:24:22.240> the<00:24:22.320> field,<00:24:22.560> but<00:24:22.640> I<00:24:22.680> don't" + }, + { + "start": 1462.83, + "duration": 0.0, + "text": "It's pervasive in the field, but I don't" + }, + { + "start": 1462.84, + "duration": 0.0, + "text": "It's pervasive in the field, but I don't like<00:24:23.040> it.<00:24:23.680> Um<00:24:24.480> and<00:24:24.560> then<00:24:24.720> you<00:24:24.880> literally<00:24:25.200> just" + }, + { + "start": 1465.39, + "duration": 0.0, + "text": "like it. Um and then you literally just" + }, + { + "start": 1465.4, + "duration": 0.0, + "text": "like it. Um and then you literally just stack<00:24:25.640> a<00:24:25.680> hundred<00:24:26.120> of<00:24:26.280> these<00:24:26.600> attention" + }, + { + "start": 1467.03, + "duration": 0.0, + "text": "stack a hundred of these attention" + }, + { + "start": 1467.04, + "duration": 0.0, + "text": "stack a hundred of these attention layers<00:24:27.720> onto<00:24:28.000> those<00:24:28.400> uh<00:24:28.480> vectors,<00:24:28.960> right?<00:24:29.160> And" + }, + { + "start": 1469.27, + "duration": 0.0, + "text": "layers onto those uh vectors, right? And" + }, + { + "start": 1469.28, + "duration": 0.0, + "text": "layers onto those uh vectors, right? And each<00:24:29.400> of<00:24:29.480> these<00:24:29.600> attention<00:24:29.960> layers<00:24:30.200> every" + }, + { + "start": 1470.43, + "duration": 0.0, + "text": "each of these attention layers every" + }, + { + "start": 1470.44, + "duration": 0.0, + "text": "each of these attention layers every single<00:24:30.680> time,<00:24:31.280> every<00:24:31.480> single<00:24:31.720> one<00:24:31.880> of<00:24:31.920> those" + }, + { + "start": 1472.11, + "duration": 0.0, + "text": "single time, every single one of those" + }, + { + "start": 1472.12, + "duration": 0.0, + "text": "single time, every single one of those tokens,<00:24:32.560> right?<00:24:33.080> It's<00:24:33.280> basically<00:24:33.920> looking<00:24:34.200> at" + }, + { + "start": 1474.31, + "duration": 0.0, + "text": "tokens, right? It's basically looking at" + }, + { + "start": 1474.32, + "duration": 0.0, + "text": "tokens, right? It's basically looking at all<00:24:34.440> of<00:24:34.520> the<00:24:34.600> past<00:24:35.000> tokens<00:24:35.360> and<00:24:35.440> enriching<00:24:35.840> the" + }, + { + "start": 1475.91, + "duration": 0.0, + "text": "all of the past tokens and enriching the" + }, + { + "start": 1475.92, + "duration": 0.0, + "text": "all of the past tokens and enriching the current<00:24:36.600> vector<00:24:37.000> with<00:24:37.240> more<00:24:37.520> information" + }, + { + "start": 1477.99, + "duration": 0.0, + "text": "current vector with more information" + }, + { + "start": 1478.0, + "duration": 0.0, + "text": "current vector with more information about<00:24:38.200> the<00:24:38.280> surrounding<00:24:38.640> tokens<00:24:39.400> and<00:24:39.480> then" + }, + { + "start": 1479.59, + "duration": 0.0, + "text": "about the surrounding tokens and then" + }, + { + "start": 1479.6, + "duration": 0.0, + "text": "about the surrounding tokens and then more<00:24:39.760> facts<00:24:40.320> potentially<00:24:40.760> about<00:24:41.000> that" + }, + { + "start": 1481.11, + "duration": 0.0, + "text": "more facts potentially about that" + }, + { + "start": 1481.12, + "duration": 0.0, + "text": "more facts potentially about that information<00:24:41.600> in<00:24:41.720> the<00:24:42.240> form<00:24:42.440> of<00:24:42.520> another" + }, + { + "start": 1482.75, + "duration": 0.0, + "text": "information in the form of another" + }, + { + "start": 1482.76, + "duration": 0.0, + "text": "information in the form of another perceptron,<00:24:43.320> right?<00:24:43.480> it's<00:24:43.560> like<00:24:43.720> learning" + }, + { + "start": 1483.99, + "duration": 0.0, + "text": "perceptron, right? it's like learning" + }, + { + "start": 1484.0, + "duration": 0.0, + "text": "perceptron, right? it's like learning what<00:24:44.200> what<00:24:44.720> facts<00:24:45.000> might<00:24:45.120> be<00:24:45.200> related<00:24:45.520> to<00:24:45.640> a" + }, + { + "start": 1485.67, + "duration": 0.0, + "text": "what what facts might be related to a" + }, + { + "start": 1485.68, + "duration": 0.0, + "text": "what what facts might be related to a word<00:24:45.880> like<00:24:46.040> house,<00:24:46.440> right?<00:24:46.600> It's<00:24:47.120> It<00:24:47.280> It" + }, + { + "start": 1487.39, + "duration": 0.0, + "text": "word like house, right? It's It It" + }, + { + "start": 1487.4, + "duration": 0.0, + "text": "word like house, right? It's It It continuously<00:24:48.120> enriches<00:24:48.640> each<00:24:48.840> one<00:24:49.000> of<00:24:49.040> these" + }, + { + "start": 1489.87, + "duration": 0.0, + "text": "continuously enriches each one of these" + }, + { + "start": 1489.88, + "duration": 0.0, + "text": "continuously enriches each one of these uh<00:24:49.960> words<00:24:50.240> with<00:24:50.360> more<00:24:50.520> and<00:24:50.600> more<00:24:50.720> context," + }, + { + "start": 1491.35, + "duration": 0.0, + "text": "uh words with more and more context," + }, + { + "start": 1491.36, + "duration": 0.0, + "text": "uh words with more and more context, right?<00:24:51.560> In<00:24:51.640> a<00:24:51.680> in<00:24:51.840> a<00:24:51.920> highly<00:24:52.200> parallel<00:24:52.600> easy<00:24:52.800> to" + }, + { + "start": 1492.87, + "duration": 0.0, + "text": "right? In a in a highly parallel easy to" + }, + { + "start": 1492.88, + "duration": 0.0, + "text": "right? In a in a highly parallel easy to train<00:24:53.120> manner.<00:24:53.920> Um<00:24:54.760> and<00:24:54.920> then<00:24:55.080> there's<00:24:55.280> just" + }, + { + "start": 1495.43, + "duration": 0.0, + "text": "train manner. Um and then there's just" + }, + { + "start": 1495.44, + "duration": 0.0, + "text": "train manner. Um and then there's just this<00:24:55.800> two-layer<00:24:56.120> perceptron<00:24:56.640> and<00:24:56.720> then<00:24:56.800> it" + }, + { + "start": 1496.83, + "duration": 0.0, + "text": "this two-layer perceptron and then it" + }, + { + "start": 1496.84, + "duration": 0.0, + "text": "this two-layer perceptron and then it gives<00:24:57.040> you<00:24:57.120> the<00:24:57.200> next<00:24:57.440> word,<00:24:57.640> and<00:24:57.760> it<00:24:57.840> does" + }, + { + "start": 1497.95, + "duration": 0.0, + "text": "gives you the next word, and it does" + }, + { + "start": 1497.96, + "duration": 0.0, + "text": "gives you the next word, and it does that<00:24:58.080> over<00:24:58.200> and<00:24:58.280> over<00:24:58.400> again,<00:24:58.600> right?<00:24:59.480> Um" + }, + { + "start": 1500.75, + "duration": 0.0, + "text": "that over and over again, right? Um" + }, + { + "start": 1500.76, + "duration": 0.0, + "text": "that over and over again, right? Um there<00:25:00.840> are<00:25:00.920> many<00:25:01.160> hyper<00:25:01.600> parameters<00:25:02.280> in<00:25:02.400> this" + }, + { + "start": 1502.55, + "duration": 0.0, + "text": "there are many hyper parameters in this" + }, + { + "start": 1502.56, + "duration": 0.0, + "text": "there are many hyper parameters in this massive<00:25:02.960> system,<00:25:03.480> right?<00:25:03.720> Many,<00:25:03.960> many,<00:25:04.200> many," + }, + { + "start": 1504.55, + "duration": 0.0, + "text": "massive system, right? Many, many, many," + }, + { + "start": 1504.56, + "duration": 0.0, + "text": "massive system, right? Many, many, many, many<00:25:04.880> things<00:25:05.160> can<00:25:05.280> go<00:25:05.440> wrong.<00:25:06.240> You<00:25:06.320> know,<00:25:06.520> it's" + }, + { + "start": 1506.63, + "duration": 0.0, + "text": "many things can go wrong. You know, it's" + }, + { + "start": 1506.64, + "duration": 0.0, + "text": "many things can go wrong. You know, it's like<00:25:06.840> what<00:25:06.960> is<00:25:07.120> the<00:25:07.240> learning<00:25:07.600> rate?<00:25:07.800> You" + }, + { + "start": 1507.87, + "duration": 0.0, + "text": "like what is the learning rate? You" + }, + { + "start": 1507.88, + "duration": 0.0, + "text": "like what is the learning rate? You know,<00:25:07.960> how<00:25:08.120> quickly<00:25:08.440> do<00:25:08.560> we<00:25:08.680> turn<00:25:08.920> these" + }, + { + "start": 1509.07, + "duration": 0.0, + "text": "know, how quickly do we turn these" + }, + { + "start": 1509.08, + "duration": 0.0, + "text": "know, how quickly do we turn these parameters<00:25:09.520> to,<00:25:09.800> you<00:25:09.880> know,<00:25:09.960> make<00:25:10.120> it<00:25:10.200> better," + }, + { + "start": 1510.43, + "duration": 0.0, + "text": "parameters to, you know, make it better," + }, + { + "start": 1510.44, + "duration": 0.0, + "text": "parameters to, you know, make it better, right?<00:25:11.040> It's<00:25:11.200> like<00:25:11.320> a<00:25:11.400> massive<00:25:11.760> problem," + }, + { + "start": 1512.15, + "duration": 0.0, + "text": "right? It's like a massive problem," + }, + { + "start": 1512.16, + "duration": 0.0, + "text": "right? It's like a massive problem, actually,<00:25:13.040> um<00:25:13.240> to<00:25:13.360> know<00:25:13.640> what<00:25:14.160> These<00:25:14.400> are" + }, + { + "start": 1514.47, + "duration": 0.0, + "text": "actually, um to know what These are" + }, + { + "start": 1514.48, + "duration": 0.0, + "text": "actually, um to know what These are called<00:25:14.640> hyper<00:25:14.920> parameters<00:25:15.440> because<00:25:15.800> unlike" + }, + { + "start": 1516.03, + "duration": 0.0, + "text": "called hyper parameters because unlike" + }, + { + "start": 1516.04, + "duration": 0.0, + "text": "called hyper parameters because unlike parameters,<00:25:16.560> like<00:25:17.400> they're<00:25:17.520> not" + }, + { + "start": 1517.75, + "duration": 0.0, + "text": "parameters, like they're not" + }, + { + "start": 1517.76, + "duration": 0.0, + "text": "parameters, like they're not automatically<00:25:18.240> learned<00:25:18.600> in<00:25:18.720> the<00:25:19.080> core" + }, + { + "start": 1519.31, + "duration": 0.0, + "text": "automatically learned in the core" + }, + { + "start": 1519.32, + "duration": 0.0, + "text": "automatically learned in the core algorithm<00:25:19.680> of<00:25:19.760> deep<00:25:19.960> learning,<00:25:20.240> right?<00:25:20.560> You" + }, + { + "start": 1520.67, + "duration": 0.0, + "text": "algorithm of deep learning, right? You" + }, + { + "start": 1520.68, + "duration": 0.0, + "text": "algorithm of deep learning, right? You can<00:25:21.200> or<00:25:21.440> you<00:25:21.560> can<00:25:21.760> try<00:25:21.880> to<00:25:21.960> figure<00:25:22.120> out<00:25:22.160> what" + }, + { + "start": 1522.27, + "duration": 0.0, + "text": "can or you can try to figure out what" + }, + { + "start": 1522.28, + "duration": 0.0, + "text": "can or you can try to figure out what they<00:25:22.400> are<00:25:23.200> uh<00:25:23.280> with<00:25:23.480> a<00:25:23.520> search<00:25:23.800> algorithm." + }, + { + "start": 1524.23, + "duration": 0.0, + "text": "they are uh with a search algorithm." + }, + { + "start": 1524.24, + "duration": 0.0, + "text": "they are uh with a search algorithm. That's<00:25:24.520> very<00:25:24.760> costly<00:25:25.240> if<00:25:25.360> you're<00:25:25.520> OpenAI<00:25:26.040> and" + }, + { + "start": 1526.11, + "duration": 0.0, + "text": "That's very costly if you're OpenAI and" + }, + { + "start": 1526.12, + "duration": 0.0, + "text": "That's very costly if you're OpenAI and each<00:25:26.280> training<00:25:26.520> run<00:25:26.640> costs<00:25:26.840> $10<00:25:27.160> million.<00:25:28.120> Um" + }, + { + "start": 1528.59, + "duration": 0.0, + "text": "each training run costs $10 million. Um" + }, + { + "start": 1528.6, + "duration": 0.0, + "text": "each training run costs $10 million. Um and<00:25:28.760> as<00:25:28.880> it<00:25:28.960> turns<00:25:29.280> out,<00:25:29.560> we<00:25:30.040> heard<00:25:30.280> from<00:25:30.440> an" + }, + { + "start": 1530.55, + "duration": 0.0, + "text": "and as it turns out, we heard from an" + }, + { + "start": 1530.56, + "duration": 0.0, + "text": "and as it turns out, we heard from an insider<00:25:31.000> here<00:25:31.680> that<00:25:32.040> most<00:25:32.320> of<00:25:32.400> the<00:25:32.520> big<00:25:32.920> gains" + }, + { + "start": 1533.51, + "duration": 0.0, + "text": "insider here that most of the big gains" + }, + { + "start": 1533.52, + "duration": 0.0, + "text": "insider here that most of the big gains from<00:25:33.960> these,<00:25:34.360> you<00:25:34.400> know," + }, + { + "start": 1535.35, + "duration": 0.0, + "text": "from these, you know," + }, + { + "start": 1535.36, + "duration": 0.0, + "text": "from these, you know, most<00:25:35.520> of<00:25:35.600> the<00:25:35.640> state-of-the-art<00:25:36.120> models<00:25:36.400> that" + }, + { + "start": 1536.47, + "duration": 0.0, + "text": "most of the state-of-the-art models that" + }, + { + "start": 1536.48, + "duration": 0.0, + "text": "most of the state-of-the-art models that come<00:25:36.640> out<00:25:36.680> of<00:25:36.800> OpenAI,<00:25:37.200> they<00:25:37.320> came<00:25:37.480> from" + }, + { + "start": 1537.63, + "duration": 0.0, + "text": "come out of OpenAI, they came from" + }, + { + "start": 1537.64, + "duration": 0.0, + "text": "come out of OpenAI, they came from so-called<00:25:38.120> YOLO<00:25:38.480> runs,<00:25:39.000> where<00:25:39.560> they" + }, + { + "start": 1539.71, + "duration": 0.0, + "text": "so-called YOLO runs, where they" + }, + { + "start": 1539.72, + "duration": 0.0, + "text": "so-called YOLO runs, where they basically<00:25:40.640> they<00:25:40.760> basically<00:25:41.240> asked<00:25:41.520> Karpathy" + }, + { + "start": 1542.03, + "duration": 0.0, + "text": "basically they basically asked Karpathy" + }, + { + "start": 1542.04, + "duration": 0.0, + "text": "basically they basically asked Karpathy or<00:25:42.200> one<00:25:42.320> of<00:25:42.400> these,<00:25:42.640> you<00:25:42.760> know,<00:25:43.200> like<00:25:43.360> deep" + }, + { + "start": 1543.55, + "duration": 0.0, + "text": "or one of these, you know, like deep" + }, + { + "start": 1543.56, + "duration": 0.0, + "text": "or one of these, you know, like deep learning<00:25:43.760> generals,<00:25:44.120> like,<00:25:44.280> \"Hey," + }, + { + "start": 1545.15, + "duration": 0.0, + "text": "learning generals, like, \"Hey," + }, + { + "start": 1545.16, + "duration": 0.0, + "text": "learning generals, like, \"Hey, we're<00:25:45.240> going<00:25:45.360> to<00:25:45.400> give<00:25:45.560> you<00:25:45.680> $20<00:25:46.120> million<00:25:46.400> to" + }, + { + "start": 1546.51, + "duration": 0.0, + "text": "we're going to give you $20 million to" + }, + { + "start": 1546.52, + "duration": 0.0, + "text": "we're going to give you $20 million to just<00:25:46.680> try<00:25:46.920> something,<00:25:47.240> right?<00:25:47.400> You<00:25:47.440> just<00:25:47.640> pick" + }, + { + "start": 1547.83, + "duration": 0.0, + "text": "just try something, right? You just pick" + }, + { + "start": 1547.84, + "duration": 0.0, + "text": "just try something, right? You just pick numbers,<00:25:48.240> pick<00:25:48.400> hyper<00:25:48.640> parameters,<00:25:49.080> pick" + }, + { + "start": 1549.23, + "duration": 0.0, + "text": "numbers, pick hyper parameters, pick" + }, + { + "start": 1549.24, + "duration": 0.0, + "text": "numbers, pick hyper parameters, pick whatever<00:25:49.480> you<00:25:49.600> want,<00:25:50.280> and<00:25:50.440> we'll<00:25:50.520> see<00:25:50.640> what" + }, + { + "start": 1550.71, + "duration": 0.0, + "text": "whatever you want, and we'll see what" + }, + { + "start": 1550.72, + "duration": 0.0, + "text": "whatever you want, and we'll see what happens.\"<00:25:51.600> And<00:25:51.960> state-of-the-art<00:25:52.360> model," + }, + { + "start": 1552.91, + "duration": 0.0, + "text": "happens.\" And state-of-the-art model," + }, + { + "start": 1552.92, + "duration": 0.0, + "text": "happens.\" And state-of-the-art model, right?<00:25:53.400> It's<00:25:53.480> basically<00:25:53.880> vibe<00:25:54.160> engineering." + }, + { + "start": 1555.03, + "duration": 0.0, + "text": "right? It's basically vibe engineering." + }, + { + "start": 1555.04, + "duration": 0.0, + "text": "right? It's basically vibe engineering. Vibe<00:25:55.280> engineering<00:25:55.760> created<00:25:56.080> the<00:25:56.160> vibe<00:25:56.360> coding" + }, + { + "start": 1556.59, + "duration": 0.0, + "text": "Vibe engineering created the vibe coding" + }, + { + "start": 1556.6, + "duration": 0.0, + "text": "Vibe engineering created the vibe coding machines,<00:25:57.480> and<00:25:57.600> everybody's<00:25:58.000> vibing,<00:25:58.440> and" + }, + { + "start": 1558.51, + "duration": 0.0, + "text": "machines, and everybody's vibing, and" + }, + { + "start": 1558.52, + "duration": 0.0, + "text": "machines, and everybody's vibing, and then,<00:25:58.720> you<00:25:58.800> know,<00:25:58.960> we're<00:25:59.080> being<00:25:59.280> wireheaded," + }, + { + "start": 1559.75, + "duration": 0.0, + "text": "then, you know, we're being wireheaded," + }, + { + "start": 1559.76, + "duration": 0.0, + "text": "then, you know, we're being wireheaded, and<00:25:59.840> then<00:25:59.960> it's<00:26:00.160> over." + }, + { + "start": 1561.31, + "duration": 0.0, + "text": "and then it's over." + }, + { + "start": 1561.32, + "duration": 0.0, + "text": "and then it's over. Um" + }, + { + "start": 1565.44, + "duration": 0.0, + "text": "Well," + }, + { + "start": 1566.75, + "duration": 0.0, + "text": "Well," + }, + { + "start": 1566.76, + "duration": 0.0, + "text": "Well, I<00:26:06.880> I<00:26:07.040> I<00:26:07.200> I<00:26:07.400> I<00:26:07.520> think<00:26:07.800> that<00:26:08.160> it's<00:26:08.480> a<00:26:08.560> very" + }, + { + "start": 1568.91, + "duration": 0.0, + "text": "I I I I I think that it's a very" + }, + { + "start": 1568.92, + "duration": 0.0, + "text": "I I I I I think that it's a very interesting<00:26:09.360> question,<00:26:09.760> can<00:26:09.880> these<00:26:10.040> LLMs" + }, + { + "start": 1570.47, + "duration": 0.0, + "text": "interesting question, can these LLMs" + }, + { + "start": 1570.48, + "duration": 0.0, + "text": "interesting question, can these LLMs reason,<00:26:10.880> right?<00:26:11.440> So,<00:26:11.560> they<00:26:11.680> can't<00:26:11.920> reason<00:26:12.160> cuz" + }, + { + "start": 1572.31, + "duration": 0.0, + "text": "reason, right? So, they can't reason cuz" + }, + { + "start": 1572.32, + "duration": 0.0, + "text": "reason, right? So, they can't reason cuz they're<00:26:12.440> not<00:26:12.600> human.<00:26:12.920> I<00:26:12.960> think<00:26:13.120> that's<00:26:13.280> a<00:26:13.320> good" + }, + { + "start": 1574.11, + "duration": 0.0, + "text": "they're not human. I think that's a good" + }, + { + "start": 1574.12, + "duration": 0.0, + "text": "they're not human. I think that's a good way<00:26:14.280> to<00:26:14.360> use<00:26:14.480> the<00:26:14.560> definition<00:26:15.040> of<00:26:15.160> reason.<00:26:15.920> Um" + }, + { + "start": 1576.03, + "duration": 0.0, + "text": "way to use the definition of reason. Um" + }, + { + "start": 1576.04, + "duration": 0.0, + "text": "way to use the definition of reason. Um but<00:26:16.200> let's<00:26:16.360> say<00:26:16.520> we<00:26:16.640> want<00:26:16.800> to<00:26:16.880> figure<00:26:17.080> out<00:26:17.200> if" + }, + { + "start": 1577.23, + "duration": 0.0, + "text": "but let's say we want to figure out if" + }, + { + "start": 1577.24, + "duration": 0.0, + "text": "but let's say we want to figure out if these<00:26:17.440> LLMs<00:26:17.880> can,<00:26:18.360> you<00:26:18.440> know,<00:26:18.680> output<00:26:19.080> text<00:26:19.320> as" + }, + { + "start": 1579.43, + "duration": 0.0, + "text": "these LLMs can, you know, output text as" + }, + { + "start": 1579.44, + "duration": 0.0, + "text": "these LLMs can, you know, output text as if<00:26:19.560> they<00:26:19.680> were<00:26:19.800> reasoning,<00:26:20.600> right?<00:26:20.920> I<00:26:20.960> think" + }, + { + "start": 1581.19, + "duration": 0.0, + "text": "if they were reasoning, right? I think" + }, + { + "start": 1581.2, + "duration": 0.0, + "text": "if they were reasoning, right? I think that<00:26:21.360> this<00:26:21.600> benchmark<00:26:22.080> ARC-AGI<00:26:22.760> is<00:26:23.000> the<00:26:23.560> It's" + }, + { + "start": 1583.75, + "duration": 0.0, + "text": "that this benchmark ARC-AGI is the It's" + }, + { + "start": 1583.76, + "duration": 0.0, + "text": "that this benchmark ARC-AGI is the It's a<00:26:23.840> stellar<00:26:24.280> benchmark<00:26:24.640> for<00:26:24.760> this<00:26:24.880> task," + }, + { + "start": 1585.19, + "duration": 0.0, + "text": "a stellar benchmark for this task," + }, + { + "start": 1585.2, + "duration": 0.0, + "text": "a stellar benchmark for this task, right?<00:26:25.360> If<00:26:25.520> you<00:26:25.600> take<00:26:25.800> a<00:26:25.840> look<00:26:26.000> at<00:26:26.120> this" + }, + { + "start": 1586.27, + "duration": 0.0, + "text": "right? If you take a look at this" + }, + { + "start": 1586.28, + "duration": 0.0, + "text": "right? If you take a look at this diagram<00:26:26.680> here,<00:26:27.000> right?<00:26:27.640> That's<00:26:28.000> uh<00:26:28.160> one" + }, + { + "start": 1588.35, + "duration": 0.0, + "text": "diagram here, right? That's uh one" + }, + { + "start": 1588.36, + "duration": 0.0, + "text": "diagram here, right? That's uh one example,<00:26:28.880> and<00:26:29.000> that's<00:26:29.120> another<00:26:29.360> example,<00:26:29.720> and" + }, + { + "start": 1589.79, + "duration": 0.0, + "text": "example, and that's another example, and" + }, + { + "start": 1589.8, + "duration": 0.0, + "text": "example, and that's another example, and you<00:26:29.840> have<00:26:29.920> to<00:26:30.000> predict<00:26:30.400> what<00:26:30.520> happens<00:26:30.800> here." + }, + { + "start": 1591.67, + "duration": 0.0, + "text": "you have to predict what happens here." + }, + { + "start": 1591.68, + "duration": 0.0, + "text": "you have to predict what happens here. If<00:26:31.880> you're<00:26:32.080> human,<00:26:32.560> which<00:26:32.680> I'm<00:26:32.800> assuming<00:26:33.160> all" + }, + { + "start": 1593.23, + "duration": 0.0, + "text": "If you're human, which I'm assuming all" + }, + { + "start": 1593.24, + "duration": 0.0, + "text": "If you're human, which I'm assuming all of<00:26:33.320> you<00:26:33.440> are,<00:26:34.080> it<00:26:34.160> should<00:26:34.280> be<00:26:34.400> pretty<00:26:34.800> obvious" + }, + { + "start": 1595.23, + "duration": 0.0, + "text": "of you are, it should be pretty obvious" + }, + { + "start": 1595.24, + "duration": 0.0, + "text": "of you are, it should be pretty obvious to<00:26:35.360> you<00:26:35.520> that<00:26:35.640> you<00:26:35.720> basically<00:26:36.200> just<00:26:36.360> put<00:26:36.560> the" + }, + { + "start": 1596.63, + "duration": 0.0, + "text": "to you that you basically just put the" + }, + { + "start": 1596.64, + "duration": 0.0, + "text": "to you that you basically just put the yellow<00:26:36.880> dots<00:26:37.320> like<00:26:37.480> where<00:26:37.560> the<00:26:37.640> little<00:26:37.800> corner" + }, + { + "start": 1598.03, + "duration": 0.0, + "text": "yellow dots like where the little corner" + }, + { + "start": 1598.04, + "duration": 0.0, + "text": "yellow dots like where the little corner thing<00:26:38.200> goes.<00:26:39.040> And<00:26:39.240> ARC-AGI<00:26:39.760> is<00:26:39.840> the<00:26:39.920> kind<00:26:40.080> of" + }, + { + "start": 1600.15, + "duration": 0.0, + "text": "thing goes. And ARC-AGI is the kind of" + }, + { + "start": 1600.16, + "duration": 0.0, + "text": "thing goes. And ARC-AGI is the kind of benchmark<00:26:40.520> where<00:26:40.600> they<00:26:40.720> basically<00:26:41.560> take" + }, + { + "start": 1601.75, + "duration": 0.0, + "text": "benchmark where they basically take" + }, + { + "start": 1601.76, + "duration": 0.0, + "text": "benchmark where they basically take random<00:26:42.000> people<00:26:42.280> and<00:26:42.360> they're<00:26:42.440> like,<00:26:42.560> \"All" + }, + { + "start": 1602.59, + "duration": 0.0, + "text": "random people and they're like, \"All" + }, + { + "start": 1602.6, + "duration": 0.0, + "text": "random people and they're like, \"All right,<00:26:42.680> we'll<00:26:42.760> give<00:26:42.880> you<00:26:42.920> a<00:26:42.960> $15<00:26:43.520> Starbucks" + }, + { + "start": 1603.87, + "duration": 0.0, + "text": "right, we'll give you a $15 Starbucks" + }, + { + "start": 1603.88, + "duration": 0.0, + "text": "right, we'll give you a $15 Starbucks gift<00:26:44.080> card<00:26:44.240> if<00:26:44.360> you<00:26:44.400> solve<00:26:44.560> this<00:26:44.680> problem.\"" + }, + { + "start": 1605.39, + "duration": 0.0, + "text": "gift card if you solve this problem.\"" + }, + { + "start": 1605.4, + "duration": 0.0, + "text": "gift card if you solve this problem.\" And<00:26:45.520> they<00:26:45.640> verify<00:26:46.120> that<00:26:46.280> like<00:26:46.840> people<00:26:47.120> can" + }, + { + "start": 1607.27, + "duration": 0.0, + "text": "And they verify that like people can" + }, + { + "start": 1607.28, + "duration": 0.0, + "text": "And they verify that like people can just<00:26:47.400> figure<00:26:47.600> it<00:26:47.720> out<00:26:47.880> in<00:26:47.960> a<00:26:48.000> context-free" + }, + { + "start": 1608.55, + "duration": 0.0, + "text": "just figure it out in a context-free" + }, + { + "start": 1608.56, + "duration": 0.0, + "text": "just figure it out in a context-free manner,<00:26:48.840> right?<00:26:49.120> And<00:26:49.200> so,<00:26:49.800> the<00:26:49.920> whole<00:26:50.080> idea" + }, + { + "start": 1610.31, + "duration": 0.0, + "text": "manner, right? And so, the whole idea" + }, + { + "start": 1610.32, + "duration": 0.0, + "text": "manner, right? And so, the whole idea with<00:26:50.440> this<00:26:50.600> is<00:26:50.680> they're<00:26:50.800> trying<00:26:51.000> to<00:26:51.120> find" + }, + { + "start": 1611.35, + "duration": 0.0, + "text": "with this is they're trying to find" + }, + { + "start": 1611.36, + "duration": 0.0, + "text": "with this is they're trying to find problems<00:26:51.880> that<00:26:52.040> like<00:26:52.280> average<00:26:52.560> people<00:26:52.840> can" + }, + { + "start": 1612.99, + "duration": 0.0, + "text": "problems that like average people can" + }, + { + "start": 1613.0, + "duration": 0.0, + "text": "problems that like average people can do,<00:26:53.200> but<00:26:53.360> like,<00:26:53.920> you<00:26:54.000> know,<00:26:54.120> these<00:26:54.280> LLMs" + }, + { + "start": 1614.59, + "duration": 0.0, + "text": "do, but like, you know, these LLMs" + }, + { + "start": 1614.6, + "duration": 0.0, + "text": "do, but like, you know, these LLMs they're<00:26:55.120> they<00:26:55.240> can't<00:26:55.440> do<00:26:55.560> it<00:26:55.640> at<00:26:55.760> all.<00:26:56.240> Um" + }, + { + "start": 1617.27, + "duration": 0.0, + "text": "they're they can't do it at all. Um" + }, + { + "start": 1617.28, + "duration": 0.0, + "text": "they're they can't do it at all. Um And<00:26:57.480> ARK<00:26:57.640> G<00:26:57.880> I<00:26:58.480> they<00:26:58.960> were<00:26:59.240> originally,<00:26:59.960> you" + }, + { + "start": 1620.03, + "duration": 0.0, + "text": "And ARK G I they were originally, you" + }, + { + "start": 1620.04, + "duration": 0.0, + "text": "And ARK G I they were originally, you know,<00:27:00.120> totally<00:27:00.400> defeated<00:27:00.800> by<00:27:00.960> OpenAI's" + }, + { + "start": 1621.75, + "duration": 0.0, + "text": "know, totally defeated by OpenAI's" + }, + { + "start": 1621.76, + "duration": 0.0, + "text": "know, totally defeated by OpenAI's latest<00:27:02.120> reasoning<00:27:02.400> model." + }, + { + "start": 1623.39, + "duration": 0.0, + "text": "latest reasoning model." + }, + { + "start": 1623.4, + "duration": 0.0, + "text": "latest reasoning model. Um<00:27:03.480> it<00:27:03.800> it<00:27:03.920> got<00:27:04.200> human<00:27:04.440> baseline<00:27:04.880> on<00:27:05.000> a<00:27:05.080> huge" + }, + { + "start": 1625.67, + "duration": 0.0, + "text": "Um it it got human baseline on a huge" + }, + { + "start": 1625.68, + "duration": 0.0, + "text": "Um it it got human baseline on a huge suite<00:27:06.040> of<00:27:06.160> these<00:27:06.360> that<00:27:06.480> are<00:27:06.560> private,<00:27:07.120> not" + }, + { + "start": 1627.35, + "duration": 0.0, + "text": "suite of these that are private, not" + }, + { + "start": 1627.36, + "duration": 0.0, + "text": "suite of these that are private, not anywhere<00:27:07.640> existing<00:27:08.000> on<00:27:08.080> the<00:27:08.160> internet.<00:27:09.240> Um" + }, + { + "start": 1629.35, + "duration": 0.0, + "text": "anywhere existing on the internet. Um" + }, + { + "start": 1629.36, + "duration": 0.0, + "text": "anywhere existing on the internet. Um you<00:27:09.440> know,<00:27:09.520> generic<00:27:10.160> Each<00:27:10.400> each<00:27:10.720> ARK<00:27:10.920> G<00:27:11.040> I" + }, + { + "start": 1631.15, + "duration": 0.0, + "text": "you know, generic Each each ARK G I" + }, + { + "start": 1631.16, + "duration": 0.0, + "text": "you know, generic Each each ARK G I problem<00:27:11.520> requires<00:27:11.920> a<00:27:11.960> new<00:27:12.160> unique<00:27:12.480> insight." + }, + { + "start": 1632.83, + "duration": 0.0, + "text": "problem requires a new unique insight." + }, + { + "start": 1632.84, + "duration": 0.0, + "text": "problem requires a new unique insight. They're<00:27:12.960> basically<00:27:13.320> puzzle<00:27:13.600> game<00:27:13.760> problems," + }, + { + "start": 1634.19, + "duration": 0.0, + "text": "They're basically puzzle game problems," + }, + { + "start": 1634.2, + "duration": 0.0, + "text": "They're basically puzzle game problems, right?<00:27:14.680> They<00:27:14.760> require<00:27:15.000> some<00:27:15.240> kind<00:27:15.480> of<00:27:15.600> like," + }, + { + "start": 1636.15, + "duration": 0.0, + "text": "right? They require some kind of like," + }, + { + "start": 1636.16, + "duration": 0.0, + "text": "right? They require some kind of like, \"Oh,<00:27:16.360> that's<00:27:16.600> how<00:27:16.720> it<00:27:16.840> works,\"<00:27:17.080> right?<00:27:17.280> And" + }, + { + "start": 1637.83, + "duration": 0.0, + "text": "\"Oh, that's how it works,\" right? And" + }, + { + "start": 1637.84, + "duration": 0.0, + "text": "\"Oh, that's how it works,\" right? And LLMs<00:27:18.200> have<00:27:18.320> already<00:27:18.560> beaten<00:27:18.800> the<00:27:18.920> first<00:27:19.360> ARK<00:27:19.520> G" + }, + { + "start": 1639.67, + "duration": 0.0, + "text": "LLMs have already beaten the first ARK G" + }, + { + "start": 1639.68, + "duration": 0.0, + "text": "LLMs have already beaten the first ARK G I<00:27:20.080> benchmark.<00:27:20.520> Which<00:27:20.640> that<00:27:20.760> was<00:27:20.880> quite" + }, + { + "start": 1641.07, + "duration": 0.0, + "text": "I benchmark. Which that was quite" + }, + { + "start": 1641.08, + "duration": 0.0, + "text": "I benchmark. Which that was quite significant.<00:27:21.520> I<00:27:21.560> didn't<00:27:21.720> think<00:27:21.840> that<00:27:21.920> would" + }, + { + "start": 1641.99, + "duration": 0.0, + "text": "significant. I didn't think that would" + }, + { + "start": 1642.0, + "duration": 0.0, + "text": "significant. I didn't think that would happen.<00:27:22.680> And<00:27:22.800> neither<00:27:23.000> did<00:27:23.120> the<00:27:23.200> author<00:27:23.400> of" + }, + { + "start": 1643.43, + "duration": 0.0, + "text": "happen. And neither did the author of" + }, + { + "start": 1643.44, + "duration": 0.0, + "text": "happen. And neither did the author of the<00:27:23.520> benchmark.<00:27:23.880> He's<00:27:24.000> a<00:27:24.040> massive<00:27:24.360> skeptic" + }, + { + "start": 1644.71, + "duration": 0.0, + "text": "the benchmark. He's a massive skeptic" + }, + { + "start": 1644.72, + "duration": 0.0, + "text": "the benchmark. He's a massive skeptic and<00:27:24.800> he<00:27:24.840> thinks<00:27:25.040> it's<00:27:25.160> a<00:27:25.200> bubble.<00:27:25.880> And<00:27:26.080> then" + }, + { + "start": 1646.31, + "duration": 0.0, + "text": "and he thinks it's a bubble. And then" + }, + { + "start": 1646.32, + "duration": 0.0, + "text": "and he thinks it's a bubble. And then OpenAI<00:27:26.680> pulled<00:27:26.920> them<00:27:27.080> on<00:27:27.200> stream<00:27:27.520> and<00:27:27.720> he<00:27:27.720> was" + }, + { + "start": 1647.83, + "duration": 0.0, + "text": "OpenAI pulled them on stream and he was" + }, + { + "start": 1647.84, + "duration": 0.0, + "text": "OpenAI pulled them on stream and he was like,<00:27:28.000> \"My<00:27:28.120> benchmark<00:27:28.520> is<00:27:28.800> defeated.\"<00:27:29.760> Um<00:27:30.400> so" + }, + { + "start": 1650.51, + "duration": 0.0, + "text": "like, \"My benchmark is defeated.\" Um so" + }, + { + "start": 1650.52, + "duration": 0.0, + "text": "like, \"My benchmark is defeated.\" Um so we<00:27:30.560> made<00:27:30.680> a<00:27:30.720> new<00:27:30.840> benchmark<00:27:31.280> that<00:27:31.440> they're<00:27:31.520> all" + }, + { + "start": 1651.63, + "duration": 0.0, + "text": "we made a new benchmark that they're all" + }, + { + "start": 1651.64, + "duration": 0.0, + "text": "we made a new benchmark that they're all failing<00:27:31.960> at<00:27:32.040> again." + }, + { + "start": 1653.03, + "duration": 0.0, + "text": "failing at again." + }, + { + "start": 1653.04, + "duration": 0.0, + "text": "failing at again. And<00:27:33.760> that's<00:27:34.080> Grok<00:27:34.400> 4.<00:27:34.960> A<00:27:35.000> few<00:27:35.160> days<00:27:35.360> ago<00:27:35.560> they" + }, + { + "start": 1655.67, + "duration": 0.0, + "text": "And that's Grok 4. A few days ago they" + }, + { + "start": 1655.68, + "duration": 0.0, + "text": "And that's Grok 4. A few days ago they just<00:27:36.000> doubled<00:27:36.320> the<00:27:36.400> state<00:27:36.600> of<00:27:36.640> the<00:27:36.760> art<00:27:36.880> and" + }, + { + "start": 1656.95, + "duration": 0.0, + "text": "just doubled the state of the art and" + }, + { + "start": 1656.96, + "duration": 0.0, + "text": "just doubled the state of the art and got<00:27:37.120> 15%<00:27:38.120> the<00:27:38.240> 16%<00:27:38.840> of<00:27:38.920> the<00:27:39.000> questions<00:27:39.360> right." + }, + { + "start": 1660.51, + "duration": 0.0, + "text": "got 15% the 16% of the questions right." + }, + { + "start": 1660.52, + "duration": 0.0, + "text": "got 15% the 16% of the questions right. Um<00:27:41.440> So<00:27:41.520> just<00:27:41.720> circle<00:27:41.960> back<00:27:42.160> on<00:27:42.240> some" + }, + { + "start": 1662.39, + "duration": 0.0, + "text": "Um So just circle back on some" + }, + { + "start": 1662.4, + "duration": 0.0, + "text": "Um So just circle back on some applications<00:27:42.920> to<00:27:43.000> games<00:27:43.320> really<00:27:43.480> quick<00:27:43.960> as<00:27:44.080> I" + }, + { + "start": 1664.11, + "duration": 0.0, + "text": "applications to games really quick as I" + }, + { + "start": 1664.12, + "duration": 0.0, + "text": "applications to games really quick as I kind<00:27:44.320> of" + }, + { + "start": 1665.03, + "duration": 0.0, + "text": "kind of" + }, + { + "start": 1665.04, + "duration": 0.0, + "text": "kind of close<00:27:45.280> up<00:27:45.440> on<00:27:45.520> this<00:27:45.760> kind<00:27:45.960> of<00:27:46.000> like<00:27:46.200> lightning" + }, + { + "start": 1666.51, + "duration": 0.0, + "text": "close up on this kind of like lightning" + }, + { + "start": 1666.52, + "duration": 0.0, + "text": "close up on this kind of like lightning overview<00:27:46.880> of<00:27:46.960> this,<00:27:47.240> right?<00:27:48.000> Any<00:27:48.200> problem" + }, + { + "start": 1668.55, + "duration": 0.0, + "text": "overview of this, right? Any problem" + }, + { + "start": 1668.56, + "duration": 0.0, + "text": "overview of this, right? Any problem where<00:27:48.680> you<00:27:48.760> have<00:27:48.960> input<00:27:49.240> to<00:27:49.320> output<00:27:49.640> data<00:27:50.000> and" + }, + { + "start": 1670.07, + "duration": 0.0, + "text": "where you have input to output data and" + }, + { + "start": 1670.08, + "duration": 0.0, + "text": "where you have input to output data and it's<00:27:50.200> harder<00:27:50.480> to<00:27:50.560> find<00:27:50.800> programmatically" + }, + { + "start": 1671.47, + "duration": 0.0, + "text": "it's harder to find programmatically" + }, + { + "start": 1671.48, + "duration": 0.0, + "text": "it's harder to find programmatically what<00:27:51.640> that<00:27:51.800> mapping<00:27:52.160> is,<00:27:52.760> that's<00:27:53.000> where<00:27:53.120> you" + }, + { + "start": 1673.23, + "duration": 0.0, + "text": "what that mapping is, that's where you" + }, + { + "start": 1673.24, + "duration": 0.0, + "text": "what that mapping is, that's where you would<00:27:53.360> throw<00:27:53.520> something<00:27:53.840> like<00:27:54.360> deep<00:27:54.560> learning" + }, + { + "start": 1674.79, + "duration": 0.0, + "text": "would throw something like deep learning" + }, + { + "start": 1674.8, + "duration": 0.0, + "text": "would throw something like deep learning at<00:27:54.880> the<00:27:54.960> problem<00:27:55.360> and<00:27:55.440> try<00:27:55.600> to<00:27:55.840> automatically" + }, + { + "start": 1676.55, + "duration": 0.0, + "text": "at the problem and try to automatically" + }, + { + "start": 1676.56, + "duration": 0.0, + "text": "at the problem and try to automatically program<00:27:57.080> that,<00:27:57.560> you<00:27:57.640> know,<00:27:58.000> uh" + }, + { + "start": 1678.35, + "duration": 0.0, + "text": "program that, you know, uh" + }, + { + "start": 1678.36, + "duration": 0.0, + "text": "program that, you know, uh transformation." + }, + { + "start": 1679.99, + "duration": 0.0, + "text": "transformation." + }, + { + "start": 1680.0, + "duration": 0.0, + "text": "transformation. Um<00:28:00.160> this<00:28:00.400> is<00:28:00.560> Nerf<00:28:00.840> rendering.<00:28:01.200> This<00:28:01.320> is<00:28:01.400> a" + }, + { + "start": 1681.43, + "duration": 0.0, + "text": "Um this is Nerf rendering. This is a" + }, + { + "start": 1681.44, + "duration": 0.0, + "text": "Um this is Nerf rendering. This is a very<00:28:01.640> interesting<00:28:02.000> application<00:28:02.640> of<00:28:02.880> deep" + }, + { + "start": 1683.07, + "duration": 0.0, + "text": "very interesting application of deep" + }, + { + "start": 1683.08, + "duration": 0.0, + "text": "very interesting application of deep learning<00:28:03.360> to<00:28:03.480> the<00:28:03.560> problem<00:28:03.800> of<00:28:03.880> rendering,<00:28:04.240> I" + }, + { + "start": 1684.31, + "duration": 0.0, + "text": "learning to the problem of rendering, I" + }, + { + "start": 1684.32, + "duration": 0.0, + "text": "learning to the problem of rendering, I think.<00:28:04.880> Um<00:28:05.200> it's<00:28:05.320> very<00:28:05.560> old<00:28:06.160> and<00:28:06.280> there's<00:28:06.400> much" + }, + { + "start": 1686.55, + "duration": 0.0, + "text": "think. Um it's very old and there's much" + }, + { + "start": 1686.56, + "duration": 0.0, + "text": "think. Um it's very old and there's much better<00:28:06.760> techniques<00:28:07.360> like<00:28:07.560> uh<00:28:07.720> Gaussian" + }, + { + "start": 1687.99, + "duration": 0.0, + "text": "better techniques like uh Gaussian" + }, + { + "start": 1688.0, + "duration": 0.0, + "text": "better techniques like uh Gaussian splatting." + }, + { + "start": 1689.03, + "duration": 0.0, + "text": "splatting." + }, + { + "start": 1689.04, + "duration": 0.0, + "text": "splatting. Um<00:28:09.120> but<00:28:09.360> Nerf<00:28:09.800> rendering<00:28:10.120> is<00:28:10.200> basically<00:28:10.920> given" + }, + { + "start": 1691.19, + "duration": 0.0, + "text": "Um but Nerf rendering is basically given" + }, + { + "start": 1691.2, + "duration": 0.0, + "text": "Um but Nerf rendering is basically given a<00:28:11.240> point<00:28:11.520> in<00:28:11.600> the<00:28:11.680> scene,<00:28:12.000> a<00:28:12.040> deep<00:28:12.280> learning" + }, + { + "start": 1692.51, + "duration": 0.0, + "text": "a point in the scene, a deep learning" + }, + { + "start": 1692.52, + "duration": 0.0, + "text": "a point in the scene, a deep learning model<00:28:12.840> learns<00:28:13.160> how<00:28:13.280> dense<00:28:13.560> it<00:28:13.680> is<00:28:14.280> and<00:28:14.400> then" + }, + { + "start": 1694.51, + "duration": 0.0, + "text": "model learns how dense it is and then" + }, + { + "start": 1694.52, + "duration": 0.0, + "text": "model learns how dense it is and then it,<00:28:14.640> you<00:28:14.720> know,<00:28:14.800> kind<00:28:14.960> of<00:28:15.040> traces<00:28:15.360> rays" + }, + { + "start": 1695.59, + "duration": 0.0, + "text": "it, you know, kind of traces rays" + }, + { + "start": 1695.6, + "duration": 0.0, + "text": "it, you know, kind of traces rays through<00:28:15.920> that<00:28:16.160> learned<00:28:16.360> density<00:28:16.680> to<00:28:16.800> render." + }, + { + "start": 1697.71, + "duration": 0.0, + "text": "through that learned density to render." + }, + { + "start": 1697.72, + "duration": 0.0, + "text": "through that learned density to render. Um" + }, + { + "start": 1698.67, + "duration": 0.0, + "text": "Um" + }, + { + "start": 1698.68, + "duration": 0.0, + "text": "Um And<00:28:19.000> just<00:28:19.200> as<00:28:19.280> a<00:28:19.320> little<00:28:19.680> interesting<00:28:20.080> tidbit" + }, + { + "start": 1700.63, + "duration": 0.0, + "text": "And just as a little interesting tidbit" + }, + { + "start": 1700.64, + "duration": 0.0, + "text": "And just as a little interesting tidbit more<00:28:21.080> vibe<00:28:21.760> uh<00:28:21.920> engineering,<00:28:22.960> uh<00:28:23.080> basically" + }, + { + "start": 1703.79, + "duration": 0.0, + "text": "more vibe uh engineering, uh basically" + }, + { + "start": 1703.8, + "duration": 0.0, + "text": "more vibe uh engineering, uh basically deep<00:28:24.000> learning<00:28:24.200> models<00:28:24.440> are<00:28:24.520> very<00:28:24.720> bad<00:28:24.960> at" + }, + { + "start": 1705.07, + "duration": 0.0, + "text": "deep learning models are very bad at" + }, + { + "start": 1705.08, + "duration": 0.0, + "text": "deep learning models are very bad at learning<00:28:25.400> so-called<00:28:25.760> high-frequency<00:28:27.120> um" + }, + { + "start": 1708.47, + "duration": 0.0, + "text": "learning so-called high-frequency um" + }, + { + "start": 1708.48, + "duration": 0.0, + "text": "learning so-called high-frequency um high-frequency<00:28:29.080> functions,<00:28:29.480> right?<00:28:29.680> So<00:28:29.840> a" + }, + { + "start": 1709.91, + "duration": 0.0, + "text": "high-frequency functions, right? So a" + }, + { + "start": 1709.92, + "duration": 0.0, + "text": "high-frequency functions, right? So a little<00:28:30.160> tiny<00:28:30.440> thin<00:28:30.640> metal<00:28:30.880> pole<00:28:31.120> is<00:28:31.200> a<00:28:31.240> very" + }, + { + "start": 1711.43, + "duration": 0.0, + "text": "little tiny thin metal pole is a very" + }, + { + "start": 1711.44, + "duration": 0.0, + "text": "little tiny thin metal pole is a very high-frequency<00:28:32.040> function<00:28:32.520> density<00:28:32.840> function" + }, + { + "start": 1713.11, + "duration": 0.0, + "text": "high-frequency function density function" + }, + { + "start": 1713.12, + "duration": 0.0, + "text": "high-frequency function density function for<00:28:33.240> it<00:28:33.280> to<00:28:33.360> learn.<00:28:33.960> It's<00:28:34.040> like<00:28:34.200> nothing" + }, + { + "start": 1714.43, + "duration": 0.0, + "text": "for it to learn. It's like nothing" + }, + { + "start": 1714.44, + "duration": 0.0, + "text": "for it to learn. It's like nothing nothing<00:28:34.680> nothing<00:28:34.960> bam<00:28:35.360> really<00:28:35.560> thin<00:28:35.760> nothing" + }, + { + "start": 1715.99, + "duration": 0.0, + "text": "nothing nothing bam really thin nothing" + }, + { + "start": 1716.0, + "duration": 0.0, + "text": "nothing nothing bam really thin nothing nothing<00:28:36.200> nothing.<00:28:36.400> It's<00:28:36.520> very<00:28:36.760> like<00:28:37.280> tight" + }, + { + "start": 1717.55, + "duration": 0.0, + "text": "nothing nothing. It's very like tight" + }, + { + "start": 1717.56, + "duration": 0.0, + "text": "nothing nothing. It's very like tight and<00:28:37.640> so<00:28:37.720> it's<00:28:37.840> hard<00:28:38.080> for<00:28:38.280> a<00:28:38.360> deep<00:28:38.520> learning" + }, + { + "start": 1718.67, + "duration": 0.0, + "text": "and so it's hard for a deep learning" + }, + { + "start": 1718.68, + "duration": 0.0, + "text": "and so it's hard for a deep learning model<00:28:38.840> to<00:28:38.960> pick<00:28:39.160> up<00:28:39.280> on<00:28:39.400> that<00:28:39.920> cuz<00:28:40.040> it<00:28:40.080> would" + }, + { + "start": 1720.19, + "duration": 0.0, + "text": "model to pick up on that cuz it would" + }, + { + "start": 1720.2, + "duration": 0.0, + "text": "model to pick up on that cuz it would require<00:28:40.560> a<00:28:40.600> lot<00:28:40.800> of<00:28:40.880> non-linearity.<00:28:41.480> It<00:28:41.520> would" + }, + { + "start": 1721.59, + "duration": 0.0, + "text": "require a lot of non-linearity. It would" + }, + { + "start": 1721.6, + "duration": 0.0, + "text": "require a lot of non-linearity. It would require<00:28:41.760> a<00:28:41.800> lot<00:28:41.960> of<00:28:42.040> layers,<00:28:42.520> right?<00:28:43.280> Um<00:28:43.560> and" + }, + { + "start": 1723.67, + "duration": 0.0, + "text": "require a lot of layers, right? Um and" + }, + { + "start": 1723.68, + "duration": 0.0, + "text": "require a lot of layers, right? Um and so<00:28:43.760> what<00:28:43.880> they<00:28:44.000> did<00:28:44.400> in<00:28:44.520> Nerf<00:28:44.840> to,<00:28:45.240> you<00:28:45.320> know," + }, + { + "start": 1725.43, + "duration": 0.0, + "text": "so what they did in Nerf to, you know," + }, + { + "start": 1725.44, + "duration": 0.0, + "text": "so what they did in Nerf to, you know, achieve<00:28:45.680> state-of-the-art<00:28:46.240> results<00:28:46.600> for" + }, + { + "start": 1726.95, + "duration": 0.0, + "text": "achieve state-of-the-art results for" + }, + { + "start": 1726.96, + "duration": 0.0, + "text": "achieve state-of-the-art results for learning<00:28:47.200> based<00:28:47.400> rendering,<00:28:48.160> is<00:28:48.360> they<00:28:48.480> pass" + }, + { + "start": 1728.91, + "duration": 0.0, + "text": "learning based rendering, is they pass" + }, + { + "start": 1728.92, + "duration": 0.0, + "text": "learning based rendering, is they pass that<00:28:49.200> position<00:28:49.760> into" + }, + { + "start": 1731.27, + "duration": 0.0, + "text": "that position into" + }, + { + "start": 1731.28, + "duration": 0.0, + "text": "that position into into<00:28:51.480> a<00:28:51.560> bunch<00:28:51.720> of<00:28:51.800> different<00:28:52.040> sine<00:28:52.280> waves<00:28:52.760> and" + }, + { + "start": 1733.07, + "duration": 0.0, + "text": "into a bunch of different sine waves and" + }, + { + "start": 1733.08, + "duration": 0.0, + "text": "into a bunch of different sine waves and they<00:28:53.200> basically<00:28:53.600> just<00:28:53.760> made<00:28:53.920> it<00:28:54.000> easier<00:28:54.240> for" + }, + { + "start": 1734.31, + "duration": 0.0, + "text": "they basically just made it easier for" + }, + { + "start": 1734.32, + "duration": 0.0, + "text": "they basically just made it easier for the<00:28:54.400> deep<00:28:54.560> learning<00:28:54.760> model<00:28:54.960> to<00:28:55.040> pick<00:28:55.280> up<00:28:55.440> on" + }, + { + "start": 1735.99, + "duration": 0.0, + "text": "the deep learning model to pick up on" + }, + { + "start": 1736.0, + "duration": 0.0, + "text": "the deep learning model to pick up on higher<00:28:56.240> frequency<00:28:57.000> spatial<00:28:57.320> information.<00:28:58.160> I" + }, + { + "start": 1738.19, + "duration": 0.0, + "text": "higher frequency spatial information. I" + }, + { + "start": 1738.2, + "duration": 0.0, + "text": "higher frequency spatial information. I thought<00:28:58.360> that<00:28:58.440> was<00:28:58.560> kind<00:28:58.680> of<00:28:58.760> funny.<00:28:59.040> She's" + }, + { + "start": 1739.15, + "duration": 0.0, + "text": "thought that was kind of funny. She's" + }, + { + "start": 1739.16, + "duration": 0.0, + "text": "thought that was kind of funny. She's like,<00:28:59.240> pass<00:28:59.440> it<00:28:59.520> into<00:28:59.640> the<00:28:59.720> sine<00:28:59.920> waves<00:29:00.080> and" + }, + { + "start": 1740.19, + "duration": 0.0, + "text": "like, pass it into the sine waves and" + }, + { + "start": 1740.2, + "duration": 0.0, + "text": "like, pass it into the sine waves and then<00:29:00.280> it's<00:29:00.360> better." + }, + { + "start": 1741.23, + "duration": 0.0, + "text": "then it's better." + }, + { + "start": 1741.24, + "duration": 0.0, + "text": "then it's better. Um<00:29:01.600> this<00:29:01.760> is<00:29:01.840> called<00:29:02.040> Ash<00:29:02.360> Animatable" + }, + { + "start": 1742.83, + "duration": 0.0, + "text": "Um this is called Ash Animatable" + }, + { + "start": 1742.84, + "duration": 0.0, + "text": "Um this is called Ash Animatable Gaussian<00:29:03.240> Splatting.<00:29:03.640> Let<00:29:03.720> me<00:29:03.760> show<00:29:03.840> you<00:29:03.920> how" + }, + { + "start": 1743.95, + "duration": 0.0, + "text": "Gaussian Splatting. Let me show you how" + }, + { + "start": 1743.96, + "duration": 0.0, + "text": "Gaussian Splatting. Let me show you how the<00:29:04.040> video<00:29:04.920> works<00:29:05.200> here.<00:29:05.800> This<00:29:06.000> is<00:29:06.120> basically" + }, + { + "start": 1746.55, + "duration": 0.0, + "text": "the video works here. This is basically" + }, + { + "start": 1746.56, + "duration": 0.0, + "text": "the video works here. This is basically photorealistic<00:29:07.280> human<00:29:07.560> rendering<00:29:07.920> with<00:29:08.120> with" + }, + { + "start": 1748.31, + "duration": 0.0, + "text": "photorealistic human rendering with with" + }, + { + "start": 1748.32, + "duration": 0.0, + "text": "photorealistic human rendering with with accurate<00:29:08.640> physics<00:29:08.920> on<00:29:09.040> their<00:29:09.120> clothes<00:29:09.480> done" + }, + { + "start": 1749.67, + "duration": 0.0, + "text": "accurate physics on their clothes done" + }, + { + "start": 1749.68, + "duration": 0.0, + "text": "accurate physics on their clothes done in<00:29:09.840> real<00:29:10.040> time.<00:29:11.000> Um<00:29:11.520> it's<00:29:11.640> just" + }, + { + "start": 1751.75, + "duration": 0.0, + "text": "in real time. Um it's just" + }, + { + "start": 1751.76, + "duration": 0.0, + "text": "in real time. Um it's just unfortunately,<00:29:12.360> in<00:29:12.440> order<00:29:12.600> to,<00:29:13.480> you<00:29:13.560> know," + }, + { + "start": 1753.79, + "duration": 0.0, + "text": "unfortunately, in order to, you know," + }, + { + "start": 1753.8, + "duration": 0.0, + "text": "unfortunately, in order to, you know, scan<00:29:14.160> somebody,<00:29:14.440> you<00:29:14.520> need<00:29:14.640> about<00:29:14.800> 100" + }, + { + "start": 1755.19, + "duration": 0.0, + "text": "scan somebody, you need about 100" + }, + { + "start": 1755.2, + "duration": 0.0, + "text": "scan somebody, you need about 100 cameras<00:29:15.600> synchronized<00:29:16.160> and<00:29:16.240> it's<00:29:16.560> very" + }, + { + "start": 1756.79, + "duration": 0.0, + "text": "cameras synchronized and it's very" + }, + { + "start": 1756.8, + "duration": 0.0, + "text": "cameras synchronized and it's very expensive.<00:29:17.720> So,<00:29:17.880> maybe<00:29:18.080> they'll<00:29:18.200> figure<00:29:18.400> out" + }, + { + "start": 1758.51, + "duration": 0.0, + "text": "expensive. So, maybe they'll figure out" + }, + { + "start": 1758.52, + "duration": 0.0, + "text": "expensive. So, maybe they'll figure out how<00:29:18.560> to<00:29:18.680> make<00:29:18.800> that<00:29:18.920> cheaper<00:29:19.160> someday,<00:29:19.520> but<00:29:19.840> in" + }, + { + "start": 1759.95, + "duration": 0.0, + "text": "how to make that cheaper someday, but in" + }, + { + "start": 1759.96, + "duration": 0.0, + "text": "how to make that cheaper someday, but in this<00:29:20.080> specific<00:29:20.480> paper,<00:29:21.200> a<00:29:21.280> deep<00:29:21.480> learning" + }, + { + "start": 1761.67, + "duration": 0.0, + "text": "this specific paper, a deep learning" + }, + { + "start": 1761.68, + "duration": 0.0, + "text": "this specific paper, a deep learning model<00:29:21.920> is<00:29:22.040> used<00:29:22.280> to<00:29:22.400> go<00:29:22.520> from<00:29:22.960> a<00:29:23.080> texture<00:29:23.520> that" + }, + { + "start": 1763.67, + "duration": 0.0, + "text": "model is used to go from a texture that" + }, + { + "start": 1763.68, + "duration": 0.0, + "text": "model is used to go from a texture that describes<00:29:24.120> the<00:29:24.240> skeletal<00:29:24.800> deformation<00:29:25.520> of" + }, + { + "start": 1765.79, + "duration": 0.0, + "text": "describes the skeletal deformation of" + }, + { + "start": 1765.8, + "duration": 0.0, + "text": "describes the skeletal deformation of what<00:29:25.960> you<00:29:26.080> want<00:29:26.240> the<00:29:26.320> person<00:29:26.600> to<00:29:26.680> be<00:29:26.840> in<00:29:27.640> into" + }, + { + "start": 1768.19, + "duration": 0.0, + "text": "what you want the person to be in into" + }, + { + "start": 1768.2, + "duration": 0.0, + "text": "what you want the person to be in into the<00:29:28.480> the<00:29:28.600> view<00:29:28.760> dependent<00:29:29.160> color<00:29:29.480> that<00:29:29.640> you" + }, + { + "start": 1769.71, + "duration": 0.0, + "text": "the the view dependent color that you" + }, + { + "start": 1769.72, + "duration": 0.0, + "text": "the the view dependent color that you would<00:29:29.880> see" + }, + { + "start": 1770.83, + "duration": 0.0, + "text": "would see" + }, + { + "start": 1770.84, + "duration": 0.0, + "text": "would see um<00:29:30.960> from<00:29:31.480> that<00:29:31.680> skeletal<00:29:31.960> deformation.<00:29:32.440> And" + }, + { + "start": 1772.59, + "duration": 0.0, + "text": "um from that skeletal deformation. And" + }, + { + "start": 1772.6, + "duration": 0.0, + "text": "um from that skeletal deformation. And and<00:29:32.920> essentially,<00:29:33.360> you<00:29:33.440> need<00:29:33.560> to<00:29:33.640> train<00:29:34.000> that" + }, + { + "start": 1774.15, + "duration": 0.0, + "text": "and essentially, you need to train that" + }, + { + "start": 1774.16, + "duration": 0.0, + "text": "and essentially, you need to train that deep<00:29:34.280> learning<00:29:34.480> model<00:29:34.680> to<00:29:34.800> learn<00:29:34.960> that" + }, + { + "start": 1775.07, + "duration": 0.0, + "text": "deep learning model to learn that" + }, + { + "start": 1775.08, + "duration": 0.0, + "text": "deep learning model to learn that relationship<00:29:35.640> between<00:29:35.920> like<00:29:36.400> how<00:29:36.560> their<00:29:36.720> arm" + }, + { + "start": 1776.91, + "duration": 0.0, + "text": "relationship between like how their arm" + }, + { + "start": 1776.92, + "duration": 0.0, + "text": "relationship between like how their arm is<00:29:37.040> bending<00:29:37.440> and<00:29:37.720> what<00:29:37.880> their<00:29:38.000> clothes<00:29:38.280> look" + }, + { + "start": 1778.43, + "duration": 0.0, + "text": "is bending and what their clothes look" + }, + { + "start": 1778.44, + "duration": 0.0, + "text": "is bending and what their clothes look like,<00:29:39.280> um<00:29:39.560> you<00:29:39.640> know,<00:29:39.760> per<00:29:39.920> person.<00:29:40.760> Um" + }, + { + "start": 1781.55, + "duration": 0.0, + "text": "like, um you know, per person. Um" + }, + { + "start": 1781.56, + "duration": 0.0, + "text": "like, um you know, per person. Um And<00:29:41.640> then<00:29:41.720> of<00:29:41.840> course,<00:29:42.040> I'm<00:29:42.160> sure<00:29:42.640> probably" + }, + { + "start": 1783.03, + "duration": 0.0, + "text": "And then of course, I'm sure probably" + }, + { + "start": 1783.04, + "duration": 0.0, + "text": "And then of course, I'm sure probably all<00:29:43.120> of<00:29:43.200> you<00:29:43.240> guys<00:29:43.400> already<00:29:43.600> know,<00:29:44.320> ray" + }, + { + "start": 1784.47, + "duration": 0.0, + "text": "all of you guys already know, ray" + }, + { + "start": 1784.48, + "duration": 0.0, + "text": "all of you guys already know, ray tracing<00:29:44.960> is<00:29:45.400> is<00:29:45.760> is<00:29:46.040> greatly<00:29:46.400> benefited<00:29:46.880> by<00:29:47.720> um" + }, + { + "start": 1787.91, + "duration": 0.0, + "text": "tracing is is is greatly benefited by um" + }, + { + "start": 1787.92, + "duration": 0.0, + "text": "tracing is is is greatly benefited by um deep<00:29:48.120> learning<00:29:48.440> because<00:29:48.720> it's<00:29:48.880> used<00:29:49.080> to" + }, + { + "start": 1789.15, + "duration": 0.0, + "text": "deep learning because it's used to" + }, + { + "start": 1789.16, + "duration": 0.0, + "text": "deep learning because it's used to denoise,<00:29:49.960> right?<00:29:50.840> Um" + }, + { + "start": 1792.15, + "duration": 0.0, + "text": "denoise, right? Um" + }, + { + "start": 1792.16, + "duration": 0.0, + "text": "denoise, right? Um Deep<00:29:52.320> learning<00:29:52.560> is<00:29:52.680> used<00:29:53.000> in<00:29:53.240> a<00:29:53.640> algorithm" + }, + { + "start": 1793.99, + "duration": 0.0, + "text": "Deep learning is used in a algorithm" + }, + { + "start": 1794.0, + "duration": 0.0, + "text": "Deep learning is used in a algorithm called<00:29:54.200> AlphaZero<00:29:54.840> to<00:29:54.960> master<00:29:55.320> any<00:29:55.480> discrete" + }, + { + "start": 1795.91, + "duration": 0.0, + "text": "called AlphaZero to master any discrete" + }, + { + "start": 1795.92, + "duration": 0.0, + "text": "called AlphaZero to master any discrete perfect<00:29:56.200> information<00:29:56.640> game.<00:29:57.520> Basically,<00:29:58.040> a<00:29:58.400> a" + }, + { + "start": 1798.43, + "duration": 0.0, + "text": "perfect information game. Basically, a a" + }, + { + "start": 1798.44, + "duration": 0.0, + "text": "perfect information game. Basically, a a neural<00:29:58.640> network<00:29:58.880> goes<00:29:59.120> from<00:29:59.240> the<00:29:59.360> board<00:29:59.600> state" + }, + { + "start": 1799.83, + "duration": 0.0, + "text": "neural network goes from the board state" + }, + { + "start": 1799.84, + "duration": 0.0, + "text": "neural network goes from the board state to<00:29:59.960> a<00:30:00.000> likely<00:30:00.280> next<00:30:00.480> move" + }, + { + "start": 1801.51, + "duration": 0.0, + "text": "to a likely next move" + }, + { + "start": 1801.52, + "duration": 0.0, + "text": "to a likely next move um<00:30:01.560> and<00:30:01.680> in<00:30:01.880> a<00:30:02.120> win<00:30:02.320> probability.<00:30:03.120> And<00:30:03.240> then" + }, + { + "start": 1803.47, + "duration": 0.0, + "text": "um and in a win probability. And then" + }, + { + "start": 1803.48, + "duration": 0.0, + "text": "um and in a win probability. And then another<00:30:03.800> algorithm<00:30:04.440> manipulates<00:30:05.040> that<00:30:05.680> those" + }, + { + "start": 1805.87, + "duration": 0.0, + "text": "another algorithm manipulates that those" + }, + { + "start": 1805.88, + "duration": 0.0, + "text": "another algorithm manipulates that those probabilities<00:30:06.600> and<00:30:06.920> you<00:30:07.000> know," + }, + { + "start": 1807.79, + "duration": 0.0, + "text": "probabilities and you know," + }, + { + "start": 1807.8, + "duration": 0.0, + "text": "probabilities and you know, beats<00:30:07.960> any<00:30:08.120> human<00:30:08.320> in<00:30:08.400> any<00:30:08.520> game<00:30:08.680> we've<00:30:08.800> ever" + }, + { + "start": 1808.95, + "duration": 0.0, + "text": "beats any human in any game we've ever" + }, + { + "start": 1808.96, + "duration": 0.0, + "text": "beats any human in any game we've ever invented,<00:30:09.280> unfortunately." + }, + { + "start": 1810.55, + "duration": 0.0, + "text": "invented, unfortunately." + }, + { + "start": 1810.56, + "duration": 0.0, + "text": "invented, unfortunately. Um" + }, + { + "start": 1811.35, + "duration": 0.0, + "text": "Um" + }, + { + "start": 1811.36, + "duration": 0.0, + "text": "Um Automatic<00:30:11.720> game<00:30:11.920> AI,<00:30:12.200> so<00:30:12.400> you<00:30:12.480> can<00:30:12.600> make<00:30:12.720> a" + }, + { + "start": 1812.79, + "duration": 0.0, + "text": "Automatic game AI, so you can make a" + }, + { + "start": 1812.8, + "duration": 0.0, + "text": "Automatic game AI, so you can make a style<00:30:13.040> transfer,<00:30:13.560> you<00:30:13.680> know," + }, + { + "start": 1815.15, + "duration": 0.0, + "text": "style transfer, you know," + }, + { + "start": 1815.16, + "duration": 0.0, + "text": "style transfer, you know, you<00:30:15.280> can<00:30:15.400> make<00:30:15.520> it<00:30:15.640> some<00:30:15.760> kind<00:30:15.880> of<00:30:15.960> style" + }, + { + "start": 1816.07, + "duration": 0.0, + "text": "you can make it some kind of style" + }, + { + "start": 1816.08, + "duration": 0.0, + "text": "you can make it some kind of style transfer<00:30:16.400> for<00:30:16.520> something<00:30:16.720> like<00:30:16.800> this<00:30:16.960> for" + }, + { + "start": 1817.07, + "duration": 0.0, + "text": "transfer for something like this for" + }, + { + "start": 1817.08, + "duration": 0.0, + "text": "transfer for something like this for your<00:30:17.280> video<00:30:17.560> game<00:30:17.880> where<00:30:18.080> maybe<00:30:18.880> like<00:30:19.280> you" + }, + { + "start": 1819.39, + "duration": 0.0, + "text": "your video game where maybe like you" + }, + { + "start": 1819.4, + "duration": 0.0, + "text": "your video game where maybe like you want<00:30:19.560> it<00:30:19.640> to<00:30:19.880> act<00:30:20.080> more<00:30:20.200> like<00:30:20.320> the<00:30:20.400> human" + }, + { + "start": 1820.63, + "duration": 0.0, + "text": "want it to act more like the human" + }, + { + "start": 1820.64, + "duration": 0.0, + "text": "want it to act more like the human people.<00:30:21.040> So,<00:30:21.560> once<00:30:21.760> it<00:30:21.880> masters<00:30:22.240> the<00:30:22.360> game," + }, + { + "start": 1822.99, + "duration": 0.0, + "text": "people. So, once it masters the game," + }, + { + "start": 1823.0, + "duration": 0.0, + "text": "people. So, once it masters the game, you<00:30:23.120> then<00:30:23.680> train<00:30:23.960> it<00:30:24.040> to<00:30:24.200> act<00:30:24.440> like<00:30:24.560> a<00:30:24.600> person" + }, + { + "start": 1825.07, + "duration": 0.0, + "text": "you then train it to act like a person" + }, + { + "start": 1825.08, + "duration": 0.0, + "text": "you then train it to act like a person and<00:30:25.200> then<00:30:25.840> it<00:30:25.960> would<00:30:26.080> be<00:30:26.280> a<00:30:26.360> kind<00:30:26.480> of" + }, + { + "start": 1826.59, + "duration": 0.0, + "text": "and then it would be a kind of" + }, + { + "start": 1826.6, + "duration": 0.0, + "text": "and then it would be a kind of interesting<00:30:26.920> mix<00:30:27.120> between<00:30:27.360> a<00:30:27.400> good<00:30:27.600> player" + }, + { + "start": 1827.95, + "duration": 0.0, + "text": "interesting mix between a good player" + }, + { + "start": 1827.96, + "duration": 0.0, + "text": "interesting mix between a good player and<00:30:28.200> oh,<00:30:28.280> it<00:30:28.400> plays<00:30:28.600> like<00:30:28.720> this<00:30:28.840> specific" + }, + { + "start": 1829.43, + "duration": 0.0, + "text": "and oh, it plays like this specific" + }, + { + "start": 1829.44, + "duration": 0.0, + "text": "and oh, it plays like this specific chess<00:30:29.680> player<00:30:29.840> or<00:30:29.880> whatever." + }, + { + "start": 1830.79, + "duration": 0.0, + "text": "chess player or whatever." + }, + { + "start": 1830.8, + "duration": 0.0, + "text": "chess player or whatever. Um<00:30:31.400> or<00:30:31.560> you<00:30:31.640> could<00:30:31.760> even<00:30:31.920> have<00:30:32.120> it<00:30:32.200> validate" + }, + { + "start": 1832.59, + "duration": 0.0, + "text": "Um or you could even have it validate" + }, + { + "start": 1832.6, + "duration": 0.0, + "text": "Um or you could even have it validate that<00:30:32.720> your<00:30:32.840> game<00:30:33.240> is<00:30:34.160> uh<00:30:34.240> balanced<00:30:34.640> and<00:30:34.720> stuff" + }, + { + "start": 1834.87, + "duration": 0.0, + "text": "that your game is uh balanced and stuff" + }, + { + "start": 1834.88, + "duration": 0.0, + "text": "that your game is uh balanced and stuff like<00:30:35.040> that,<00:30:35.240> right?<00:30:35.400> I<00:30:35.440> think<00:30:35.640> this<00:30:35.800> actually" + }, + { + "start": 1836.07, + "duration": 0.0, + "text": "like that, right? I think this actually" + }, + { + "start": 1836.08, + "duration": 0.0, + "text": "like that, right? I think this actually is<00:30:36.200> very<00:30:36.680> it's<00:30:36.800> quite<00:30:37.000> green<00:30:37.240> field<00:30:37.480> for<00:30:37.600> video" + }, + { + "start": 1837.79, + "duration": 0.0, + "text": "is very it's quite green field for video" + }, + { + "start": 1837.8, + "duration": 0.0, + "text": "is very it's quite green field for video games<00:30:38.040> in<00:30:38.120> general<00:30:38.400> for<00:30:38.520> how<00:30:38.600> to<00:30:38.720> apply<00:30:39.760> uh" + }, + { + "start": 1839.83, + "duration": 0.0, + "text": "games in general for how to apply uh" + }, + { + "start": 1839.84, + "duration": 0.0, + "text": "games in general for how to apply uh deep<00:30:40.000> learning<00:30:40.160> models<00:30:40.400> in<00:30:40.480> a<00:30:40.520> lot<00:30:40.640> of<00:30:40.760> places." + }, + { + "start": 1841.87, + "duration": 0.0, + "text": "deep learning models in a lot of places." + }, + { + "start": 1841.88, + "duration": 0.0, + "text": "deep learning models in a lot of places. Um" + }, + { + "start": 1842.75, + "duration": 0.0, + "text": "Um" + }, + { + "start": 1842.76, + "duration": 0.0, + "text": "Um And<00:30:43.000> you<00:30:43.080> know,<00:30:43.200> for<00:30:43.440> automating<00:30:43.840> like<00:30:43.960> a<00:30:44.040> call" + }, + { + "start": 1844.23, + "duration": 0.0, + "text": "And you know, for automating like a call" + }, + { + "start": 1844.24, + "duration": 0.0, + "text": "And you know, for automating like a call center,<00:30:44.600> I<00:30:44.680> think<00:30:44.880> that<00:30:45.040> there's<00:30:45.200> a<00:30:45.240> lot<00:30:45.440> of" + }, + { + "start": 1845.51, + "duration": 0.0, + "text": "center, I think that there's a lot of" + }, + { + "start": 1845.52, + "duration": 0.0, + "text": "center, I think that there's a lot of these<00:30:45.680> call<00:30:45.880> center<00:30:46.200> AI<00:30:46.400> startups.<00:30:47.280> You<00:30:47.360> know," + }, + { + "start": 1847.47, + "duration": 0.0, + "text": "these call center AI startups. You know," + }, + { + "start": 1847.48, + "duration": 0.0, + "text": "these call center AI startups. You know, it<00:30:47.640> seems<00:30:47.880> easy<00:30:48.080> on<00:30:48.160> the<00:30:48.240> surface,<00:30:48.600> but" + }, + { + "start": 1848.75, + "duration": 0.0, + "text": "it seems easy on the surface, but" + }, + { + "start": 1848.76, + "duration": 0.0, + "text": "it seems easy on the surface, but usually<00:30:49.040> people<00:30:49.320> call<00:30:49.480> support<00:30:49.880> because<00:30:50.120> they" + }, + { + "start": 1850.23, + "duration": 0.0, + "text": "usually people call support because they" + }, + { + "start": 1850.24, + "duration": 0.0, + "text": "usually people call support because they want<00:30:50.520> to<00:30:50.600> talk<00:30:50.760> to<00:30:50.880> a<00:30:50.920> person<00:30:51.320> and<00:30:51.400> not<00:30:51.520> a" + }, + { + "start": 1851.55, + "duration": 0.0, + "text": "want to talk to a person and not a" + }, + { + "start": 1851.56, + "duration": 0.0, + "text": "want to talk to a person and not a machine.<00:30:52.400> So,<00:30:52.520> I<00:30:52.560> think<00:30:52.760> that<00:30:52.880> this<00:30:53.040> is" + }, + { + "start": 1853.15, + "duration": 0.0, + "text": "machine. So, I think that this is" + }, + { + "start": 1853.16, + "duration": 0.0, + "text": "machine. So, I think that this is probably<00:30:53.520> going<00:30:53.720> to<00:30:53.840> be<00:30:54.120> a<00:30:54.440> a<00:30:54.480> giant<00:30:54.720> disaster." + }, + { + "start": 1855.83, + "duration": 0.0, + "text": "probably going to be a a giant disaster." + }, + { + "start": 1855.84, + "duration": 0.0, + "text": "probably going to be a a giant disaster. Um<00:30:56.280> and,<00:30:56.480> you<00:30:56.520> know,<00:30:56.640> of<00:30:56.720> course,<00:30:57.040> automating" + }, + { + "start": 1857.47, + "duration": 0.0, + "text": "Um and, you know, of course, automating" + }, + { + "start": 1857.48, + "duration": 0.0, + "text": "Um and, you know, of course, automating something<00:30:57.640> like<00:30:57.760> engineering<00:30:58.160> is<00:30:58.320> way<00:30:58.520> more" + }, + { + "start": 1858.67, + "duration": 0.0, + "text": "something like engineering is way more" + }, + { + "start": 1858.68, + "duration": 0.0, + "text": "something like engineering is way more complicated<00:30:59.600> than<00:30:59.760> just<00:30:59.960> that,<00:31:00.320> which<00:31:00.560> isn't" + }, + { + "start": 1860.71, + "duration": 0.0, + "text": "complicated than just that, which isn't" + }, + { + "start": 1860.72, + "duration": 0.0, + "text": "complicated than just that, which isn't even<00:31:00.880> working.<00:31:01.200> So," + }, + { + "start": 1862.11, + "duration": 0.0, + "text": "even working. So," + }, + { + "start": 1862.12, + "duration": 0.0, + "text": "even working. So, you<00:31:02.280> know,<00:31:02.360> just<00:31:02.480> something<00:31:02.640> to<00:31:02.720> think<00:31:02.880> about." + }, + { + "start": 1863.31, + "duration": 0.0, + "text": "you know, just something to think about." + }, + { + "start": 1863.32, + "duration": 0.0, + "text": "you know, just something to think about. Um<00:31:03.840> and<00:31:03.920> a<00:31:03.960> quick<00:31:04.160> note<00:31:04.360> on<00:31:04.400> computer<00:31:04.720> vision" + }, + { + "start": 1865.03, + "duration": 0.0, + "text": "Um and a quick note on computer vision" + }, + { + "start": 1865.04, + "duration": 0.0, + "text": "Um and a quick note on computer vision cuz<00:31:05.160> I<00:31:05.200> didn't<00:31:05.400> really<00:31:05.560> get<00:31:05.720> to<00:31:05.840> flush<00:31:06.080> that" + }, + { + "start": 1866.23, + "duration": 0.0, + "text": "cuz I didn't really get to flush that" + }, + { + "start": 1866.24, + "duration": 0.0, + "text": "cuz I didn't really get to flush that out<00:31:06.320> that<00:31:06.440> much<00:31:06.640> in<00:31:06.720> this<00:31:06.840> talk.<00:31:07.600> But,<00:31:08.120> the" + }, + { + "start": 1868.19, + "duration": 0.0, + "text": "out that much in this talk. But, the" + }, + { + "start": 1868.2, + "duration": 0.0, + "text": "out that much in this talk. But, the basic<00:31:08.720> idea<00:31:09.040> I<00:31:09.120> had<00:31:09.480> was<00:31:10.160> for<00:31:10.320> computer" + }, + { + "start": 1870.71, + "duration": 0.0, + "text": "basic idea I had was for computer" + }, + { + "start": 1870.72, + "duration": 0.0, + "text": "basic idea I had was for computer vision,<00:31:11.160> you're<00:31:11.280> essentially<00:31:11.680> trying<00:31:11.920> to" + }, + { + "start": 1872.15, + "duration": 0.0, + "text": "vision, you're essentially trying to" + }, + { + "start": 1872.16, + "duration": 0.0, + "text": "vision, you're essentially trying to take<00:31:12.360> in<00:31:12.440> a<00:31:12.480> bunch<00:31:12.640> of<00:31:12.760> images<00:31:13.320> in<00:31:13.480> something" + }, + { + "start": 1873.67, + "duration": 0.0, + "text": "take in a bunch of images in something" + }, + { + "start": 1873.68, + "duration": 0.0, + "text": "take in a bunch of images in something called<00:31:13.840> monocular<00:31:14.280> slam<00:31:15.040> and<00:31:15.840> get<00:31:16.040> some<00:31:16.200> kind" + }, + { + "start": 1876.47, + "duration": 0.0, + "text": "called monocular slam and get some kind" + }, + { + "start": 1876.48, + "duration": 0.0, + "text": "called monocular slam and get some kind of<00:31:16.720> representation<00:31:17.400> of<00:31:17.480> the<00:31:17.600> world<00:31:18.000> that<00:31:18.160> it's" + }, + { + "start": 1878.31, + "duration": 0.0, + "text": "of representation of the world that it's" + }, + { + "start": 1878.32, + "duration": 0.0, + "text": "of representation of the world that it's seeing<00:31:19.200> um<00:31:19.360> automatically,<00:31:19.920> right?<00:31:20.200> It's<00:31:20.320> a" + }, + { + "start": 1880.39, + "duration": 0.0, + "text": "seeing um automatically, right? It's a" + }, + { + "start": 1880.4, + "duration": 0.0, + "text": "seeing um automatically, right? It's a bit<00:31:20.560> like<00:31:20.680> deep<00:31:20.880> learning<00:31:21.120> where<00:31:21.240> you<00:31:21.360> have" + }, + { + "start": 1881.51, + "duration": 0.0, + "text": "bit like deep learning where you have" + }, + { + "start": 1881.52, + "duration": 0.0, + "text": "bit like deep learning where you have some<00:31:21.720> automatic<00:31:22.200> optimization<00:31:22.760> process<00:31:23.600> that" + }, + { + "start": 1883.75, + "duration": 0.0, + "text": "some automatic optimization process that" + }, + { + "start": 1883.76, + "duration": 0.0, + "text": "some automatic optimization process that operates<00:31:24.080> on<00:31:24.160> a<00:31:24.200> lot<00:31:24.360> of<00:31:24.440> data<00:31:24.760> to<00:31:24.880> uncover" + }, + { + "start": 1885.19, + "duration": 0.0, + "text": "operates on a lot of data to uncover" + }, + { + "start": 1885.2, + "duration": 0.0, + "text": "operates on a lot of data to uncover something<00:31:25.640> interesting<00:31:26.000> about<00:31:26.160> the<00:31:26.240> world." + }, + { + "start": 1887.23, + "duration": 0.0, + "text": "something interesting about the world." + }, + { + "start": 1887.24, + "duration": 0.0, + "text": "something interesting about the world. Um<00:31:27.360> hopefully<00:31:27.680> interesting.<00:31:28.680> Um<00:31:29.200> and<00:31:29.640> it<00:31:30.440> this" + }, + { + "start": 1890.67, + "duration": 0.0, + "text": "Um hopefully interesting. Um and it this" + }, + { + "start": 1890.68, + "duration": 0.0, + "text": "Um hopefully interesting. Um and it this that<00:31:30.920> that<00:31:31.120> visual<00:31:31.480> slam<00:31:31.680> pipeline<00:31:32.080> is" + }, + { + "start": 1892.15, + "duration": 0.0, + "text": "that that visual slam pipeline is" + }, + { + "start": 1892.16, + "duration": 0.0, + "text": "that that visual slam pipeline is basically<00:31:32.520> the<00:31:32.720> opposite<00:31:33.280> of<00:31:33.680> game" + }, + { + "start": 1893.91, + "duration": 0.0, + "text": "basically the opposite of game" + }, + { + "start": 1893.92, + "duration": 0.0, + "text": "basically the opposite of game rendering.<00:31:34.640> And<00:31:34.760> it<00:31:34.840> involves<00:31:35.280> somewhat" + }, + { + "start": 1895.55, + "duration": 0.0, + "text": "rendering. And it involves somewhat" + }, + { + "start": 1895.56, + "duration": 0.0, + "text": "rendering. And it involves somewhat similar<00:31:36.000> linear<00:31:36.280> algebra,<00:31:36.640> right?<00:31:37.440> Um<00:31:38.040> you<00:31:38.160> go" + }, + { + "start": 1898.19, + "duration": 0.0, + "text": "similar linear algebra, right? Um you go" + }, + { + "start": 1898.2, + "duration": 0.0, + "text": "similar linear algebra, right? Um you go from<00:31:38.360> an<00:31:38.400> image<00:31:38.720> to<00:31:38.840> easy<00:31:39.040> to<00:31:39.120> track<00:31:39.280> key" + }, + { + "start": 1899.39, + "duration": 0.0, + "text": "from an image to easy to track key" + }, + { + "start": 1899.4, + "duration": 0.0, + "text": "from an image to easy to track key points,<00:31:40.000> pose<00:31:40.280> estimate<00:31:40.680> for<00:31:40.800> the<00:31:40.880> camera," + }, + { + "start": 1901.47, + "duration": 0.0, + "text": "points, pose estimate for the camera," + }, + { + "start": 1901.48, + "duration": 0.0, + "text": "points, pose estimate for the camera, some<00:31:41.640> representation<00:31:42.160> of<00:31:42.240> the<00:31:42.320> scene<00:31:42.720> for<00:31:42.880> the" + }, + { + "start": 1902.95, + "duration": 0.0, + "text": "some representation of the scene for the" + }, + { + "start": 1902.96, + "duration": 0.0, + "text": "some representation of the scene for the robot<00:31:43.240> to<00:31:43.320> drive<00:31:43.480> around<00:31:43.800> in,<00:31:44.280> right?<00:31:45.040> Yeah," + }, + { + "start": 1905.27, + "duration": 0.0, + "text": "robot to drive around in, right? Yeah," + }, + { + "start": 1905.28, + "duration": 0.0, + "text": "robot to drive around in, right? Yeah, game<00:31:45.520> rendering,<00:31:45.840> you<00:31:45.920> know,<00:31:46.040> it's<00:31:46.160> like" + }, + { + "start": 1906.43, + "duration": 0.0, + "text": "game rendering, you know, it's like" + }, + { + "start": 1906.44, + "duration": 0.0, + "text": "game rendering, you know, it's like scene<00:31:46.640> representation,<00:31:47.160> triangles,<00:31:47.520> pixels." + }, + { + "start": 1907.95, + "duration": 0.0, + "text": "scene representation, triangles, pixels." + }, + { + "start": 1907.96, + "duration": 0.0, + "text": "scene representation, triangles, pixels. I<00:31:48.040> thought<00:31:48.240> that<00:31:48.320> was<00:31:48.400> interesting." + }, + { + "start": 1909.47, + "duration": 0.0, + "text": "I thought that was interesting." + }, + { + "start": 1909.48, + "duration": 0.0, + "text": "I thought that was interesting. So,<00:31:49.600> will<00:31:49.880> LLMs<00:31:50.240> replace<00:31:50.720> knowledge<00:31:51.000> work<00:31:51.160> en" + }, + { + "start": 1911.27, + "duration": 0.0, + "text": "So, will LLMs replace knowledge work en" + }, + { + "start": 1911.28, + "duration": 0.0, + "text": "So, will LLMs replace knowledge work en masse<00:31:51.560> someday,<00:31:51.960> right?" + }, + { + "start": 1912.95, + "duration": 0.0, + "text": "masse someday, right?" + }, + { + "start": 1912.96, + "duration": 0.0, + "text": "masse someday, right? I<00:31:53.120> put<00:31:53.280> a<00:31:53.320> steelman<00:31:53.680> yes<00:31:53.840> and<00:31:53.960> a<00:31:54.000> steelman<00:31:54.360> no" + }, + { + "start": 1914.51, + "duration": 0.0, + "text": "I put a steelman yes and a steelman no" + }, + { + "start": 1914.52, + "duration": 0.0, + "text": "I put a steelman yes and a steelman no here." + }, + { + "start": 1915.55, + "duration": 0.0, + "text": "here." + }, + { + "start": 1915.56, + "duration": 0.0, + "text": "here. Um<00:31:55.760> really,<00:31:56.560> I<00:31:56.640> think<00:31:56.840> it's<00:31:56.960> probably" + }, + { + "start": 1917.39, + "duration": 0.0, + "text": "Um really, I think it's probably" + }, + { + "start": 1917.4, + "duration": 0.0, + "text": "Um really, I think it's probably probably<00:31:57.800> not.<00:31:58.440> But," + }, + { + "start": 1919.59, + "duration": 0.0, + "text": "probably not. But," + }, + { + "start": 1919.6, + "duration": 0.0, + "text": "probably not. But, there's<00:31:59.760> like<00:31:59.880> a<00:31:59.920> lot<00:32:00.160> of<00:32:00.440> investors<00:32:00.840> who" + }, + { + "start": 1920.91, + "duration": 0.0, + "text": "there's like a lot of investors who" + }, + { + "start": 1920.92, + "duration": 0.0, + "text": "there's like a lot of investors who disagree<00:32:01.200> with<00:32:01.360> me,<00:32:01.440> so<00:32:01.560> I<00:32:01.600> don't<00:32:01.720> know." + }, + { + "start": 1922.59, + "duration": 0.0, + "text": "disagree with me, so I don't know." + }, + { + "start": 1922.6, + "duration": 0.0, + "text": "disagree with me, so I don't know. Um" + }, + { + "start": 1924.07, + "duration": 0.0, + "text": "Um" + }, + { + "start": 1924.08, + "duration": 0.0, + "text": "Um new<00:32:04.240> knowledge<00:32:04.600> can't<00:32:04.760> be<00:32:04.840> made<00:32:04.960> by<00:32:05.080> machines," + }, + { + "start": 1925.75, + "duration": 0.0, + "text": "new knowledge can't be made by machines," + }, + { + "start": 1925.76, + "duration": 0.0, + "text": "new knowledge can't be made by machines, right?<00:32:06.400> This<00:32:06.560> is<00:32:06.640> a<00:32:06.680> very<00:32:06.880> funny<00:32:07.080> joke<00:32:07.440> made<00:32:07.560> by" + }, + { + "start": 1927.67, + "duration": 0.0, + "text": "right? This is a very funny joke made by" + }, + { + "start": 1927.68, + "duration": 0.0, + "text": "right? This is a very funny joke made by a<00:32:07.720> philosopher<00:32:08.080> called<00:32:08.240> David<00:32:08.520> Deutsch.<00:32:09.480> Um" + }, + { + "start": 1929.87, + "duration": 0.0, + "text": "a philosopher called David Deutsch. Um" + }, + { + "start": 1929.88, + "duration": 0.0, + "text": "a philosopher called David Deutsch. Um maybe<00:32:10.080> this<00:32:10.280> is<00:32:10.360> what<00:32:10.480> we're<00:32:10.560> going<00:32:10.680> to<00:32:10.720> have" + }, + { + "start": 1930.79, + "duration": 0.0, + "text": "maybe this is what we're going to have" + }, + { + "start": 1930.8, + "duration": 0.0, + "text": "maybe this is what we're going to have to<00:32:10.840> do<00:32:10.960> someday.<00:32:11.200> You're<00:32:11.280> going<00:32:11.400> to<00:32:11.480> have<00:32:11.560> to" + }, + { + "start": 1931.63, + "duration": 0.0, + "text": "to do someday. You're going to have to" + }, + { + "start": 1931.64, + "duration": 0.0, + "text": "to do someday. You're going to have to invent<00:32:11.960> new<00:32:12.120> scientific<00:32:12.560> knowledge<00:32:12.920> to<00:32:13.000> prove" + }, + { + "start": 1933.19, + "duration": 0.0, + "text": "invent new scientific knowledge to prove" + }, + { + "start": 1933.2, + "duration": 0.0, + "text": "invent new scientific knowledge to prove that<00:32:13.320> you're<00:32:13.400> not<00:32:13.520> a<00:32:13.600> robot.<00:32:14.440> Cuz<00:32:14.640> maybe" + }, + { + "start": 1934.79, + "duration": 0.0, + "text": "that you're not a robot. Cuz maybe" + }, + { + "start": 1934.8, + "duration": 0.0, + "text": "that you're not a robot. Cuz maybe that's<00:32:14.960> the<00:32:15.040> one<00:32:15.280> thing<00:32:15.640> the<00:32:15.760> one<00:32:16.000> thing<00:32:16.120> that" + }, + { + "start": 1936.23, + "duration": 0.0, + "text": "that's the one thing the one thing that" + }, + { + "start": 1936.24, + "duration": 0.0, + "text": "that's the one thing the one thing that humans<00:32:16.600> can<00:32:16.760> always<00:32:17.000> do<00:32:17.120> is<00:32:17.200> create<00:32:17.520> new" + }, + { + "start": 1937.71, + "duration": 0.0, + "text": "humans can always do is create new" + }, + { + "start": 1937.72, + "duration": 0.0, + "text": "humans can always do is create new knowledge<00:32:18.080> cuz<00:32:18.240> they're<00:32:18.360> the<00:32:18.480> only<00:32:18.720> ones<00:32:18.920> that" + }, + { + "start": 1939.03, + "duration": 0.0, + "text": "knowledge cuz they're the only ones that" + }, + { + "start": 1939.04, + "duration": 0.0, + "text": "knowledge cuz they're the only ones that can<00:32:19.200> officially<00:32:19.640> reason." + }, + { + "start": 1940.63, + "duration": 0.0, + "text": "can officially reason." + }, + { + "start": 1940.64, + "duration": 0.0, + "text": "can officially reason. Um" + }, + { + "start": 1943.11, + "duration": 0.0, + "text": "Um" + }, + { + "start": 1943.12, + "duration": 0.0, + "text": "Um But,<00:32:23.360> if,<00:32:23.760> you<00:32:23.840> know,<00:32:24.600> you<00:32:24.720> have<00:32:24.880> to<00:32:24.960> have<00:32:25.080> a" + }, + { + "start": 1945.15, + "duration": 0.0, + "text": "But, if, you know, you have to have a" + }, + { + "start": 1945.16, + "duration": 0.0, + "text": "But, if, you know, you have to have a human<00:32:25.520> know<00:32:25.760> if<00:32:25.880> you<00:32:25.960> did<00:32:26.120> new<00:32:26.240> knowledge,<00:32:26.560> I" + }, + { + "start": 1946.59, + "duration": 0.0, + "text": "human know if you did new knowledge, I" + }, + { + "start": 1946.6, + "duration": 0.0, + "text": "human know if you did new knowledge, I don't<00:32:26.720> know<00:32:26.800> how<00:32:26.920> a<00:32:26.960> system<00:32:27.280> would<00:32:27.400> detect<00:32:27.760> if" + }, + { + "start": 1947.87, + "duration": 0.0, + "text": "don't know how a system would detect if" + }, + { + "start": 1947.88, + "duration": 0.0, + "text": "don't know how a system would detect if it's<00:32:28.000> new<00:32:28.120> knowledge<00:32:28.440> or<00:32:28.480> not.<00:32:29.200> Maybe<00:32:29.400> that's" + }, + { + "start": 1949.55, + "duration": 0.0, + "text": "it's new knowledge or not. Maybe that's" + }, + { + "start": 1949.56, + "duration": 0.0, + "text": "it's new knowledge or not. Maybe that's like<00:32:29.720> P<00:32:29.920> and<00:32:30.080> NP<00:32:30.360> or<00:32:30.440> something,<00:32:30.680> I<00:32:30.720> don't" + }, + { + "start": 1950.83, + "duration": 0.0, + "text": "like P and NP or something, I don't" + }, + { + "start": 1950.84, + "duration": 0.0, + "text": "like P and NP or something, I don't know.<00:32:31.560> Um" + }, + { + "start": 1952.67, + "duration": 0.0, + "text": "know. Um" + }, + { + "start": 1952.68, + "duration": 0.0, + "text": "know. Um Wait<00:32:32.880> until<00:32:33.080> RKG<00:32:33.560> I<00:32:33.640> continues<00:32:34.000> to<00:32:34.080> be<00:32:34.160> beaten," + }, + { + "start": 1954.47, + "duration": 0.0, + "text": "Wait until RKG I continues to be beaten," + }, + { + "start": 1954.48, + "duration": 0.0, + "text": "Wait until RKG I continues to be beaten, I<00:32:34.520> think,<00:32:34.760> is<00:32:35.280> how<00:32:35.400> you<00:32:35.840> monitor<00:32:36.240> this" + }, + { + "start": 1956.35, + "duration": 0.0, + "text": "I think, is how you monitor this" + }, + { + "start": 1956.36, + "duration": 0.0, + "text": "I think, is how you monitor this problem." + }, + { + "start": 1957.39, + "duration": 0.0, + "text": "problem." + }, + { + "start": 1957.4, + "duration": 0.0, + "text": "problem. Um<00:32:37.640> cuz<00:32:37.840> if<00:32:38.040> that<00:32:38.160> continues<00:32:38.480> to<00:32:38.560> be<00:32:38.680> beaten,<00:32:38.960> I" + }, + { + "start": 1959.03, + "duration": 0.0, + "text": "Um cuz if that continues to be beaten, I" + }, + { + "start": 1959.04, + "duration": 0.0, + "text": "Um cuz if that continues to be beaten, I think<00:32:39.840> uh<00:32:40.360> you<00:32:40.480> know,<00:32:40.680> that's<00:32:41.040> crazy." + }, + { + "start": 1962.87, + "duration": 0.0, + "text": "think uh you know, that's crazy." + }, + { + "start": 1962.88, + "duration": 0.0, + "text": "think uh you know, that's crazy. Thanks<00:32:43.080> for<00:32:43.160> listening.<00:32:43.440> Don't<00:32:43.560> forget<00:32:43.720> to" + }, + { + "start": 1963.75, + "duration": 0.0, + "text": "Thanks for listening. Don't forget to" + }, + { + "start": 1963.76, + "duration": 0.0, + "text": "Thanks for listening. Don't forget to check<00:32:44.000> out<00:32:44.200> and<00:32:45.000> work<00:32:45.280> on<00:32:45.480> github.com<00:32:46.520> /ai." + }, + { + "start": 1967.87, + "duration": 0.0, + "text": "check out and work on github.com /ai." + }, + { + "start": 1967.88, + "duration": 0.0, + "text": "check out and work on github.com /ai. I<00:32:47.920> want<00:32:48.120> to<00:32:48.200> make<00:32:48.400> ai.c<00:32:49.160> so<00:32:49.280> that<00:32:49.400> it's<00:32:49.520> like" + }, + { + "start": 1969.67, + "duration": 0.0, + "text": "I want to make ai.c so that it's like" + }, + { + "start": 1969.68, + "duration": 0.0, + "text": "I want to make ai.c so that it's like fun<00:32:49.880> to<00:32:49.960> mess<00:32:50.120> around<00:32:50.360> with<00:32:50.440> these<00:32:50.600> models<00:32:50.880> and" + }, + { + "start": 1970.99, + "duration": 0.0, + "text": "fun to mess around with these models and" + }, + { + "start": 1971.0, + "duration": 0.0, + "text": "fun to mess around with these models and techniques<00:32:51.480> and<00:32:51.600> games." + }, + { + "start": 1972.99, + "duration": 0.0, + "text": "techniques and games." + }, + { + "start": 1973.0, + "duration": 0.0, + "text": "techniques and games. No<00:32:53.120> dependencies,<00:32:53.800> no<00:32:54.000> python,<00:32:54.520> no<00:32:54.640> nothing." + }, + { + "start": 1974.99, + "duration": 0.0, + "text": "No dependencies, no python, no nothing." + }, + { + "start": 1975.0, + "duration": 0.0, + "text": "No dependencies, no python, no nothing. I<00:32:55.280> you<00:32:55.360> know,<00:32:55.480> I<00:32:55.560> just<00:32:56.000> it's<00:32:56.120> like<00:32:56.880> it's<00:32:57.080> not" + }, + { + "start": 1977.23, + "duration": 0.0, + "text": "I you know, I just it's like it's not" + }, + { + "start": 1977.24, + "duration": 0.0, + "text": "I you know, I just it's like it's not good<00:32:57.400> right<00:32:57.560> now.<00:32:57.760> It<00:32:57.840> needs<00:32:58.080> to<00:32:58.120> be<00:32:58.200> rewritten" + }, + { + "start": 1978.55, + "duration": 0.0, + "text": "good right now. It needs to be rewritten" + }, + { + "start": 1978.56, + "duration": 0.0, + "text": "good right now. It needs to be rewritten like<00:32:58.720> 10<00:32:58.920> times.<00:32:59.240> I've<00:32:59.280> never<00:32:59.480> made<00:32:59.640> a<00:32:59.960> deep" + }, + { + "start": 1980.11, + "duration": 0.0, + "text": "like 10 times. I've never made a deep" + }, + { + "start": 1980.12, + "duration": 0.0, + "text": "like 10 times. I've never made a deep learning<00:33:00.320> framework<00:33:00.600> before,<00:33:00.920> but" + }, + { + "start": 1982.15, + "duration": 0.0, + "text": "learning framework before, but" + }, + { + "start": 1982.16, + "duration": 0.0, + "text": "learning framework before, but Thanks<00:33:02.360> for<00:33:02.440> listening<00:33:02.640> to<00:33:02.680> my<00:33:02.760> talk." + }, + { + "start": 1992.28, + "duration": 0.0, + "text": "Sam." + }, + { + "start": 1997.8, + "duration": 0.0, + "text": "Okay,<00:33:18.040> that's<00:33:18.280> a<00:33:18.360> bit<00:33:18.600> similar<00:33:18.880> to<00:33:19.000> the<00:33:19.120> psych" + }, + { + "start": 1999.47, + "duration": 0.0, + "text": "Okay, that's a bit similar to the psych" + }, + { + "start": 1999.48, + "duration": 0.0, + "text": "Okay, that's a bit similar to the psych stuff." + }, + { + "start": 2000.47, + "duration": 0.0, + "text": "stuff." + }, + { + "start": 2000.48, + "duration": 0.0, + "text": "stuff. It's<00:33:20.720> used<00:33:20.960> by<00:33:21.080> cookie<00:33:21.400> people<00:33:21.680> for<00:33:21.880> AI,<00:33:22.200> but" + }, + { + "start": 2002.35, + "duration": 0.0, + "text": "It's used by cookie people for AI, but" + }, + { + "start": 2002.36, + "duration": 0.0, + "text": "It's used by cookie people for AI, but not<00:33:22.520> serious<00:33:22.800> contenders." + }, + { + "start": 2005.64, + "duration": 0.0, + "text": "Hello." + }, + { + "start": 2006.55, + "duration": 0.0, + "text": "Hello." + }, + { + "start": 2006.56, + "duration": 0.0, + "text": "Hello. What's<00:33:26.760> up?" + }, + { + "start": 2007.11, + "duration": 0.0, + "text": "What's up?" + }, + { + "start": 2007.12, + "duration": 0.0, + "text": "What's up? >> Good." + }, + { + "start": 2008.59, + "duration": 0.0, + "text": ">> Good." + }, + { + "start": 2008.6, + "duration": 0.0, + "text": ">> Good. Awesome.<00:33:29.520> Can<00:33:29.640> we<00:33:29.720> get<00:33:29.880> another<00:33:30.160> round<00:33:30.400> of" + }, + { + "start": 2010.47, + "duration": 0.0, + "text": "Awesome. Can we get another round of" + }, + { + "start": 2010.48, + "duration": 0.0, + "text": "Awesome. Can we get another round of applause<00:33:30.880> for<00:33:31.000> Krakey?" + }, + { + "start": 2017.2, + "duration": 0.0, + "text": "Subscribe<00:33:37.640> to<00:33:37.680> my<00:33:37.800> YouTube.<00:33:38.360> Subscribe<00:33:38.880> to" + }, + { + "start": 2018.95, + "duration": 0.0, + "text": "Subscribe to my YouTube. Subscribe to" + }, + { + "start": 2018.96, + "duration": 0.0, + "text": "Subscribe to my YouTube. Subscribe to his<00:33:39.120> YouTube.<00:33:39.560> Watch" + }, + { + "start": 2021.31, + "duration": 0.0, + "text": "his YouTube. Watch" + }, + { + "start": 2021.32, + "duration": 0.0, + "text": "his YouTube. Watch eye<00:33:41.480> contact<00:33:42.040> practice.<00:33:42.680> It's<00:33:42.800> the<00:33:42.880> best<00:33:43.120> one." + }, + { + "start": 2027.2, + "duration": 0.0, + "text": "I<00:33:47.280> mean,<00:33:47.440> that<00:33:47.600> talk<00:33:47.800> was<00:33:48.280> definitely<00:33:48.640> like<00:33:48.840> it" + }, + { + "start": 2028.91, + "duration": 0.0, + "text": "I mean, that talk was definitely like it" + }, + { + "start": 2028.92, + "duration": 0.0, + "text": "I mean, that talk was definitely like it was<00:33:49.000> like<00:33:49.120> 15%<00:33:49.840> skits,<00:33:50.160> so" + }, + { + "start": 2031.03, + "duration": 0.0, + "text": "was like 15% skits, so" + }, + { + "start": 2031.04, + "duration": 0.0, + "text": "was like 15% skits, so but<00:33:51.200> that's<00:33:51.440> why<00:33:51.560> it's<00:33:51.720> entertaining<00:33:52.480> for" + }, + { + "start": 2032.59, + "duration": 0.0, + "text": "but that's why it's entertaining for" + }, + { + "start": 2032.6, + "duration": 0.0, + "text": "but that's why it's entertaining for sure." + }, + { + "start": 2034.15, + "duration": 0.0, + "text": "sure." + }, + { + "start": 2034.16, + "duration": 0.0, + "text": "sure. I<00:33:54.240> think<00:33:54.400> you're<00:33:54.520> like<00:33:54.720> part<00:33:55.040> LLM." + }, + { + "start": 2036.99, + "duration": 0.0, + "text": "I think you're like part LLM." + }, + { + "start": 2037.0, + "duration": 0.0, + "text": "I think you're like part LLM. Like<00:33:57.200> anyone<00:33:57.520> that's<00:33:57.680> talked<00:33:57.960> to<00:33:58.080> Krakey,<00:33:58.480> you" + }, + { + "start": 2038.55, + "duration": 0.0, + "text": "Like anyone that's talked to Krakey, you" + }, + { + "start": 2038.56, + "duration": 0.0, + "text": "Like anyone that's talked to Krakey, you could<00:33:58.680> just<00:33:58.840> give<00:33:59.040> him<00:33:59.560> a<00:33:59.640> prompt<00:34:00.360> and<00:34:00.480> he<00:34:00.600> just" + }, + { + "start": 2041.19, + "duration": 0.0, + "text": "could just give him a prompt and he just" + }, + { + "start": 2041.2, + "duration": 0.0, + "text": "could just give him a prompt and he just [ __ ]<00:34:01.480> spits<00:34:01.840> out<00:34:02.680> infinite<00:34:03.280> like" + }, + { + "start": 2043.79, + "duration": 0.0, + "text": "[ __ ] spits out infinite like" + }, + { + "start": 2043.8, + "duration": 0.0, + "text": "[ __ ] spits out infinite like sometimes<00:34:04.280> he<00:34:04.360> hallucinates.<00:34:05.000> Yeah," + }, + { + "start": 2045.63, + "duration": 0.0, + "text": "sometimes he hallucinates. Yeah," + }, + { + "start": 2045.64, + "duration": 0.0, + "text": "sometimes he hallucinates. Yeah, sometimes<00:34:06.040> it's<00:34:06.160> garbage,<00:34:06.640> but<00:34:07.440> most<00:34:07.680> of<00:34:07.760> the" + }, + { + "start": 2047.83, + "duration": 0.0, + "text": "sometimes it's garbage, but most of the" + }, + { + "start": 2047.84, + "duration": 0.0, + "text": "sometimes it's garbage, but most of the time<00:34:08.040> it's<00:34:08.159> kind<00:34:08.320> of<00:34:08.440> right.<00:34:09.080> Okay." + }, + { + "start": 2050.23, + "duration": 0.0, + "text": "time it's kind of right. Okay." + }, + { + "start": 2050.24, + "duration": 0.0, + "text": "time it's kind of right. Okay. That<00:34:10.560> is<00:34:10.679> what<00:34:10.879> that's<00:34:11.080> kind<00:34:11.200> of<00:34:11.280> what<00:34:11.399> LLM's" + }, + { + "start": 2051.71, + "duration": 0.0, + "text": "That is what that's kind of what LLM's" + }, + { + "start": 2051.72, + "duration": 0.0, + "text": "That is what that's kind of what LLM's do,<00:34:11.879> yeah.<00:34:12.399> Yeah,<00:34:12.600> exactly.<00:34:13.240> So<00:34:13.360> you're<00:34:13.720> it's" + }, + { + "start": 2053.869, + "duration": 0.0, + "text": "do, yeah. Yeah, exactly. So you're it's" + }, + { + "start": 2053.879, + "duration": 0.0, + "text": "do, yeah. Yeah, exactly. So you're it's in<00:34:14.040> there." + }, + { + "start": 2054.83, + "duration": 0.0, + "text": "in there." + }, + { + "start": 2054.84, + "duration": 0.0, + "text": "in there. It's<00:34:15.000> interesting." + }, + { + "start": 2056.389, + "duration": 0.0, + "text": "It's interesting." + }, + { + "start": 2056.399, + "duration": 0.0, + "text": "It's interesting. So<00:34:16.760> everyone<00:34:17.200> here<00:34:17.919> at<00:34:18.040> Better<00:34:18.360> Software" + }, + { + "start": 2058.83, + "duration": 0.0, + "text": "So everyone here at Better Software" + }, + { + "start": 2058.84, + "duration": 0.0, + "text": "So everyone here at Better Software Conference" + }, + { + "start": 2060.43, + "duration": 0.0, + "text": "Conference" + }, + { + "start": 2060.44, + "duration": 0.0, + "text": "Conference wants<00:34:20.760> to<00:34:20.919> write<00:34:21.159> better<00:34:21.440> software<00:34:22.040> and<00:34:22.200> wants" + }, + { + "start": 2062.389, + "duration": 0.0, + "text": "wants to write better software and wants" + }, + { + "start": 2062.399, + "duration": 0.0, + "text": "wants to write better software and wants to<00:34:22.520> live<00:34:22.679> in<00:34:22.800> a<00:34:22.840> world<00:34:23.760> where<00:34:23.879> there<00:34:24.080> is<00:34:24.320> better" + }, + { + "start": 2064.55, + "duration": 0.0, + "text": "to live in a world where there is better" + }, + { + "start": 2064.56, + "duration": 0.0, + "text": "to live in a world where there is better software." + }, + { + "start": 2066.59, + "duration": 0.0, + "text": "software." + }, + { + "start": 2066.6, + "duration": 0.0, + "text": "software. So<00:34:26.840> will<00:34:27.159> AI<00:34:28.040> cause<00:34:28.320> the<00:34:28.399> destruction<00:34:29.159> of" + }, + { + "start": 2069.71, + "duration": 0.0, + "text": "So will AI cause the destruction of" + }, + { + "start": 2069.72, + "duration": 0.0, + "text": "So will AI cause the destruction of better<00:34:29.960> software?" + }, + { + "start": 2076.639, + "duration": 0.0, + "text": "Um" + }, + { + "start": 2078.59, + "duration": 0.0, + "text": "Um" + }, + { + "start": 2078.6, + "duration": 0.0, + "text": "Um So,<00:34:38.840> okay,<00:34:39.600> one<00:34:39.800> of<00:34:39.840> the<00:34:39.960> biggest<00:34:40.320> like<00:34:40.520> hot" + }, + { + "start": 2080.71, + "duration": 0.0, + "text": "So, okay, one of the biggest like hot" + }, + { + "start": 2080.72, + "duration": 0.0, + "text": "So, okay, one of the biggest like hot topics<00:34:41.120> for<00:34:41.360> for<00:34:41.679> AI<00:34:41.919> startups<00:34:42.320> right<00:34:42.480> now<00:34:42.760> is," + }, + { + "start": 2083.23, + "duration": 0.0, + "text": "topics for for AI startups right now is," + }, + { + "start": 2083.24, + "duration": 0.0, + "text": "topics for for AI startups right now is, well,<00:34:43.360> if<00:34:43.520> AI<00:34:43.720> writes<00:34:43.960> the<00:34:44.040> programs,<00:34:44.600> right," + }, + { + "start": 2084.869, + "duration": 0.0, + "text": "well, if AI writes the programs, right," + }, + { + "start": 2084.879, + "duration": 0.0, + "text": "well, if AI writes the programs, right, why<00:34:44.960> do<00:34:45.080> we<00:34:45.159> even<00:34:45.360> need<00:34:45.480> an<00:34:45.560> app<00:34:45.720> store?<00:34:46.040> Why<00:34:46.120> do" + }, + { + "start": 2086.19, + "duration": 0.0, + "text": "why do we even need an app store? Why do" + }, + { + "start": 2086.2, + "duration": 0.0, + "text": "why do we even need an app store? Why do we<00:34:46.280> even<00:34:46.440> need<00:34:46.600> software<00:34:46.879> engineers?<00:34:47.679> You" + }, + { + "start": 2087.75, + "duration": 0.0, + "text": "we even need software engineers? You" + }, + { + "start": 2087.76, + "duration": 0.0, + "text": "we even need software engineers? You should<00:34:47.879> basically<00:34:48.200> tell<00:34:48.360> your<00:34:48.480> computer<00:34:48.840> the" + }, + { + "start": 2088.95, + "duration": 0.0, + "text": "should basically tell your computer the" + }, + { + "start": 2088.96, + "duration": 0.0, + "text": "should basically tell your computer the app<00:34:49.159> you<00:34:49.240> want<00:34:49.520> and<00:34:49.600> then<00:34:49.679> it<00:34:49.760> does<00:34:49.960> it<00:34:50.200> and" + }, + { + "start": 2090.31, + "duration": 0.0, + "text": "app you want and then it does it and" + }, + { + "start": 2090.32, + "duration": 0.0, + "text": "app you want and then it does it and then<00:34:50.440> it<00:34:50.520> gives<00:34:50.639> you<00:34:50.720> the<00:34:50.840> interface<00:34:51.200> and<00:34:51.280> then" + }, + { + "start": 2091.35, + "duration": 0.0, + "text": "then it gives you the interface and then" + }, + { + "start": 2091.36, + "duration": 0.0, + "text": "then it gives you the interface and then the<00:34:51.440> data.<00:34:52.360> Um" + }, + { + "start": 2093.47, + "duration": 0.0, + "text": "the data. Um" + }, + { + "start": 2093.48, + "duration": 0.0, + "text": "the data. Um and<00:34:53.600> it<00:34:53.679> doesn't<00:34:53.879> work<00:34:54.000> at<00:34:54.080> all<00:34:54.200> right<00:34:54.360> now<00:34:54.600> and" + }, + { + "start": 2094.71, + "duration": 0.0, + "text": "and it doesn't work at all right now and" + }, + { + "start": 2094.72, + "duration": 0.0, + "text": "and it doesn't work at all right now and so,<00:34:54.919> I<00:34:55.280> you<00:34:55.399> know,<00:34:55.879> it<00:34:56.000> would<00:34:56.560> We're<00:34:56.639> going<00:34:56.760> to" + }, + { + "start": 2096.79, + "duration": 0.0, + "text": "so, I you know, it would We're going to" + }, + { + "start": 2096.8, + "duration": 0.0, + "text": "so, I you know, it would We're going to have<00:34:56.879> to<00:34:56.919> see<00:34:57.080> some<00:34:57.480> new<00:34:57.640> developments,<00:34:58.280> I" + }, + { + "start": 2098.31, + "duration": 0.0, + "text": "have to see some new developments, I" + }, + { + "start": 2098.32, + "duration": 0.0, + "text": "have to see some new developments, I think,<00:34:58.560> to<00:34:58.680> judge<00:34:59.640> if<00:34:59.760> that<00:34:59.880> will<00:35:00.000> happen<00:35:00.280> or" + }, + { + "start": 2100.35, + "duration": 0.0, + "text": "think, to judge if that will happen or" + }, + { + "start": 2100.36, + "duration": 0.0, + "text": "think, to judge if that will happen or not." + }, + { + "start": 2101.23, + "duration": 0.0, + "text": "not." + }, + { + "start": 2101.24, + "duration": 0.0, + "text": "not. I'm<00:35:01.600> basically<00:35:02.080> 90%<00:35:03.440> it<00:35:03.600> won't<00:35:03.840> cause<00:35:04.200> the" + }, + { + "start": 2104.27, + "duration": 0.0, + "text": "I'm basically 90% it won't cause the" + }, + { + "start": 2104.28, + "duration": 0.0, + "text": "I'm basically 90% it won't cause the destruction<00:35:04.720> of<00:35:04.800> better<00:35:05.000> software<00:35:05.400> for<00:35:05.520> at" + }, + { + "start": 2105.63, + "duration": 0.0, + "text": "destruction of better software for at" + }, + { + "start": 2105.64, + "duration": 0.0, + "text": "destruction of better software for at least<00:35:05.840> a<00:35:05.880> decade," + }, + { + "start": 2107.15, + "duration": 0.0, + "text": "least a decade," + }, + { + "start": 2107.16, + "duration": 0.0, + "text": "least a decade, but<00:35:07.360> 10%,<00:35:08.720> you<00:35:08.840> know,<00:35:09.000> there's<00:35:09.160> always" + }, + { + "start": 2109.35, + "duration": 0.0, + "text": "but 10%, you know, there's always" + }, + { + "start": 2109.36, + "duration": 0.0, + "text": "but 10%, you know, there's always something<00:35:09.840> around<00:35:10.000> the<00:35:10.080> corner.<00:35:10.360> Grok<00:35:10.640> 4<00:35:10.920> is" + }, + { + "start": 2111.07, + "duration": 0.0, + "text": "something around the corner. Grok 4 is" + }, + { + "start": 2111.08, + "duration": 0.0, + "text": "something around the corner. Grok 4 is like<00:35:11.360> much<00:35:11.560> better<00:35:12.040> after<00:35:12.240> some<00:35:12.400> reason." + }, + { + "start": 2113.11, + "duration": 0.0, + "text": "like much better after some reason." + }, + { + "start": 2113.12, + "duration": 0.0, + "text": "like much better after some reason. Okay." + }, + { + "start": 2114.31, + "duration": 0.0, + "text": "Okay." + }, + { + "start": 2114.32, + "duration": 0.0, + "text": "Okay. So,<00:35:14.440> in<00:35:14.560> 10<00:35:14.800> years,<00:35:15.240> like" + }, + { + "start": 2116.59, + "duration": 0.0, + "text": "So, in 10 years, like" + }, + { + "start": 2116.6, + "duration": 0.0, + "text": "So, in 10 years, like is<00:35:16.720> software<00:35:17.040> engineering<00:35:17.440> a<00:35:17.480> thing?<00:35:17.800> Like," + }, + { + "start": 2117.91, + "duration": 0.0, + "text": "is software engineering a thing? Like," + }, + { + "start": 2117.92, + "duration": 0.0, + "text": "is software engineering a thing? Like, when<00:35:18.040> you<00:35:18.160> walk<00:35:18.400> into<00:35:18.680> a<00:35:18.760> room<00:35:19.200> or<00:35:19.280> people<00:35:20.040> or" + }, + { + "start": 2120.19, + "duration": 0.0, + "text": "when you walk into a room or people or" + }, + { + "start": 2120.2, + "duration": 0.0, + "text": "when you walk into a room or people or software<00:35:20.520> engineers<00:35:20.920> going<00:35:21.040> to<00:35:21.120> be<00:35:21.720> crafting" + }, + { + "start": 2122.23, + "duration": 0.0, + "text": "software engineers going to be crafting" + }, + { + "start": 2122.24, + "duration": 0.0, + "text": "software engineers going to be crafting prompts?<00:35:23.120> So,<00:35:23.360> I've<00:35:23.560> been<00:35:23.720> thinking<00:35:24.040> about" + }, + { + "start": 2124.15, + "duration": 0.0, + "text": "prompts? So, I've been thinking about" + }, + { + "start": 2124.16, + "duration": 0.0, + "text": "prompts? So, I've been thinking about this<00:35:24.320> a<00:35:24.360> lot.<00:35:24.800> I<00:35:24.840> mean,<00:35:25.000> obviously,<00:35:25.360> prompt" + }, + { + "start": 2125.63, + "duration": 0.0, + "text": "this a lot. I mean, obviously, prompt" + }, + { + "start": 2125.64, + "duration": 0.0, + "text": "this a lot. I mean, obviously, prompt engineering<00:35:26.000> is<00:35:26.120> not<00:35:26.240> going<00:35:26.360> to<00:35:26.400> be<00:35:26.480> a<00:35:26.520> real" + }, + { + "start": 2126.75, + "duration": 0.0, + "text": "engineering is not going to be a real" + }, + { + "start": 2126.76, + "duration": 0.0, + "text": "engineering is not going to be a real thing<00:35:27.000> because<00:35:27.360> if<00:35:27.480> you<00:35:27.600> do<00:35:27.760> believe<00:35:28.000> these<00:35:28.200> AI" + }, + { + "start": 2128.35, + "duration": 0.0, + "text": "thing because if you do believe these AI" + }, + { + "start": 2128.36, + "duration": 0.0, + "text": "thing because if you do believe these AI systems<00:35:28.720> will<00:35:28.840> get<00:35:28.960> to<00:35:29.040> the<00:35:29.120> point<00:35:29.280> where<00:35:29.360> they" + }, + { + "start": 2129.43, + "duration": 0.0, + "text": "systems will get to the point where they" + }, + { + "start": 2129.44, + "duration": 0.0, + "text": "systems will get to the point where they can<00:35:29.560> automate<00:35:29.800> software<00:35:30.080> engineering,<00:35:30.400> they" + }, + { + "start": 2130.47, + "duration": 0.0, + "text": "can automate software engineering, they" + }, + { + "start": 2130.48, + "duration": 0.0, + "text": "can automate software engineering, they can<00:35:30.600> do<00:35:30.680> the<00:35:30.760> prompting,<00:35:31.080> too.<00:35:31.320> I<00:35:31.360> don't<00:35:31.480> think" + }, + { + "start": 2131.63, + "duration": 0.0, + "text": "can do the prompting, too. I don't think" + }, + { + "start": 2131.64, + "duration": 0.0, + "text": "can do the prompting, too. I don't think it's<00:35:31.760> that<00:35:31.880> crazy." + }, + { + "start": 2132.91, + "duration": 0.0, + "text": "it's that crazy." + }, + { + "start": 2132.92, + "duration": 0.0, + "text": "it's that crazy. Um<00:35:33.000> but<00:35:33.240> in<00:35:33.320> any<00:35:33.480> case," + }, + { + "start": 2134.67, + "duration": 0.0, + "text": "Um but in any case," + }, + { + "start": 2134.68, + "duration": 0.0, + "text": "Um but in any case, I<00:35:34.760> think<00:35:35.000> that<00:35:35.840> the<00:35:36.000> effect<00:35:36.360> of<00:35:36.400> this<00:35:36.560> on" + }, + { + "start": 2136.67, + "duration": 0.0, + "text": "I think that the effect of this on" + }, + { + "start": 2136.68, + "duration": 0.0, + "text": "I think that the effect of this on software<00:35:37.240> I<00:35:37.440> Well,<00:35:37.920> I<00:35:37.960> I've<00:35:38.320> I've<00:35:38.440> been" + }, + { + "start": 2138.63, + "duration": 0.0, + "text": "software I Well, I I've I've been" + }, + { + "start": 2138.64, + "duration": 0.0, + "text": "software I Well, I I've I've been thinking<00:35:38.920> a<00:35:38.960> lot<00:35:39.080> about<00:35:39.240> this<00:35:39.360> question<00:35:40.040> and" + }, + { + "start": 2140.19, + "duration": 0.0, + "text": "thinking a lot about this question and" + }, + { + "start": 2140.2, + "duration": 0.0, + "text": "thinking a lot about this question and I'm<00:35:40.400> I'm<00:35:40.520> still<00:35:40.680> very<00:35:40.920> uncertain,<00:35:41.360> but<00:35:41.960> if" + }, + { + "start": 2142.23, + "duration": 0.0, + "text": "I'm I'm still very uncertain, but if" + }, + { + "start": 2142.24, + "duration": 0.0, + "text": "I'm I'm still very uncertain, but if software<00:35:42.560> engineers<00:35:42.840> are<00:35:42.920> basically<00:35:43.320> making" + }, + { + "start": 2143.63, + "duration": 0.0, + "text": "software engineers are basically making" + }, + { + "start": 2143.64, + "duration": 0.0, + "text": "software engineers are basically making these<00:35:43.800> like<00:35:44.000> virtual<00:35:44.520> machines<00:35:45.040> that<00:35:45.160> do<00:35:45.280> some" + }, + { + "start": 2145.39, + "duration": 0.0, + "text": "these like virtual machines that do some" + }, + { + "start": 2145.4, + "duration": 0.0, + "text": "these like virtual machines that do some kind<00:35:45.600> of<00:35:45.800> processing<00:35:46.280> or<00:35:46.360> transformation" + }, + { + "start": 2147.19, + "duration": 0.0, + "text": "kind of processing or transformation" + }, + { + "start": 2147.2, + "duration": 0.0, + "text": "kind of processing or transformation that's<00:35:47.360> valuable<00:35:47.680> to<00:35:47.760> people<00:35:48.480> and<00:35:48.880> these<00:35:49.160> LLMs" + }, + { + "start": 2149.67, + "duration": 0.0, + "text": "that's valuable to people and these LLMs" + }, + { + "start": 2149.68, + "duration": 0.0, + "text": "that's valuable to people and these LLMs learned<00:35:49.960> the<00:35:50.040> patterns" + }, + { + "start": 2151.23, + "duration": 0.0, + "text": "learned the patterns" + }, + { + "start": 2151.24, + "duration": 0.0, + "text": "learned the patterns uh<00:35:51.360> to<00:35:51.560> to<00:35:51.800> to<00:35:52.000> create<00:35:52.400> some<00:35:52.560> of<00:35:52.640> these" + }, + { + "start": 2152.79, + "duration": 0.0, + "text": "uh to to to create some of these" + }, + { + "start": 2152.8, + "duration": 0.0, + "text": "uh to to to create some of these machines<00:35:53.280> as<00:35:53.440> it<00:35:53.520> relates<00:35:53.840> to<00:35:53.960> the<00:35:54.520> to<00:35:54.600> what" + }, + { + "start": 2154.71, + "duration": 0.0, + "text": "machines as it relates to the to what" + }, + { + "start": 2154.72, + "duration": 0.0, + "text": "machines as it relates to the to what you're<00:35:54.840> asking<00:35:55.200> of<00:35:55.360> it<00:35:56.000> and<00:35:56.680> I<00:35:56.760> think<00:35:56.960> people" + }, + { + "start": 2157.19, + "duration": 0.0, + "text": "you're asking of it and I think people" + }, + { + "start": 2157.2, + "duration": 0.0, + "text": "you're asking of it and I think people who<00:35:57.320> are<00:35:57.400> making<00:35:57.880> machines<00:35:58.440> that<00:35:58.560> are<00:35:58.680> very" + }, + { + "start": 2159.27, + "duration": 0.0, + "text": "who are making machines that are very" + }, + { + "start": 2159.28, + "duration": 0.0, + "text": "who are making machines that are very same-ish,<00:36:00.240> software<00:36:00.520> engineers<00:36:00.840> who<00:36:00.920> are" + }, + { + "start": 2161.87, + "duration": 0.0, + "text": "same-ish, software engineers who are" + }, + { + "start": 2161.88, + "duration": 0.0, + "text": "same-ish, software engineers who are data<00:36:02.120> transformations<00:36:02.760> that<00:36:02.880> are<00:36:03.000> similar<00:36:03.360> to" + }, + { + "start": 2163.47, + "duration": 0.0, + "text": "data transformations that are similar to" + }, + { + "start": 2163.48, + "duration": 0.0, + "text": "data transformations that are similar to what<00:36:03.680> exist<00:36:04.800> are<00:36:04.880> going<00:36:05.080> to<00:36:05.160> probably<00:36:05.520> be" + }, + { + "start": 2165.63, + "duration": 0.0, + "text": "what exist are going to probably be" + }, + { + "start": 2165.64, + "duration": 0.0, + "text": "what exist are going to probably be screwed.<00:36:06.600> Um<00:36:07.280> but<00:36:07.480> if<00:36:07.640> you're<00:36:07.800> making<00:36:08.080> unique" + }, + { + "start": 2168.39, + "duration": 0.0, + "text": "screwed. Um but if you're making unique" + }, + { + "start": 2168.4, + "duration": 0.0, + "text": "screwed. Um but if you're making unique data<00:36:08.600> transformations<00:36:09.280> that<00:36:09.400> are<00:36:09.520> unlike" + }, + { + "start": 2169.91, + "duration": 0.0, + "text": "data transformations that are unlike" + }, + { + "start": 2169.92, + "duration": 0.0, + "text": "data transformations that are unlike what<00:36:10.000> have<00:36:10.120> been<00:36:10.240> made<00:36:10.400> before,<00:36:10.720> kind<00:36:10.880> of" + }, + { + "start": 2170.95, + "duration": 0.0, + "text": "what have been made before, kind of" + }, + { + "start": 2170.96, + "duration": 0.0, + "text": "what have been made before, kind of unique<00:36:11.240> machines,<00:36:12.080> I<00:36:12.160> don't<00:36:12.360> see<00:36:12.480> how<00:36:12.680> these" + }, + { + "start": 2172.87, + "duration": 0.0, + "text": "unique machines, I don't see how these" + }, + { + "start": 2172.88, + "duration": 0.0, + "text": "unique machines, I don't see how these models<00:36:13.280> could<00:36:13.720> really<00:36:13.960> create<00:36:14.240> new<00:36:14.360> knowledge" + }, + { + "start": 2174.67, + "duration": 0.0, + "text": "models could really create new knowledge" + }, + { + "start": 2174.68, + "duration": 0.0, + "text": "models could really create new knowledge like<00:36:14.800> that<00:36:15.040> yet.<00:36:15.280> I<00:36:15.320> haven't<00:36:15.520> seen<00:36:15.640> evidence" + }, + { + "start": 2175.87, + "duration": 0.0, + "text": "like that yet. I haven't seen evidence" + }, + { + "start": 2175.88, + "duration": 0.0, + "text": "like that yet. I haven't seen evidence that<00:36:16.000> they<00:36:16.080> can<00:36:16.200> do<00:36:16.320> that." + }, + { + "start": 2177.35, + "duration": 0.0, + "text": "that they can do that." + }, + { + "start": 2177.36, + "duration": 0.0, + "text": "that they can do that. Okay." + }, + { + "start": 2178.67, + "duration": 0.0, + "text": "Okay." + }, + { + "start": 2178.68, + "duration": 0.0, + "text": "Okay. Um<00:36:19.240> you<00:36:19.360> pointed<00:36:19.680> out<00:36:19.800> something<00:36:20.560> uh" + }, + { + "start": 2181.23, + "duration": 0.0, + "text": "Um you pointed out something uh" + }, + { + "start": 2181.24, + "duration": 0.0, + "text": "Um you pointed out something uh like<00:36:21.400> learning<00:36:21.720> is<00:36:21.920> different<00:36:22.320> from" + }, + { + "start": 2183.47, + "duration": 0.0, + "text": "like learning is different from" + }, + { + "start": 2183.48, + "duration": 0.0, + "text": "like learning is different from optimization.<00:36:24.520> And<00:36:24.680> that's<00:36:24.920> that's<00:36:25.120> more<00:36:25.240> of" + }, + { + "start": 2185.31, + "duration": 0.0, + "text": "optimization. And that's that's more of" + }, + { + "start": 2185.32, + "duration": 0.0, + "text": "optimization. And that's that's more of a<00:36:25.600> philosophical<00:36:26.320> thing,<00:36:26.520> I<00:36:26.560> think." + }, + { + "start": 2186.99, + "duration": 0.0, + "text": "a philosophical thing, I think." + }, + { + "start": 2187.0, + "duration": 0.0, + "text": "a philosophical thing, I think. >> Like,<00:36:27.040> when<00:36:27.160> a<00:36:27.240> human<00:36:27.600> is<00:36:27.760> born,<00:36:28.200> right?" + }, + { + "start": 2188.43, + "duration": 0.0, + "text": ">> Like, when a human is born, right?" + }, + { + "start": 2188.44, + "duration": 0.0, + "text": ">> Like, when a human is born, right? They're<00:36:28.760> they're<00:36:28.920> born<00:36:29.200> like<00:36:29.360> extremely" + }, + { + "start": 2189.83, + "duration": 0.0, + "text": "They're they're born like extremely" + }, + { + "start": 2189.84, + "duration": 0.0, + "text": "They're they're born like extremely inefficient<00:36:30.360> and<00:36:30.440> they<00:36:30.520> just<00:36:30.680> get<00:36:30.800> more" + }, + { + "start": 2190.99, + "duration": 0.0, + "text": "inefficient and they just get more" + }, + { + "start": 2191.0, + "duration": 0.0, + "text": "inefficient and they just get more optimized<00:36:31.640> as<00:36:31.800> they<00:36:32.120> grow.<00:36:32.400> Is<00:36:32.520> that<00:36:32.640> not<00:36:32.800> the" + }, + { + "start": 2192.87, + "duration": 0.0, + "text": "optimized as they grow. Is that not the" + }, + { + "start": 2192.88, + "duration": 0.0, + "text": "optimized as they grow. Is that not the same?<00:36:33.560> It's<00:36:34.000> not<00:36:34.240> I<00:36:34.320> think<00:36:34.560> it's<00:36:34.720> not<00:36:34.880> the<00:36:35.000> same" + }, + { + "start": 2195.39, + "duration": 0.0, + "text": "same? It's not I think it's not the same" + }, + { + "start": 2195.4, + "duration": 0.0, + "text": "same? It's not I think it's not the same because<00:36:35.800> whatever<00:36:36.080> human<00:36:36.320> being<00:36:36.560> is<00:36:36.680> doing" + }, + { + "start": 2197.43, + "duration": 0.0, + "text": "because whatever human being is doing" + }, + { + "start": 2197.44, + "duration": 0.0, + "text": "because whatever human being is doing can<00:36:37.680> create<00:36:38.400> human<00:36:38.720> reasoning<00:36:39.400> and<00:36:39.880> I<00:36:39.960> don't" + }, + { + "start": 2200.19, + "duration": 0.0, + "text": "can create human reasoning and I don't" + }, + { + "start": 2200.2, + "duration": 0.0, + "text": "can create human reasoning and I don't think<00:36:40.440> it's" + }, + { + "start": 2201.35, + "duration": 0.0, + "text": "think it's" + }, + { + "start": 2201.36, + "duration": 0.0, + "text": "think it's I<00:36:41.440> don't<00:36:41.600> think<00:36:41.840> it's<00:36:42.560> it's<00:36:42.760> it<00:36:42.840> would<00:36:42.920> be<00:36:43.040> very" + }, + { + "start": 2203.19, + "duration": 0.0, + "text": "I don't think it's it's it would be very" + }, + { + "start": 2203.2, + "duration": 0.0, + "text": "I don't think it's it's it would be very hard<00:36:43.360> to<00:36:43.440> make<00:36:43.560> the<00:36:43.640> claim<00:36:43.880> that<00:36:43.960> these<00:36:44.120> things" + }, + { + "start": 2204.27, + "duration": 0.0, + "text": "hard to make the claim that these things" + }, + { + "start": 2204.28, + "duration": 0.0, + "text": "hard to make the claim that these things are<00:36:44.360> doing<00:36:44.520> human<00:36:44.760> reasoning<00:36:45.040> right<00:36:45.200> now,<00:36:45.360> I" + }, + { + "start": 2205.43, + "duration": 0.0, + "text": "are doing human reasoning right now, I" + }, + { + "start": 2205.44, + "duration": 0.0, + "text": "are doing human reasoning right now, I think,<00:36:45.720> for<00:36:45.840> a<00:36:45.880> lot<00:36:46.040> of<00:36:46.160> reasons.<00:36:47.160> Um<00:36:48.040> you" + }, + { + "start": 2208.11, + "duration": 0.0, + "text": "think, for a lot of reasons. Um you" + }, + { + "start": 2208.12, + "duration": 0.0, + "text": "think, for a lot of reasons. Um you know,<00:36:48.240> oftentimes<00:36:48.680> the<00:36:48.760> reasoning<00:36:49.040> doesn't" + }, + { + "start": 2209.23, + "duration": 0.0, + "text": "know, oftentimes the reasoning doesn't" + }, + { + "start": 2209.24, + "duration": 0.0, + "text": "know, oftentimes the reasoning doesn't correspond<00:36:49.640> to<00:36:49.680> what<00:36:49.800> they<00:36:49.880> say.<00:36:50.240> It's<00:36:50.400> like" + }, + { + "start": 2210.55, + "duration": 0.0, + "text": "correspond to what they say. It's like" + }, + { + "start": 2210.56, + "duration": 0.0, + "text": "correspond to what they say. It's like not<00:36:51.560> logically<00:36:51.960> coherent,<00:36:52.560> right?<00:36:52.840> It<00:36:53.320> you" + }, + { + "start": 2213.43, + "duration": 0.0, + "text": "not logically coherent, right? It you" + }, + { + "start": 2213.44, + "duration": 0.0, + "text": "not logically coherent, right? It you know," + }, + { + "start": 2214.23, + "duration": 0.0, + "text": "know," + }, + { + "start": 2214.24, + "duration": 0.0, + "text": "know, fundamentally<00:36:54.800> from<00:36:55.000> the<00:36:55.160> I<00:36:55.200> mean,<00:36:55.360> you<00:36:55.520> kind" + }, + { + "start": 2215.67, + "duration": 0.0, + "text": "fundamentally from the I mean, you kind" + }, + { + "start": 2215.68, + "duration": 0.0, + "text": "fundamentally from the I mean, you kind of<00:36:55.760> understand<00:36:56.360> what's<00:36:56.560> happening<00:36:56.840> here," + }, + { + "start": 2217.07, + "duration": 0.0, + "text": "of understand what's happening here," + }, + { + "start": 2217.08, + "duration": 0.0, + "text": "of understand what's happening here, right?<00:36:57.280> It's<00:36:58.080> it's<00:36:58.520> optimizing<00:36:59.120> and<00:36:59.400> and" + }, + { + "start": 2219.55, + "duration": 0.0, + "text": "right? It's it's optimizing and and" + }, + { + "start": 2219.56, + "duration": 0.0, + "text": "right? It's it's optimizing and and creating<00:37:00.000> an<00:37:00.160> automatically<00:37:01.200> created" + }, + { + "start": 2221.63, + "duration": 0.0, + "text": "creating an automatically created" + }, + { + "start": 2221.64, + "duration": 0.0, + "text": "creating an automatically created program" + }, + { + "start": 2222.87, + "duration": 0.0, + "text": "program" + }, + { + "start": 2222.88, + "duration": 0.0, + "text": "program um<00:37:03.040> that<00:37:03.200> might<00:37:03.320> have<00:37:03.440> nothing<00:37:03.680> to<00:37:03.800> do<00:37:04.040> with" + }, + { + "start": 2224.47, + "duration": 0.0, + "text": "um that might have nothing to do with" + }, + { + "start": 2224.48, + "duration": 0.0, + "text": "um that might have nothing to do with the<00:37:04.760> the<00:37:05.120> the<00:37:05.240> universal<00:37:05.840> causal" + }, + { + "start": 2226.31, + "duration": 0.0, + "text": "the the the universal causal" + }, + { + "start": 2226.32, + "duration": 0.0, + "text": "the the the universal causal relationship<00:37:06.880> between<00:37:07.160> like,<00:37:07.800> I<00:37:07.880> don't<00:37:08.000> know," + }, + { + "start": 2228.11, + "duration": 0.0, + "text": "relationship between like, I don't know," + }, + { + "start": 2228.12, + "duration": 0.0, + "text": "relationship between like, I don't know, you<00:37:08.240> know,<00:37:08.320> whatever's<00:37:08.640> happening<00:37:08.920> in<00:37:08.960> your" + }, + { + "start": 2229.07, + "duration": 0.0, + "text": "you know, whatever's happening in your" + }, + { + "start": 2229.08, + "duration": 0.0, + "text": "you know, whatever's happening in your brain<00:37:09.440> or<00:37:09.520> maybe<00:37:09.760> it's<00:37:10.240> if<00:37:10.400> you're<00:37:10.520> religious," + }, + { + "start": 2230.87, + "duration": 0.0, + "text": "brain or maybe it's if you're religious," + }, + { + "start": 2230.88, + "duration": 0.0, + "text": "brain or maybe it's if you're religious, then<00:37:11.040> it's<00:37:11.200> not<00:37:11.360> even<00:37:11.880> a<00:37:11.960> question." + }, + { + "start": 2242.12, + "duration": 0.0, + "text": "What's<00:37:22.280> up?" + }, + { + "start": 2243.71, + "duration": 0.0, + "text": "What's up?" + }, + { + "start": 2243.72, + "duration": 0.0, + "text": "What's up? Like<00:37:24.000> isn't<00:37:24.320> the<00:37:24.480> real<00:37:25.040> problem<00:37:25.520> like<00:37:25.760> not" + }, + { + "start": 2245.95, + "duration": 0.0, + "text": "Like isn't the real problem like not" + }, + { + "start": 2245.96, + "duration": 0.0, + "text": "Like isn't the real problem like not that<00:37:26.440> AI<00:37:26.800> replaces<00:37:27.560> software<00:37:27.880> engineers<00:37:28.320> and" + }, + { + "start": 2248.39, + "duration": 0.0, + "text": "that AI replaces software engineers and" + }, + { + "start": 2248.4, + "duration": 0.0, + "text": "that AI replaces software engineers and that<00:37:28.560> destroys<00:37:29.000> better<00:37:29.160> software,<00:37:29.520> but<00:37:29.640> that" + }, + { + "start": 2250.23, + "duration": 0.0, + "text": "that destroys better software, but that" + }, + { + "start": 2250.24, + "duration": 0.0, + "text": "that destroys better software, but that software<00:37:30.560> engineers<00:37:31.080> use<00:37:31.320> AI<00:37:31.720> and<00:37:31.880> that" + }, + { + "start": 2252.07, + "duration": 0.0, + "text": "software engineers use AI and that" + }, + { + "start": 2252.08, + "duration": 0.0, + "text": "software engineers use AI and that destroys<00:37:32.600> better<00:37:32.760> software<00:37:33.600> because<00:37:33.760> that's" + }, + { + "start": 2253.95, + "duration": 0.0, + "text": "destroys better software because that's" + }, + { + "start": 2253.96, + "duration": 0.0, + "text": "destroys better software because that's what's<00:37:34.080> happening<00:37:35.040> pretty<00:37:35.240> much<00:37:35.400> already." + }, + { + "start": 2255.95, + "duration": 0.0, + "text": "what's happening pretty much already." + }, + { + "start": 2255.96, + "duration": 0.0, + "text": "what's happening pretty much already. Well,<00:37:36.520> if<00:37:36.680> you<00:37:36.800> believe<00:37:37.000> that<00:37:37.120> society" + }, + { + "start": 2257.63, + "duration": 0.0, + "text": "Well, if you believe that society" + }, + { + "start": 2257.64, + "duration": 0.0, + "text": "Well, if you believe that society functions,<00:37:38.200> then<00:37:38.840> people<00:37:39.160> will<00:37:39.320> choose<00:37:39.720> the" + }, + { + "start": 2259.79, + "duration": 0.0, + "text": "functions, then people will choose the" + }, + { + "start": 2259.8, + "duration": 0.0, + "text": "functions, then people will choose the products<00:37:40.200> that<00:37:40.320> are<00:37:40.400> better." + }, + { + "start": 2261.55, + "duration": 0.0, + "text": "products that are better." + }, + { + "start": 2261.56, + "duration": 0.0, + "text": "products that are better. So," + }, + { + "start": 2262.51, + "duration": 0.0, + "text": "So," + }, + { + "start": 2262.52, + "duration": 0.0, + "text": "So, in<00:37:42.600> the<00:37:42.680> long<00:37:42.920> run<00:37:43.080> it,<00:37:43.320> you<00:37:43.400> know," + }, + { + "start": 2264.75, + "duration": 0.0, + "text": "in the long run it, you know," + }, + { + "start": 2264.76, + "duration": 0.0, + "text": "in the long run it, you know, if<00:37:45.360> yeah,<00:37:45.920> I<00:37:46.000> don't<00:37:46.120> know.<00:37:46.680> I'm<00:37:46.800> not<00:37:46.960> sure." + }, + { + "start": 2267.51, + "duration": 0.0, + "text": "if yeah, I don't know. I'm not sure." + }, + { + "start": 2267.52, + "duration": 0.0, + "text": "if yeah, I don't know. I'm not sure. >> you<00:37:47.640> pointed<00:37:48.040> out<00:37:48.480> uh" + }, + { + "start": 2270.03, + "duration": 0.0, + "text": ">> you pointed out uh" + }, + { + "start": 2270.04, + "duration": 0.0, + "text": ">> you pointed out uh the<00:37:50.440> like<00:37:50.680> Grok<00:37:51.080> had<00:37:51.240> a<00:37:51.320> massive<00:37:51.800> jump" + }, + { + "start": 2272.11, + "duration": 0.0, + "text": "the like Grok had a massive jump" + }, + { + "start": 2272.12, + "duration": 0.0, + "text": "the like Grok had a massive jump recently.<00:37:52.520> Like<00:37:52.640> did<00:37:52.800> you<00:37:52.880> look<00:37:53.080> into<00:37:53.280> that<00:37:53.440> at" + }, + { + "start": 2273.55, + "duration": 0.0, + "text": "recently. Like did you look into that at" + }, + { + "start": 2273.56, + "duration": 0.0, + "text": "recently. Like did you look into that at all?" + }, + { + "start": 2274.59, + "duration": 0.0, + "text": "all?" + }, + { + "start": 2274.6, + "duration": 0.0, + "text": "all? On<00:37:55.000> this<00:37:55.400> specifically<00:37:55.880> on<00:37:56.000> the<00:37:56.160> Arc<00:37:56.480> AGI" + }, + { + "start": 2276.87, + "duration": 0.0, + "text": "On this specifically on the Arc AGI" + }, + { + "start": 2276.88, + "duration": 0.0, + "text": "On this specifically on the Arc AGI test,<00:37:57.240> which<00:37:58.040> interestingly" + }, + { + "start": 2279.79, + "duration": 0.0, + "text": "test, which interestingly" + }, + { + "start": 2279.8, + "duration": 0.0, + "text": "test, which interestingly there's<00:38:00.040> another<00:38:00.360> Arc.<00:38:00.600> I<00:38:00.640> don't<00:38:00.760> know<00:38:00.800> if" + }, + { + "start": 2280.91, + "duration": 0.0, + "text": "there's another Arc. I don't know if" + }, + { + "start": 2280.92, + "duration": 0.0, + "text": "there's another Arc. I don't know if Taha<00:38:01.240> point<00:38:01.600> pointed<00:38:01.880> that<00:38:02.080> out.<00:38:02.400> I<00:38:02.480> missed" + }, + { + "start": 2282.79, + "duration": 0.0, + "text": "Taha point pointed that out. I missed" + }, + { + "start": 2282.8, + "duration": 0.0, + "text": "Taha point pointed that out. I missed it." + }, + { + "start": 2283.59, + "duration": 0.0, + "text": "it." + }, + { + "start": 2283.6, + "duration": 0.0, + "text": "it. Everything<00:38:03.960> is<00:38:04.120> Arc." + }, + { + "start": 2285.67, + "duration": 0.0, + "text": "Everything is Arc." + }, + { + "start": 2285.68, + "duration": 0.0, + "text": "Everything is Arc. So,<00:38:05.880> yeah,<00:38:06.320> Grok<00:38:07.160> might<00:38:07.560> have<00:38:07.800> performed" + }, + { + "start": 2288.31, + "duration": 0.0, + "text": "So, yeah, Grok might have performed" + }, + { + "start": 2288.32, + "duration": 0.0, + "text": "So, yeah, Grok might have performed better<00:38:08.600> because<00:38:08.920> of<00:38:09.000> its<00:38:09.200> utter<00:38:09.400> lack<00:38:09.640> of" + }, + { + "start": 2289.75, + "duration": 0.0, + "text": "better because of its utter lack of" + }, + { + "start": 2289.76, + "duration": 0.0, + "text": "better because of its utter lack of safety<00:38:10.080> training.<00:38:10.600> That<00:38:10.760> is<00:38:10.920> a<00:38:11.000> rumor<00:38:11.320> I<00:38:11.360> kind" + }, + { + "start": 2291.55, + "duration": 0.0, + "text": "safety training. That is a rumor I kind" + }, + { + "start": 2291.56, + "duration": 0.0, + "text": "safety training. That is a rumor I kind of<00:38:11.640> saw<00:38:12.000> and<00:38:12.120> it<00:38:12.240> seemed<00:38:12.400> to<00:38:12.480> be<00:38:12.600> somewhat" + }, + { + "start": 2292.91, + "duration": 0.0, + "text": "of saw and it seemed to be somewhat" + }, + { + "start": 2292.92, + "duration": 0.0, + "text": "of saw and it seemed to be somewhat confirmed<00:38:13.440> by,<00:38:14.200> you<00:38:14.280> know,<00:38:14.400> another<00:38:14.760> times" + }, + { + "start": 2295.19, + "duration": 0.0, + "text": "confirmed by, you know, another times" + }, + { + "start": 2295.2, + "duration": 0.0, + "text": "confirmed by, you know, another times it's<00:38:15.320> like<00:38:15.440> they<00:38:15.520> did<00:38:15.720> safety<00:38:16.040> training<00:38:16.360> and" + }, + { + "start": 2296.43, + "duration": 0.0, + "text": "it's like they did safety training and" + }, + { + "start": 2296.44, + "duration": 0.0, + "text": "it's like they did safety training and the<00:38:16.520> model<00:38:16.800> got<00:38:17.200> much<00:38:17.440> less,<00:38:18.160> you<00:38:18.280> know," + }, + { + "start": 2298.75, + "duration": 0.0, + "text": "the model got much less, you know," + }, + { + "start": 2298.76, + "duration": 0.0, + "text": "the model got much less, you know, reasonable<00:38:19.360> in<00:38:19.480> quotes." + }, + { + "start": 2300.59, + "duration": 0.0, + "text": "reasonable in quotes." + }, + { + "start": 2300.6, + "duration": 0.0, + "text": "reasonable in quotes. Um<00:38:20.920> it<00:38:21.080> might<00:38:21.280> be<00:38:21.400> the<00:38:21.480> case<00:38:21.760> that<00:38:22.240> the<00:38:22.400> safety" + }, + { + "start": 2302.71, + "duration": 0.0, + "text": "Um it might be the case that the safety" + }, + { + "start": 2302.72, + "duration": 0.0, + "text": "Um it might be the case that the safety training<00:38:23.160> is<00:38:23.320> not<00:38:23.560> intelligent<00:38:24.080> cuz<00:38:24.240> it's<00:38:24.360> so" + }, + { + "start": 2304.47, + "duration": 0.0, + "text": "training is not intelligent cuz it's so" + }, + { + "start": 2304.48, + "duration": 0.0, + "text": "training is not intelligent cuz it's so directly<00:38:24.880> refusing<00:38:25.400> to<00:38:25.640> to<00:38:25.760> say<00:38:25.960> something," + }, + { + "start": 2306.35, + "duration": 0.0, + "text": "directly refusing to to say something," + }, + { + "start": 2306.36, + "duration": 0.0, + "text": "directly refusing to to say something, right?" + }, + { + "start": 2307.35, + "duration": 0.0, + "text": "right?" + }, + { + "start": 2307.36, + "duration": 0.0, + "text": "right? Um,<00:38:27.560> there's<00:38:27.720> all<00:38:27.800> sorts<00:38:28.040> of<00:38:28.120> hypotheses<00:38:28.760> and" + }, + { + "start": 2308.83, + "duration": 0.0, + "text": "Um, there's all sorts of hypotheses and" + }, + { + "start": 2308.84, + "duration": 0.0, + "text": "Um, there's all sorts of hypotheses and there's<00:38:29.000> no<00:38:29.240> real,<00:38:29.600> again,<00:38:29.880> rigor<00:38:30.160> in<00:38:30.240> the" + }, + { + "start": 2310.59, + "duration": 0.0, + "text": "there's no real, again, rigor in the" + }, + { + "start": 2310.6, + "duration": 0.0, + "text": "there's no real, again, rigor in the It's<00:38:30.760> not<00:38:31.000> really<00:38:31.240> it's<00:38:31.760> They're<00:38:31.880> just<00:38:32.040> trying" + }, + { + "start": 2312.19, + "duration": 0.0, + "text": "It's not really it's They're just trying" + }, + { + "start": 2312.2, + "duration": 0.0, + "text": "It's not really it's They're just trying to<00:38:32.240> figure<00:38:32.480> it<00:38:32.600> out.<00:38:32.720> They<00:38:32.760> don't<00:38:33.120> Nobody" + }, + { + "start": 2313.43, + "duration": 0.0, + "text": "to figure it out. They don't Nobody" + }, + { + "start": 2313.44, + "duration": 0.0, + "text": "to figure it out. They don't Nobody Nobody<00:38:33.680> really<00:38:33.920> knows." + }, + { + "start": 2314.51, + "duration": 0.0, + "text": "Nobody really knows." + }, + { + "start": 2314.52, + "duration": 0.0, + "text": "Nobody really knows. >> like<00:38:34.680> broccolicing<00:38:35.240> like<00:38:35.440> some<00:38:35.600> schizo<00:38:36.160> like" + }, + { + "start": 2316.59, + "duration": 0.0, + "text": ">> like broccolicing like some schizo like" + }, + { + "start": 2316.6, + "duration": 0.0, + "text": ">> like broccolicing like some schizo like >> It's<00:38:36.680> not<00:38:36.880> good.<00:38:37.040> I<00:38:37.080> don't<00:38:37.240> want<00:38:37.360> to<00:38:37.640> repeat<00:38:37.960> it" + }, + { + "start": 2318.23, + "duration": 0.0, + "text": ">> It's not good. I don't want to repeat it" + }, + { + "start": 2318.24, + "duration": 0.0, + "text": ">> It's not good. I don't want to repeat it out<00:38:38.280> loud.<00:38:38.400> You<00:38:38.440> can<00:38:38.560> Google<00:38:38.880> it<00:38:38.960> if<00:38:39.080> you<00:38:39.200> want." + }, + { + "start": 2319.55, + "duration": 0.0, + "text": "out loud. You can Google it if you want." + }, + { + "start": 2319.56, + "duration": 0.0, + "text": "out loud. You can Google it if you want. It's<00:38:39.680> not<00:38:39.920> good.<00:38:40.200> It's<00:38:40.520> It's" + }, + { + "start": 2321.47, + "duration": 0.0, + "text": "It's not good. It's It's" + }, + { + "start": 2321.48, + "duration": 0.0, + "text": "It's not good. It's It's It's<00:38:41.680> completely<00:38:42.240> out<00:38:42.360> of<00:38:42.440> control,<00:38:43.040> but<00:38:43.280> it" + }, + { + "start": 2323.47, + "duration": 0.0, + "text": "It's completely out of control, but it" + }, + { + "start": 2323.48, + "duration": 0.0, + "text": "It's completely out of control, but it is<00:38:43.960> smarter.<00:38:44.600> So," + }, + { + "start": 2326.15, + "duration": 0.0, + "text": "is smarter. So," + }, + { + "start": 2326.16, + "duration": 0.0, + "text": "is smarter. So, well,<00:38:46.720> I<00:38:46.760> don't<00:38:46.920> know." + }, + { + "start": 2328.71, + "duration": 0.0, + "text": "well, I don't know." + }, + { + "start": 2328.72, + "duration": 0.0, + "text": "well, I don't know. Okay.<00:38:49.480> Um,<00:38:50.400> you<00:38:50.520> didn't<00:38:50.680> point<00:38:50.920> out<00:38:51.040> either" + }, + { + "start": 2331.27, + "duration": 0.0, + "text": "Okay. Um, you didn't point out either" + }, + { + "start": 2331.28, + "duration": 0.0, + "text": "Okay. Um, you didn't point out either like<00:38:51.440> what<00:38:51.960> caused<00:38:52.320> the" + }, + { + "start": 2333.71, + "duration": 0.0, + "text": "like what caused the" + }, + { + "start": 2333.72, + "duration": 0.0, + "text": "like what caused the this<00:38:54.280> AI<00:38:54.760> the<00:38:54.920> last<00:38:55.280> AI<00:38:55.560> winter<00:38:56.000> to<00:38:56.200> end," + }, + { + "start": 2336.47, + "duration": 0.0, + "text": "this AI the last AI winter to end," + }, + { + "start": 2336.48, + "duration": 0.0, + "text": "this AI the last AI winter to end, right?<00:38:56.720> Like<00:38:56.880> what<00:38:57.000> was<00:38:57.080> the<00:38:57.400> It<00:38:57.440> specifically" + }, + { + "start": 2338.19, + "duration": 0.0, + "text": "right? Like what was the It specifically" + }, + { + "start": 2338.2, + "duration": 0.0, + "text": "right? Like what was the It specifically happened<00:38:58.720> So,<00:38:58.880> transformers<00:38:59.560> were<00:38:59.680> made<00:38:59.960> kind" + }, + { + "start": 2340.15, + "duration": 0.0, + "text": "happened So, transformers were made kind" + }, + { + "start": 2340.16, + "duration": 0.0, + "text": "happened So, transformers were made kind of<00:39:00.200> I<00:39:00.280> think<00:39:00.440> like<00:39:00.560> 4<00:39:00.840> years<00:39:01.080> before<00:39:01.800> they<00:39:01.920> were" + }, + { + "start": 2342.03, + "duration": 0.0, + "text": "of I think like 4 years before they were" + }, + { + "start": 2342.04, + "duration": 0.0, + "text": "of I think like 4 years before they were really<00:39:02.280> used<00:39:02.480> to<00:39:02.600> generate<00:39:02.960> text<00:39:03.320> in<00:39:03.440> a<00:39:03.640> very" + }, + { + "start": 2344.51, + "duration": 0.0, + "text": "really used to generate text in a very" + }, + { + "start": 2344.52, + "duration": 0.0, + "text": "really used to generate text in a very in<00:39:04.680> a<00:39:04.760> very<00:39:05.080> parallel<00:39:05.520> large,<00:39:06.360> you<00:39:06.440> know,<00:39:06.760> mega" + }, + { + "start": 2347.19, + "duration": 0.0, + "text": "in a very parallel large, you know, mega" + }, + { + "start": 2347.2, + "duration": 0.0, + "text": "in a very parallel large, you know, mega data<00:39:07.560> scale<00:39:07.880> way.<00:39:08.680> Um,<00:39:09.520> and<00:39:09.880> it<00:39:10.080> happened<00:39:10.360> when" + }, + { + "start": 2350.51, + "duration": 0.0, + "text": "data scale way. Um, and it happened when" + }, + { + "start": 2350.52, + "duration": 0.0, + "text": "data scale way. Um, and it happened when OpenAI<00:39:11.000> I<00:39:11.040> think<00:39:11.280> released<00:39:11.720> GPT-2<00:39:12.400> and<00:39:12.480> they" + }, + { + "start": 2352.55, + "duration": 0.0, + "text": "OpenAI I think released GPT-2 and they" + }, + { + "start": 2352.56, + "duration": 0.0, + "text": "OpenAI I think released GPT-2 and they pointed<00:39:12.880> out<00:39:12.960> like,<00:39:13.120> \"Hey,<00:39:13.840> we<00:39:13.960> keep<00:39:14.200> just" + }, + { + "start": 2354.39, + "duration": 0.0, + "text": "pointed out like, \"Hey, we keep just" + }, + { + "start": 2354.4, + "duration": 0.0, + "text": "pointed out like, \"Hey, we keep just making<00:39:14.640> these<00:39:14.760> models<00:39:15.040> bigger<00:39:15.480> and<00:39:15.640> it<00:39:15.760> seems" + }, + { + "start": 2355.95, + "duration": 0.0, + "text": "making these models bigger and it seems" + }, + { + "start": 2355.96, + "duration": 0.0, + "text": "making these models bigger and it seems like<00:39:16.120> the<00:39:16.240> text<00:39:16.480> is<00:39:16.600> getting<00:39:16.840> more<00:39:16.960> and<00:39:17.040> more" + }, + { + "start": 2357.15, + "duration": 0.0, + "text": "like the text is getting more and more" + }, + { + "start": 2357.16, + "duration": 0.0, + "text": "like the text is getting more and more coherent.<00:39:18.000> So,<00:39:18.160> if<00:39:18.280> you<00:39:18.360> just<00:39:18.520> complete<00:39:18.800> the" + }, + { + "start": 2358.91, + "duration": 0.0, + "text": "coherent. So, if you just complete the" + }, + { + "start": 2358.92, + "duration": 0.0, + "text": "coherent. So, if you just complete the line,<00:39:19.360> like<00:39:19.480> someday<00:39:19.760> we<00:39:20.040> get<00:39:20.040> to<00:39:20.160> the<00:39:20.280> thing" + }, + { + "start": 2360.43, + "duration": 0.0, + "text": "line, like someday we get to the thing" + }, + { + "start": 2360.44, + "duration": 0.0, + "text": "line, like someday we get to the thing that<00:39:20.600> makes<00:39:20.760> everybody,<00:39:21.320> you<00:39:21.440> know," + }, + { + "start": 2362.07, + "duration": 0.0, + "text": "that makes everybody, you know," + }, + { + "start": 2362.08, + "duration": 0.0, + "text": "that makes everybody, you know, Pop-Tarts<00:39:22.520> for<00:39:22.640> free<00:39:22.840> or<00:39:22.920> something.\"<00:39:23.280> Like" + }, + { + "start": 2363.99, + "duration": 0.0, + "text": "Pop-Tarts for free or something.\" Like" + }, + { + "start": 2364.0, + "duration": 0.0, + "text": "Pop-Tarts for free or something.\" Like And<00:39:24.160> then<00:39:24.680> I<00:39:24.760> don't<00:39:24.920> know.<00:39:25.080> It's<00:39:25.320> That's<00:39:25.560> what" + }, + { + "start": 2365.67, + "duration": 0.0, + "text": "And then I don't know. It's That's what" + }, + { + "start": 2365.68, + "duration": 0.0, + "text": "And then I don't know. It's That's what started<00:39:26.000> this<00:39:26.120> massive<00:39:26.360> speculation<00:39:26.880> bubble." + }, + { + "start": 2367.15, + "duration": 0.0, + "text": "started this massive speculation bubble." + }, + { + "start": 2367.16, + "duration": 0.0, + "text": "started this massive speculation bubble. It's<00:39:27.280> like,<00:39:27.400> \"Wait<00:39:27.560> a<00:39:27.600> second.<00:39:27.960> We<00:39:28.040> might<00:39:28.280> have" + }, + { + "start": 2368.47, + "duration": 0.0, + "text": "It's like, \"Wait a second. We might have" + }, + { + "start": 2368.48, + "duration": 0.0, + "text": "It's like, \"Wait a second. We might have a<00:39:29.040> path<00:39:29.400> to<00:39:29.560> completely<00:39:30.040> removing" + }, + { + "start": 2371.43, + "duration": 0.0, + "text": "a path to completely removing" + }, + { + "start": 2371.44, + "duration": 0.0, + "text": "a path to completely removing uh<00:39:31.960> humanity<00:39:32.560> from<00:39:32.800> what<00:39:33.400> from<00:39:33.560> satisfying" + }, + { + "start": 2374.11, + "duration": 0.0, + "text": "uh humanity from what from satisfying" + }, + { + "start": 2374.12, + "duration": 0.0, + "text": "uh humanity from what from satisfying our<00:39:34.280> our<00:39:34.560> our<00:39:35.000> desires<00:39:35.400> in<00:39:35.480> the<00:39:35.560> physical" + }, + { + "start": 2375.87, + "duration": 0.0, + "text": "our our our desires in the physical" + }, + { + "start": 2375.88, + "duration": 0.0, + "text": "our our our desires in the physical realm.\"" + }, + { + "start": 2376.99, + "duration": 0.0, + "text": "realm.\"" + }, + { + "start": 2377.0, + "duration": 0.0, + "text": "realm.\" So,<00:39:37.120> it<00:39:37.200> was<00:39:37.280> just<00:39:37.480> they<00:39:37.800> they<00:39:37.880> just<00:39:38.040> scaled<00:39:38.360> it" + }, + { + "start": 2378.47, + "duration": 0.0, + "text": "So, it was just they they just scaled it" + }, + { + "start": 2378.48, + "duration": 0.0, + "text": "So, it was just they they just scaled it up.<00:39:39.000> Yeah,<00:39:39.160> they<00:39:39.280> scaled<00:39:39.560> it<00:39:39.680> up<00:39:39.800> and<00:39:39.880> it" + }, + { + "start": 2379.95, + "duration": 0.0, + "text": "up. Yeah, they scaled it up and it" + }, + { + "start": 2379.96, + "duration": 0.0, + "text": "up. Yeah, they scaled it up and it seemed<00:39:40.120> like<00:39:40.240> the<00:39:40.320> scaling<00:39:41.240> wouldn't<00:39:41.480> end." + }, + { + "start": 2382.71, + "duration": 0.0, + "text": "seemed like the scaling wouldn't end." + }, + { + "start": 2382.72, + "duration": 0.0, + "text": "seemed like the scaling wouldn't end. Is<00:39:42.800> that<00:39:42.960> what<00:39:43.080> they're<00:39:43.440> Is<00:39:43.560> that<00:39:43.680> what<00:39:43.800> people" + }, + { + "start": 2383.99, + "duration": 0.0, + "text": "Is that what they're Is that what people" + }, + { + "start": 2384.0, + "duration": 0.0, + "text": "Is that what they're Is that what people are<00:39:44.080> still<00:39:44.240> doing?<00:39:44.520> They<00:39:44.600> just<00:39:44.760> keep<00:39:44.920> scaling" + }, + { + "start": 2385.27, + "duration": 0.0, + "text": "are still doing? They just keep scaling" + }, + { + "start": 2385.28, + "duration": 0.0, + "text": "are still doing? They just keep scaling things." + }, + { + "start": 2385.67, + "duration": 0.0, + "text": "things." + }, + { + "start": 2385.68, + "duration": 0.0, + "text": "things. >> buying<00:39:45.920> GPUs.<00:39:46.320> I<00:39:46.360> mean,<00:39:46.480> they're<00:39:46.560> not<00:39:46.720> scaling" + }, + { + "start": 2387.03, + "duration": 0.0, + "text": ">> buying GPUs. I mean, they're not scaling" + }, + { + "start": 2387.04, + "duration": 0.0, + "text": ">> buying GPUs. I mean, they're not scaling in<00:39:47.080> the<00:39:47.160> same<00:39:47.440> way.<00:39:47.640> They're<00:39:47.720> not<00:39:47.880> just<00:39:48.000> making" + }, + { + "start": 2388.23, + "duration": 0.0, + "text": "in the same way. They're not just making" + }, + { + "start": 2388.24, + "duration": 0.0, + "text": "in the same way. They're not just making the<00:39:48.320> models<00:39:48.600> bigger.<00:39:48.840> They're<00:39:48.960> trying<00:39:49.160> to" + }, + { + "start": 2389.23, + "duration": 0.0, + "text": "the models bigger. They're trying to" + }, + { + "start": 2389.24, + "duration": 0.0, + "text": "the models bigger. They're trying to scale<00:39:50.240> I<00:39:50.680> the<00:39:50.800> so-called<00:39:51.160> reasoning<00:39:51.480> models." + }, + { + "start": 2392.67, + "duration": 0.0, + "text": "scale I the so-called reasoning models." + }, + { + "start": 2392.68, + "duration": 0.0, + "text": "scale I the so-called reasoning models. Um,<00:39:53.040> they're<00:39:53.160> basically<00:39:53.560> trying<00:39:53.760> to<00:39:53.840> make<00:39:54.000> the" + }, + { + "start": 2394.07, + "duration": 0.0, + "text": "Um, they're basically trying to make the" + }, + { + "start": 2394.08, + "duration": 0.0, + "text": "Um, they're basically trying to make the models<00:39:54.440> more<00:39:54.640> and<00:39:54.720> more<00:39:54.840> correct<00:39:55.280> instead<00:39:55.560> of" + }, + { + "start": 2395.63, + "duration": 0.0, + "text": "models more and more correct instead of" + }, + { + "start": 2395.64, + "duration": 0.0, + "text": "models more and more correct instead of just<00:39:55.840> making<00:39:56.120> them<00:39:56.240> remember<00:39:56.560> more<00:39:56.680> and<00:39:56.760> more" + }, + { + "start": 2396.87, + "duration": 0.0, + "text": "just making them remember more and more" + }, + { + "start": 2396.88, + "duration": 0.0, + "text": "just making them remember more and more facts,<00:39:57.240> basically." + }, + { + "start": 2398.55, + "duration": 0.0, + "text": "facts, basically." + }, + { + "start": 2398.56, + "duration": 0.0, + "text": "facts, basically. Okay." + }, + { + "start": 2399.99, + "duration": 0.0, + "text": "Okay." + }, + { + "start": 2400.0, + "duration": 0.0, + "text": "Okay. Um," + }, + { + "start": 2401.11, + "duration": 0.0, + "text": "Um," + }, + { + "start": 2401.12, + "duration": 0.0, + "text": "Um, I<00:40:01.160> don't<00:40:01.280> want<00:40:01.400> to<00:40:01.480> run<00:40:01.600> out<00:40:01.720> of<00:40:01.800> time<00:40:02.000> again," + }, + { + "start": 2402.23, + "duration": 0.0, + "text": "I don't want to run out of time again," + }, + { + "start": 2402.24, + "duration": 0.0, + "text": "I don't want to run out of time again, so<00:40:02.320> I'm<00:40:02.360> going<00:40:02.480> to<00:40:02.560> open" + }, + { + "start": 2403.55, + "duration": 0.0, + "text": "so I'm going to open" + }, + { + "start": 2403.56, + "duration": 0.0, + "text": "so I'm going to open I'm<00:40:03.600> going<00:40:03.720> to<00:40:03.760> do<00:40:03.880> questions<00:40:04.360> right<00:40:04.480> now." + }, + { + "start": 2405.27, + "duration": 0.0, + "text": "I'm going to do questions right now." + }, + { + "start": 2405.28, + "duration": 0.0, + "text": "I'm going to do questions right now. What's<00:40:05.440> up?<00:40:06.200> Uh<00:40:06.440> so,<00:40:06.760> Cameron,<00:40:07.200> a<00:40:07.240> bunch<00:40:07.520> of<00:40:07.640> us" + }, + { + "start": 2407.83, + "duration": 0.0, + "text": "What's up? Uh so, Cameron, a bunch of us" + }, + { + "start": 2407.84, + "duration": 0.0, + "text": "What's up? Uh so, Cameron, a bunch of us have<00:40:08.000> seen<00:40:08.320> your<00:40:08.840> uh<00:40:08.920> Dante<00:40:09.360> demo<00:40:09.720> just" + }, + { + "start": 2409.95, + "duration": 0.0, + "text": "have seen your uh Dante demo just" + }, + { + "start": 2409.96, + "duration": 0.0, + "text": "have seen your uh Dante demo just floating<00:40:10.200> around.<00:40:10.880> Um,<00:40:11.640> can<00:40:11.800> you<00:40:11.920> get<00:40:12.080> into" + }, + { + "start": 2412.35, + "duration": 0.0, + "text": "floating around. Um, can you get into" + }, + { + "start": 2412.36, + "duration": 0.0, + "text": "floating around. Um, can you get into why<00:40:12.920> it's<00:40:13.160> flawed<00:40:13.680> and<00:40:13.840> why<00:40:14.040> Yeah,<00:40:14.800> so<00:40:15.520> my<00:40:15.920> my" + }, + { + "start": 2416.03, + "duration": 0.0, + "text": "why it's flawed and why Yeah, so my my" + }, + { + "start": 2416.04, + "duration": 0.0, + "text": "why it's flawed and why Yeah, so my my whole<00:40:16.240> thinking<00:40:16.560> with<00:40:16.680> Dante's<00:40:17.120> Inferno<00:40:17.560> is<00:40:18.200> I" + }, + { + "start": 2418.27, + "duration": 0.0, + "text": "whole thinking with Dante's Inferno is I" + }, + { + "start": 2418.28, + "duration": 0.0, + "text": "whole thinking with Dante's Inferno is I think<00:40:18.480> when<00:40:18.600> I<00:40:18.680> saw<00:40:19.160> GPT-4<00:40:20.280> first<00:40:20.680> like<00:40:20.840> on<00:40:20.920> the" + }, + { + "start": 2420.99, + "duration": 0.0, + "text": "think when I saw GPT-4 first like on the" + }, + { + "start": 2421.0, + "duration": 0.0, + "text": "think when I saw GPT-4 first like on the API,<00:40:21.600> I<00:40:21.680> was<00:40:21.800> like<00:40:21.960> certain<00:40:22.840> that<00:40:22.960> this<00:40:23.160> would" + }, + { + "start": 2423.35, + "duration": 0.0, + "text": "API, I was like certain that this would" + }, + { + "start": 2423.36, + "duration": 0.0, + "text": "API, I was like certain that this would enable<00:40:23.720> a<00:40:23.800> huge<00:40:24.200> new<00:40:24.360> category<00:40:24.920> of<00:40:25.200> of<00:40:25.360> video" + }, + { + "start": 2425.63, + "duration": 0.0, + "text": "enable a huge new category of of video" + }, + { + "start": 2425.64, + "duration": 0.0, + "text": "enable a huge new category of of video games." + }, + { + "start": 2426.75, + "duration": 0.0, + "text": "games." + }, + { + "start": 2426.76, + "duration": 0.0, + "text": "games. Um<00:40:26.840> and<00:40:26.960> so<00:40:27.120> I<00:40:27.200> immediately<00:40:27.560> began<00:40:27.880> making" + }, + { + "start": 2428.11, + "duration": 0.0, + "text": "Um and so I immediately began making" + }, + { + "start": 2428.12, + "duration": 0.0, + "text": "Um and so I immediately began making like<00:40:28.280> an<00:40:28.400> RPG<00:40:28.880> where<00:40:29.040> all<00:40:29.160> the<00:40:29.240> characters" + }, + { + "start": 2429.59, + "duration": 0.0, + "text": "like an RPG where all the characters" + }, + { + "start": 2429.6, + "duration": 0.0, + "text": "like an RPG where all the characters were<00:40:29.680> controlled<00:40:30.120> by<00:40:30.280> an<00:40:30.400> LLM,<00:40:31.080> not<00:40:31.240> just" + }, + { + "start": 2431.39, + "duration": 0.0, + "text": "were controlled by an LLM, not just" + }, + { + "start": 2431.4, + "duration": 0.0, + "text": "were controlled by an LLM, not just their<00:40:31.480> speech<00:40:31.760> but<00:40:31.880> their<00:40:32.040> actions,<00:40:32.520> right?" + }, + { + "start": 2433.07, + "duration": 0.0, + "text": "their speech but their actions, right?" + }, + { + "start": 2433.08, + "duration": 0.0, + "text": "their speech but their actions, right? And<00:40:33.200> the<00:40:33.240> idea<00:40:33.480> was<00:40:33.600> like,<00:40:33.760> \"Whoa,<00:40:33.920> this<00:40:34.080> thing" + }, + { + "start": 2434.23, + "duration": 0.0, + "text": "And the idea was like, \"Whoa, this thing" + }, + { + "start": 2434.24, + "duration": 0.0, + "text": "And the idea was like, \"Whoa, this thing is<00:40:34.360> basically<00:40:34.680> like<00:40:34.880> prototype<00:40:35.520> AGI<00:40:36.000> or" + }, + { + "start": 2436.07, + "duration": 0.0, + "text": "is basically like prototype AGI or" + }, + { + "start": 2436.08, + "duration": 0.0, + "text": "is basically like prototype AGI or whatever,<00:40:36.360> right?\"<00:40:36.560> So,<00:40:36.760> it'll<00:40:37.000> just<00:40:37.120> kind<00:40:37.280> of" + }, + { + "start": 2437.35, + "duration": 0.0, + "text": "whatever, right?\" So, it'll just kind of" + }, + { + "start": 2437.36, + "duration": 0.0, + "text": "whatever, right?\" So, it'll just kind of operate<00:40:37.840> and<00:40:37.920> it'll<00:40:38.040> be<00:40:38.160> fun<00:40:38.400> and<00:40:38.480> like<00:40:39.280> and<00:40:39.520> I" + }, + { + "start": 2439.71, + "duration": 0.0, + "text": "operate and it'll be fun and like and I" + }, + { + "start": 2439.72, + "duration": 0.0, + "text": "operate and it'll be fun and like and I went<00:40:39.880> through<00:40:40.040> so<00:40:40.200> many<00:40:40.360> iterations<00:40:40.880> of<00:40:40.960> that" + }, + { + "start": 2441.23, + "duration": 0.0, + "text": "went through so many iterations of that" + }, + { + "start": 2441.24, + "duration": 0.0, + "text": "went through so many iterations of that idea.<00:40:41.800> I<00:40:41.920> made<00:40:42.120> it<00:40:42.200> so<00:40:42.240> that<00:40:42.360> like<00:40:42.480> they<00:40:42.600> could" + }, + { + "start": 2442.75, + "duration": 0.0, + "text": "idea. I made it so that like they could" + }, + { + "start": 2442.76, + "duration": 0.0, + "text": "idea. I made it so that like they could join<00:40:42.960> your<00:40:43.080> party<00:40:43.400> and<00:40:43.520> trade<00:40:43.840> items<00:40:44.120> and<00:40:44.240> all" + }, + { + "start": 2444.31, + "duration": 0.0, + "text": "join your party and trade items and all" + }, + { + "start": 2444.32, + "duration": 0.0, + "text": "join your party and trade items and all this<00:40:44.440> stuff<00:40:44.720> and<00:40:44.840> like<00:40:44.960> I<00:40:45.000> played<00:40:45.240> with<00:40:45.360> it<00:40:45.400> a" + }, + { + "start": 2445.47, + "duration": 0.0, + "text": "this stuff and like I played with it a" + }, + { + "start": 2445.48, + "duration": 0.0, + "text": "this stuff and like I played with it a bunch,<00:40:46.120> but<00:40:46.240> it<00:40:46.360> never<00:40:46.600> felt<00:40:46.800> like<00:40:46.920> there<00:40:47.000> was" + }, + { + "start": 2447.19, + "duration": 0.0, + "text": "bunch, but it never felt like there was" + }, + { + "start": 2447.2, + "duration": 0.0, + "text": "bunch, but it never felt like there was any<00:40:47.400> meat<00:40:47.680> to<00:40:47.760> the<00:40:47.840> game.<00:40:48.280> The<00:40:48.400> problem<00:40:48.840> is" + }, + { + "start": 2449.43, + "duration": 0.0, + "text": "any meat to the game. The problem is" + }, + { + "start": 2449.44, + "duration": 0.0, + "text": "any meat to the game. The problem is games<00:40:49.720> are<00:40:49.800> about<00:40:50.480> predicting<00:40:50.880> and" + }, + { + "start": 2450.99, + "duration": 0.0, + "text": "games are about predicting and" + }, + { + "start": 2451.0, + "duration": 0.0, + "text": "games are about predicting and understanding<00:40:51.800> systems<00:40:52.160> almost,<00:40:52.480> right?<00:40:52.720> And" + }, + { + "start": 2452.83, + "duration": 0.0, + "text": "understanding systems almost, right? And" + }, + { + "start": 2452.84, + "duration": 0.0, + "text": "understanding systems almost, right? And an<00:40:52.960> LLM<00:40:53.280> is<00:40:53.400> an<00:40:53.520> unpredictable<00:40:54.520> black<00:40:54.800> box," + }, + { + "start": 2455.43, + "duration": 0.0, + "text": "an LLM is an unpredictable black box," + }, + { + "start": 2455.44, + "duration": 0.0, + "text": "an LLM is an unpredictable black box, sometimes<00:40:55.880> acts<00:40:56.040> like<00:40:56.160> a<00:40:56.200> person,<00:40:56.480> sometimes" + }, + { + "start": 2456.83, + "duration": 0.0, + "text": "sometimes acts like a person, sometimes" + }, + { + "start": 2456.84, + "duration": 0.0, + "text": "sometimes acts like a person, sometimes doesn't.<00:40:57.200> It's<00:40:57.320> not<00:40:57.480> fun<00:40:57.680> to<00:40:57.800> interact<00:40:58.200> with" + }, + { + "start": 2458.35, + "duration": 0.0, + "text": "doesn't. It's not fun to interact with" + }, + { + "start": 2458.36, + "duration": 0.0, + "text": "doesn't. It's not fun to interact with in<00:40:58.440> a<00:40:58.480> way<00:40:58.640> where<00:40:58.760> you're<00:40:58.840> trying<00:40:59.040> to<00:40:59.120> get<00:40:59.240> it" + }, + { + "start": 2459.35, + "duration": 0.0, + "text": "in a way where you're trying to get it" + }, + { + "start": 2459.36, + "duration": 0.0, + "text": "in a way where you're trying to get it to<00:40:59.440> do<00:40:59.600> something.<00:41:00.360> At<00:41:00.520> best,<00:41:00.800> you're" + }, + { + "start": 2460.87, + "duration": 0.0, + "text": "to do something. At best, you're" + }, + { + "start": 2460.88, + "duration": 0.0, + "text": "to do something. At best, you're basically<00:41:01.240> like,<00:41:01.400> \"Hey<00:41:01.640> trader,<00:41:02.680> I'll" + }, + { + "start": 2462.79, + "duration": 0.0, + "text": "basically like, \"Hey trader, I'll" + }, + { + "start": 2462.8, + "duration": 0.0, + "text": "basically like, \"Hey trader, I'll totally<00:41:03.120> give<00:41:03.320> you<00:41:03.760> a<00:41:03.840> million<00:41:04.080> dollars<00:41:04.360> if" + }, + { + "start": 2464.43, + "duration": 0.0, + "text": "totally give you a million dollars if" + }, + { + "start": 2464.44, + "duration": 0.0, + "text": "totally give you a million dollars if you<00:41:04.600> drop<00:41:04.840> your<00:41:05.000> item.\"<00:41:05.320> And<00:41:05.400> then<00:41:05.520> it's<00:41:05.640> like," + }, + { + "start": 2465.83, + "duration": 0.0, + "text": "you drop your item.\" And then it's like," + }, + { + "start": 2465.84, + "duration": 0.0, + "text": "you drop your item.\" And then it's like, \"This<00:41:06.000> is<00:41:06.120> not<00:41:06.320> fun.<00:41:06.640> This<00:41:06.760> is<00:41:06.920> dumb.\"" + }, + { + "start": 2467.95, + "duration": 0.0, + "text": "\"This is not fun. This is dumb.\"" + }, + { + "start": 2467.96, + "duration": 0.0, + "text": "\"This is not fun. This is dumb.\" It's<00:41:08.080> funny<00:41:08.320> It's<00:41:08.520> funny<00:41:08.840> like<00:41:09.040> for<00:41:09.200> one<00:41:09.440> time," + }, + { + "start": 2469.79, + "duration": 0.0, + "text": "It's funny It's funny like for one time," + }, + { + "start": 2469.8, + "duration": 0.0, + "text": "It's funny It's funny like for one time, but<00:41:09.920> it's<00:41:10.040> not<00:41:10.200> like<00:41:10.320> a<00:41:10.400> game<00:41:10.640> you<00:41:10.720> could" + }, + { + "start": 2470.87, + "duration": 0.0, + "text": "but it's not like a game you could" + }, + { + "start": 2470.88, + "duration": 0.0, + "text": "but it's not like a game you could master." + }, + { + "start": 2472.11, + "duration": 0.0, + "text": "master." + }, + { + "start": 2472.12, + "duration": 0.0, + "text": "master. It's<00:41:12.240> like<00:41:12.400> the<00:41:12.480> vending<00:41:12.880> machine<00:41:13.480> at<00:41:13.680> Claude." + }, + { + "start": 2474.27, + "duration": 0.0, + "text": "It's like the vending machine at Claude." + }, + { + "start": 2474.28, + "duration": 0.0, + "text": "It's like the vending machine at Claude. Okay,<00:41:14.480> there's<00:41:15.040> Right<00:41:15.240> now,<00:41:15.400> there's<00:41:15.720> some<00:41:15.960> AI" + }, + { + "start": 2476.15, + "duration": 0.0, + "text": "Okay, there's Right now, there's some AI" + }, + { + "start": 2476.16, + "duration": 0.0, + "text": "Okay, there's Right now, there's some AI companies<00:41:16.520> are<00:41:16.600> trying<00:41:16.880> to<00:41:17.200> first<00:41:17.520> deploy" + }, + { + "start": 2477.91, + "duration": 0.0, + "text": "companies are trying to first deploy" + }, + { + "start": 2477.92, + "duration": 0.0, + "text": "companies are trying to first deploy these<00:41:18.440> these<00:41:18.640> LLMs<00:41:19.080> in<00:41:19.160> the<00:41:19.240> real<00:41:19.440> world<00:41:19.640> to" + }, + { + "start": 2479.75, + "duration": 0.0, + "text": "these these LLMs in the real world to" + }, + { + "start": 2479.76, + "duration": 0.0, + "text": "these these LLMs in the real world to start<00:41:20.000> to," + }, + { + "start": 2480.79, + "duration": 0.0, + "text": "start to," + }, + { + "start": 2480.8, + "duration": 0.0, + "text": "start to, you<00:41:20.840> know,<00:41:21.000> create<00:41:21.280> a<00:41:21.400> utopian<00:41:22.000> post-scarcity" + }, + { + "start": 2482.71, + "duration": 0.0, + "text": "you know, create a utopian post-scarcity" + }, + { + "start": 2482.72, + "duration": 0.0, + "text": "you know, create a utopian post-scarcity world<00:41:23.000> where<00:41:23.120> machines<00:41:23.600> make<00:41:23.800> everything" + }, + { + "start": 2484.11, + "duration": 0.0, + "text": "world where machines make everything" + }, + { + "start": 2484.12, + "duration": 0.0, + "text": "world where machines make everything automatically<00:41:24.680> and<00:41:24.760> they're<00:41:24.840> like,<00:41:25.360> \"Step" + }, + { + "start": 2485.63, + "duration": 0.0, + "text": "automatically and they're like, \"Step" + }, + { + "start": 2485.64, + "duration": 0.0, + "text": "automatically and they're like, \"Step number<00:41:25.840> one,<00:41:26.160> put<00:41:26.320> the<00:41:26.440> LLM<00:41:26.760> in<00:41:26.840> a<00:41:26.880> vending" + }, + { + "start": 2487.19, + "duration": 0.0, + "text": "number one, put the LLM in a vending" + }, + { + "start": 2487.2, + "duration": 0.0, + "text": "number one, put the LLM in a vending machine<00:41:27.560> and<00:41:27.680> have<00:41:27.840> it<00:41:28.120> automatically<00:41:28.600> order" + }, + { + "start": 2488.83, + "duration": 0.0, + "text": "machine and have it automatically order" + }, + { + "start": 2488.84, + "duration": 0.0, + "text": "machine and have it automatically order and<00:41:28.920> interface<00:41:29.280> with<00:41:29.360> the<00:41:29.440> people<00:41:29.680> or" + }, + { + "start": 2489.75, + "duration": 0.0, + "text": "and interface with the people or" + }, + { + "start": 2489.76, + "duration": 0.0, + "text": "and interface with the people or whatever.\"<00:41:30.480> And<00:41:30.640> right<00:41:30.800> now,<00:41:30.960> people<00:41:31.200> are" + }, + { + "start": 2491.27, + "duration": 0.0, + "text": "whatever.\" And right now, people are" + }, + { + "start": 2491.28, + "duration": 0.0, + "text": "whatever.\" And right now, people are basically<00:41:31.600> convincing<00:41:32.080> these<00:41:32.320> AI<00:41:32.520> vending" + }, + { + "start": 2492.75, + "duration": 0.0, + "text": "basically convincing these AI vending" + }, + { + "start": 2492.76, + "duration": 0.0, + "text": "basically convincing these AI vending machines<00:41:33.160> to<00:41:33.680> stock<00:41:34.040> many<00:41:34.320> tungsten<00:41:34.680> cubes" + }, + { + "start": 2494.91, + "duration": 0.0, + "text": "machines to stock many tungsten cubes" + }, + { + "start": 2494.92, + "duration": 0.0, + "text": "machines to stock many tungsten cubes and<00:41:35.000> sell<00:41:35.120> them<00:41:35.240> at<00:41:35.320> a<00:41:35.360> loss<00:41:35.680> and<00:41:35.760> then<00:41:35.880> they<00:41:35.920> go" + }, + { + "start": 2496.03, + "duration": 0.0, + "text": "and sell them at a loss and then they go" + }, + { + "start": 2496.04, + "duration": 0.0, + "text": "and sell them at a loss and then they go bankrupt." + }, + { + "start": 2497.35, + "duration": 0.0, + "text": "bankrupt." + }, + { + "start": 2497.36, + "duration": 0.0, + "text": "bankrupt. Um<00:41:37.600> they've<00:41:37.840> published<00:41:38.160> many<00:41:38.360> reports<00:41:38.720> about" + }, + { + "start": 2498.91, + "duration": 0.0, + "text": "Um they've published many reports about" + }, + { + "start": 2498.92, + "duration": 0.0, + "text": "Um they've published many reports about this<00:41:39.120> and<00:41:39.240> they're<00:41:39.320> like,<00:41:39.560> \"You<00:41:39.640> know,<00:41:39.760> we're" + }, + { + "start": 2499.99, + "duration": 0.0, + "text": "this and they're like, \"You know, we're" + }, + { + "start": 2500.0, + "duration": 0.0, + "text": "this and they're like, \"You know, we're we're<00:41:40.320> you<00:41:40.400> know,<00:41:40.480> we're<00:41:40.560> doing<00:41:40.720> some<00:41:40.840> prompt" + }, + { + "start": 2501.11, + "duration": 0.0, + "text": "we're you know, we're doing some prompt" + }, + { + "start": 2501.12, + "duration": 0.0, + "text": "we're you know, we're doing some prompt engineering<00:41:41.440> to<00:41:41.520> fix<00:41:41.680> the<00:41:41.760> business<00:41:42.120> model" + }, + { + "start": 2502.43, + "duration": 0.0, + "text": "engineering to fix the business model" + }, + { + "start": 2502.44, + "duration": 0.0, + "text": "engineering to fix the business model and<00:41:42.560> we're<00:41:42.640> trying<00:41:42.800> to<00:41:42.880> make<00:41:43.040> it<00:41:43.200> refuse" + }, + { + "start": 2503.59, + "duration": 0.0, + "text": "and we're trying to make it refuse" + }, + { + "start": 2503.6, + "duration": 0.0, + "text": "and we're trying to make it refuse tungsten<00:41:43.920> cubes<00:41:44.160> more.\"<00:41:44.960> Um<00:41:45.640> but" + }, + { + "start": 2506.63, + "duration": 0.0, + "text": "tungsten cubes more.\" Um but" + }, + { + "start": 2506.64, + "duration": 0.0, + "text": "tungsten cubes more.\" Um but really,<00:41:46.880> I<00:41:46.960> think<00:41:47.120> there<00:41:47.240> might<00:41:47.400> be<00:41:47.480> no<00:41:47.680> end<00:41:47.840> to" + }, + { + "start": 2507.91, + "duration": 0.0, + "text": "really, I think there might be no end to" + }, + { + "start": 2507.92, + "duration": 0.0, + "text": "really, I think there might be no end to humanity's<00:41:48.480> ingenuity<00:41:48.960> as<00:41:49.080> we<00:41:49.200> battle" + }, + { + "start": 2509.47, + "duration": 0.0, + "text": "humanity's ingenuity as we battle" + }, + { + "start": 2509.48, + "duration": 0.0, + "text": "humanity's ingenuity as we battle against<00:41:49.720> the<00:41:49.800> machines<00:41:50.200> for<00:41:50.280> the<00:41:50.360> next<00:41:50.520> few" + }, + { + "start": 2510.63, + "duration": 0.0, + "text": "against the machines for the next few" + }, + { + "start": 2510.64, + "duration": 0.0, + "text": "against the machines for the next few decades.<00:41:51.440> So." + }, + { + "start": 2513.11, + "duration": 0.0, + "text": "decades. So." + }, + { + "start": 2513.12, + "duration": 0.0, + "text": "decades. So. That's<00:41:53.280> probably<00:41:53.560> fine,<00:41:53.880> right?<00:41:54.080> It's" + }, + { + "start": 2515.59, + "duration": 0.0, + "text": "That's probably fine, right? It's" + }, + { + "start": 2515.6, + "duration": 0.0, + "text": "That's probably fine, right? It's Everything's<00:41:55.960> probably<00:41:56.440> fine,<00:41:56.880> but" + }, + { + "start": 2516.95, + "duration": 0.0, + "text": "Everything's probably fine, but" + }, + { + "start": 2516.96, + "duration": 0.0, + "text": "Everything's probably fine, but everything's<00:41:57.800> probably<00:41:58.280> fine,<00:41:58.600> but" + }, + { + "start": 2518.83, + "duration": 0.0, + "text": "everything's probably fine, but" + }, + { + "start": 2518.84, + "duration": 0.0, + "text": "everything's probably fine, but sometimes<00:41:59.280> bad<00:41:59.520> things<00:41:59.720> happen.<00:42:00.640> I<00:42:00.680> don't" + }, + { + "start": 2520.83, + "duration": 0.0, + "text": "sometimes bad things happen. I don't" + }, + { + "start": 2520.84, + "duration": 0.0, + "text": "sometimes bad things happen. I don't know.<00:42:01.560> What's<00:42:02.000> up?" + }, + { + "start": 2523.11, + "duration": 0.0, + "text": "know. What's up?" + }, + { + "start": 2523.12, + "duration": 0.0, + "text": "know. What's up? Do<00:42:03.160> you<00:42:03.200> think<00:42:03.360> there<00:42:03.600> So,<00:42:03.760> I<00:42:03.960> I<00:42:04.000> feel<00:42:04.120> like" + }, + { + "start": 2524.23, + "duration": 0.0, + "text": "Do you think there So, I I feel like" + }, + { + "start": 2524.24, + "duration": 0.0, + "text": "Do you think there So, I I feel like there<00:42:04.360> is<00:42:04.440> a<00:42:04.480> field<00:42:04.680> of<00:42:04.720> research,<00:42:05.000> I<00:42:05.040> just" + }, + { + "start": 2525.23, + "duration": 0.0, + "text": "there is a field of research, I just" + }, + { + "start": 2525.24, + "duration": 0.0, + "text": "there is a field of research, I just don't<00:42:05.400> know<00:42:05.480> the<00:42:05.560> words<00:42:05.800> for<00:42:05.960> it<00:42:06.040> or<00:42:06.120> how<00:42:06.240> big" + }, + { + "start": 2526.43, + "duration": 0.0, + "text": "don't know the words for it or how big" + }, + { + "start": 2526.44, + "duration": 0.0, + "text": "don't know the words for it or how big it<00:42:06.560> is,<00:42:06.720> but<00:42:06.840> do<00:42:06.920> you<00:42:06.960> think<00:42:07.120> there's<00:42:07.280> a<00:42:07.360> way" + }, + { + "start": 2527.51, + "duration": 0.0, + "text": "it is, but do you think there's a way" + }, + { + "start": 2527.52, + "duration": 0.0, + "text": "it is, but do you think there's a way that<00:42:07.680> we<00:42:07.840> can<00:42:08.080> reverse<00:42:08.400> engineer<00:42:09.000> Yeah," + }, + { + "start": 2529.15, + "duration": 0.0, + "text": "that we can reverse engineer Yeah," + }, + { + "start": 2529.16, + "duration": 0.0, + "text": "that we can reverse engineer Yeah, interpretability<00:42:09.880> is<00:42:10.040> big.<00:42:10.560> A<00:42:10.640> lot<00:42:10.840> of<00:42:10.960> Yeah," + }, + { + "start": 2531.23, + "duration": 0.0, + "text": "interpretability is big. A lot of Yeah," + }, + { + "start": 2531.24, + "duration": 0.0, + "text": "interpretability is big. A lot of Yeah, Anthropic<00:42:11.720> is<00:42:11.920> the<00:42:12.080> leader<00:42:12.320> in" + }, + { + "start": 2532.63, + "duration": 0.0, + "text": "Anthropic is the leader in" + }, + { + "start": 2532.64, + "duration": 0.0, + "text": "Anthropic is the leader in interpretability,<00:42:13.280> I<00:42:13.360> think.<00:42:14.040> Um<00:42:14.120> they" + }, + { + "start": 2534.23, + "duration": 0.0, + "text": "interpretability, I think. Um they" + }, + { + "start": 2534.24, + "duration": 0.0, + "text": "interpretability, I think. Um they basically<00:42:14.640> do<00:42:14.800> data<00:42:15.000> science<00:42:15.320> on<00:42:15.440> the<00:42:15.520> neurons" + }, + { + "start": 2535.91, + "duration": 0.0, + "text": "basically do data science on the neurons" + }, + { + "start": 2535.92, + "duration": 0.0, + "text": "basically do data science on the neurons themselves,<00:42:16.600> and<00:42:16.680> they<00:42:16.760> try<00:42:16.960> to<00:42:17.040> make<00:42:17.240> claims" + }, + { + "start": 2537.55, + "duration": 0.0, + "text": "themselves, and they try to make claims" + }, + { + "start": 2537.56, + "duration": 0.0, + "text": "themselves, and they try to make claims about,<00:42:17.960> you<00:42:18.040> know,<00:42:18.200> how<00:42:18.320> the<00:42:18.440> data<00:42:18.640> is<00:42:18.760> flowing" + }, + { + "start": 2538.99, + "duration": 0.0, + "text": "about, you know, how the data is flowing" + }, + { + "start": 2539.0, + "duration": 0.0, + "text": "about, you know, how the data is flowing through<00:42:19.120> the<00:42:19.240> model.<00:42:20.080> Um<00:42:20.280> you<00:42:20.400> can<00:42:20.520> look<00:42:20.640> at" + }, + { + "start": 2540.71, + "duration": 0.0, + "text": "through the model. Um you can look at" + }, + { + "start": 2540.72, + "duration": 0.0, + "text": "through the model. Um you can look at their<00:42:20.840> Golden<00:42:21.080> Gate<00:42:21.240> model.<00:42:21.440> They<00:42:21.480> have<00:42:21.600> a<00:42:21.680> lot" + }, + { + "start": 2541.87, + "duration": 0.0, + "text": "their Golden Gate model. They have a lot" + }, + { + "start": 2541.88, + "duration": 0.0, + "text": "their Golden Gate model. They have a lot of<00:42:22.080> really<00:42:22.360> excellent<00:42:22.680> stuff<00:42:22.920> you<00:42:23.000> can<00:42:23.360> read" + }, + { + "start": 2543.59, + "duration": 0.0, + "text": "of really excellent stuff you can read" + }, + { + "start": 2543.6, + "duration": 0.0, + "text": "of really excellent stuff you can read forever.<00:42:23.920> It's<00:42:24.040> really<00:42:24.240> cool.<00:42:25.080> Um<00:42:25.640> Do<00:42:25.720> you" + }, + { + "start": 2545.79, + "duration": 0.0, + "text": "forever. It's really cool. Um Do you" + }, + { + "start": 2545.8, + "duration": 0.0, + "text": "forever. It's really cool. Um Do you know<00:42:25.960> much<00:42:26.200> about<00:42:26.520> it?<00:42:26.600> Like,<00:42:26.760> do<00:42:26.840> you<00:42:26.920> feel" + }, + { + "start": 2547.03, + "duration": 0.0, + "text": "know much about it? Like, do you feel" + }, + { + "start": 2547.04, + "duration": 0.0, + "text": "know much about it? Like, do you feel like<00:42:27.200> that<00:42:27.440> is<00:42:27.600> the<00:42:27.680> way<00:42:27.840> forward,<00:42:28.280> or<00:42:28.400> is<00:42:28.480> that" + }, + { + "start": 2548.63, + "duration": 0.0, + "text": "like that is the way forward, or is that" + }, + { + "start": 2548.64, + "duration": 0.0, + "text": "like that is the way forward, or is that just<00:42:28.840> a<00:42:28.960> byproduct<00:42:29.560> of<00:42:29.760> the<00:42:29.880> craziness?<00:42:30.400> I" + }, + { + "start": 2550.47, + "duration": 0.0, + "text": "just a byproduct of the craziness? I" + }, + { + "start": 2550.48, + "duration": 0.0, + "text": "just a byproduct of the craziness? I don't<00:42:30.720> think<00:42:30.920> there<00:42:31.000> will<00:42:31.120> be<00:42:31.320> any<00:42:31.560> value" + }, + { + "start": 2551.91, + "duration": 0.0, + "text": "don't think there will be any value" + }, + { + "start": 2551.92, + "duration": 0.0, + "text": "don't think there will be any value created<00:42:32.440> from<00:42:32.600> interpretability<00:42:33.240> research," + }, + { + "start": 2553.71, + "duration": 0.0, + "text": "created from interpretability research," + }, + { + "start": 2553.72, + "duration": 0.0, + "text": "created from interpretability research, honestly,<00:42:34.320> other<00:42:34.560> than<00:42:34.720> maybe<00:42:35.720> um" + }, + { + "start": 2557.07, + "duration": 0.0, + "text": "honestly, other than maybe um" + }, + { + "start": 2557.08, + "duration": 0.0, + "text": "honestly, other than maybe um optimizing<00:42:37.640> the<00:42:37.760> training<00:42:38.360> process<00:42:38.800> somehow." + }, + { + "start": 2559.15, + "duration": 0.0, + "text": "optimizing the training process somehow." + }, + { + "start": 2559.16, + "duration": 0.0, + "text": "optimizing the training process somehow. I'm<00:42:39.280> honestly<00:42:39.520> not<00:42:39.680> really<00:42:39.880> sure.<00:42:40.160> I<00:42:40.240> think" + }, + { + "start": 2560.39, + "duration": 0.0, + "text": "I'm honestly not really sure. I think" + }, + { + "start": 2560.4, + "duration": 0.0, + "text": "I'm honestly not really sure. I think they're<00:42:40.520> mostly<00:42:40.760> doing<00:42:41.040> it<00:42:41.160> because<00:42:41.800> they" + }, + { + "start": 2561.91, + "duration": 0.0, + "text": "they're mostly doing it because they" + }, + { + "start": 2561.92, + "duration": 0.0, + "text": "they're mostly doing it because they have<00:42:42.160> to,<00:42:42.400> because<00:42:42.760> they<00:42:42.880> want<00:42:43.040> to<00:42:43.120> be<00:42:43.200> the" + }, + { + "start": 2563.31, + "duration": 0.0, + "text": "have to, because they want to be the" + }, + { + "start": 2563.32, + "duration": 0.0, + "text": "have to, because they want to be the safe<00:42:43.560> people<00:42:43.840> or<00:42:43.920> something." + }, + { + "start": 2564.95, + "duration": 0.0, + "text": "safe people or something." + }, + { + "start": 2564.96, + "duration": 0.0, + "text": "safe people or something. Um<00:42:45.080> and<00:42:45.280> it's<00:42:45.480> cool,<00:42:45.960> but<00:42:46.080> it<00:42:46.160> doesn't<00:42:46.360> seem" + }, + { + "start": 2566.91, + "duration": 0.0, + "text": "Um and it's cool, but it doesn't seem" + }, + { + "start": 2566.92, + "duration": 0.0, + "text": "Um and it's cool, but it doesn't seem valuable<00:42:47.440> to<00:42:47.560> me,<00:42:47.720> but<00:42:47.840> I<00:42:48.200> I'm<00:42:48.280> not<00:42:48.400> like<00:42:48.560> a<00:42:48.680> big" + }, + { + "start": 2569.39, + "duration": 0.0, + "text": "valuable to me, but I I'm not like a big" + }, + { + "start": 2569.4, + "duration": 0.0, + "text": "valuable to me, but I I'm not like a big deep<00:42:49.600> learning<00:42:49.840> research.<00:42:50.200> I'm<00:42:50.240> like<00:42:50.360> a" + }, + { + "start": 2570.39, + "duration": 0.0, + "text": "deep learning research. I'm like a" + }, + { + "start": 2570.4, + "duration": 0.0, + "text": "deep learning research. I'm like a hobbyist,<00:42:50.880> like<00:42:51.000> I<00:42:51.080> said<00:42:51.280> earlier,<00:42:51.560> so<00:42:51.640> I'm" + }, + { + "start": 2571.67, + "duration": 0.0, + "text": "hobbyist, like I said earlier, so I'm" + }, + { + "start": 2571.68, + "duration": 0.0, + "text": "hobbyist, like I said earlier, so I'm not<00:42:51.800> really<00:42:51.960> sure." + }, + { + "start": 2572.99, + "duration": 0.0, + "text": "not really sure." + }, + { + "start": 2573.0, + "duration": 0.0, + "text": "not really sure. Uh<00:42:53.200> follow-up<00:42:53.560> to<00:42:53.680> that,<00:42:54.320> is<00:42:54.520> it" + }, + { + "start": 2574.67, + "duration": 0.0, + "text": "Uh follow-up to that, is it" + }, + { + "start": 2574.68, + "duration": 0.0, + "text": "Uh follow-up to that, is it interpretability" + }, + { + "start": 2577.03, + "duration": 0.0, + "text": "interpretability" + }, + { + "start": 2577.04, + "duration": 0.0, + "text": "interpretability that<00:42:57.800> made<00:42:58.000> people<00:42:58.280> figure<00:42:58.520> out<00:42:58.640> you<00:42:58.760> could" + }, + { + "start": 2578.91, + "duration": 0.0, + "text": "that made people figure out you could" + }, + { + "start": 2578.92, + "duration": 0.0, + "text": "that made people figure out you could remove<00:42:59.320> the<00:42:59.440> safety<00:42:59.800> neuron?<00:43:00.680> You<00:43:00.760> can.<00:43:01.120> They" + }, + { + "start": 2581.19, + "duration": 0.0, + "text": "remove the safety neuron? You can. They" + }, + { + "start": 2581.2, + "duration": 0.0, + "text": "remove the safety neuron? You can. They already<00:43:01.400> figured<00:43:01.600> that<00:43:01.760> out.<00:43:01.840> It's<00:43:01.960> called" + }, + { + "start": 2582.07, + "duration": 0.0, + "text": "already figured that out. It's called" + }, + { + "start": 2582.08, + "duration": 0.0, + "text": "already figured that out. It's called ablation.<00:43:03.080> So,<00:43:03.240> any<00:43:03.440> single<00:43:03.760> open-source" + }, + { + "start": 2584.15, + "duration": 0.0, + "text": "ablation. So, any single open-source" + }, + { + "start": 2584.16, + "duration": 0.0, + "text": "ablation. So, any single open-source model,<00:43:04.560> you<00:43:04.640> can<00:43:04.760> basically<00:43:05.040> find<00:43:05.200> what's" + }, + { + "start": 2585.31, + "duration": 0.0, + "text": "model, you can basically find what's" + }, + { + "start": 2585.32, + "duration": 0.0, + "text": "model, you can basically find what's called<00:43:05.560> an<00:43:05.640> ablated<00:43:06.080> model,<00:43:06.400> where<00:43:06.480> they<00:43:06.600> just" + }, + { + "start": 2586.75, + "duration": 0.0, + "text": "called an ablated model, where they just" + }, + { + "start": 2586.76, + "duration": 0.0, + "text": "called an ablated model, where they just basically<00:43:07.640> subtract<00:43:08.160> the<00:43:08.280> safety<00:43:08.600> training," + }, + { + "start": 2588.95, + "duration": 0.0, + "text": "basically subtract the safety training," + }, + { + "start": 2588.96, + "duration": 0.0, + "text": "basically subtract the safety training, and<00:43:09.040> then<00:43:09.160> it<00:43:09.360> does<00:43:09.520> whatever.<00:43:10.200> But,<00:43:10.360> it's" + }, + { + "start": 2590.55, + "duration": 0.0, + "text": "and then it does whatever. But, it's" + }, + { + "start": 2590.56, + "duration": 0.0, + "text": "and then it does whatever. But, it's But,<00:43:10.800> you<00:43:10.920> told<00:43:11.120> me<00:43:11.200> it's<00:43:11.480> like<00:43:11.760> one<00:43:12.040> node" + }, + { + "start": 2592.75, + "duration": 0.0, + "text": "But, you told me it's like one node" + }, + { + "start": 2592.76, + "duration": 0.0, + "text": "But, you told me it's like one node that's<00:43:12.920> like<00:43:13.040> this.<00:43:13.280> Yeah,<00:43:13.560> they<00:43:14.160> I<00:43:14.200> don't" + }, + { + "start": 2594.35, + "duration": 0.0, + "text": "that's like this. Yeah, they I don't" + }, + { + "start": 2594.36, + "duration": 0.0, + "text": "that's like this. Yeah, they I don't remember<00:43:14.680> exactly<00:43:15.080> how<00:43:15.200> it<00:43:15.280> goes,<00:43:15.640> but<00:43:16.200> it's" + }, + { + "start": 2596.35, + "duration": 0.0, + "text": "remember exactly how it goes, but it's" + }, + { + "start": 2596.36, + "duration": 0.0, + "text": "remember exactly how it goes, but it's not<00:43:16.520> a<00:43:16.560> node,<00:43:16.840> but<00:43:16.960> it's<00:43:17.040> like<00:43:17.200> one<00:43:17.560> vector" + }, + { + "start": 2598.03, + "duration": 0.0, + "text": "not a node, but it's like one vector" + }, + { + "start": 2598.04, + "duration": 0.0, + "text": "not a node, but it's like one vector kind<00:43:18.240> of<00:43:18.360> represents<00:43:18.840> the<00:43:18.960> safety," + }, + { + "start": 2600.19, + "duration": 0.0, + "text": "kind of represents the safety," + }, + { + "start": 2600.2, + "duration": 0.0, + "text": "kind of represents the safety, uh<00:43:20.280> the<00:43:20.360> refusal.<00:43:20.920> Like,<00:43:21.320> when<00:43:21.480> whenever<00:43:21.720> an" + }, + { + "start": 2601.79, + "duration": 0.0, + "text": "uh the refusal. Like, when whenever an" + }, + { + "start": 2601.8, + "duration": 0.0, + "text": "uh the refusal. Like, when whenever an LLM<00:43:22.160> starts<00:43:22.400> to<00:43:22.480> refuse,<00:43:22.920> it<00:43:23.000> always<00:43:23.240> seems<00:43:23.560> to" + }, + { + "start": 2603.79, + "duration": 0.0, + "text": "LLM starts to refuse, it always seems to" + }, + { + "start": 2603.8, + "duration": 0.0, + "text": "LLM starts to refuse, it always seems to output<00:43:24.160> words<00:43:24.520> in<00:43:24.600> a<00:43:24.640> very,<00:43:25.120> you<00:43:25.200> know,<00:43:25.680> this" + }, + { + "start": 2605.91, + "duration": 0.0, + "text": "output words in a very, you know, this" + }, + { + "start": 2605.92, + "duration": 0.0, + "text": "output words in a very, you know, this kind<00:43:26.120> of<00:43:26.200> sounds<00:43:26.400> like<00:43:26.520> the<00:43:26.600> same<00:43:26.840> thing,<00:43:27.120> and" + }, + { + "start": 2607.55, + "duration": 0.0, + "text": "kind of sounds like the same thing, and" + }, + { + "start": 2607.56, + "duration": 0.0, + "text": "kind of sounds like the same thing, and it<00:43:27.680> basically<00:43:28.000> you<00:43:28.080> can<00:43:28.280> find<00:43:28.760> that<00:43:29.080> inside<00:43:29.480> of" + }, + { + "start": 2609.51, + "duration": 0.0, + "text": "it basically you can find that inside of" + }, + { + "start": 2609.52, + "duration": 0.0, + "text": "it basically you can find that inside of the<00:43:29.640> model<00:43:30.040> and<00:43:30.240> then<00:43:30.360> nullify<00:43:30.760> it<00:43:30.960> is<00:43:31.360> what" + }, + { + "start": 2611.55, + "duration": 0.0, + "text": "the model and then nullify it is what" + }, + { + "start": 2611.56, + "duration": 0.0, + "text": "the model and then nullify it is what ablation<00:43:31.960> is,<00:43:32.280> from<00:43:32.440> what<00:43:32.560> I<00:43:32.600> remember.<00:43:32.960> It's" + }, + { + "start": 2613.07, + "duration": 0.0, + "text": "ablation is, from what I remember. It's" + }, + { + "start": 2613.08, + "duration": 0.0, + "text": "ablation is, from what I remember. It's been<00:43:33.200> a<00:43:33.240> while." + }, + { + "start": 2616.96, + "duration": 0.0, + "text": "What's<00:43:37.120> up?" + }, + { + "start": 2618.11, + "duration": 0.0, + "text": "What's up?" + }, + { + "start": 2618.12, + "duration": 0.0, + "text": "What's up? Uh<00:43:38.480> as<00:43:38.640> you<00:43:38.800> said<00:43:39.160> that<00:43:39.760> AI<00:43:40.000> is<00:43:40.080> basically" + }, + { + "start": 2620.59, + "duration": 0.0, + "text": "Uh as you said that AI is basically" + }, + { + "start": 2620.6, + "duration": 0.0, + "text": "Uh as you said that AI is basically metaprogramming.<00:43:41.800> Yes.<00:43:42.400> So,<00:43:42.680> and<00:43:43.240> Julia<00:43:43.840> is" + }, + { + "start": 2624.07, + "duration": 0.0, + "text": "metaprogramming. Yes. So, and Julia is" + }, + { + "start": 2624.08, + "duration": 0.0, + "text": "metaprogramming. Yes. So, and Julia is very<00:43:44.320> famous<00:43:44.680> for<00:43:44.800> its<00:43:44.920> metaprogramming." + }, + { + "start": 2626.03, + "duration": 0.0, + "text": "very famous for its metaprogramming." + }, + { + "start": 2626.04, + "duration": 0.0, + "text": "very famous for its metaprogramming. Will<00:43:46.240> it<00:43:46.320> replace<00:43:47.080> Python?<00:43:47.440> Like,<00:43:47.680> will<00:43:47.840> we" + }, + { + "start": 2627.95, + "duration": 0.0, + "text": "Will it replace Python? Like, will we" + }, + { + "start": 2627.96, + "duration": 0.0, + "text": "Will it replace Python? Like, will we have<00:43:48.600> Julia<00:43:49.080> AI<00:43:49.680> or<00:43:49.760> like<00:43:50.000> So,<00:43:50.680> maybe" + }, + { + "start": 2630.99, + "duration": 0.0, + "text": "have Julia AI or like So, maybe" + }, + { + "start": 2631.0, + "duration": 0.0, + "text": "have Julia AI or like So, maybe metaprogramming<00:43:51.680> is<00:43:51.800> a<00:43:51.840> bad<00:43:52.480> thing,<00:43:52.800> cuz<00:43:53.280> it's" + }, + { + "start": 2633.43, + "duration": 0.0, + "text": "metaprogramming is a bad thing, cuz it's" + }, + { + "start": 2633.44, + "duration": 0.0, + "text": "metaprogramming is a bad thing, cuz it's not<00:43:53.600> really<00:43:53.760> comparable.<00:43:54.520> I<00:43:54.640> was<00:43:54.840> moon<00:43:55.280> I<00:43:55.360> was" + }, + { + "start": 2635.55, + "duration": 0.0, + "text": "not really comparable. I was moon I was" + }, + { + "start": 2635.56, + "duration": 0.0, + "text": "not really comparable. I was moon I was mostly<00:43:56.320> It's<00:43:56.600> like<00:43:56.920> metaprogramming<00:43:57.760> only<00:43:58.120> to" + }, + { + "start": 2638.23, + "duration": 0.0, + "text": "mostly It's like metaprogramming only to" + }, + { + "start": 2638.24, + "duration": 0.0, + "text": "mostly It's like metaprogramming only to the<00:43:58.360> extent<00:43:58.800> that<00:43:58.960> you're<00:43:59.440> producing<00:44:00.000> a" + }, + { + "start": 2640.03, + "duration": 0.0, + "text": "the extent that you're producing a" + }, + { + "start": 2640.04, + "duration": 0.0, + "text": "the extent that you're producing a program<00:44:00.520> with<00:44:00.720> a<00:44:00.760> program,<00:44:01.640> but<00:44:01.800> it's<00:44:01.960> not" + }, + { + "start": 2642.15, + "duration": 0.0, + "text": "program with a program, but it's not" + }, + { + "start": 2642.16, + "duration": 0.0, + "text": "program with a program, but it's not like<00:44:02.880> it's<00:44:03.160> automatic<00:44:03.680> metaprogramming." + }, + { + "start": 2644.47, + "duration": 0.0, + "text": "like it's automatic metaprogramming." + }, + { + "start": 2644.48, + "duration": 0.0, + "text": "like it's automatic metaprogramming. It's<00:44:04.640> not,<00:44:05.080> you<00:44:05.200> know,<00:44:05.480> mechanistic<00:44:06.120> symbolic" + }, + { + "start": 2646.63, + "duration": 0.0, + "text": "It's not, you know, mechanistic symbolic" + }, + { + "start": 2646.64, + "duration": 0.0, + "text": "It's not, you know, mechanistic symbolic metaprogramming,<00:44:07.360> where<00:44:07.480> you're<00:44:08.200> yeah.<00:44:08.680> So," + }, + { + "start": 2648.91, + "duration": 0.0, + "text": "metaprogramming, where you're yeah. So," + }, + { + "start": 2648.92, + "duration": 0.0, + "text": "metaprogramming, where you're yeah. So, it<00:44:09.000> would<00:44:09.080> not<00:44:09.200> be<00:44:09.320> a<00:44:09.360> good<00:44:09.520> application<00:44:10.120> for" + }, + { + "start": 2650.19, + "duration": 0.0, + "text": "it would not be a good application for" + }, + { + "start": 2650.2, + "duration": 0.0, + "text": "it would not be a good application for this<00:44:10.360> language,<00:44:10.840> next<00:44:11.080> one.<00:44:11.520> Huh?" + }, + { + "start": 2652.67, + "duration": 0.0, + "text": "this language, next one. Huh?" + }, + { + "start": 2652.68, + "duration": 0.0, + "text": "this language, next one. Huh? It's<00:44:12.920> not<00:44:13.080> a<00:44:13.120> good<00:44:13.240> application<00:44:13.920> for<00:44:14.520> jail." + }, + { + "start": 2654.99, + "duration": 0.0, + "text": "It's not a good application for jail." + }, + { + "start": 2655.0, + "duration": 0.0, + "text": "It's not a good application for jail. >> No.<00:44:15.680> No." + }, + { + "start": 2656.95, + "duration": 0.0, + "text": ">> No. No." + }, + { + "start": 2656.96, + "duration": 0.0, + "text": ">> No. No. No." + }, + { + "start": 2658.51, + "duration": 0.0, + "text": "No." + }, + { + "start": 2658.52, + "duration": 0.0, + "text": "No. Uh" + }, + { + "start": 2660.71, + "duration": 0.0, + "text": "Uh" + }, + { + "start": 2660.72, + "duration": 0.0, + "text": "Uh rough<00:44:21.320> To<00:44:21.760> bring<00:44:22.000> more<00:44:22.160> into<00:44:22.320> the<00:44:22.400> concrete" + }, + { + "start": 2662.79, + "duration": 0.0, + "text": "rough To bring more into the concrete" + }, + { + "start": 2662.8, + "duration": 0.0, + "text": "rough To bring more into the concrete realm,<00:44:23.040> you<00:44:23.120> mentioned<00:44:23.480> SLAM.<00:44:24.200> Yes.<00:44:24.720> It's" + }, + { + "start": 2664.79, + "duration": 0.0, + "text": "realm, you mentioned SLAM. Yes. It's" + }, + { + "start": 2664.8, + "duration": 0.0, + "text": "realm, you mentioned SLAM. Yes. It's been<00:44:25.000> a<00:44:25.040> decade<00:44:25.520> since<00:44:25.720> I've<00:44:25.800> looked<00:44:25.920> at<00:44:26.040> SLAM." + }, + { + "start": 2666.39, + "duration": 0.0, + "text": "been a decade since I've looked at SLAM." + }, + { + "start": 2666.4, + "duration": 0.0, + "text": "been a decade since I've looked at SLAM. Like,<00:44:26.640> is<00:44:26.760> deep<00:44:27.000> learning<00:44:27.320> the<00:44:27.440> current<00:44:27.720> state" + }, + { + "start": 2667.95, + "duration": 0.0, + "text": "Like, is deep learning the current state" + }, + { + "start": 2667.96, + "duration": 0.0, + "text": "Like, is deep learning the current state of<00:44:28.040> the<00:44:28.560> achieving<00:44:29.440> SLAM<00:44:29.920> algorithms,<00:44:30.520> or<00:44:30.640> are" + }, + { + "start": 2670.67, + "duration": 0.0, + "text": "of the achieving SLAM algorithms, or are" + }, + { + "start": 2670.68, + "duration": 0.0, + "text": "of the achieving SLAM algorithms, or are they<00:44:30.760> still" + }, + { + "start": 2671.15, + "duration": 0.0, + "text": "they still" + }, + { + "start": 2671.16, + "duration": 0.0, + "text": "they still >> I<00:44:31.360> I<00:44:31.480> just<00:44:31.800> recently<00:44:32.160> asked<00:44:32.360> a<00:44:32.400> SLAM" + }, + { + "start": 2672.71, + "duration": 0.0, + "text": ">> I I just recently asked a SLAM" + }, + { + "start": 2672.72, + "duration": 0.0, + "text": ">> I I just recently asked a SLAM researcher<00:44:33.120> that,<00:44:33.360> and<00:44:33.440> they<00:44:33.560> basically<00:44:34.000> said" + }, + { + "start": 2674.43, + "duration": 0.0, + "text": "researcher that, and they basically said" + }, + { + "start": 2674.44, + "duration": 0.0, + "text": "researcher that, and they basically said deep<00:44:34.680> learning<00:44:35.000> is<00:44:35.120> like<00:44:35.320> used<00:44:35.640> sometimes," + }, + { + "start": 2676.15, + "duration": 0.0, + "text": "deep learning is like used sometimes," + }, + { + "start": 2676.16, + "duration": 0.0, + "text": "deep learning is like used sometimes, but<00:44:36.280> it's<00:44:36.440> not<00:44:36.600> really<00:44:36.800> necessary<00:44:37.240> to<00:44:37.320> solve" + }, + { + "start": 2677.51, + "duration": 0.0, + "text": "but it's not really necessary to solve" + }, + { + "start": 2677.52, + "duration": 0.0, + "text": "but it's not really necessary to solve the<00:44:37.640> problem<00:44:38.000> though.<00:44:38.680> Um" + }, + { + "start": 2679.99, + "duration": 0.0, + "text": "the problem though. Um" + }, + { + "start": 2680.0, + "duration": 0.0, + "text": "the problem though. Um you<00:44:40.160> know,<00:44:40.240> they<00:44:40.360> found<00:44:40.600> good<00:44:40.760> heuristics<00:44:41.200> to" + }, + { + "start": 2681.31, + "duration": 0.0, + "text": "you know, they found good heuristics to" + }, + { + "start": 2681.32, + "duration": 0.0, + "text": "you know, they found good heuristics to find<00:44:41.520> the<00:44:41.600> key<00:44:41.800> points<00:44:42.040> that<00:44:42.160> are<00:44:42.240> trackable." + }, + { + "start": 2683.27, + "duration": 0.0, + "text": "find the key points that are trackable." + }, + { + "start": 2683.28, + "duration": 0.0, + "text": "find the key points that are trackable. Um<00:44:43.480> and<00:44:43.560> like<00:44:43.720> neural<00:44:44.200> I<00:44:44.240> think<00:44:44.440> that<00:44:44.600> neural" + }, + { + "start": 2684.83, + "duration": 0.0, + "text": "Um and like neural I think that neural" + }, + { + "start": 2684.84, + "duration": 0.0, + "text": "Um and like neural I think that neural SLAM<00:44:45.200> seems<00:44:45.400> like<00:44:45.560> it's<00:44:45.680> something<00:44:45.880> they've" + }, + { + "start": 2686.47, + "duration": 0.0, + "text": "SLAM seems like it's something they've" + }, + { + "start": 2686.48, + "duration": 0.0, + "text": "SLAM seems like it's something they've kind<00:44:46.640> of<00:44:46.720> been<00:44:46.800> like<00:44:47.000> rumor<00:44:47.760> you<00:44:47.840> know,<00:44:47.960> to" + }, + { + "start": 2688.03, + "duration": 0.0, + "text": "kind of been like rumor you know, to" + }, + { + "start": 2688.04, + "duration": 0.0, + "text": "kind of been like rumor you know, to kind<00:44:48.200> of<00:44:48.240> take<00:44:48.400> over<00:44:48.600> a<00:44:48.640> state<00:44:48.840> of<00:44:48.880> the<00:44:49.000> art," + }, + { + "start": 2689.11, + "duration": 0.0, + "text": "kind of take over a state of the art," + }, + { + "start": 2689.12, + "duration": 0.0, + "text": "kind of take over a state of the art, but<00:44:49.240> like<00:44:49.640> you<00:44:49.720> know,<00:44:49.800> the<00:44:49.880> Quest<00:44:50.320> doesn't<00:44:50.680> do" + }, + { + "start": 2690.75, + "duration": 0.0, + "text": "but like you know, the Quest doesn't do" + }, + { + "start": 2690.76, + "duration": 0.0, + "text": "but like you know, the Quest doesn't do any<00:44:50.920> neural<00:44:51.120> SLAM<00:44:51.680> as<00:44:51.840> far<00:44:51.960> as<00:44:52.080> I<00:44:52.120> know." + }, + { + "start": 2693.39, + "duration": 0.0, + "text": "any neural SLAM as far as I know." + }, + { + "start": 2693.4, + "duration": 0.0, + "text": "any neural SLAM as far as I know. Um" + }, + { + "start": 2695.39, + "duration": 0.0, + "text": "Um" + }, + { + "start": 2695.4, + "duration": 0.0, + "text": "Um You<00:44:55.640> mentioned<00:44:56.360> using<00:44:57.440> uh<00:44:57.680> AI<00:44:57.920> models<00:44:58.320> for" + }, + { + "start": 2698.91, + "duration": 0.0, + "text": "You mentioned using uh AI models for" + }, + { + "start": 2698.92, + "duration": 0.0, + "text": "You mentioned using uh AI models for balancing<00:44:59.600> games.<00:45:00.360> Yes." + }, + { + "start": 2701.19, + "duration": 0.0, + "text": "balancing games. Yes." + }, + { + "start": 2701.2, + "duration": 0.0, + "text": "balancing games. Yes. >> and" + }, + { + "start": 2702.71, + "duration": 0.0, + "text": ">> and" + }, + { + "start": 2702.72, + "duration": 0.0, + "text": ">> and let's<00:45:03.040> say<00:45:03.240> you<00:45:03.360> then<00:45:03.600> train<00:45:03.840> a<00:45:03.920> model<00:45:04.240> to<00:45:04.360> play" + }, + { + "start": 2704.51, + "duration": 0.0, + "text": "let's say you then train a model to play" + }, + { + "start": 2704.52, + "duration": 0.0, + "text": "let's say you then train a model to play StarCraft<00:45:05.000> II,<00:45:05.040> and<00:45:05.120> you<00:45:05.320> you<00:45:05.360> finally<00:45:05.720> do" + }, + { + "start": 2705.91, + "duration": 0.0, + "text": "StarCraft II, and you you finally do" + }, + { + "start": 2705.92, + "duration": 0.0, + "text": "StarCraft II, and you you finally do definitively<00:45:06.520> prove<00:45:06.840> that<00:45:07.080> Protoss<00:45:07.640> should" + }, + { + "start": 2707.79, + "duration": 0.0, + "text": "definitively prove that Protoss should" + }, + { + "start": 2707.8, + "duration": 0.0, + "text": "definitively prove that Protoss should be<00:45:07.920> nerfed." + }, + { + "start": 2708.99, + "duration": 0.0, + "text": "be nerfed." + }, + { + "start": 2709.0, + "duration": 0.0, + "text": "be nerfed. Uh" + }, + { + "start": 2709.95, + "duration": 0.0, + "text": "Uh" + }, + { + "start": 2709.96, + "duration": 0.0, + "text": "Uh yes." + }, + { + "start": 2712.11, + "duration": 0.0, + "text": "yes." + }, + { + "start": 2712.12, + "duration": 0.0, + "text": "yes. God,<00:45:12.480> go<00:45:12.640> ahead<00:45:12.800> and<00:45:12.880> start<00:45:13.120> recording." + }, + { + "start": 2714.15, + "duration": 0.0, + "text": "God, go ahead and start recording." + }, + { + "start": 2714.16, + "duration": 0.0, + "text": "God, go ahead and start recording. Uh<00:45:14.520> once<00:45:14.840> you<00:45:15.000> nerf<00:45:15.280> Protoss,<00:45:16.280> do<00:45:16.440> you<00:45:16.520> then<00:45:16.800> go" + }, + { + "start": 2716.83, + "duration": 0.0, + "text": "Uh once you nerf Protoss, do you then go" + }, + { + "start": 2716.84, + "duration": 0.0, + "text": "Uh once you nerf Protoss, do you then go like,<00:45:17.040> \"Okay,<00:45:17.360> cool.<00:45:17.720> Now<00:45:17.880> we<00:45:18.000> do<00:45:18.200> a<00:45:18.240> $3" + }, + { + "start": 2718.55, + "duration": 0.0, + "text": "like, \"Okay, cool. Now we do a $3" + }, + { + "start": 2718.56, + "duration": 0.0, + "text": "like, \"Okay, cool. Now we do a $3 million<00:45:19.240> run<00:45:19.960> again?\"<00:45:20.680> I<00:45:20.800> don't<00:45:20.960> think<00:45:21.120> it" + }, + { + "start": 2721.15, + "duration": 0.0, + "text": "million run again?\" I don't think it" + }, + { + "start": 2721.16, + "duration": 0.0, + "text": "million run again?\" I don't think it would<00:45:21.240> be<00:45:21.360> that<00:45:21.520> expensive<00:45:22.440> to<00:45:22.600> do<00:45:22.800> it<00:45:22.960> on<00:45:23.080> a" + }, + { + "start": 2723.15, + "duration": 0.0, + "text": "would be that expensive to do it on a" + }, + { + "start": 2723.16, + "duration": 0.0, + "text": "would be that expensive to do it on a game.<00:45:23.520> It's<00:45:24.120> the<00:45:24.320> cost<00:45:24.800> of<00:45:25.120> AI<00:45:25.400> right<00:45:25.600> now<00:45:25.840> is" + }, + { + "start": 2725.95, + "duration": 0.0, + "text": "game. It's the cost of AI right now is" + }, + { + "start": 2725.96, + "duration": 0.0, + "text": "game. It's the cost of AI right now is because<00:45:26.800> to<00:45:26.960> make<00:45:27.240> these<00:45:27.480> large<00:45:27.720> language" + }, + { + "start": 2728.03, + "duration": 0.0, + "text": "because to make these large language" + }, + { + "start": 2728.04, + "duration": 0.0, + "text": "because to make these large language models<00:45:29.000> learn,<00:45:29.440> you<00:45:29.520> know,<00:45:29.720> reasoning,<00:45:30.720> you" + }, + { + "start": 2730.95, + "duration": 0.0, + "text": "models learn, you know, reasoning, you" + }, + { + "start": 2730.96, + "duration": 0.0, + "text": "models learn, you know, reasoning, you have<00:45:31.120> to<00:45:31.320> basically<00:45:31.840> shuffle<00:45:32.200> the<00:45:32.320> entire" + }, + { + "start": 2732.83, + "duration": 0.0, + "text": "have to basically shuffle the entire" + }, + { + "start": 2732.84, + "duration": 0.0, + "text": "have to basically shuffle the entire internet<00:45:33.280> in<00:45:33.400> and<00:45:33.560> out<00:45:33.720> of<00:45:34.200> um<00:45:34.720> multiple" + }, + { + "start": 2735.11, + "duration": 0.0, + "text": "internet in and out of um multiple" + }, + { + "start": 2735.12, + "duration": 0.0, + "text": "internet in and out of um multiple accumulate<00:45:35.560> ads.<00:45:35.840> It's<00:45:35.960> really<00:45:36.200> not<00:45:36.360> so<00:45:36.440> much" + }, + { + "start": 2736.59, + "duration": 0.0, + "text": "accumulate ads. It's really not so much" + }, + { + "start": 2736.6, + "duration": 0.0, + "text": "accumulate ads. It's really not so much data<00:45:36.760> they<00:45:36.840> have<00:45:36.960> to<00:45:37.080> process.<00:45:37.920> For<00:45:38.000> something" + }, + { + "start": 2738.19, + "duration": 0.0, + "text": "data they have to process. For something" + }, + { + "start": 2738.2, + "duration": 0.0, + "text": "data they have to process. For something like<00:45:38.280> StarCraft<00:45:38.640> II<00:45:38.720> or<00:45:38.840> AlphaZero,<00:45:39.400> it's" + }, + { + "start": 2739.55, + "duration": 0.0, + "text": "like StarCraft II or AlphaZero, it's" + }, + { + "start": 2739.56, + "duration": 0.0, + "text": "like StarCraft II or AlphaZero, it's really<00:45:39.840> not<00:45:40.120> on<00:45:40.240> that<00:45:40.400> scale<00:45:40.640> of<00:45:40.760> cost.<00:45:41.640> Um<00:45:42.320> I" + }, + { + "start": 2742.39, + "duration": 0.0, + "text": "really not on that scale of cost. Um I" + }, + { + "start": 2742.4, + "duration": 0.0, + "text": "really not on that scale of cost. Um I could<00:45:42.560> easily<00:45:42.840> see<00:45:43.000> many<00:45:43.320> many<00:45:43.520> many<00:45:43.760> games" + }, + { + "start": 2744.11, + "duration": 0.0, + "text": "could easily see many many many games" + }, + { + "start": 2744.12, + "duration": 0.0, + "text": "could easily see many many many games being<00:45:44.280> able<00:45:44.400> to<00:45:44.480> do<00:45:44.640> this,<00:45:44.960> and<00:45:45.840> for<00:45:46.000> simpler" + }, + { + "start": 2746.27, + "duration": 0.0, + "text": "being able to do this, and for simpler" + }, + { + "start": 2746.28, + "duration": 0.0, + "text": "being able to do this, and for simpler games,<00:45:46.480> you<00:45:46.560> could<00:45:46.640> probably<00:45:46.800> do<00:45:46.880> it<00:45:46.960> on<00:45:47.040> your" + }, + { + "start": 2747.15, + "duration": 0.0, + "text": "games, you could probably do it on your" + }, + { + "start": 2747.16, + "duration": 0.0, + "text": "games, you could probably do it on your laptop<00:45:47.520> even<00:45:47.720> maybe.<00:45:48.400> Um" + }, + { + "start": 2749.59, + "duration": 0.0, + "text": "laptop even maybe. Um" + }, + { + "start": 2749.6, + "duration": 0.0, + "text": "laptop even maybe. Um But<00:45:49.880> yeah,<00:45:50.160> like<00:45:50.320> I<00:45:50.360> think<00:45:50.560> that<00:45:50.880> for" + }, + { + "start": 2751.43, + "duration": 0.0, + "text": "But yeah, like I think that for" + }, + { + "start": 2751.44, + "duration": 0.0, + "text": "But yeah, like I think that for something<00:45:51.680> like<00:45:51.840> StarCraft<00:45:52.320> II,<00:45:52.480> it's<00:45:52.600> like<00:45:52.760> I" + }, + { + "start": 2752.79, + "duration": 0.0, + "text": "something like StarCraft II, it's like I" + }, + { + "start": 2752.8, + "duration": 0.0, + "text": "something like StarCraft II, it's like I could<00:45:53.000> see<00:45:53.160> them<00:45:53.400> doing<00:45:53.720> these<00:45:53.920> kind<00:45:54.040> of" + }, + { + "start": 2754.11, + "duration": 0.0, + "text": "could see them doing these kind of" + }, + { + "start": 2754.12, + "duration": 0.0, + "text": "could see them doing these kind of studies<00:45:54.600> if<00:45:54.680> they<00:45:54.800> have<00:45:55.000> the<00:45:55.360> the<00:45:55.440> cap<00:45:55.680> backs" + }, + { + "start": 2755.83, + "duration": 0.0, + "text": "studies if they have the the cap backs" + }, + { + "start": 2755.84, + "duration": 0.0, + "text": "studies if they have the the cap backs and<00:45:55.920> they<00:45:55.960> believe<00:45:56.160> it's" + }, + { + "start": 2757.03, + "duration": 0.0, + "text": "and they believe it's" + }, + { + "start": 2757.04, + "duration": 0.0, + "text": "and they believe it's you<00:45:57.160> know,<00:45:57.280> it's<00:45:57.640> it's<00:45:57.840> valuable<00:45:58.240> for<00:45:58.360> the" + }, + { + "start": 2758.55, + "duration": 0.0, + "text": "you know, it's it's valuable for the" + }, + { + "start": 2758.56, + "duration": 0.0, + "text": "you know, it's it's valuable for the players<00:45:58.960> to<00:45:59.760> have<00:46:00.000> have<00:46:00.120> truly<00:46:00.480> be<00:46:00.600> officially" + }, + { + "start": 2761.11, + "duration": 0.0, + "text": "players to have have truly be officially" + }, + { + "start": 2761.12, + "duration": 0.0, + "text": "players to have have truly be officially like<00:46:01.880> even<00:46:02.240> the<00:46:02.600> genius<00:46:03.000> master<00:46:03.480> AI" + }, + { + "start": 2764.47, + "duration": 0.0, + "text": "like even the genius master AI" + }, + { + "start": 2764.48, + "duration": 0.0, + "text": "like even the genius master AI uh<00:46:04.600> players<00:46:05.000> of<00:46:05.080> this<00:46:05.200> game<00:46:05.560> it's" + }, + { + "start": 2766.51, + "duration": 0.0, + "text": "uh players of this game it's" + }, + { + "start": 2766.52, + "duration": 0.0, + "text": "uh players of this game it's even<00:46:06.880> right<00:46:07.160> that<00:46:07.240> might<00:46:07.360> be<00:46:07.440> a<00:46:07.480> good<00:46:07.640> thing<00:46:07.760> to" + }, + { + "start": 2767.87, + "duration": 0.0, + "text": "even right that might be a good thing to" + }, + { + "start": 2767.88, + "duration": 0.0, + "text": "even right that might be a good thing to prove." + }, + { + "start": 2770.03, + "duration": 0.0, + "text": "prove." + }, + { + "start": 2770.04, + "duration": 0.0, + "text": "prove. Uh<00:46:10.560> yeah.<00:46:11.200> Uh<00:46:11.320> would<00:46:11.480> you<00:46:11.560> say<00:46:11.800> that<00:46:12.000> like" + }, + { + "start": 2772.19, + "duration": 0.0, + "text": "Uh yeah. Uh would you say that like" + }, + { + "start": 2772.2, + "duration": 0.0, + "text": "Uh yeah. Uh would you say that like training<00:46:12.680> is<00:46:12.840> where<00:46:13.040> like<00:46:13.240> mostly<00:46:13.560> cost<00:46:13.960> is<00:46:14.280> or" + }, + { + "start": 2775.43, + "duration": 0.0, + "text": "training is where like mostly cost is or" + }, + { + "start": 2775.44, + "duration": 0.0, + "text": "training is where like mostly cost is or Um<00:46:16.360> actually<00:46:17.320> well<00:46:17.640> opening<00:46:17.880> I<00:46:17.960> says<00:46:18.200> that" + }, + { + "start": 2778.31, + "duration": 0.0, + "text": "Um actually well opening I says that" + }, + { + "start": 2778.32, + "duration": 0.0, + "text": "Um actually well opening I says that inference<00:46:18.640> is<00:46:18.760> their<00:46:18.880> biggest<00:46:19.280> cost<00:46:19.600> right" + }, + { + "start": 2779.75, + "duration": 0.0, + "text": "inference is their biggest cost right" + }, + { + "start": 2779.76, + "duration": 0.0, + "text": "inference is their biggest cost right now" + }, + { + "start": 2780.79, + "duration": 0.0, + "text": "now" + }, + { + "start": 2780.8, + "duration": 0.0, + "text": "now but<00:46:21.400> there's<00:46:21.600> a<00:46:21.680> massive<00:46:22.640> push<00:46:22.920> to<00:46:23.040> make<00:46:23.560> you" + }, + { + "start": 2783.63, + "duration": 0.0, + "text": "but there's a massive push to make you" + }, + { + "start": 2783.64, + "duration": 0.0, + "text": "but there's a massive push to make you know<00:46:24.160> on<00:46:24.480> on<00:46:24.640> device<00:46:25.080> inference<00:46:25.520> chips<00:46:26.440> so<00:46:26.560> I" + }, + { + "start": 2786.59, + "duration": 0.0, + "text": "know on on device inference chips so I" + }, + { + "start": 2786.6, + "duration": 0.0, + "text": "know on on device inference chips so I think<00:46:26.800> that's<00:46:26.960> kind<00:46:27.120> of<00:46:27.200> temporary<00:46:27.840> and<00:46:28.000> it's" + }, + { + "start": 2788.27, + "duration": 0.0, + "text": "think that's kind of temporary and it's" + }, + { + "start": 2788.28, + "duration": 0.0, + "text": "think that's kind of temporary and it's just<00:46:28.480> because<00:46:28.800> like<00:46:28.960> consumer<00:46:29.720> my" + }, + { + "start": 2789.83, + "duration": 0.0, + "text": "just because like consumer my" + }, + { + "start": 2789.84, + "duration": 0.0, + "text": "just because like consumer my understanding<00:46:30.440> is<00:46:31.200> I<00:46:31.240> don't<00:46:31.360> know<00:46:31.400> if<00:46:31.520> this<00:46:31.640> is" + }, + { + "start": 2791.71, + "duration": 0.0, + "text": "understanding is I don't know if this is" + }, + { + "start": 2791.72, + "duration": 0.0, + "text": "understanding is I don't know if this is exactly<00:46:32.080> true<00:46:32.240> but<00:46:32.360> I've<00:46:32.440> seen<00:46:32.600> this<00:46:32.720> floating" + }, + { + "start": 2792.95, + "duration": 0.0, + "text": "exactly true but I've seen this floating" + }, + { + "start": 2792.96, + "duration": 0.0, + "text": "exactly true but I've seen this floating around<00:46:33.360> that<00:46:33.920> Nvidia<00:46:34.320> cards<00:46:34.680> don't<00:46:34.840> have" + }, + { + "start": 2794.95, + "duration": 0.0, + "text": "around that Nvidia cards don't have" + }, + { + "start": 2794.96, + "duration": 0.0, + "text": "around that Nvidia cards don't have enough<00:46:35.160> VRM<00:46:35.480> to<00:46:35.600> run<00:46:35.720> these<00:46:35.880> models<00:46:36.240> simply" + }, + { + "start": 2796.47, + "duration": 0.0, + "text": "enough VRM to run these models simply" + }, + { + "start": 2796.48, + "duration": 0.0, + "text": "enough VRM to run these models simply because<00:46:36.840> it's<00:46:37.000> not<00:46:37.120> something<00:46:37.360> consumers" + }, + { + "start": 2797.83, + "duration": 0.0, + "text": "because it's not something consumers" + }, + { + "start": 2797.84, + "duration": 0.0, + "text": "because it's not something consumers wanted<00:46:38.160> and<00:46:38.240> they<00:46:38.320> kind<00:46:38.480> of<00:46:38.560> want<00:46:38.680> to<00:46:38.720> nickel" + }, + { + "start": 2798.91, + "duration": 0.0, + "text": "wanted and they kind of want to nickel" + }, + { + "start": 2798.92, + "duration": 0.0, + "text": "wanted and they kind of want to nickel and<00:46:39.000> dime<00:46:39.200> you<00:46:39.320> for<00:46:39.440> VRAM" + }, + { + "start": 2800.87, + "duration": 0.0, + "text": "and dime you for VRAM" + }, + { + "start": 2800.88, + "duration": 0.0, + "text": "and dime you for VRAM and<00:46:41.040> so<00:46:41.160> this<00:46:41.320> seems<00:46:41.480> like<00:46:41.600> the<00:46:41.680> kind<00:46:41.840> of<00:46:41.920> thing" + }, + { + "start": 2802.11, + "duration": 0.0, + "text": "and so this seems like the kind of thing" + }, + { + "start": 2802.12, + "duration": 0.0, + "text": "and so this seems like the kind of thing where<00:46:42.400> you<00:46:42.480> know<00:46:42.640> if<00:46:42.840> it's<00:46:43.000> a<00:46:43.080> healthy<00:46:43.320> market" + }, + { + "start": 2803.75, + "duration": 0.0, + "text": "where you know if it's a healthy market" + }, + { + "start": 2803.76, + "duration": 0.0, + "text": "where you know if it's a healthy market and<00:46:44.000> you<00:46:44.040> know<00:46:44.160> it's<00:46:44.280> not<00:46:44.400> there's<00:46:44.520> not<00:46:44.640> like" + }, + { + "start": 2804.75, + "duration": 0.0, + "text": "and you know it's not there's not like" + }, + { + "start": 2804.76, + "duration": 0.0, + "text": "and you know it's not there's not like an<00:46:44.840> Nvidia<00:46:45.160> monopoly<00:46:45.600> or<00:46:45.680> whatever" + }, + { + "start": 2807.23, + "duration": 0.0, + "text": "an Nvidia monopoly or whatever" + }, + { + "start": 2807.24, + "duration": 0.0, + "text": "an Nvidia monopoly or whatever which<00:46:47.400> there<00:46:47.560> is<00:46:47.640> I'm<00:46:47.760> joking<00:46:48.080> it's<00:46:48.280> not<00:46:48.440> going" + }, + { + "start": 2808.55, + "duration": 0.0, + "text": "which there is I'm joking it's not going" + }, + { + "start": 2808.56, + "duration": 0.0, + "text": "which there is I'm joking it's not going to<00:46:48.600> get<00:46:48.800> better<00:46:49.040> but" + }, + { + "start": 2810.07, + "duration": 0.0, + "text": "to get better but" + }, + { + "start": 2810.08, + "duration": 0.0, + "text": "to get better but um" + }, + { + "start": 2812.03, + "duration": 0.0, + "text": "um" + }, + { + "start": 2812.04, + "duration": 0.0, + "text": "um ideally<00:46:52.440> I<00:46:52.480> think<00:46:52.680> it's<00:46:52.920> totally<00:46:53.240> feasible" + }, + { + "start": 2813.55, + "duration": 0.0, + "text": "ideally I think it's totally feasible" + }, + { + "start": 2813.56, + "duration": 0.0, + "text": "ideally I think it's totally feasible that<00:46:53.680> these<00:46:54.000> the<00:46:54.120> inference<00:46:54.440> cost<00:46:54.680> is<00:46:54.800> zero" + }, + { + "start": 2815.03, + "duration": 0.0, + "text": "that these the inference cost is zero" + }, + { + "start": 2815.04, + "duration": 0.0, + "text": "that these the inference cost is zero for<00:46:55.160> the<00:46:55.240> people<00:46:55.520> who<00:46:55.640> are" + }, + { + "start": 2816.55, + "duration": 0.0, + "text": "for the people who are" + }, + { + "start": 2816.56, + "duration": 0.0, + "text": "for the people who are making<00:46:56.800> these<00:46:56.920> models<00:46:57.280> it<00:46:57.560> it<00:46:57.680> should<00:46:57.800> be" + }, + { + "start": 2817.91, + "duration": 0.0, + "text": "making these models it it should be" + }, + { + "start": 2817.92, + "duration": 0.0, + "text": "making these models it it should be zero." + }, + { + "start": 2819.47, + "duration": 0.0, + "text": "zero." + }, + { + "start": 2819.48, + "duration": 0.0, + "text": "zero. Uh<00:46:59.640> yeah.<00:47:00.280> Yeah<00:47:00.960> in<00:47:01.080> your<00:47:01.280> little<00:47:01.520> experiment" + }, + { + "start": 2821.99, + "duration": 0.0, + "text": "Uh yeah. Yeah in your little experiment" + }, + { + "start": 2822.0, + "duration": 0.0, + "text": "Uh yeah. Yeah in your little experiment game<00:47:02.880> how<00:47:03.000> did<00:47:03.200> you<00:47:03.320> translate<00:47:03.840> basically<00:47:04.240> the" + }, + { + "start": 2824.39, + "duration": 0.0, + "text": "game how did you translate basically the" + }, + { + "start": 2824.4, + "duration": 0.0, + "text": "game how did you translate basically the LM<00:47:04.840> output<00:47:05.240> to<00:47:05.400> actions<00:47:06.000> in<00:47:06.200> your<00:47:06.400> world?<00:47:06.920> So" + }, + { + "start": 2826.99, + "duration": 0.0, + "text": "LM output to actions in your world? So" + }, + { + "start": 2827.0, + "duration": 0.0, + "text": "LM output to actions in your world? So basically<00:47:07.520> you<00:47:07.600> can<00:47:07.760> view<00:47:08.240> putting<00:47:08.520> LMs<00:47:08.880> in<00:47:09.000> a" + }, + { + "start": 2829.03, + "duration": 0.0, + "text": "basically you can view putting LMs in a" + }, + { + "start": 2829.04, + "duration": 0.0, + "text": "basically you can view putting LMs in a video<00:47:09.280> game<00:47:09.720> as<00:47:10.280> so<00:47:10.360> you<00:47:10.400> have<00:47:10.480> your<00:47:10.600> game" + }, + { + "start": 2830.87, + "duration": 0.0, + "text": "video game as so you have your game" + }, + { + "start": 2830.88, + "duration": 0.0, + "text": "video game as so you have your game state<00:47:11.360> and<00:47:11.480> then<00:47:11.640> there's<00:47:11.840> like<00:47:12.640> the<00:47:13.280> um" + }, + { + "start": 2834.31, + "duration": 0.0, + "text": "state and then there's like the um" + }, + { + "start": 2834.32, + "duration": 0.0, + "text": "state and then there's like the um physics" + }, + { + "start": 2835.39, + "duration": 0.0, + "text": "physics" + }, + { + "start": 2835.4, + "duration": 0.0, + "text": "physics simulation<00:47:16.440> kind<00:47:16.680> of<00:47:16.800> like<00:47:17.200> you<00:47:17.320> know<00:47:17.440> human" + }, + { + "start": 2837.95, + "duration": 0.0, + "text": "simulation kind of like you know human" + }, + { + "start": 2837.96, + "duration": 0.0, + "text": "simulation kind of like you know human in<00:47:18.080> the<00:47:18.200> loop<00:47:18.920> like<00:47:19.080> you're<00:47:19.480> inputting" + }, + { + "start": 2840.03, + "duration": 0.0, + "text": "in the loop like you're inputting" + }, + { + "start": 2840.04, + "duration": 0.0, + "text": "in the loop like you're inputting actions<00:47:20.480> and<00:47:20.840> state<00:47:21.160> 60<00:47:21.480> hertz<00:47:21.840> or<00:47:21.920> whatever" + }, + { + "start": 2842.79, + "duration": 0.0, + "text": "actions and state 60 hertz or whatever" + }, + { + "start": 2842.8, + "duration": 0.0, + "text": "actions and state 60 hertz or whatever and<00:47:23.080> there's<00:47:23.240> a<00:47:23.280> text<00:47:23.600> adventure<00:47:24.200> way<00:47:24.400> to" + }, + { + "start": 2844.55, + "duration": 0.0, + "text": "and there's a text adventure way to" + }, + { + "start": 2844.56, + "duration": 0.0, + "text": "and there's a text adventure way to interact<00:47:24.840> with<00:47:24.960> the<00:47:25.000> game<00:47:25.240> simultaneously" + }, + { + "start": 2846.03, + "duration": 0.0, + "text": "interact with the game simultaneously" + }, + { + "start": 2846.04, + "duration": 0.0, + "text": "interact with the game simultaneously right<00:47:26.680> and<00:47:26.800> a<00:47:26.840> good<00:47:27.080> way<00:47:27.280> to<00:47:27.480> check<00:47:27.720> that<00:47:27.880> like" + }, + { + "start": 2848.03, + "duration": 0.0, + "text": "right and a good way to check that like" + }, + { + "start": 2848.04, + "duration": 0.0, + "text": "right and a good way to check that like the<00:47:28.160> game<00:47:28.880> was<00:47:29.040> properly<00:47:30.000> you<00:47:30.040> know" + }, + { + "start": 2850.11, + "duration": 0.0, + "text": "the game was properly you know" + }, + { + "start": 2850.12, + "duration": 0.0, + "text": "the game was properly you know presenting<00:47:30.520> itself<00:47:30.720> to<00:47:30.800> LMs<00:47:31.160> is<00:47:31.280> if<00:47:31.400> you<00:47:31.480> could" + }, + { + "start": 2851.63, + "duration": 0.0, + "text": "presenting itself to LMs is if you could" + }, + { + "start": 2851.64, + "duration": 0.0, + "text": "presenting itself to LMs is if you could flip<00:47:31.920> it<00:47:32.360> and<00:47:32.480> as<00:47:32.600> a<00:47:32.640> human<00:47:33.000> I<00:47:33.040> could<00:47:33.240> play<00:47:33.880> the" + }, + { + "start": 2853.99, + "duration": 0.0, + "text": "flip it and as a human I could play the" + }, + { + "start": 2854.0, + "duration": 0.0, + "text": "flip it and as a human I could play the text<00:47:34.240> adventure<00:47:34.520> version<00:47:34.840> of<00:47:34.920> it<00:47:35.040> right" + }, + { + "start": 2855.55, + "duration": 0.0, + "text": "text adventure version of it right" + }, + { + "start": 2855.56, + "duration": 0.0, + "text": "text adventure version of it right that's<00:47:35.760> how<00:47:35.800> you<00:47:35.920> would<00:47:36.000> know<00:47:36.120> that<00:47:36.240> it's" + }, + { + "start": 2856.35, + "duration": 0.0, + "text": "that's how you would know that it's" + }, + { + "start": 2856.36, + "duration": 0.0, + "text": "that's how you would know that it's being<00:47:36.520> represented<00:47:37.000> well<00:47:37.160> in<00:47:37.240> text<00:47:37.560> and" + }, + { + "start": 2857.63, + "duration": 0.0, + "text": "being represented well in text and" + }, + { + "start": 2857.64, + "duration": 0.0, + "text": "being represented well in text and that's<00:47:37.800> basically<00:47:38.120> what<00:47:38.240> I<00:47:38.280> did<00:47:38.520> is<00:47:38.600> like<00:47:38.720> a" + }, + { + "start": 2858.79, + "duration": 0.0, + "text": "that's basically what I did is like a" + }, + { + "start": 2858.8, + "duration": 0.0, + "text": "that's basically what I did is like a structured<00:47:39.200> text<00:47:39.400> adventure<00:47:40.080> presented<00:47:40.520> to" + }, + { + "start": 2860.55, + "duration": 0.0, + "text": "structured text adventure presented to" + }, + { + "start": 2860.56, + "duration": 0.0, + "text": "structured text adventure presented to the<00:47:40.680> LMs<00:47:41.680> um" + }, + { + "start": 2862.27, + "duration": 0.0, + "text": "the LMs um" + }, + { + "start": 2862.28, + "duration": 0.0, + "text": "the LMs um describing<00:47:42.640> what's<00:47:42.760> around<00:47:43.000> it<00:47:43.080> describing" + }, + { + "start": 2863.39, + "duration": 0.0, + "text": "describing what's around it describing" + }, + { + "start": 2863.4, + "duration": 0.0, + "text": "describing what's around it describing what<00:47:43.480> you're<00:47:43.600> wearing<00:47:43.840> whatever<00:47:44.120> yeah." + }, + { + "start": 2865.91, + "duration": 0.0, + "text": "what you're wearing whatever yeah." + }, + { + "start": 2865.92, + "duration": 0.0, + "text": "what you're wearing whatever yeah. Yeah.<00:47:46.560> So<00:47:47.200> right<00:47:47.400> now<00:47:47.600> when<00:47:47.800> I<00:47:47.920> use<00:47:48.280> some" + }, + { + "start": 2868.51, + "duration": 0.0, + "text": "Yeah. So right now when I use some" + }, + { + "start": 2868.52, + "duration": 0.0, + "text": "Yeah. So right now when I use some programs<00:47:49.200> I<00:47:49.760> go<00:47:50.120> oh<00:47:50.360> I<00:47:50.440> want<00:47:50.640> to<00:47:50.720> do<00:47:50.920> kind<00:47:51.160> of" + }, + { + "start": 2871.27, + "duration": 0.0, + "text": "programs I go oh I want to do kind of" + }, + { + "start": 2871.28, + "duration": 0.0, + "text": "programs I go oh I want to do kind of this<00:47:51.520> thing<00:47:52.000> and<00:47:52.120> it<00:47:52.240> used<00:47:52.520> to<00:47:52.600> be<00:47:52.880> that<00:47:53.080> I<00:47:53.200> was" + }, + { + "start": 2873.39, + "duration": 0.0, + "text": "this thing and it used to be that I was" + }, + { + "start": 2873.4, + "duration": 0.0, + "text": "this thing and it used to be that I was able<00:47:53.560> to<00:47:53.640> Google<00:47:54.720> how<00:47:54.920> do<00:47:55.120> then<00:47:55.400> X<00:47:55.680> in<00:47:55.800> program" + }, + { + "start": 2876.35, + "duration": 0.0, + "text": "able to Google how do then X in program" + }, + { + "start": 2876.36, + "duration": 0.0, + "text": "able to Google how do then X in program Y" + }, + { + "start": 2877.39, + "duration": 0.0, + "text": "Y" + }, + { + "start": 2877.4, + "duration": 0.0, + "text": "Y and<00:47:58.120> some<00:47:58.360> of<00:47:58.440> it<00:47:58.560> will<00:47:58.680> be<00:47:58.800> like<00:47:59.000> a<00:47:59.080> forum" + }, + { + "start": 2879.43, + "duration": 0.0, + "text": "and some of it will be like a forum" + }, + { + "start": 2879.44, + "duration": 0.0, + "text": "and some of it will be like a forum post.<00:48:00.280> But<00:48:00.400> now<00:48:01.240> you<00:48:01.360> can't<00:48:01.600> find<00:48:01.800> those" + }, + { + "start": 2881.99, + "duration": 0.0, + "text": "post. But now you can't find those" + }, + { + "start": 2882.0, + "duration": 0.0, + "text": "post. But now you can't find those anymore<00:48:02.600> or<00:48:03.280> I'm<00:48:03.440> doing<00:48:03.600> difficult<00:48:04.000> stuff.<00:48:04.680> So" + }, + { + "start": 2884.79, + "duration": 0.0, + "text": "anymore or I'm doing difficult stuff. So" + }, + { + "start": 2884.8, + "duration": 0.0, + "text": "anymore or I'm doing difficult stuff. So my<00:48:04.920> question<00:48:05.280> is" + }, + { + "start": 2886.27, + "duration": 0.0, + "text": "my question is" + }, + { + "start": 2886.28, + "duration": 0.0, + "text": "my question is what<00:48:06.800> year<00:48:07.200> will<00:48:07.440> I<00:48:07.520> be<00:48:07.680> able<00:48:07.920> to<00:48:08.120> put<00:48:08.480> like<00:48:08.720> a" + }, + { + "start": 2888.87, + "duration": 0.0, + "text": "what year will I be able to put like a" + }, + { + "start": 2888.88, + "duration": 0.0, + "text": "what year will I be able to put like a good<00:48:09.280> version<00:48:09.640> of<00:48:09.760> Clippy<00:48:10.800> in<00:48:11.000> my<00:48:11.160> programs?" + }, + { + "start": 2892.23, + "duration": 0.0, + "text": "good version of Clippy in my programs?" + }, + { + "start": 2892.24, + "duration": 0.0, + "text": "good version of Clippy in my programs? So<00:48:12.360> like<00:48:12.520> there's<00:48:12.680> a<00:48:12.760> little<00:48:13.000> dude<00:48:13.760> you<00:48:13.920> can" + }, + { + "start": 2894.03, + "duration": 0.0, + "text": "So like there's a little dude you can" + }, + { + "start": 2894.04, + "duration": 0.0, + "text": "So like there's a little dude you can actually<00:48:14.320> describe<00:48:14.800> like<00:48:14.960> hey<00:48:15.160> I'm<00:48:15.320> trying<00:48:15.600> to" + }, + { + "start": 2895.79, + "duration": 0.0, + "text": "actually describe like hey I'm trying to" + }, + { + "start": 2895.8, + "duration": 0.0, + "text": "actually describe like hey I'm trying to rotate<00:48:16.240> the<00:48:16.320> mesh<00:48:16.720> in<00:48:16.880> this<00:48:17.080> way<00:48:17.400> what" + }, + { + "start": 2898.75, + "duration": 0.0, + "text": "rotate the mesh in this way what" + }, + { + "start": 2898.76, + "duration": 0.0, + "text": "rotate the mesh in this way what like<00:48:19.200> how<00:48:19.320> does<00:48:19.480> the<00:48:19.600> sub<00:48:19.840> stuff<00:48:20.120> know<00:48:20.400> what?" + }, + { + "start": 2900.71, + "duration": 0.0, + "text": "like how does the sub stuff know what?" + }, + { + "start": 2900.72, + "duration": 0.0, + "text": "like how does the sub stuff know what? There's<00:48:20.840> so<00:48:20.960> many<00:48:21.120> startups<00:48:21.600> trying<00:48:21.760> to<00:48:21.840> do" + }, + { + "start": 2902.03, + "duration": 0.0, + "text": "There's so many startups trying to do" + }, + { + "start": 2902.04, + "duration": 0.0, + "text": "There's so many startups trying to do that.<00:48:22.200> It's<00:48:22.480> Yeah,<00:48:22.600> it's<00:48:22.760> like<00:48:22.920> cursor," + }, + { + "start": 2903.35, + "duration": 0.0, + "text": "that. It's Yeah, it's like cursor," + }, + { + "start": 2903.36, + "duration": 0.0, + "text": "that. It's Yeah, it's like cursor, right?" + }, + { + "start": 2904.31, + "duration": 0.0, + "text": "right?" + }, + { + "start": 2904.32, + "duration": 0.0, + "text": "right? No,<00:48:24.480> well<00:48:24.600> it's<00:48:24.720> not<00:48:24.840> even<00:48:24.960> cursor<00:48:25.360> for<00:48:25.560> X." + }, + { + "start": 2905.87, + "duration": 0.0, + "text": "No, well it's not even cursor for X." + }, + { + "start": 2905.88, + "duration": 0.0, + "text": "No, well it's not even cursor for X. It's<00:48:26.000> like<00:48:26.800> there's<00:48:26.960> so<00:48:27.040> many<00:48:27.200> people<00:48:27.440> who<00:48:27.520> put" + }, + { + "start": 2907.67, + "duration": 0.0, + "text": "It's like there's so many people who put" + }, + { + "start": 2907.68, + "duration": 0.0, + "text": "It's like there's so many people who put the<00:48:27.760> little<00:48:27.920> chat<00:48:28.160> window<00:48:28.360> on<00:48:28.440> the<00:48:28.520> thing<00:48:28.800> and" + }, + { + "start": 2908.87, + "duration": 0.0, + "text": "the little chat window on the thing and" + }, + { + "start": 2908.88, + "duration": 0.0, + "text": "the little chat window on the thing and it's<00:48:28.960> like<00:48:29.120> ask<00:48:29.320> it<00:48:29.400> a<00:48:29.440> question.<00:48:29.840> I<00:48:29.880> just<00:48:30.000> feel" + }, + { + "start": 2910.07, + "duration": 0.0, + "text": "it's like ask it a question. I just feel" + }, + { + "start": 2910.08, + "duration": 0.0, + "text": "it's like ask it a question. I just feel like<00:48:30.200> they<00:48:30.280> never<00:48:30.520> work.<00:48:30.920> Like<00:48:31.320> they<00:48:31.440> still" + }, + { + "start": 2911.63, + "duration": 0.0, + "text": "like they never work. Like they still" + }, + { + "start": 2911.64, + "duration": 0.0, + "text": "like they never work. Like they still are<00:48:31.760> just<00:48:31.920> so<00:48:32.120> bad<00:48:32.480> most<00:48:32.640> of<00:48:32.720> the<00:48:32.760> time.<00:48:33.080> Like<00:48:33.800> I" + }, + { + "start": 2913.83, + "duration": 0.0, + "text": "are just so bad most of the time. Like I" + }, + { + "start": 2913.84, + "duration": 0.0, + "text": "are just so bad most of the time. Like I don't<00:48:33.960> know<00:48:34.080> when<00:48:34.200> it<00:48:34.280> will<00:48:34.360> get<00:48:34.560> good,<00:48:35.080> but" + }, + { + "start": 2915.83, + "duration": 0.0, + "text": "don't know when it will get good, but" + }, + { + "start": 2915.84, + "duration": 0.0, + "text": "don't know when it will get good, but honestly<00:48:36.480> I<00:48:36.720> like" + }, + { + "start": 2917.75, + "duration": 0.0, + "text": "honestly I like" + }, + { + "start": 2917.76, + "duration": 0.0, + "text": "honestly I like I<00:48:37.960> find<00:48:38.320> that<00:48:38.440> when<00:48:38.680> I<00:48:38.840> use<00:48:39.320> LLMs<00:48:39.960> to" + }, + { + "start": 2920.11, + "duration": 0.0, + "text": "I find that when I use LLMs to" + }, + { + "start": 2920.12, + "duration": 0.0, + "text": "I find that when I use LLMs to understand<00:48:40.800> how<00:48:41.400> software<00:48:41.840> works,<00:48:42.480> it<00:48:42.640> gives" + }, + { + "start": 2922.83, + "duration": 0.0, + "text": "understand how software works, it gives" + }, + { + "start": 2922.84, + "duration": 0.0, + "text": "understand how software works, it gives me<00:48:42.960> an<00:48:43.080> incomplete<00:48:44.360> understanding<00:48:45.000> because" + }, + { + "start": 2925.31, + "duration": 0.0, + "text": "me an incomplete understanding because" + }, + { + "start": 2925.32, + "duration": 0.0, + "text": "me an incomplete understanding because I'm<00:48:45.400> not<00:48:45.560> really<00:48:45.760> like<00:48:45.960> traversing<00:48:46.520> through" + }, + { + "start": 2926.87, + "duration": 0.0, + "text": "I'm not really like traversing through" + }, + { + "start": 2926.88, + "duration": 0.0, + "text": "I'm not really like traversing through the<00:48:47.000> whole<00:48:47.200> set<00:48:47.520> of<00:48:47.640> knowledge,<00:48:48.160> right?<00:48:48.400> I'm" + }, + { + "start": 2928.55, + "duration": 0.0, + "text": "the whole set of knowledge, right? I'm" + }, + { + "start": 2928.56, + "duration": 0.0, + "text": "the whole set of knowledge, right? I'm just<00:48:49.040> specifically<00:48:49.680> getting<00:48:49.880> the<00:48:50.000> answer<00:48:50.280> I" + }, + { + "start": 2930.31, + "duration": 0.0, + "text": "just specifically getting the answer I" + }, + { + "start": 2930.32, + "duration": 0.0, + "text": "just specifically getting the answer I need<00:48:50.520> and<00:48:50.640> implementing<00:48:51.080> it<00:48:51.200> and<00:48:51.320> moving<00:48:51.560> on" + }, + { + "start": 2931.67, + "duration": 0.0, + "text": "need and implementing it and moving on" + }, + { + "start": 2931.68, + "duration": 0.0, + "text": "need and implementing it and moving on with<00:48:51.760> my<00:48:51.880> life.<00:48:52.480> Whereas<00:48:52.720> when<00:48:52.840> I<00:48:52.880> would<00:48:53.160> when" + }, + { + "start": 2933.31, + "duration": 0.0, + "text": "with my life. Whereas when I would when" + }, + { + "start": 2933.32, + "duration": 0.0, + "text": "with my life. Whereas when I would when I<00:48:53.400> read<00:48:53.640> documentation<00:48:54.520> and<00:48:54.640> like<00:48:54.840> actually" + }, + { + "start": 2935.11, + "duration": 0.0, + "text": "I read documentation and like actually" + }, + { + "start": 2935.12, + "duration": 0.0, + "text": "I read documentation and like actually go<00:48:55.280> line<00:48:55.400> by<00:48:55.520> line<00:48:55.840> and<00:48:55.960> sit<00:48:56.080> down<00:48:56.240> and" + }, + { + "start": 2936.31, + "duration": 0.0, + "text": "go line by line and sit down and" + }, + { + "start": 2936.32, + "duration": 0.0, + "text": "go line by line and sit down and understand<00:48:56.760> it,<00:48:57.240> I<00:48:57.360> basically<00:48:58.320> I<00:48:58.440> just<00:48:58.680> know" + }, + { + "start": 2939.03, + "duration": 0.0, + "text": "understand it, I basically I just know" + }, + { + "start": 2939.04, + "duration": 0.0, + "text": "understand it, I basically I just know the<00:48:59.560> unknown<00:48:59.920> unknowns,<00:49:00.520> right?<00:49:00.800> Because<00:49:01.040> I" + }, + { + "start": 2941.43, + "duration": 0.0, + "text": "the unknown unknowns, right? Because I" + }, + { + "start": 2941.44, + "duration": 0.0, + "text": "the unknown unknowns, right? Because I explored.<00:49:02.480> And<00:49:02.640> so<00:49:03.400> LLMs<00:49:03.840> to<00:49:04.200> say<00:49:04.400> how<00:49:04.520> do<00:49:04.640> I<00:49:04.680> do" + }, + { + "start": 2944.79, + "duration": 0.0, + "text": "explored. And so LLMs to say how do I do" + }, + { + "start": 2944.8, + "duration": 0.0, + "text": "explored. And so LLMs to say how do I do this<00:49:04.920> in<00:49:05.000> this<00:49:05.120> program?<00:49:05.520> I<00:49:05.760> I'm<00:49:05.920> actually<00:49:06.120> not" + }, + { + "start": 2946.23, + "duration": 0.0, + "text": "this in this program? I I'm actually not" + }, + { + "start": 2946.24, + "duration": 0.0, + "text": "this in this program? I I'm actually not even<00:49:06.400> sure<00:49:06.520> that's<00:49:06.720> good<00:49:06.880> UX,<00:49:07.400> basically.<00:49:08.080> But" + }, + { + "start": 2948.15, + "duration": 0.0, + "text": "even sure that's good UX, basically. But" + }, + { + "start": 2948.16, + "duration": 0.0, + "text": "even sure that's good UX, basically. But if<00:49:08.240> if<00:49:08.440> you<00:49:08.520> can<00:49:08.680> like<00:49:08.880> search<00:49:09.200> the" + }, + { + "start": 2949.31, + "duration": 0.0, + "text": "if if you can like search the" + }, + { + "start": 2949.32, + "duration": 0.0, + "text": "if if you can like search the documentation<00:49:10.720> for<00:49:11.040> So<00:49:11.200> like<00:49:11.400> you<00:49:11.520> need<00:49:12.080> to" + }, + { + "start": 2952.15, + "duration": 0.0, + "text": "documentation for So like you need to" + }, + { + "start": 2952.16, + "duration": 0.0, + "text": "documentation for So like you need to take<00:49:12.360> it<00:49:12.440> put<00:49:12.640> the<00:49:12.680> documentation<00:49:13.320> is<00:49:13.440> like<00:49:13.840> a" + }, + { + "start": 2953.95, + "duration": 0.0, + "text": "take it put the documentation is like a" + }, + { + "start": 2953.96, + "duration": 0.0, + "text": "take it put the documentation is like a hundred<00:49:14.520> page<00:49:14.880> PDF<00:49:15.400> documentation<00:49:15.960> of<00:49:16.040> this" + }, + { + "start": 2956.19, + "duration": 0.0, + "text": "hundred page PDF documentation of this" + }, + { + "start": 2956.2, + "duration": 0.0, + "text": "hundred page PDF documentation of this program.<00:49:16.440> I'm<00:49:16.520> not<00:49:16.640> going<00:49:16.760> to<00:49:16.840> read<00:49:17.080> that." + }, + { + "start": 2957.71, + "duration": 0.0, + "text": "program. I'm not going to read that." + }, + { + "start": 2957.72, + "duration": 0.0, + "text": "program. I'm not going to read that. Yeah,<00:49:18.080> it<00:49:18.280> is<00:49:18.400> good<00:49:18.600> for<00:49:18.720> that.<00:49:19.120> Like<00:49:19.320> LLMs<00:49:20.040> are" + }, + { + "start": 2960.71, + "duration": 0.0, + "text": "Yeah, it is good for that. Like LLMs are" + }, + { + "start": 2960.72, + "duration": 0.0, + "text": "Yeah, it is good for that. Like LLMs are almost<00:49:21.080> like<00:49:21.240> a<00:49:21.360> very<00:49:22.320> it's<00:49:22.480> like<00:49:22.600> a<00:49:22.640> better" + }, + { + "start": 2963.35, + "duration": 0.0, + "text": "almost like a very it's like a better" + }, + { + "start": 2963.36, + "duration": 0.0, + "text": "almost like a very it's like a better Google<00:49:23.760> almost." + }, + { + "start": 2965.27, + "duration": 0.0, + "text": "Google almost." + }, + { + "start": 2965.28, + "duration": 0.0, + "text": "Google almost. Which<00:49:25.440> is<00:49:25.520> why<00:49:25.640> they<00:49:25.760> pivoted<00:49:26.080> their<00:49:26.160> entire" + }, + { + "start": 2966.39, + "duration": 0.0, + "text": "Which is why they pivoted their entire" + }, + { + "start": 2966.4, + "duration": 0.0, + "text": "Which is why they pivoted their entire company<00:49:26.720> to<00:49:26.840> LLMs.<00:49:27.920> I<00:49:28.000> might<00:49:28.160> be<00:49:28.280> able<00:49:28.400> to<00:49:28.760> So" + }, + { + "start": 2969.03, + "duration": 0.0, + "text": "company to LLMs. I might be able to So" + }, + { + "start": 2969.04, + "duration": 0.0, + "text": "company to LLMs. I might be able to So at<00:49:29.160> what<00:49:29.480> year<00:49:29.840> can<00:49:30.040> I<00:49:30.120> expect<00:49:30.600> to<00:49:31.360> like<00:49:31.800> put" + }, + { + "start": 2971.99, + "duration": 0.0, + "text": "at what year can I expect to like put" + }, + { + "start": 2972.0, + "duration": 0.0, + "text": "at what year can I expect to like put one<00:49:32.160> of<00:49:32.200> these<00:49:32.440> in<00:49:32.640> my<00:49:32.840> C<00:49:33.080> program?" + }, + { + "start": 2974.31, + "duration": 0.0, + "text": "one of these in my C program?" + }, + { + "start": 2974.32, + "duration": 0.0, + "text": "one of these in my C program? Where<00:49:34.520> it's<00:49:34.720> like<00:49:34.960> running<00:49:35.320> locally?<00:49:36.040> Yeah," + }, + { + "start": 2976.23, + "duration": 0.0, + "text": "Where it's like running locally? Yeah," + }, + { + "start": 2976.24, + "duration": 0.0, + "text": "Where it's like running locally? Yeah, AI.C.<00:49:36.760> Yeah,<00:49:37.000> cuz<00:49:37.120> I<00:49:37.160> want<00:49:37.280> to<00:49:37.480> Yes,<00:49:37.680> my<00:49:37.880> AI.C" + }, + { + "start": 2978.51, + "duration": 0.0, + "text": "AI.C. Yeah, cuz I want to Yes, my AI.C" + }, + { + "start": 2978.52, + "duration": 0.0, + "text": "AI.C. Yeah, cuz I want to Yes, my AI.C project.<00:49:39.280> To<00:49:39.320> get<00:49:39.480> the<00:49:39.600> guy<00:49:40.040> in<00:49:40.560> who<00:49:40.680> like" + }, + { + "start": 2980.87, + "duration": 0.0, + "text": "project. To get the guy in who like" + }, + { + "start": 2980.88, + "duration": 0.0, + "text": "project. To get the guy in who like sometimes<00:49:41.280> complains<00:49:41.800> Um" + }, + { + "start": 2982.75, + "duration": 0.0, + "text": "sometimes complains Um" + }, + { + "start": 2982.76, + "duration": 0.0, + "text": "sometimes complains Um Let's<00:49:42.920> say<00:49:43.120> let's<00:49:43.320> say<00:49:43.880> you<00:49:43.960> know<00:49:44.080> smaller" + }, + { + "start": 2984.43, + "duration": 0.0, + "text": "Let's say let's say you know smaller" + }, + { + "start": 2984.44, + "duration": 0.0, + "text": "Let's say let's say you know smaller models<00:49:44.760> are<00:49:44.840> always<00:49:45.000> getting<00:49:45.240> better<00:49:45.440> and" + }, + { + "start": 2985.51, + "duration": 0.0, + "text": "models are always getting better and" + }, + { + "start": 2985.52, + "duration": 0.0, + "text": "models are always getting better and better.<00:49:45.880> Let's<00:49:46.400> let's<00:49:46.600> put<00:49:46.760> a<00:49:47.160> let's<00:49:47.320> put<00:49:47.440> a" + }, + { + "start": 2987.47, + "duration": 0.0, + "text": "better. Let's let's put a let's put a" + }, + { + "start": 2987.48, + "duration": 0.0, + "text": "better. Let's let's put a let's put a release<00:49:47.760> date<00:49:47.960> on<00:49:48.080> it<00:49:48.160> and<00:49:48.240> say<00:49:48.360> five<00:49:48.720> years." + }, + { + "start": 2991.03, + "duration": 0.0, + "text": "release date on it and say five years." + }, + { + "start": 2991.04, + "duration": 0.0, + "text": "release date on it and say five years. Okay." + }, + { + "start": 2992.43, + "duration": 0.0, + "text": "Okay." + }, + { + "start": 2992.44, + "duration": 0.0, + "text": "Okay. So<00:49:53.000> yeah,<00:49:53.160> so<00:49:53.760> have<00:49:54.000> you<00:49:54.280> seen<00:49:54.840> like<00:49:55.040> a<00:49:55.120> model" + }, + { + "start": 2995.51, + "duration": 0.0, + "text": "So yeah, so have you seen like a model" + }, + { + "start": 2995.52, + "duration": 0.0, + "text": "So yeah, so have you seen like a model that<00:49:55.760> sort<00:49:56.000> of<00:49:56.120> can<00:49:56.960> work<00:49:57.280> outside<00:49:57.800> of<00:49:57.920> the" + }, + { + "start": 2997.99, + "duration": 0.0, + "text": "that sort of can work outside of the" + }, + { + "start": 2998.0, + "duration": 0.0, + "text": "that sort of can work outside of the training<00:49:58.400> data?" + }, + { + "start": 2999.55, + "duration": 0.0, + "text": "training data?" + }, + { + "start": 2999.56, + "duration": 0.0, + "text": "training data? It<00:49:59.680> feels<00:49:59.920> like<00:50:00.080> they're<00:50:00.320> always<00:50:00.680> like" + }, + { + "start": 3000.87, + "duration": 0.0, + "text": "It feels like they're always like" + }, + { + "start": 3000.88, + "duration": 0.0, + "text": "It feels like they're always like weighted<00:50:01.280> averages<00:50:01.920> of<00:50:02.080> the<00:50:02.200> training<00:50:02.520> data." + }, + { + "start": 3003.07, + "duration": 0.0, + "text": "weighted averages of the training data." + }, + { + "start": 3003.08, + "duration": 0.0, + "text": "weighted averages of the training data. It's<00:50:03.280> It's<00:50:03.480> not<00:50:03.720> a<00:50:03.760> weighted<00:50:04.160> average<00:50:04.440> of<00:50:04.520> the" + }, + { + "start": 3004.59, + "duration": 0.0, + "text": "It's It's not a weighted average of the" + }, + { + "start": 3004.6, + "duration": 0.0, + "text": "It's It's not a weighted average of the training<00:50:04.920> data.<00:50:05.280> A<00:50:05.360> better<00:50:05.600> way<00:50:05.680> to<00:50:05.800> describe" + }, + { + "start": 3006.47, + "duration": 0.0, + "text": "training data. A better way to describe" + }, + { + "start": 3006.48, + "duration": 0.0, + "text": "training data. A better way to describe how<00:50:06.640> they<00:50:07.000> process<00:50:07.520> and<00:50:07.640> hold<00:50:07.920> and<00:50:08.480> how<00:50:08.760> our" + }, + { + "start": 3008.95, + "duration": 0.0, + "text": "how they process and hold and how our" + }, + { + "start": 3008.96, + "duration": 0.0, + "text": "how they process and hold and how our models<00:50:09.280> impacted<00:50:09.640> by<00:50:09.720> training<00:50:10.000> data<00:50:10.360> is" + }, + { + "start": 3011.47, + "duration": 0.0, + "text": "models impacted by training data is" + }, + { + "start": 3011.48, + "duration": 0.0, + "text": "models impacted by training data is they<00:50:11.760> find<00:50:12.400> patterns<00:50:13.160> and<00:50:13.320> then<00:50:13.640> we<00:50:13.880> run<00:50:14.120> tests" + }, + { + "start": 3014.43, + "duration": 0.0, + "text": "they find patterns and then we run tests" + }, + { + "start": 3014.44, + "duration": 0.0, + "text": "they find patterns and then we run tests on<00:50:14.520> those<00:50:14.640> patterns<00:50:15.000> to<00:50:15.080> make<00:50:15.240> sure<00:50:15.600> they" + }, + { + "start": 3015.83, + "duration": 0.0, + "text": "on those patterns to make sure they" + }, + { + "start": 3015.84, + "duration": 0.0, + "text": "on those patterns to make sure they generalize<00:50:16.400> and<00:50:16.480> they're<00:50:16.600> not<00:50:16.880> specific<00:50:17.400> to" + }, + { + "start": 3017.47, + "duration": 0.0, + "text": "generalize and they're not specific to" + }, + { + "start": 3017.48, + "duration": 0.0, + "text": "generalize and they're not specific to the<00:50:17.600> training<00:50:17.880> data.<00:50:18.480> So,<00:50:18.600> they<00:50:18.680> find" + }, + { + "start": 3018.87, + "duration": 0.0, + "text": "the training data. So, they find" + }, + { + "start": 3018.88, + "duration": 0.0, + "text": "the training data. So, they find patterns<00:50:19.240> which<00:50:19.440> are<00:50:19.520> general<00:50:20.080> that<00:50:20.840> do<00:50:21.040> the" + }, + { + "start": 3021.15, + "duration": 0.0, + "text": "patterns which are general that do the" + }, + { + "start": 3021.16, + "duration": 0.0, + "text": "patterns which are general that do the thing<00:50:21.400> you<00:50:21.480> want<00:50:21.640> it<00:50:21.720> to<00:50:21.800> do.<00:50:22.120> ARC<00:50:22.360> AGI" + }, + { + "start": 3022.67, + "duration": 0.0, + "text": "thing you want it to do. ARC AGI" + }, + { + "start": 3022.68, + "duration": 0.0, + "text": "thing you want it to do. ARC AGI specifically<00:50:23.400> is<00:50:23.520> a<00:50:23.600> benchmark<00:50:24.080> for<00:50:24.720> generic" + }, + { + "start": 3025.19, + "duration": 0.0, + "text": "specifically is a benchmark for generic" + }, + { + "start": 3025.2, + "duration": 0.0, + "text": "specifically is a benchmark for generic never-before-seen<00:50:26.040> on<00:50:26.160> the<00:50:26.280> internet" + }, + { + "start": 3026.63, + "duration": 0.0, + "text": "never-before-seen on the internet" + }, + { + "start": 3026.64, + "duration": 0.0, + "text": "never-before-seen on the internet patterns<00:50:27.000> that<00:50:27.440> fail<00:50:28.120> LLMs,<00:50:28.720> all<00:50:28.880> of<00:50:28.960> them." + }, + { + "start": 3029.35, + "duration": 0.0, + "text": "patterns that fail LLMs, all of them." + }, + { + "start": 3029.36, + "duration": 0.0, + "text": "patterns that fail LLMs, all of them. They're<00:50:29.480> private.<00:50:29.840> They're<00:50:29.960> never<00:50:30.240> leaked<00:50:30.560> or" + }, + { + "start": 3030.63, + "duration": 0.0, + "text": "They're private. They're never leaked or" + }, + { + "start": 3030.64, + "duration": 0.0, + "text": "They're private. They're never leaked or exposed.<00:50:31.120> You<00:50:31.240> can't<00:50:31.480> ever<00:50:32.120> You<00:50:32.240> can't<00:50:32.520> build" + }, + { + "start": 3033.51, + "duration": 0.0, + "text": "exposed. You can't ever You can't build" + }, + { + "start": 3033.52, + "duration": 0.0, + "text": "exposed. You can't ever You can't build an<00:50:33.640> algorithm<00:50:33.960> that<00:50:34.080> interpolates<00:50:34.520> from" + }, + { + "start": 3034.63, + "duration": 0.0, + "text": "an algorithm that interpolates from" + }, + { + "start": 3034.64, + "duration": 0.0, + "text": "an algorithm that interpolates from existing<00:50:34.960> solutions<00:50:35.400> to<00:50:35.520> solve<00:50:35.720> ARC<00:50:35.880> AGI" + }, + { + "start": 3036.23, + "duration": 0.0, + "text": "existing solutions to solve ARC AGI" + }, + { + "start": 3036.24, + "duration": 0.0, + "text": "existing solutions to solve ARC AGI ever.<00:50:36.640> And<00:50:37.120> LLMs<00:50:37.520> did<00:50:37.720> that." + }, + { + "start": 3038.59, + "duration": 0.0, + "text": "ever. And LLMs did that." + }, + { + "start": 3038.6, + "duration": 0.0, + "text": "ever. And LLMs did that. Yeah,<00:50:38.840> so<00:50:39.040> so<00:50:39.320> yeah,<00:50:39.600> I<00:50:39.680> was<00:50:39.880> asking<00:50:40.200> sort<00:50:40.360> of" + }, + { + "start": 3040.67, + "duration": 0.0, + "text": "Yeah, so so yeah, I was asking sort of" + }, + { + "start": 3040.68, + "duration": 0.0, + "text": "Yeah, so so yeah, I was asking sort of about<00:50:40.880> the<00:50:41.040> answers<00:50:41.640> you<00:50:41.720> get<00:50:42.520> because<00:50:42.840> it" + }, + { + "start": 3043.15, + "duration": 0.0, + "text": "about the answers you get because it" + }, + { + "start": 3043.16, + "duration": 0.0, + "text": "about the answers you get because it feels<00:50:43.520> like<00:50:43.720> you<00:50:43.800> get<00:50:44.040> sort<00:50:44.240> of<00:50:44.960> you<00:50:45.040> know,<00:50:45.280> you" + }, + { + "start": 3045.39, + "duration": 0.0, + "text": "feels like you get sort of you know, you" + }, + { + "start": 3045.4, + "duration": 0.0, + "text": "feels like you get sort of you know, you cannot<00:50:45.720> go<00:50:45.880> outside<00:50:46.320> of<00:50:46.440> what<00:50:46.600> it<00:50:46.720> has<00:50:46.920> been" + }, + { + "start": 3047.15, + "duration": 0.0, + "text": "cannot go outside of what it has been" + }, + { + "start": 3047.16, + "duration": 0.0, + "text": "cannot go outside of what it has been taught,<00:50:47.720> right?<00:50:48.040> Or<00:50:48.280> sort<00:50:48.520> of" + }, + { + "start": 3049.19, + "duration": 0.0, + "text": "taught, right? Or sort of" + }, + { + "start": 3049.2, + "duration": 0.0, + "text": "taught, right? Or sort of optimized<00:50:49.800> on," + }, + { + "start": 3050.59, + "duration": 0.0, + "text": "optimized on," + }, + { + "start": 3050.6, + "duration": 0.0, + "text": "optimized on, but<00:50:50.880> maybe<00:50:51.200> you<00:50:51.320> could<00:50:51.600> sort<00:50:51.760> of<00:50:52.240> extrapolate" + }, + { + "start": 3053.99, + "duration": 0.0, + "text": "but maybe you could sort of extrapolate" + }, + { + "start": 3054.0, + "duration": 0.0, + "text": "but maybe you could sort of extrapolate some,<00:50:54.480> you<00:50:54.560> know,<00:50:54.800> outside<00:50:55.280> of<00:50:55.400> it.<00:50:56.120> But<00:50:56.760> is" + }, + { + "start": 3056.91, + "duration": 0.0, + "text": "some, you know, outside of it. But is" + }, + { + "start": 3056.92, + "duration": 0.0, + "text": "some, you know, outside of it. But is that<00:50:57.120> happening<00:50:57.600> like<00:50:58.280> Yeah,<00:50:58.680> I<00:50:58.800> I<00:50:58.920> Well,<00:50:59.040> I" + }, + { + "start": 3059.11, + "duration": 0.0, + "text": "that happening like Yeah, I I Well, I" + }, + { + "start": 3059.12, + "duration": 0.0, + "text": "that happening like Yeah, I I Well, I think<00:50:59.360> that<00:50:59.520> like<00:50:59.720> when<00:51:00.200> you<00:51:00.320> get<00:51:00.480> to<00:51:00.600> these" + }, + { + "start": 3060.75, + "duration": 0.0, + "text": "think that like when you get to these" + }, + { + "start": 3060.76, + "duration": 0.0, + "text": "think that like when you get to these absurd<00:51:01.400> levels<00:51:01.760> of<00:51:01.880> scale,<00:51:02.920> um" + }, + { + "start": 3064.67, + "duration": 0.0, + "text": "absurd levels of scale, um" + }, + { + "start": 3064.68, + "duration": 0.0, + "text": "absurd levels of scale, um there's<00:51:04.920> almost<00:51:05.160> nothing<00:51:05.480> outside<00:51:05.840> the<00:51:05.920> box" + }, + { + "start": 3066.27, + "duration": 0.0, + "text": "there's almost nothing outside the box" + }, + { + "start": 3066.28, + "duration": 0.0, + "text": "there's almost nothing outside the box if<00:51:06.440> it's<00:51:06.760> predicting<00:51:07.200> the<00:51:07.280> entire<00:51:07.640> internet." + }, + { + "start": 3068.63, + "duration": 0.0, + "text": "if it's predicting the entire internet." + }, + { + "start": 3068.64, + "duration": 0.0, + "text": "if it's predicting the entire internet. Um<00:51:09.160> that's<00:51:09.360> basically<00:51:09.800> like<00:51:10.000> a" + }, + { + "start": 3070.83, + "duration": 0.0, + "text": "Um that's basically like a" + }, + { + "start": 3070.84, + "duration": 0.0, + "text": "Um that's basically like a a<00:51:10.920> huge<00:51:11.320> amount<00:51:11.640> of<00:51:11.720> examples<00:51:12.400> of<00:51:13.320> uh<00:51:13.840> the" + }, + { + "start": 3073.95, + "duration": 0.0, + "text": "a huge amount of examples of uh the" + }, + { + "start": 3073.96, + "duration": 0.0, + "text": "a huge amount of examples of uh the dynamics<00:51:14.360> of<00:51:14.480> human<00:51:14.680> intelligence<00:51:15.280> recorded" + }, + { + "start": 3076.27, + "duration": 0.0, + "text": "dynamics of human intelligence recorded" + }, + { + "start": 3076.28, + "duration": 0.0, + "text": "dynamics of human intelligence recorded um" + }, + { + "start": 3077.31, + "duration": 0.0, + "text": "um" + }, + { + "start": 3077.32, + "duration": 0.0, + "text": "um in<00:51:17.520> into<00:51:17.800> written<00:51:18.040> form<00:51:18.320> that<00:51:18.440> you<00:51:18.560> then,<00:51:19.120> you" + }, + { + "start": 3079.19, + "duration": 0.0, + "text": "in into written form that you then, you" + }, + { + "start": 3079.2, + "duration": 0.0, + "text": "in into written form that you then, you know,<00:51:19.360> process<00:51:19.800> and<00:51:19.920> find<00:51:20.160> patterns<00:51:20.680> which<00:51:21.560> uh" + }, + { + "start": 3081.95, + "duration": 0.0, + "text": "know, process and find patterns which uh" + }, + { + "start": 3081.96, + "duration": 0.0, + "text": "know, process and find patterns which uh apply<00:51:22.320> across<00:51:22.800> many<00:51:23.080> examples.<00:51:23.640> That's" + }, + { + "start": 3083.87, + "duration": 0.0, + "text": "apply across many examples. That's" + }, + { + "start": 3083.88, + "duration": 0.0, + "text": "apply across many examples. That's That's<00:51:24.080> what<00:51:24.200> they're<00:51:24.280> doing.<00:51:24.600> They" + }, + { + "start": 3085.71, + "duration": 0.0, + "text": "That's what they're doing. They" + }, + { + "start": 3085.72, + "duration": 0.0, + "text": "That's what they're doing. They Deep<00:51:25.920> learning<00:51:26.120> systems<00:51:26.480> usually<00:51:26.760> start<00:51:26.920> with" + }, + { + "start": 3087.03, + "duration": 0.0, + "text": "Deep learning systems usually start with" + }, + { + "start": 3087.04, + "duration": 0.0, + "text": "Deep learning systems usually start with very<00:51:27.280> low-frequency<00:51:28.200> generic<00:51:28.720> patterns" + }, + { + "start": 3089.19, + "duration": 0.0, + "text": "very low-frequency generic patterns" + }, + { + "start": 3089.2, + "duration": 0.0, + "text": "very low-frequency generic patterns like,<00:51:29.600> \"Oh,<00:51:29.720> there's<00:51:29.920> a<00:51:29.960> period<00:51:30.360> every" + }, + { + "start": 3090.55, + "duration": 0.0, + "text": "like, \"Oh, there's a period every" + }, + { + "start": 3090.56, + "duration": 0.0, + "text": "like, \"Oh, there's a period every so-and-so<00:51:30.920> words.\"<00:51:31.240> And<00:51:31.320> then<00:51:31.400> they<00:51:31.480> start<00:51:31.640> to" + }, + { + "start": 3091.71, + "duration": 0.0, + "text": "so-and-so words.\" And then they start to" + }, + { + "start": 3091.72, + "duration": 0.0, + "text": "so-and-so words.\" And then they start to get<00:51:31.840> to<00:51:31.920> higher-frequency<00:51:32.480> patterns<00:51:32.760> and" + }, + { + "start": 3092.83, + "duration": 0.0, + "text": "get to higher-frequency patterns and" + }, + { + "start": 3092.84, + "duration": 0.0, + "text": "get to higher-frequency patterns and more<00:51:32.960> specific,<00:51:33.760> more<00:51:33.920> nuanced<00:51:34.240> things,<00:51:34.560> and" + }, + { + "start": 3094.63, + "duration": 0.0, + "text": "more specific, more nuanced things, and" + }, + { + "start": 3094.64, + "duration": 0.0, + "text": "more specific, more nuanced things, and then" + }, + { + "start": 3095.51, + "duration": 0.0, + "text": "then" + }, + { + "start": 3095.52, + "duration": 0.0, + "text": "then they<00:51:35.680> they<00:51:35.920> they're<00:51:36.400> they're<00:51:36.520> basically" + }, + { + "start": 3096.91, + "duration": 0.0, + "text": "they they they're they're basically" + }, + { + "start": 3096.92, + "duration": 0.0, + "text": "they they they're they're basically pattern<00:51:37.240> finders.<00:51:38.040> Patterns<00:51:38.440> which<00:51:39.080> apply<00:51:39.760> in" + }, + { + "start": 3099.99, + "duration": 0.0, + "text": "pattern finders. Patterns which apply in" + }, + { + "start": 3100.0, + "duration": 0.0, + "text": "pattern finders. Patterns which apply in on<00:51:40.080> not<00:51:40.320> just<00:51:40.520> the<00:51:40.600> training<00:51:40.920> data.<00:51:41.120> That's" + }, + { + "start": 3101.27, + "duration": 0.0, + "text": "on not just the training data. That's" + }, + { + "start": 3101.28, + "duration": 0.0, + "text": "on not just the training data. That's the<00:51:41.360> whole<00:51:41.520> point<00:51:41.760> of<00:51:41.840> being<00:51:42.480> like<00:51:42.600> a<00:51:42.680> deep" + }, + { + "start": 3102.87, + "duration": 0.0, + "text": "the whole point of being like a deep" + }, + { + "start": 3102.88, + "duration": 0.0, + "text": "the whole point of being like a deep learning<00:51:43.160> person<00:51:43.960> is<00:51:44.200> you<00:51:44.320> make<00:51:44.520> sure<00:51:44.720> that" + }, + { + "start": 3104.87, + "duration": 0.0, + "text": "learning person is you make sure that" + }, + { + "start": 3104.88, + "duration": 0.0, + "text": "learning person is you make sure that the<00:51:44.960> model<00:51:45.520> does<00:51:45.640> not<00:51:45.800> just<00:51:45.960> apply<00:51:46.160> on<00:51:46.240> the" + }, + { + "start": 3106.31, + "duration": 0.0, + "text": "the model does not just apply on the" + }, + { + "start": 3106.32, + "duration": 0.0, + "text": "the model does not just apply on the training<00:51:46.600> data.<00:51:46.920> That's<00:51:47.120> like<00:51:47.440> the<00:51:47.560> point<00:51:47.760> of" + }, + { + "start": 3107.83, + "duration": 0.0, + "text": "training data. That's like the point of" + }, + { + "start": 3107.84, + "duration": 0.0, + "text": "training data. That's like the point of all<00:51:47.920> the<00:51:48.040> architectures,<00:51:48.680> that's<00:51:48.840> the<00:51:48.920> point" + }, + { + "start": 3109.15, + "duration": 0.0, + "text": "all the architectures, that's the point" + }, + { + "start": 3109.16, + "duration": 0.0, + "text": "all the architectures, that's the point of<00:51:49.240> like,<00:51:50.000> you<00:51:50.040> know,<00:51:50.160> doing<00:51:50.400> everything" + }, + { + "start": 3110.67, + "duration": 0.0, + "text": "of like, you know, doing everything" + }, + { + "start": 3110.68, + "duration": 0.0, + "text": "of like, you know, doing everything you're<00:51:50.800> doing<00:51:51.080> is<00:51:51.160> making<00:51:51.400> sure<00:51:51.520> that<00:51:51.640> it's" + }, + { + "start": 3111.75, + "duration": 0.0, + "text": "you're doing is making sure that it's" + }, + { + "start": 3111.76, + "duration": 0.0, + "text": "you're doing is making sure that it's not<00:51:51.920> just<00:51:52.080> memorizing<00:51:52.480> the<00:51:52.560> training<00:51:52.800> data," + }, + { + "start": 3113.07, + "duration": 0.0, + "text": "not just memorizing the training data," + }, + { + "start": 3113.08, + "duration": 0.0, + "text": "not just memorizing the training data, that<00:51:53.200> it's<00:51:53.400> finding<00:51:53.920> patterns<00:51:54.320> in<00:51:54.400> the" + }, + { + "start": 3114.47, + "duration": 0.0, + "text": "that it's finding patterns in the" + }, + { + "start": 3114.48, + "duration": 0.0, + "text": "that it's finding patterns in the training<00:51:54.720> data<00:51:54.920> that<00:51:55.080> will<00:51:55.200> apply<00:51:55.640> in<00:51:55.880> your" + }, + { + "start": 3116.03, + "duration": 0.0, + "text": "training data that will apply in your" + }, + { + "start": 3116.04, + "duration": 0.0, + "text": "training data that will apply in your specific<00:51:56.440> task<00:51:56.680> that<00:51:56.800> you<00:51:56.880> need<00:51:56.960> the<00:51:57.000> model" + }, + { + "start": 3117.23, + "duration": 0.0, + "text": "specific task that you need the model" + }, + { + "start": 3117.24, + "duration": 0.0, + "text": "specific task that you need the model for." + }, + { + "start": 3118.43, + "duration": 0.0, + "text": "for." + }, + { + "start": 3118.44, + "duration": 0.0, + "text": "for. So,<00:51:59.080> so<00:51:59.520> sort<00:51:59.720> of<00:51:59.800> like<00:52:00.040> say<00:52:00.240> that<00:52:00.440> you" + }, + { + "start": 3121.03, + "duration": 0.0, + "text": "So, so sort of like say that you" + }, + { + "start": 3121.04, + "duration": 0.0, + "text": "So, so sort of like say that you recognize<00:52:01.600> images," + }, + { + "start": 3123.15, + "duration": 0.0, + "text": "recognize images," + }, + { + "start": 3123.16, + "duration": 0.0, + "text": "recognize images, like<00:52:03.440> you<00:52:03.760> can<00:52:03.960> recognize<00:52:04.400> cat<00:52:04.720> and<00:52:04.920> dog." + }, + { + "start": 3125.39, + "duration": 0.0, + "text": "like you can recognize cat and dog." + }, + { + "start": 3125.4, + "duration": 0.0, + "text": "like you can recognize cat and dog. Yeah.<00:52:05.920> But<00:52:06.120> like<00:52:06.440> a<00:52:06.480> model<00:52:06.760> that<00:52:06.920> can<00:52:07.080> do<00:52:07.680> train" + }, + { + "start": 3127.99, + "duration": 0.0, + "text": "Yeah. But like a model that can do train" + }, + { + "start": 3128.0, + "duration": 0.0, + "text": "Yeah. But like a model that can do train on<00:52:08.120> cat<00:52:08.360> and<00:52:08.480> dog<00:52:08.720> images,<00:52:09.520> but<00:52:09.760> also<00:52:10.120> like" + }, + { + "start": 3130.51, + "duration": 0.0, + "text": "on cat and dog images, but also like" + }, + { + "start": 3130.52, + "duration": 0.0, + "text": "on cat and dog images, but also like this<00:52:10.800> is<00:52:10.920> something<00:52:11.720> else.<00:52:12.480> Like<00:52:13.280> without<00:52:13.600> you" + }, + { + "start": 3133.71, + "duration": 0.0, + "text": "this is something else. Like without you" + }, + { + "start": 3133.72, + "duration": 0.0, + "text": "this is something else. Like without you showing<00:52:14.280> something<00:52:14.600> else." + }, + { + "start": 3135.31, + "duration": 0.0, + "text": "showing something else." + }, + { + "start": 3135.32, + "duration": 0.0, + "text": "showing something else. Yes,<00:52:15.680> I<00:52:15.760> mean<00:52:15.920> obviously<00:52:16.280> that<00:52:16.440> model<00:52:16.720> is<00:52:16.880> not" + }, + { + "start": 3137.39, + "duration": 0.0, + "text": "Yes, I mean obviously that model is not" + }, + { + "start": 3137.4, + "duration": 0.0, + "text": "Yes, I mean obviously that model is not um<00:52:18.080> learned<00:52:18.360> patterns<00:52:18.720> which<00:52:18.840> would" + }, + { + "start": 3138.95, + "duration": 0.0, + "text": "um learned patterns which would" + }, + { + "start": 3138.96, + "duration": 0.0, + "text": "um learned patterns which would generalize<00:52:19.480> to<00:52:19.640> is<00:52:19.760> it<00:52:19.880> a<00:52:19.920> turtle<00:52:20.200> or<00:52:20.280> not.<00:52:20.520> It" + }, + { + "start": 3140.59, + "duration": 0.0, + "text": "generalize to is it a turtle or not. It" + }, + { + "start": 3140.6, + "duration": 0.0, + "text": "generalize to is it a turtle or not. It would<00:52:20.720> try<00:52:20.840> to<00:52:20.920> bucket<00:52:21.240> into<00:52:21.400> cat<00:52:21.560> or<00:52:21.640> dog,<00:52:22.400> cuz" + }, + { + "start": 3142.51, + "duration": 0.0, + "text": "would try to bucket into cat or dog, cuz" + }, + { + "start": 3142.52, + "duration": 0.0, + "text": "would try to bucket into cat or dog, cuz that's<00:52:22.680> what<00:52:22.800> that," + }, + { + "start": 3143.71, + "duration": 0.0, + "text": "that's what that," + }, + { + "start": 3143.72, + "duration": 0.0, + "text": "that's what that, you<00:52:23.800> know,<00:52:23.920> system<00:52:24.240> has<00:52:24.360> done,<00:52:24.680> but<00:52:25.560> you<00:52:25.720> can" + }, + { + "start": 3145.95, + "duration": 0.0, + "text": "you know, system has done, but you can" + }, + { + "start": 3145.96, + "duration": 0.0, + "text": "you know, system has done, but you can feed<00:52:26.280> any<00:52:26.480> image<00:52:26.760> into<00:52:27.360> an<00:52:27.560> LM<00:52:27.880> right<00:52:28.040> now<00:52:28.160> and" + }, + { + "start": 3148.23, + "duration": 0.0, + "text": "feed any image into an LM right now and" + }, + { + "start": 3148.24, + "duration": 0.0, + "text": "feed any image into an LM right now and it'll<00:52:28.400> tell<00:52:28.520> you<00:52:28.600> what<00:52:28.720> it<00:52:28.840> is.<00:52:29.360> Almost<00:52:29.600> like" + }, + { + "start": 3149.79, + "duration": 0.0, + "text": "it'll tell you what it is. Almost like" + }, + { + "start": 3149.8, + "duration": 0.0, + "text": "it'll tell you what it is. Almost like exactly<00:52:30.200> like<00:52:30.360> what<00:52:30.480> a<00:52:30.520> person<00:52:30.800> would<00:52:30.920> say." + }, + { + "start": 3151.75, + "duration": 0.0, + "text": "exactly like what a person would say." + }, + { + "start": 3151.76, + "duration": 0.0, + "text": "exactly like what a person would say. Obviously<00:52:32.200> it's<00:52:32.320> not<00:52:32.520> perfect<00:52:32.840> and<00:52:32.920> it" + }, + { + "start": 3152.95, + "duration": 0.0, + "text": "Obviously it's not perfect and it" + }, + { + "start": 3152.96, + "duration": 0.0, + "text": "Obviously it's not perfect and it constantly<00:52:33.360> fails,<00:52:34.120> um" + }, + { + "start": 3155.07, + "duration": 0.0, + "text": "constantly fails, um" + }, + { + "start": 3155.08, + "duration": 0.0, + "text": "constantly fails, um but<00:52:35.320> that<00:52:35.560> error<00:52:35.800> is<00:52:35.920> always<00:52:36.200> decreasing<00:52:36.800> as" + }, + { + "start": 3156.95, + "duration": 0.0, + "text": "but that error is always decreasing as" + }, + { + "start": 3156.96, + "duration": 0.0, + "text": "but that error is always decreasing as they,<00:52:37.560> you<00:52:37.640> know," + }, + { + "start": 3158.51, + "duration": 0.0, + "text": "they, you know," + }, + { + "start": 3158.52, + "duration": 0.0, + "text": "they, you know, fix<00:52:38.720> bugs,<00:52:39.160> they<00:52:39.400> they<00:52:39.560> improve<00:52:39.880> the<00:52:40.080> training" + }, + { + "start": 3160.39, + "duration": 0.0, + "text": "fix bugs, they they improve the training" + }, + { + "start": 3160.4, + "duration": 0.0, + "text": "fix bugs, they they improve the training pipeline,<00:52:40.920> they<00:52:41.400> improve<00:52:41.720> the<00:52:41.800> data<00:52:42.000> they" + }, + { + "start": 3162.11, + "duration": 0.0, + "text": "pipeline, they improve the data they" + }, + { + "start": 3162.12, + "duration": 0.0, + "text": "pipeline, they improve the data they give<00:52:42.320> it,<00:52:42.480> and<00:52:42.560> then<00:52:42.640> it's<00:52:42.760> always" + }, + { + "start": 3163.55, + "duration": 0.0, + "text": "give it, and then it's always" + }, + { + "start": 3163.56, + "duration": 0.0, + "text": "give it, and then it's always approaching<00:52:44.600> something<00:52:44.960> that<00:52:45.120> seems<00:52:45.680> like<00:52:45.840> it" + }, + { + "start": 3165.95, + "duration": 0.0, + "text": "approaching something that seems like it" + }, + { + "start": 3165.96, + "duration": 0.0, + "text": "approaching something that seems like it will<00:52:46.080> apply<00:52:46.480> in<00:52:46.680> any<00:52:46.920> situation.<00:52:48.120> Um<00:52:48.880> of" + }, + { + "start": 3168.95, + "duration": 0.0, + "text": "will apply in any situation. Um of" + }, + { + "start": 3168.96, + "duration": 0.0, + "text": "will apply in any situation. Um of course<00:52:49.120> it's<00:52:49.200> very<00:52:49.360> far<00:52:49.520> from<00:52:49.640> that<00:52:49.840> right" + }, + { + "start": 3170.03, + "duration": 0.0, + "text": "course it's very far from that right" + }, + { + "start": 3170.04, + "duration": 0.0, + "text": "course it's very far from that right now.<00:52:50.200> I'm<00:52:50.280> not<00:52:50.520> meaning<00:52:50.720> to<00:52:50.800> claim<00:52:51.040> that<00:52:51.200> like" + }, + { + "start": 3171.91, + "duration": 0.0, + "text": "now. I'm not meaning to claim that like" + }, + { + "start": 3171.92, + "duration": 0.0, + "text": "now. I'm not meaning to claim that like you<00:52:51.960> could<00:52:52.040> take<00:52:52.200> any<00:52:52.360> image<00:52:52.640> and<00:52:52.720> then<00:52:52.800> it'll" + }, + { + "start": 3172.95, + "duration": 0.0, + "text": "you could take any image and then it'll" + }, + { + "start": 3172.96, + "duration": 0.0, + "text": "you could take any image and then it'll perfectly<00:52:53.360> recognize<00:52:53.720> what's<00:52:53.880> in<00:52:53.960> it.<00:52:54.040> You" + }, + { + "start": 3174.11, + "duration": 0.0, + "text": "perfectly recognize what's in it. You" + }, + { + "start": 3174.12, + "duration": 0.0, + "text": "perfectly recognize what's in it. You can<00:52:54.240> fool<00:52:54.400> them<00:52:54.560> all<00:52:54.680> the<00:52:54.760> time.<00:52:55.680> Um<00:52:56.240> it's<00:52:56.480> just" + }, + { + "start": 3176.71, + "duration": 0.0, + "text": "can fool them all the time. Um it's just" + }, + { + "start": 3176.72, + "duration": 0.0, + "text": "can fool them all the time. Um it's just that<00:52:56.880> like" + }, + { + "start": 3178.83, + "duration": 0.0, + "text": "that like" + }, + { + "start": 3178.84, + "duration": 0.0, + "text": "that like it's<00:52:59.000> the<00:52:59.080> entire<00:52:59.400> internet,<00:53:00.000> right?<00:53:00.360> I<00:53:00.440> just" + }, + { + "start": 3180.63, + "duration": 0.0, + "text": "it's the entire internet, right? I just" + }, + { + "start": 3180.64, + "duration": 0.0, + "text": "it's the entire internet, right? I just that's<00:53:00.880> the<00:53:01.080> thing<00:53:01.320> that<00:53:01.440> I<00:53:01.480> keep<00:53:01.680> thinking" + }, + { + "start": 3181.95, + "duration": 0.0, + "text": "that's the thing that I keep thinking" + }, + { + "start": 3181.96, + "duration": 0.0, + "text": "that's the thing that I keep thinking about.<00:53:02.480> You<00:53:02.600> know,<00:53:02.720> it's<00:53:03.000> that's<00:53:03.200> just<00:53:03.360> a<00:53:03.400> lot." + }, + { + "start": 3183.99, + "duration": 0.0, + "text": "about. You know, it's that's just a lot." + }, + { + "start": 3184.0, + "duration": 0.0, + "text": "about. You know, it's that's just a lot. It's<00:53:04.160> hard<00:53:04.280> for<00:53:04.360> me<00:53:04.440> to<00:53:04.520> even<00:53:04.640> comprehend<00:53:05.080> what" + }, + { + "start": 3185.15, + "duration": 0.0, + "text": "It's hard for me to even comprehend what" + }, + { + "start": 3185.16, + "duration": 0.0, + "text": "It's hard for me to even comprehend what the<00:53:05.240> limits<00:53:05.520> of<00:53:05.600> that<00:53:05.760> is." + }, + { + "start": 3187.03, + "duration": 0.0, + "text": "the limits of that is." + }, + { + "start": 3187.04, + "duration": 0.0, + "text": "the limits of that is. Um" + }, + { + "start": 3188.59, + "duration": 0.0, + "text": "Um" + }, + { + "start": 3188.6, + "duration": 0.0, + "text": "Um So,<00:53:08.880> even<00:53:09.080> if<00:53:09.160> we<00:53:09.240> don't<00:53:09.400> end<00:53:09.520> up<00:53:09.640> in<00:53:09.720> a<00:53:09.760> world" + }, + { + "start": 3189.99, + "duration": 0.0, + "text": "So, even if we don't end up in a world" + }, + { + "start": 3190.0, + "duration": 0.0, + "text": "So, even if we don't end up in a world where<00:53:10.080> the<00:53:10.200> AI<00:53:10.480> is<00:53:10.600> able<00:53:10.760> to<00:53:10.880> do<00:53:11.040> our<00:53:11.200> job<00:53:11.520> in" + }, + { + "start": 3191.71, + "duration": 0.0, + "text": "where the AI is able to do our job in" + }, + { + "start": 3191.72, + "duration": 0.0, + "text": "where the AI is able to do our job in in,<00:53:12.040> you<00:53:12.120> know,<00:53:12.240> 10<00:53:12.480> years<00:53:12.640> or<00:53:12.720> less," + }, + { + "start": 3192.91, + "duration": 0.0, + "text": "in, you know, 10 years or less," + }, + { + "start": 3192.92, + "duration": 0.0, + "text": "in, you know, 10 years or less, whatever.<00:53:13.360> Um<00:53:13.760> how<00:53:13.880> likely<00:53:14.160> do<00:53:14.280> you<00:53:14.360> think<00:53:14.520> it" + }, + { + "start": 3194.63, + "duration": 0.0, + "text": "whatever. Um how likely do you think it" + }, + { + "start": 3194.64, + "duration": 0.0, + "text": "whatever. Um how likely do you think it is<00:53:14.760> that<00:53:14.880> we're<00:53:15.000> heading<00:53:15.280> towards<00:53:15.720> a<00:53:15.840> dead" + }, + { + "start": 3196.07, + "duration": 0.0, + "text": "is that we're heading towards a dead" + }, + { + "start": 3196.08, + "duration": 0.0, + "text": "is that we're heading towards a dead internet<00:53:16.480> style<00:53:16.800> problem<00:53:17.280> where<00:53:17.440> the<00:53:17.600> noise" + }, + { + "start": 3198.07, + "duration": 0.0, + "text": "internet style problem where the noise" + }, + { + "start": 3198.08, + "duration": 0.0, + "text": "internet style problem where the noise generated<00:53:18.560> or<00:53:18.720> the<00:53:19.000> noise<00:53:19.280> created<00:53:19.520> by<00:53:19.600> these" + }, + { + "start": 3199.79, + "duration": 0.0, + "text": "generated or the noise created by these" + }, + { + "start": 3199.8, + "duration": 0.0, + "text": "generated or the noise created by these generated<00:53:20.280> AIs<00:53:20.760> is<00:53:21.040> a<00:53:21.120> problem<00:53:21.600> to<00:53:21.720> our<00:53:21.840> own" + }, + { + "start": 3201.99, + "duration": 0.0, + "text": "generated AIs is a problem to our own" + }, + { + "start": 3202.0, + "duration": 0.0, + "text": "generated AIs is a problem to our own ability<00:53:22.400> to<00:53:22.480> communicate?<00:53:23.280> I<00:53:23.360> mean<00:53:23.560> really" + }, + { + "start": 3203.83, + "duration": 0.0, + "text": "ability to communicate? I mean really" + }, + { + "start": 3203.84, + "duration": 0.0, + "text": "ability to communicate? I mean really already<00:53:24.280> it<00:53:24.360> feels<00:53:24.520> like<00:53:24.640> so<00:53:24.720> much<00:53:24.880> of<00:53:24.960> the" + }, + { + "start": 3205.07, + "duration": 0.0, + "text": "already it feels like so much of the" + }, + { + "start": 3205.08, + "duration": 0.0, + "text": "already it feels like so much of the internet<00:53:25.480> is<00:53:25.760> on<00:53:26.080> like<00:53:26.240> private<00:53:26.600> discords" + }, + { + "start": 3207.11, + "duration": 0.0, + "text": "internet is on like private discords" + }, + { + "start": 3207.12, + "duration": 0.0, + "text": "internet is on like private discords that<00:53:27.320> aren't,<00:53:27.800> you<00:53:27.840> know,<00:53:28.480> on<00:53:28.640> these<00:53:28.800> massive" + }, + { + "start": 3209.11, + "duration": 0.0, + "text": "that aren't, you know, on these massive" + }, + { + "start": 3209.12, + "duration": 0.0, + "text": "that aren't, you know, on these massive public<00:53:29.400> forums<00:53:29.920> and," + }, + { + "start": 3210.79, + "duration": 0.0, + "text": "public forums and," + }, + { + "start": 3210.8, + "duration": 0.0, + "text": "public forums and, you<00:53:30.840> know,<00:53:31.240> linked<00:53:31.520> with<00:53:31.640> each<00:53:31.800> other.<00:53:32.000> I" + }, + { + "start": 3212.07, + "duration": 0.0, + "text": "you know, linked with each other. I" + }, + { + "start": 3212.08, + "duration": 0.0, + "text": "you know, linked with each other. I think<00:53:32.240> that<00:53:32.720> people<00:53:32.960> are<00:53:33.080> already<00:53:33.400> kind<00:53:33.560> of" + }, + { + "start": 3213.67, + "duration": 0.0, + "text": "think that people are already kind of" + }, + { + "start": 3213.68, + "duration": 0.0, + "text": "think that people are already kind of receding<00:53:34.080> away<00:53:34.400> from<00:53:34.600> this<00:53:34.800> like<00:53:35.400> how<00:53:35.520> it<00:53:35.600> used" + }, + { + "start": 3215.79, + "duration": 0.0, + "text": "receding away from this like how it used" + }, + { + "start": 3215.8, + "duration": 0.0, + "text": "receding away from this like how it used to<00:53:35.880> be<00:53:36.000> where<00:53:36.080> it<00:53:36.160> was<00:53:36.280> like<00:53:36.480> a<00:53:36.880> like<00:53:37.080> a<00:53:37.840> massive" + }, + { + "start": 3218.19, + "duration": 0.0, + "text": "to be where it was like a like a massive" + }, + { + "start": 3218.2, + "duration": 0.0, + "text": "to be where it was like a like a massive graph<00:53:38.520> of<00:53:38.640> forum<00:53:38.920> posts<00:53:39.320> or<00:53:39.400> whatever,<00:53:39.680> right?" + }, + { + "start": 3219.99, + "duration": 0.0, + "text": "graph of forum posts or whatever, right?" + }, + { + "start": 3220.0, + "duration": 0.0, + "text": "graph of forum posts or whatever, right? Like<00:53:40.240> um" + }, + { + "start": 3222.19, + "duration": 0.0, + "text": "Like um" + }, + { + "start": 3222.2, + "duration": 0.0, + "text": "Like um I<00:53:42.240> don't<00:53:42.360> know<00:53:42.440> much<00:53:42.600> about<00:53:42.720> that.<00:53:42.840> I<00:53:42.960> mean" + }, + { + "start": 3223.07, + "duration": 0.0, + "text": "I don't know much about that. I mean" + }, + { + "start": 3223.08, + "duration": 0.0, + "text": "I don't know much about that. I mean this<00:53:43.320> this<00:53:43.560> is<00:53:43.720> the<00:53:43.840> internet<00:53:44.160> and<00:53:44.280> it's<00:53:44.400> not" + }, + { + "start": 3224.59, + "duration": 0.0, + "text": "this this is the internet and it's not" + }, + { + "start": 3224.6, + "duration": 0.0, + "text": "this this is the internet and it's not dead,<00:53:44.840> I<00:53:44.920> guess.<00:53:45.240> So,<00:53:45.440> I<00:53:45.640> it's<00:53:45.800> not<00:53:46.000> I<00:53:46.040> don't" + }, + { + "start": 3226.19, + "duration": 0.0, + "text": "dead, I guess. So, I it's not I don't" + }, + { + "start": 3226.2, + "duration": 0.0, + "text": "dead, I guess. So, I it's not I don't know<00:53:46.360> like<00:53:47.040> I<00:53:47.120> mean<00:53:47.800> maybe." + }, + { + "start": 3229.07, + "duration": 0.0, + "text": "know like I mean maybe." + }, + { + "start": 3229.08, + "duration": 0.0, + "text": "know like I mean maybe. People<00:53:49.360> I<00:53:49.720> actually<00:53:50.360> Well,<00:53:50.480> so<00:53:50.800> LLMs<00:53:51.280> do<00:53:51.480> seem" + }, + { + "start": 3231.67, + "duration": 0.0, + "text": "People I actually Well, so LLMs do seem" + }, + { + "start": 3231.68, + "duration": 0.0, + "text": "People I actually Well, so LLMs do seem like<00:53:51.800> they<00:53:52.040> kind<00:53:52.320> of<00:53:52.400> defeat<00:53:52.720> the<00:53:52.800> Turing<00:53:53.040> test" + }, + { + "start": 3233.79, + "duration": 0.0, + "text": "like they kind of defeat the Turing test" + }, + { + "start": 3233.8, + "duration": 0.0, + "text": "like they kind of defeat the Turing test and<00:53:54.080> I<00:53:54.400> I<00:53:54.480> seem<00:53:54.720> to<00:53:54.960> I<00:53:55.040> think<00:53:55.280> I'm<00:53:55.360> remembering" + }, + { + "start": 3235.87, + "duration": 0.0, + "text": "and I I seem to I think I'm remembering" + }, + { + "start": 3235.88, + "duration": 0.0, + "text": "and I I seem to I think I'm remembering there<00:53:55.960> was<00:53:56.080> some<00:53:56.480> news<00:53:56.760> article.<00:53:57.120> Somebody" + }, + { + "start": 3237.35, + "duration": 0.0, + "text": "there was some news article. Somebody" + }, + { + "start": 3237.36, + "duration": 0.0, + "text": "there was some news article. Somebody maybe<00:53:57.600> in<00:53:57.720> like<00:53:57.880> Twitch<00:53:58.080> chat<00:53:58.280> or<00:53:58.320> whatever" + }, + { + "start": 3238.55, + "duration": 0.0, + "text": "maybe in like Twitch chat or whatever" + }, + { + "start": 3238.56, + "duration": 0.0, + "text": "maybe in like Twitch chat or whatever can<00:53:58.720> find<00:53:58.960> it<00:53:59.520> where<00:54:00.040> I<00:54:00.120> think<00:54:00.320> some<00:54:00.640> lab" + }, + { + "start": 3241.07, + "duration": 0.0, + "text": "can find it where I think some lab" + }, + { + "start": 3241.08, + "duration": 0.0, + "text": "can find it where I think some lab basically<00:54:01.560> plugged<00:54:01.880> an<00:54:02.000> LLM<00:54:02.360> into<00:54:02.520> like<00:54:02.640> a" + }, + { + "start": 3242.67, + "duration": 0.0, + "text": "basically plugged an LLM into like a" + }, + { + "start": 3242.68, + "duration": 0.0, + "text": "basically plugged an LLM into like a subreddit<00:54:03.600> and<00:54:03.760> just<00:54:03.880> had<00:54:04.000> it<00:54:04.080> be<00:54:04.160> a<00:54:04.200> commenter" + }, + { + "start": 3244.79, + "duration": 0.0, + "text": "subreddit and just had it be a commenter" + }, + { + "start": 3244.8, + "duration": 0.0, + "text": "subreddit and just had it be a commenter and<00:54:04.880> it<00:54:04.960> got<00:54:05.080> a<00:54:05.120> lot<00:54:05.280> of<00:54:05.360> karma<00:54:05.640> and<00:54:05.760> people" + }, + { + "start": 3245.95, + "duration": 0.0, + "text": "and it got a lot of karma and people" + }, + { + "start": 3245.96, + "duration": 0.0, + "text": "and it got a lot of karma and people started<00:54:06.240> to<00:54:06.360> like<00:54:06.760> relate<00:54:07.040> to<00:54:07.160> it<00:54:07.400> and<00:54:07.520> like" + }, + { + "start": 3247.91, + "duration": 0.0, + "text": "started to like relate to it and like" + }, + { + "start": 3247.92, + "duration": 0.0, + "text": "started to like relate to it and like oh,<00:54:08.040> this<00:54:08.160> is<00:54:08.240> a<00:54:08.280> good<00:54:08.400> post<00:54:08.640> or<00:54:08.720> he's<00:54:08.800> a<00:54:08.840> good" + }, + { + "start": 3248.99, + "duration": 0.0, + "text": "oh, this is a good post or he's a good" + }, + { + "start": 3249.0, + "duration": 0.0, + "text": "oh, this is a good post or he's a good guy.<00:54:09.320> I<00:54:09.360> like<00:54:09.600> this<00:54:09.720> guy.<00:54:10.440> And<00:54:10.560> then<00:54:10.680> it<00:54:10.760> was" + }, + { + "start": 3250.87, + "duration": 0.0, + "text": "guy. I like this guy. And then it was" + }, + { + "start": 3250.88, + "duration": 0.0, + "text": "guy. I like this guy. And then it was like<00:54:11.000> uncovered<00:54:11.480> that<00:54:11.600> it<00:54:11.680> was<00:54:11.800> really<00:54:12.040> like<00:54:12.240> a" + }, + { + "start": 3252.43, + "duration": 0.0, + "text": "like uncovered that it was really like a" + }, + { + "start": 3252.44, + "duration": 0.0, + "text": "like uncovered that it was really like a like<00:54:12.560> a<00:54:12.640> false<00:54:12.960> flag<00:54:13.240> human<00:54:13.480> being<00:54:13.720> or" + }, + { + "start": 3253.75, + "duration": 0.0, + "text": "like a false flag human being or" + }, + { + "start": 3253.76, + "duration": 0.0, + "text": "like a false flag human being or whatever<00:54:14.040> and<00:54:14.160> like<00:54:14.800> they<00:54:14.920> were<00:54:15.040> they<00:54:15.520> they" + }, + { + "start": 3255.67, + "duration": 0.0, + "text": "whatever and like they were they they" + }, + { + "start": 3255.68, + "duration": 0.0, + "text": "whatever and like they were they they basically<00:54:16.120> were<00:54:16.240> doing<00:54:16.520> a<00:54:16.560> study<00:54:17.000> on<00:54:17.200> the" + }, + { + "start": 3257.39, + "duration": 0.0, + "text": "basically were doing a study on the" + }, + { + "start": 3257.4, + "duration": 0.0, + "text": "basically were doing a study on the subreddit<00:54:17.840> involuntarily.<00:54:18.560> I<00:54:18.600> think<00:54:19.000> people" + }, + { + "start": 3259.23, + "duration": 0.0, + "text": "subreddit involuntarily. I think people" + }, + { + "start": 3259.24, + "duration": 0.0, + "text": "subreddit involuntarily. I think people got<00:54:19.360> very<00:54:19.520> mad<00:54:19.720> at<00:54:19.800> that<00:54:20.000> cuz<00:54:20.120> it<00:54:20.200> was<00:54:20.320> like" + }, + { + "start": 3260.95, + "duration": 0.0, + "text": "got very mad at that cuz it was like" + }, + { + "start": 3260.96, + "duration": 0.0, + "text": "got very mad at that cuz it was like that's<00:54:21.120> probably<00:54:21.320> not<00:54:21.480> ethical<00:54:21.840> to<00:54:21.920> do<00:54:22.040> human" + }, + { + "start": 3262.27, + "duration": 0.0, + "text": "that's probably not ethical to do human" + }, + { + "start": 3262.28, + "duration": 0.0, + "text": "that's probably not ethical to do human beings.<00:54:22.960> Most<00:54:23.240> of<00:54:23.320> what's<00:54:23.480> happening<00:54:23.760> in<00:54:23.840> the" + }, + { + "start": 3263.87, + "duration": 0.0, + "text": "beings. Most of what's happening in the" + }, + { + "start": 3263.88, + "duration": 0.0, + "text": "beings. Most of what's happening in the modern<00:54:24.120> world<00:54:24.320> is<00:54:24.400> not<00:54:24.560> ethical<00:54:24.880> to<00:54:24.960> do<00:54:25.040> human" + }, + { + "start": 3265.23, + "duration": 0.0, + "text": "modern world is not ethical to do human" + }, + { + "start": 3265.24, + "duration": 0.0, + "text": "modern world is not ethical to do human beings,<00:54:25.520> honestly.<00:54:26.480> I<00:54:26.640> I<00:54:26.680> don't<00:54:26.880> think<00:54:27.160> I" + }, + { + "start": 3267.19, + "duration": 0.0, + "text": "beings, honestly. I I don't think I" + }, + { + "start": 3267.2, + "duration": 0.0, + "text": "beings, honestly. I I don't think I don't<00:54:27.320> think<00:54:27.440> LLMs<00:54:27.680> are<00:54:27.760> that<00:54:27.880> crazy<00:54:28.160> compared" + }, + { + "start": 3268.47, + "duration": 0.0, + "text": "don't think LLMs are that crazy compared" + }, + { + "start": 3268.48, + "duration": 0.0, + "text": "don't think LLMs are that crazy compared to<00:54:28.560> like<00:54:28.760> YouTube.<00:54:29.520> Like<00:54:29.800> you<00:54:29.840> just<00:54:30.280> show" + }, + { + "start": 3270.87, + "duration": 0.0, + "text": "to like YouTube. Like you just show" + }, + { + "start": 3270.88, + "duration": 0.0, + "text": "to like YouTube. Like you just show infant<00:54:31.280> pixels<00:54:31.600> to<00:54:31.680> a<00:54:31.720> person<00:54:31.960> like<00:54:32.120> this<00:54:32.240> is" + }, + { + "start": 3272.35, + "duration": 0.0, + "text": "infant pixels to a person like this is" + }, + { + "start": 3272.36, + "duration": 0.0, + "text": "infant pixels to a person like this is probably<00:54:32.560> not<00:54:32.760> good<00:54:32.920> for<00:54:33.000> them.<00:54:33.200> I<00:54:33.240> don't<00:54:33.360> know" + }, + { + "start": 3273.47, + "duration": 0.0, + "text": "probably not good for them. I don't know" + }, + { + "start": 3273.48, + "duration": 0.0, + "text": "probably not good for them. I don't know like" + }, + { + "start": 3274.39, + "duration": 0.0, + "text": "like" + }, + { + "start": 3274.4, + "duration": 0.0, + "text": "like Um<00:54:35.360> I<00:54:35.400> don't<00:54:35.560> really<00:54:36.160> you<00:54:36.240> know,<00:54:36.400> things<00:54:36.640> are" + }, + { + "start": 3276.75, + "duration": 0.0, + "text": "Um I don't really you know, things are" + }, + { + "start": 3276.76, + "duration": 0.0, + "text": "Um I don't really you know, things are already<00:54:37.040> so<00:54:37.440> so<00:54:37.760> so<00:54:38.080> far<00:54:38.360> from<00:54:38.560> what<00:54:38.680> is<00:54:39.000> like" + }, + { + "start": 3279.35, + "duration": 0.0, + "text": "already so so so far from what is like" + }, + { + "start": 3279.36, + "duration": 0.0, + "text": "already so so so far from what is like healthy<00:54:39.760> for<00:54:39.960> a<00:54:40.000> human<00:54:40.200> being.<00:54:40.520> It's<00:54:40.640> like" + }, + { + "start": 3281.79, + "duration": 0.0, + "text": "healthy for a human being. It's like" + }, + { + "start": 3281.8, + "duration": 0.0, + "text": "healthy for a human being. It's like What<00:54:41.880> are<00:54:41.960> you<00:54:42.000> going<00:54:42.120> to<00:54:42.400> What<00:54:42.480> do<00:54:42.560> you<00:54:42.600> want?" + }, + { + "start": 3282.99, + "duration": 0.0, + "text": "What are you going to What do you want?" + }, + { + "start": 3283.0, + "duration": 0.0, + "text": "What are you going to What do you want? Like<00:54:43.760> you<00:54:43.880> just<00:54:44.400> like<00:54:44.560> it's<00:54:44.680> not<00:54:45.000> that<00:54:45.160> much" + }, + { + "start": 3285.39, + "duration": 0.0, + "text": "Like you just like it's not that much" + }, + { + "start": 3285.4, + "duration": 0.0, + "text": "Like you just like it's not that much worse<00:54:45.800> than<00:54:46.480> TikTok.<00:54:46.960> I<00:54:47.000> don't<00:54:47.120> know<00:54:47.200> what<00:54:47.320> to" + }, + { + "start": 3287.39, + "duration": 0.0, + "text": "worse than TikTok. I don't know what to" + }, + { + "start": 3287.4, + "duration": 0.0, + "text": "worse than TikTok. I don't know what to say.<00:54:48.080> TikTok<00:54:48.400> is<00:54:48.520> a<00:54:48.600> transformer.<00:54:49.360> It's" + }, + { + "start": 3290.51, + "duration": 0.0, + "text": "say. TikTok is a transformer. It's" + }, + { + "start": 3290.52, + "duration": 0.0, + "text": "say. TikTok is a transformer. It's whatever." + }, + { + "start": 3293.12, + "duration": 0.0, + "text": "So,<00:54:53.600> we<00:54:53.720> talked<00:54:53.920> about<00:54:54.040> this<00:54:54.160> over<00:54:54.280> breakfast" + }, + { + "start": 3294.67, + "duration": 0.0, + "text": "So, we talked about this over breakfast" + }, + { + "start": 3294.68, + "duration": 0.0, + "text": "So, we talked about this over breakfast yesterday,<00:54:55.160> but<00:54:55.520> like<00:54:55.760> LLMs<00:54:56.480> kind<00:54:56.640> of<00:54:56.720> cool," + }, + { + "start": 3297.19, + "duration": 0.0, + "text": "yesterday, but like LLMs kind of cool," + }, + { + "start": 3297.2, + "duration": 0.0, + "text": "yesterday, but like LLMs kind of cool, transformers<00:54:57.720> are<00:54:57.800> cool,<00:54:58.160> doing<00:54:58.400> something," + }, + { + "start": 3299.11, + "duration": 0.0, + "text": "transformers are cool, doing something," + }, + { + "start": 3299.12, + "duration": 0.0, + "text": "transformers are cool, doing something, but<00:54:59.480> like<00:54:59.840> there's<00:55:00.040> a<00:55:00.080> missing<00:55:00.400> piece<00:55:00.800> or" + }, + { + "start": 3300.87, + "duration": 0.0, + "text": "but like there's a missing piece or" + }, + { + "start": 3300.88, + "duration": 0.0, + "text": "but like there's a missing piece or maybe<00:55:01.120> many.<00:55:01.440> Do<00:55:01.520> you<00:55:01.600> have<00:55:01.760> any<00:55:02.000> theories" + }, + { + "start": 3302.47, + "duration": 0.0, + "text": "maybe many. Do you have any theories" + }, + { + "start": 3302.48, + "duration": 0.0, + "text": "maybe many. Do you have any theories what<00:55:02.640> those<00:55:02.800> might<00:55:02.960> be?" + }, + { + "start": 3303.83, + "duration": 0.0, + "text": "what those might be?" + }, + { + "start": 3303.84, + "duration": 0.0, + "text": "what those might be? Me<00:55:04.040> personally<00:55:04.640> as<00:55:04.840> a<00:55:04.960> hobbyist" + }, + { + "start": 3306.35, + "duration": 0.0, + "text": "Me personally as a hobbyist" + }, + { + "start": 3306.36, + "duration": 0.0, + "text": "Me personally as a hobbyist crazy<00:55:06.680> person?<00:55:07.120> Uh" + }, + { + "start": 3307.95, + "duration": 0.0, + "text": "crazy person? Uh" + }, + { + "start": 3307.96, + "duration": 0.0, + "text": "crazy person? Uh I<00:55:08.000> don't<00:55:08.120> know.<00:55:08.960> I<00:55:09.040> feel<00:55:09.240> like<00:55:09.960> if<00:55:10.200> I<00:55:10.320> were<00:55:10.480> to" + }, + { + "start": 3310.63, + "duration": 0.0, + "text": "I don't know. I feel like if I were to" + }, + { + "start": 3310.64, + "duration": 0.0, + "text": "I don't know. I feel like if I were to Everybody<00:55:11.000> so<00:55:11.160> every<00:55:11.600> every<00:55:11.920> single<00:55:12.160> deep" + }, + { + "start": 3312.35, + "duration": 0.0, + "text": "Everybody so every every single deep" + }, + { + "start": 3312.36, + "duration": 0.0, + "text": "Everybody so every every single deep learning<00:55:12.560> research<00:55:12.800> is<00:55:12.880> like<00:55:13.000> oh,<00:55:13.040> it's<00:55:13.120> going" + }, + { + "start": 3313.23, + "duration": 0.0, + "text": "learning research is like oh, it's going" + }, + { + "start": 3313.24, + "duration": 0.0, + "text": "learning research is like oh, it's going to<00:55:13.320> be<00:55:13.360> in<00:55:13.440> a<00:55:13.480> reinforcement<00:55:14.000> learning.<00:55:14.280> Like" + }, + { + "start": 3314.67, + "duration": 0.0, + "text": "to be in a reinforcement learning. Like" + }, + { + "start": 3314.68, + "duration": 0.0, + "text": "to be in a reinforcement learning. Like you<00:55:14.760> take<00:55:14.920> that<00:55:15.080> success<00:55:15.440> of<00:55:15.520> AlphaZero<00:55:16.080> where" + }, + { + "start": 3316.19, + "duration": 0.0, + "text": "you take that success of AlphaZero where" + }, + { + "start": 3316.2, + "duration": 0.0, + "text": "you take that success of AlphaZero where it<00:55:16.240> basically<00:55:16.600> can<00:55:16.760> master<00:55:17.080> any<00:55:17.280> system.<00:55:17.640> You" + }, + { + "start": 3317.71, + "duration": 0.0, + "text": "it basically can master any system. You" + }, + { + "start": 3317.72, + "duration": 0.0, + "text": "it basically can master any system. You just<00:55:17.840> like<00:55:17.960> apply<00:55:18.280> that<00:55:18.520> to<00:55:18.600> like<00:55:18.720> soft" + }, + { + "start": 3318.95, + "duration": 0.0, + "text": "just like apply that to like soft" + }, + { + "start": 3318.96, + "duration": 0.0, + "text": "just like apply that to like soft concepts.<00:55:19.440> Like<00:55:19.840> MuZero<00:55:20.440> it's<00:55:20.560> like<00:55:21.280> you" + }, + { + "start": 3321.35, + "duration": 0.0, + "text": "concepts. Like MuZero it's like you" + }, + { + "start": 3321.36, + "duration": 0.0, + "text": "concepts. Like MuZero it's like you know,<00:55:21.520> I<00:55:21.800> everything<00:55:22.200> everybody<00:55:22.600> is<00:55:22.760> like" + }, + { + "start": 3323.03, + "duration": 0.0, + "text": "know, I everything everybody is like" + }, + { + "start": 3323.04, + "duration": 0.0, + "text": "know, I everything everybody is like thinking<00:55:23.360> along<00:55:23.560> those<00:55:23.760> lines<00:55:24.080> of<00:55:24.160> like<00:55:24.280> we" + }, + { + "start": 3324.39, + "duration": 0.0, + "text": "thinking along those lines of like we" + }, + { + "start": 3324.4, + "duration": 0.0, + "text": "thinking along those lines of like we basically<00:55:24.840> need<00:55:25.000> computers<00:55:25.520> to" + }, + { + "start": 3325.75, + "duration": 0.0, + "text": "basically need computers to" + }, + { + "start": 3325.76, + "duration": 0.0, + "text": "basically need computers to automatically<00:55:26.360> generate<00:55:26.760> and<00:55:26.880> master<00:55:27.800> uh" + }, + { + "start": 3327.95, + "duration": 0.0, + "text": "automatically generate and master uh" + }, + { + "start": 3327.96, + "duration": 0.0, + "text": "automatically generate and master uh things<00:55:28.200> that<00:55:28.320> not<00:55:28.440> even<00:55:28.600> humans<00:55:28.840> have" + }, + { + "start": 3328.95, + "duration": 0.0, + "text": "things that not even humans have" + }, + { + "start": 3328.96, + "duration": 0.0, + "text": "things that not even humans have mastered<00:55:29.280> yet,<00:55:29.520> right?<00:55:29.760> Because<00:55:30.000> if<00:55:30.160> all" + }, + { + "start": 3330.31, + "duration": 0.0, + "text": "mastered yet, right? Because if all" + }, + { + "start": 3330.32, + "duration": 0.0, + "text": "mastered yet, right? Because if all you're<00:55:30.480> doing<00:55:30.880> is<00:55:31.040> basically<00:55:32.040> making<00:55:32.320> these" + }, + { + "start": 3332.47, + "duration": 0.0, + "text": "you're doing is basically making these" + }, + { + "start": 3332.48, + "duration": 0.0, + "text": "you're doing is basically making these models<00:55:33.160> uh<00:55:33.440> try<00:55:33.640> to<00:55:34.000> optimize<00:55:34.400> and<00:55:34.520> find" + }, + { + "start": 3334.75, + "duration": 0.0, + "text": "models uh try to optimize and find" + }, + { + "start": 3334.76, + "duration": 0.0, + "text": "models uh try to optimize and find patterns<00:55:35.280> like<00:55:35.520> that<00:55:35.640> are<00:55:35.920> more<00:55:36.080> and<00:55:36.160> more" + }, + { + "start": 3336.23, + "duration": 0.0, + "text": "patterns like that are more and more" + }, + { + "start": 3336.24, + "duration": 0.0, + "text": "patterns like that are more and more closely<00:55:36.520> match<00:55:36.720> human<00:55:36.880> beings,<00:55:37.680> they'll" + }, + { + "start": 3337.75, + "duration": 0.0, + "text": "closely match human beings, they'll" + }, + { + "start": 3337.76, + "duration": 0.0, + "text": "closely match human beings, they'll never<00:55:37.960> exceed<00:55:38.240> the<00:55:38.320> human<00:55:38.520> baseline,<00:55:39.040> right?" + }, + { + "start": 3339.27, + "duration": 0.0, + "text": "never exceed the human baseline, right?" + }, + { + "start": 3339.28, + "duration": 0.0, + "text": "never exceed the human baseline, right? I<00:55:39.320> mean,<00:55:39.480> they<00:55:39.560> might<00:55:40.160> be<00:55:40.240> an<00:55:40.360> ensemble<00:55:40.880> of<00:55:41.000> all" + }, + { + "start": 3341.07, + "duration": 0.0, + "text": "I mean, they might be an ensemble of all" + }, + { + "start": 3341.08, + "duration": 0.0, + "text": "I mean, they might be an ensemble of all the<00:55:41.160> best<00:55:41.360> humans<00:55:41.920> and,<00:55:42.200> you<00:55:42.280> know,<00:55:42.360> they" + }, + { + "start": 3342.47, + "duration": 0.0, + "text": "the best humans and, you know, they" + }, + { + "start": 3342.48, + "duration": 0.0, + "text": "the best humans and, you know, they might<00:55:42.600> have<00:55:42.720> be<00:55:42.800> a<00:55:42.840> genius<00:55:43.160> in<00:55:43.240> every<00:55:43.440> single" + }, + { + "start": 3343.67, + "duration": 0.0, + "text": "might have be a genius in every single" + }, + { + "start": 3343.68, + "duration": 0.0, + "text": "might have be a genius in every single field<00:55:44.000> or<00:55:44.040> whatever,<00:55:44.320> but<00:55:44.440> they're<00:55:44.560> never" + }, + { + "start": 3344.75, + "duration": 0.0, + "text": "field or whatever, but they're never" + }, + { + "start": 3344.76, + "duration": 0.0, + "text": "field or whatever, but they're never going<00:55:44.880> to<00:55:44.960> surpass<00:55:45.480> human<00:55:45.760> genius,<00:55:46.480> which<00:55:46.760> you" + }, + { + "start": 3346.87, + "duration": 0.0, + "text": "going to surpass human genius, which you" + }, + { + "start": 3346.88, + "duration": 0.0, + "text": "going to surpass human genius, which you would<00:55:47.000> kind<00:55:47.160> of<00:55:47.280> want<00:55:47.520> if<00:55:47.640> you<00:55:47.720> wanted<00:55:47.920> to<00:55:48.000> be<00:55:48.080> a" + }, + { + "start": 3348.15, + "duration": 0.0, + "text": "would kind of want if you wanted to be a" + }, + { + "start": 3348.16, + "duration": 0.0, + "text": "would kind of want if you wanted to be a god-like<00:55:48.600> being<00:55:48.880> that<00:55:49.040> replaces<00:55:49.760> mankind," + }, + { + "start": 3350.51, + "duration": 0.0, + "text": "god-like being that replaces mankind," + }, + { + "start": 3350.52, + "duration": 0.0, + "text": "god-like being that replaces mankind, which<00:55:50.680> is<00:55:50.880> some<00:55:51.080> people's<00:55:51.400> goal,<00:55:51.560> not<00:55:51.720> mine.<00:55:52.080> I" + }, + { + "start": 3352.11, + "duration": 0.0, + "text": "which is some people's goal, not mine. I" + }, + { + "start": 3352.12, + "duration": 0.0, + "text": "which is some people's goal, not mine. I don't<00:55:52.320> I'm<00:55:52.400> not<00:55:52.560> into<00:55:52.720> that,<00:55:52.960> but<00:55:53.080> some<00:55:53.240> people" + }, + { + "start": 3353.43, + "duration": 0.0, + "text": "don't I'm not into that, but some people" + }, + { + "start": 3353.44, + "duration": 0.0, + "text": "don't I'm not into that, but some people like<00:55:53.600> that.<00:55:53.920> I<00:55:53.960> don't<00:55:54.120> know.<00:55:54.240> I<00:55:54.360> think<00:55:54.520> it's" + }, + { + "start": 3354.59, + "duration": 0.0, + "text": "like that. I don't know. I think it's" + }, + { + "start": 3354.6, + "duration": 0.0, + "text": "like that. I don't know. I think it's probably<00:55:54.800> bad." + }, + { + "start": 3355.79, + "duration": 0.0, + "text": "probably bad." + }, + { + "start": 3355.8, + "duration": 0.0, + "text": "probably bad. Um" + }, + { + "start": 3356.87, + "duration": 0.0, + "text": "Um" + }, + { + "start": 3356.88, + "duration": 0.0, + "text": "Um But<00:55:56.960> like<00:55:57.240> like" + }, + { + "start": 3358.43, + "duration": 0.0, + "text": "But like like" + }, + { + "start": 3358.44, + "duration": 0.0, + "text": "But like like random<00:55:58.960> thought<00:55:59.240> like<00:55:59.400> like<00:55:59.600> what<00:55:59.720> what<00:55:59.840> what" + }, + { + "start": 3359.99, + "duration": 0.0, + "text": "random thought like like what what what" + }, + { + "start": 3360.0, + "duration": 0.0, + "text": "random thought like like what what what what<00:56:00.120> could<00:56:00.280> that<00:56:00.480> be<00:56:00.640> like?" + }, + { + "start": 3360.91, + "duration": 0.0, + "text": "what could that be like?" + }, + { + "start": 3360.92, + "duration": 0.0, + "text": "what could that be like? >> Reinforce<00:56:01.600> you<00:56:01.720> need<00:56:01.920> to<00:56:02.000> have<00:56:02.200> some<00:56:02.480> way<00:56:02.840> of" + }, + { + "start": 3363.35, + "duration": 0.0, + "text": ">> Reinforce you need to have some way of" + }, + { + "start": 3363.36, + "duration": 0.0, + "text": ">> Reinforce you need to have some way of operating<00:56:03.800> on<00:56:03.880> these<00:56:04.040> concepts<00:56:05.080> um<00:56:05.560> in<00:56:05.680> a<00:56:05.760> way," + }, + { + "start": 3365.99, + "duration": 0.0, + "text": "operating on these concepts um in a way," + }, + { + "start": 3366.0, + "duration": 0.0, + "text": "operating on these concepts um in a way, some<00:56:06.160> kind<00:56:06.320> of<00:56:06.400> like<00:56:06.520> gym,<00:56:06.920> some<00:56:07.080> kind<00:56:07.240> of<00:56:07.320> like" + }, + { + "start": 3367.47, + "duration": 0.0, + "text": "some kind of like gym, some kind of like" + }, + { + "start": 3367.48, + "duration": 0.0, + "text": "some kind of like gym, some kind of like imaginary<00:56:08.040> like<00:56:08.720> it<00:56:08.840> needs<00:56:09.040> a<00:56:09.120> world<00:56:09.360> model," + }, + { + "start": 3369.83, + "duration": 0.0, + "text": "imaginary like it needs a world model," + }, + { + "start": 3369.84, + "duration": 0.0, + "text": "imaginary like it needs a world model, right?<00:56:10.080> That's<00:56:10.240> what<00:56:10.520> OpenAI<00:56:10.920> say<00:56:11.080> like,<00:56:11.240> \"Oh," + }, + { + "start": 3371.27, + "duration": 0.0, + "text": "right? That's what OpenAI say like, \"Oh," + }, + { + "start": 3371.28, + "duration": 0.0, + "text": "right? That's what OpenAI say like, \"Oh, we<00:56:11.360> need<00:56:11.480> a<00:56:11.560> world<00:56:11.800> model.\"<00:56:12.040> It's<00:56:12.120> like,<00:56:12.240> \"We" + }, + { + "start": 3372.31, + "duration": 0.0, + "text": "we need a world model.\" It's like, \"We" + }, + { + "start": 3372.32, + "duration": 0.0, + "text": "we need a world model.\" It's like, \"We need<00:56:12.440> some<00:56:12.560> way<00:56:13.040> We<00:56:13.120> need<00:56:13.240> a<00:56:13.280> transformer<00:56:13.760> not" + }, + { + "start": 3373.91, + "duration": 0.0, + "text": "need some way We need a transformer not" + }, + { + "start": 3373.92, + "duration": 0.0, + "text": "need some way We need a transformer not on<00:56:14.040> text.\"<00:56:14.360> That's<00:56:14.480> why<00:56:14.640> So," + }, + { + "start": 3375.55, + "duration": 0.0, + "text": "on text.\" That's why So," + }, + { + "start": 3375.56, + "duration": 0.0, + "text": "on text.\" That's why So, I<00:56:15.600> don't<00:56:15.720> know<00:56:15.760> if<00:56:15.840> you<00:56:15.920> know<00:56:16.040> this,<00:56:16.160> but<00:56:16.280> the" + }, + { + "start": 3376.39, + "duration": 0.0, + "text": "I don't know if you know this, but the" + }, + { + "start": 3376.4, + "duration": 0.0, + "text": "I don't know if you know this, but the reason<00:56:16.880> why<00:56:17.240> OpenAI<00:56:17.680> did<00:56:17.920> their<00:56:18.080> video<00:56:18.360> model" + }, + { + "start": 3378.67, + "duration": 0.0, + "text": "reason why OpenAI did their video model" + }, + { + "start": 3378.68, + "duration": 0.0, + "text": "reason why OpenAI did their video model is<00:56:18.800> because<00:56:19.400> of<00:56:19.680> they're<00:56:19.800> trying<00:56:20.000> to<00:56:20.120> solve" + }, + { + "start": 3380.31, + "duration": 0.0, + "text": "is because of they're trying to solve" + }, + { + "start": 3380.32, + "duration": 0.0, + "text": "is because of they're trying to solve this<00:56:20.440> problem.<00:56:20.760> They<00:56:20.920> want<00:56:21.080> a<00:56:21.160> world<00:56:21.480> model<00:56:22.120> so" + }, + { + "start": 3382.19, + "duration": 0.0, + "text": "this problem. They want a world model so" + }, + { + "start": 3382.2, + "duration": 0.0, + "text": "this problem. They want a world model so that<00:56:22.320> these<00:56:22.960> these<00:56:23.200> AI<00:56:23.400> systems<00:56:23.880> can<00:56:24.000> try<00:56:24.160> to" + }, + { + "start": 3384.27, + "duration": 0.0, + "text": "that these these AI systems can try to" + }, + { + "start": 3384.28, + "duration": 0.0, + "text": "that these these AI systems can try to imagine<00:56:24.640> new<00:56:24.760> scenarios<00:56:25.440> over<00:56:25.640> and<00:56:25.760> over" + }, + { + "start": 3385.91, + "duration": 0.0, + "text": "imagine new scenarios over and over" + }, + { + "start": 3385.92, + "duration": 0.0, + "text": "imagine new scenarios over and over again<00:56:26.240> and<00:56:26.440> find<00:56:26.840> new<00:56:26.960> ways<00:56:27.280> to<00:56:27.360> be<00:56:27.480> better." + }, + { + "start": 3388.43, + "duration": 0.0, + "text": "again and find new ways to be better." + }, + { + "start": 3388.44, + "duration": 0.0, + "text": "again and find new ways to be better. Um<00:56:28.760> and<00:56:28.880> then,<00:56:29.120> you<00:56:29.200> know,<00:56:29.360> if<00:56:29.640> you<00:56:29.760> can" + }, + { + "start": 3390.11, + "duration": 0.0, + "text": "Um and then, you know, if you can" + }, + { + "start": 3390.12, + "duration": 0.0, + "text": "Um and then, you know, if you can actually<00:56:30.800> even<00:56:30.960> if<00:56:31.040> it's<00:56:31.160> very<00:56:31.400> inefficient" + }, + { + "start": 3391.91, + "duration": 0.0, + "text": "actually even if it's very inefficient" + }, + { + "start": 3391.92, + "duration": 0.0, + "text": "actually even if it's very inefficient at<00:56:32.040> be<00:56:32.360> being<00:56:32.560> better,<00:56:32.800> right?<00:56:33.080> If<00:56:33.280> you<00:56:33.400> can" + }, + { + "start": 3393.59, + "duration": 0.0, + "text": "at be being better, right? If you can" + }, + { + "start": 3393.6, + "duration": 0.0, + "text": "at be being better, right? If you can somehow<00:56:34.400> correlate<00:56:35.200> the<00:56:35.320> strength<00:56:35.640> of<00:56:35.720> an<00:56:35.840> AI" + }, + { + "start": 3395.99, + "duration": 0.0, + "text": "somehow correlate the strength of an AI" + }, + { + "start": 3396.0, + "duration": 0.0, + "text": "somehow correlate the strength of an AI system<00:56:36.480> one-to-one<00:56:37.040> with<00:56:37.200> like<00:56:37.360> compute,<00:56:38.440> you" + }, + { + "start": 3398.59, + "duration": 0.0, + "text": "system one-to-one with like compute, you" + }, + { + "start": 3398.6, + "duration": 0.0, + "text": "system one-to-one with like compute, you we<00:56:38.720> can<00:56:38.840> always<00:56:39.040> get<00:56:39.200> more<00:56:39.320> compute,<00:56:39.960> right?" + }, + { + "start": 3400.15, + "duration": 0.0, + "text": "we can always get more compute, right?" + }, + { + "start": 3400.16, + "duration": 0.0, + "text": "we can always get more compute, right? We<00:56:40.240> can<00:56:40.360> always<00:56:40.560> make<00:56:40.720> Nvidia<00:56:41.080> stock<00:56:41.320> price<00:56:41.520> go" + }, + { + "start": 3401.71, + "duration": 0.0, + "text": "We can always make Nvidia stock price go" + }, + { + "start": 3401.72, + "duration": 0.0, + "text": "We can always make Nvidia stock price go up<00:56:41.880> and<00:56:42.000> like<00:56:42.120> get<00:56:42.240> more<00:56:42.360> chips<00:56:42.520> from<00:56:42.640> Taiwan." + }, + { + "start": 3403.03, + "duration": 0.0, + "text": "up and like get more chips from Taiwan." + }, + { + "start": 3403.04, + "duration": 0.0, + "text": "up and like get more chips from Taiwan. Like<00:56:43.480> it<00:56:43.600> can<00:56:43.680> always<00:56:44.120> we<00:56:44.240> can<00:56:44.360> always<00:56:44.560> keep" + }, + { + "start": 3404.71, + "duration": 0.0, + "text": "Like it can always we can always keep" + }, + { + "start": 3404.72, + "duration": 0.0, + "text": "Like it can always we can always keep ramping<00:56:45.040> this,<00:56:45.240> right?<00:56:45.400> There's<00:56:45.520> a<00:56:45.560> lot<00:56:45.720> of" + }, + { + "start": 3405.79, + "duration": 0.0, + "text": "ramping this, right? There's a lot of" + }, + { + "start": 3405.8, + "duration": 0.0, + "text": "ramping this, right? There's a lot of silicon<00:56:46.120> on<00:56:46.240> Earth<00:56:46.440> and<00:56:46.520> stuff<00:56:46.680> like<00:56:46.840> that," + }, + { + "start": 3406.99, + "duration": 0.0, + "text": "silicon on Earth and stuff like that," + }, + { + "start": 3407.0, + "duration": 0.0, + "text": "silicon on Earth and stuff like that, right?<00:56:47.120> It's<00:56:47.240> like<00:56:47.960> you<00:56:48.000> know,<00:56:48.160> you<00:56:48.360> you<00:56:48.440> can" + }, + { + "start": 3408.51, + "duration": 0.0, + "text": "right? It's like you know, you you can" + }, + { + "start": 3408.52, + "duration": 0.0, + "text": "right? It's like you know, you you can just<00:56:48.600> keep<00:56:48.760> doing<00:56:48.960> this.<00:56:49.120> You<00:56:49.200> can<00:56:49.320> scale<00:56:49.640> it" + }, + { + "start": 3409.75, + "duration": 0.0, + "text": "just keep doing this. You can scale it" + }, + { + "start": 3409.76, + "duration": 0.0, + "text": "just keep doing this. You can scale it and<00:56:49.840> then<00:56:50.200> once<00:56:50.360> you<00:56:50.480> have<00:56:50.600> the<00:56:50.680> compute,<00:56:51.040> it" + }, + { + "start": 3411.15, + "duration": 0.0, + "text": "and then once you have the compute, it" + }, + { + "start": 3411.16, + "duration": 0.0, + "text": "and then once you have the compute, it keeps<00:56:51.360> printing<00:56:52.120> it<00:56:52.200> keeps<00:56:52.360> being<00:56:52.600> smart.<00:56:53.000> And" + }, + { + "start": 3413.11, + "duration": 0.0, + "text": "keeps printing it keeps being smart. And" + }, + { + "start": 3413.12, + "duration": 0.0, + "text": "keeps printing it keeps being smart. And so,<00:56:53.440> it<00:56:53.720> will<00:56:53.880> it<00:56:54.000> is<00:56:54.120> only<00:56:54.280> a<00:56:54.320> matter<00:56:54.520> of<00:56:54.600> time" + }, + { + "start": 3414.79, + "duration": 0.0, + "text": "so, it will it is only a matter of time" + }, + { + "start": 3414.8, + "duration": 0.0, + "text": "so, it will it is only a matter of time until,<00:56:55.480> you<00:56:55.560> know,<00:56:55.680> if<00:56:55.880> you<00:56:56.040> can<00:56:56.280> actually<00:56:56.760> do" + }, + { + "start": 3417.87, + "duration": 0.0, + "text": "until, you know, if you can actually do" + }, + { + "start": 3417.88, + "duration": 0.0, + "text": "until, you know, if you can actually do self-optimizing<00:56:58.600> intelligent<00:56:59.000> processes" + }, + { + "start": 3419.51, + "duration": 0.0, + "text": "self-optimizing intelligent processes" + }, + { + "start": 3419.52, + "duration": 0.0, + "text": "self-optimizing intelligent processes with<00:57:00.040> silicon<00:57:00.400> that<00:57:00.520> it<00:57:00.640> exceeds<00:57:01.000> human" + }, + { + "start": 3421.19, + "duration": 0.0, + "text": "with silicon that it exceeds human" + }, + { + "start": 3421.2, + "duration": 0.0, + "text": "with silicon that it exceeds human baseline<00:57:01.640> and<00:57:02.080> becomes<00:57:02.400> the<00:57:02.480> god-like<00:57:02.760> figure" + }, + { + "start": 3423.03, + "duration": 0.0, + "text": "baseline and becomes the god-like figure" + }, + { + "start": 3423.04, + "duration": 0.0, + "text": "baseline and becomes the god-like figure some<00:57:03.320> people<00:57:03.720> want<00:57:03.960> to,<00:57:04.600> you<00:57:04.680> know,<00:57:04.800> build" + }, + { + "start": 3425.15, + "duration": 0.0, + "text": "some people want to, you know, build" + }, + { + "start": 3425.16, + "duration": 0.0, + "text": "some people want to, you know, build upon<00:57:05.440> the<00:57:05.600> Earth." + }, + { + "start": 3433.08, + "duration": 0.0, + "text": "Sam,<00:57:13.560> Timothy?" + }, + { + "start": 3434.67, + "duration": 0.0, + "text": "Sam, Timothy?" + }, + { + "start": 3434.68, + "duration": 0.0, + "text": "Sam, Timothy? Oh,<00:57:15.160> I<00:57:15.400> I<00:57:15.480> had<00:57:15.680> a<00:57:15.720> question<00:57:16.200> about<00:57:16.560> like<00:57:17.240> is<00:57:17.400> the" + }, + { + "start": 3437.51, + "duration": 0.0, + "text": "Oh, I I had a question about like is the" + }, + { + "start": 3437.52, + "duration": 0.0, + "text": "Oh, I I had a question about like is the future<00:57:17.880> then<00:57:18.520> like<00:57:19.280> oh<00:57:20.040> if<00:57:20.360> if<00:57:20.760> they<00:57:21.560> Is<00:57:21.720> there" + }, + { + "start": 3441.87, + "duration": 0.0, + "text": "future then like oh if if they Is there" + }, + { + "start": 3441.88, + "duration": 0.0, + "text": "future then like oh if if they Is there a<00:57:21.960> data<00:57:22.400> bottom<00:57:22.760> neck<00:57:23.000> at<00:57:23.080> some<00:57:23.280> point?<00:57:23.640> And<00:57:23.960> is" + }, + { + "start": 3444.19, + "duration": 0.0, + "text": "a data bottom neck at some point? And is" + }, + { + "start": 3444.2, + "duration": 0.0, + "text": "a data bottom neck at some point? And is the<00:57:24.400> future<00:57:24.760> of<00:57:24.840> reinforcement<00:57:25.480> learning?" + }, + { + "start": 3446.15, + "duration": 0.0, + "text": "the future of reinforcement learning?" + }, + { + "start": 3446.16, + "duration": 0.0, + "text": "the future of reinforcement learning? So,<00:57:26.640> what<00:57:26.720> I'm<00:57:26.800> thinking<00:57:27.040> about<00:57:27.240> is<00:57:27.320> like<00:57:27.440> a" + }, + { + "start": 3447.51, + "duration": 0.0, + "text": "So, what I'm thinking about is like a" + }, + { + "start": 3447.52, + "duration": 0.0, + "text": "So, what I'm thinking about is like a baby" + }, + { + "start": 3448.71, + "duration": 0.0, + "text": "baby" + }, + { + "start": 3448.72, + "duration": 0.0, + "text": "baby you<00:57:28.840> right<00:57:29.120> spends<00:57:29.400> like<00:57:29.560> a<00:57:29.600> lot<00:57:29.840> of<00:57:29.920> time" + }, + { + "start": 3450.23, + "duration": 0.0, + "text": "you right spends like a lot of time" + }, + { + "start": 3450.24, + "duration": 0.0, + "text": "you right spends like a lot of time receiving<00:57:30.680> input.<00:57:31.200> And<00:57:31.800> I<00:57:31.920> have<00:57:32.040> a<00:57:32.120> theory" + }, + { + "start": 3452.43, + "duration": 0.0, + "text": "receiving input. And I have a theory" + }, + { + "start": 3452.44, + "duration": 0.0, + "text": "receiving input. And I have a theory that<00:57:33.000> the<00:57:33.120> reason<00:57:33.400> puberty<00:57:33.840> happens,<00:57:34.520> and" + }, + { + "start": 3454.63, + "duration": 0.0, + "text": "that the reason puberty happens, and" + }, + { + "start": 3454.64, + "duration": 0.0, + "text": "that the reason puberty happens, and like<00:57:34.840> the<00:57:34.920> body<00:57:35.200> grows<00:57:35.440> so<00:57:35.520> quickly,<00:57:36.000> is" + }, + { + "start": 3456.19, + "duration": 0.0, + "text": "like the body grows so quickly, is" + }, + { + "start": 3456.2, + "duration": 0.0, + "text": "like the body grows so quickly, is because<00:57:37.040> it's<00:57:37.200> to<00:57:37.280> like<00:57:37.440> protect<00:57:37.960> the<00:57:38.080> brain" + }, + { + "start": 3458.91, + "duration": 0.0, + "text": "because it's to like protect the brain" + }, + { + "start": 3458.92, + "duration": 0.0, + "text": "because it's to like protect the brain from<00:57:39.160> injuring<00:57:39.560> itself<00:57:40.000> before<00:57:40.360> it<00:57:40.520> while" + }, + { + "start": 3460.71, + "duration": 0.0, + "text": "from injuring itself before it while" + }, + { + "start": 3460.72, + "duration": 0.0, + "text": "from injuring itself before it while it's<00:57:40.920> too<00:57:41.080> dumb." + }, + { + "start": 3462.63, + "duration": 0.0, + "text": "it's too dumb." + }, + { + "start": 3462.64, + "duration": 0.0, + "text": "it's too dumb. That<00:57:42.720> kind<00:57:42.840> of<00:57:42.920> makes<00:57:43.040> sense,<00:57:43.240> actually.<00:57:43.640> I" + }, + { + "start": 3463.67, + "duration": 0.0, + "text": "That kind of makes sense, actually. I" + }, + { + "start": 3463.68, + "duration": 0.0, + "text": "That kind of makes sense, actually. I don't<00:57:44.120> I<00:57:44.160> don't<00:57:44.320> know.<00:57:44.440> I<00:57:44.720> So<00:57:44.920> so<00:57:45.640> is<00:57:45.960> is<00:57:46.120> there" + }, + { + "start": 3466.31, + "duration": 0.0, + "text": "don't I don't know. I So so is is there" + }, + { + "start": 3466.32, + "duration": 0.0, + "text": "don't I don't know. I So so is is there a<00:57:46.360> future<00:57:46.720> where<00:57:46.880> like<00:57:47.160> you" + }, + { + "start": 3468.23, + "duration": 0.0, + "text": "a future where like you" + }, + { + "start": 3468.24, + "duration": 0.0, + "text": "a future where like you let<00:57:48.720> robots<00:57:49.360> like<00:57:49.520> with<00:57:49.680> cameras<00:57:50.320> out<00:57:50.520> in<00:57:50.640> a" + }, + { + "start": 3470.67, + "duration": 0.0, + "text": "let robots like with cameras out in a" + }, + { + "start": 3470.68, + "duration": 0.0, + "text": "let robots like with cameras out in a field,<00:57:51.200> and<00:57:51.320> then<00:57:51.440> the<00:57:51.560> learning<00:57:51.880> is<00:57:51.960> just" + }, + { + "start": 3472.11, + "duration": 0.0, + "text": "field, and then the learning is just" + }, + { + "start": 3472.12, + "duration": 0.0, + "text": "field, and then the learning is just like,<00:57:52.720> what<00:57:52.880> is<00:57:53.040> going<00:57:53.240> to<00:57:53.480> That<00:57:53.680> what's<00:57:53.840> the" + }, + { + "start": 3473.91, + "duration": 0.0, + "text": "like, what is going to That what's the" + }, + { + "start": 3473.92, + "duration": 0.0, + "text": "like, what is going to That what's the next<00:57:54.200> frame<00:57:54.520> of<00:57:54.600> the<00:57:54.680> video?<00:57:55.320> Yes,<00:57:55.720> actually" + }, + { + "start": 3476.03, + "duration": 0.0, + "text": "next frame of the video? Yes, actually" + }, + { + "start": 3476.04, + "duration": 0.0, + "text": "next frame of the video? Yes, actually exactly.<00:57:56.440> That's<00:57:56.600> like<00:57:56.760> what<00:57:56.880> Nvidia's" + }, + { + "start": 3477.35, + "duration": 0.0, + "text": "exactly. That's like what Nvidia's" + }, + { + "start": 3477.36, + "duration": 0.0, + "text": "exactly. That's like what Nvidia's doing.<00:57:57.680> That's<00:57:57.880> like<00:57:58.600> There's<00:57:58.800> already<00:57:59.080> many" + }, + { + "start": 3479.31, + "duration": 0.0, + "text": "doing. That's like There's already many" + }, + { + "start": 3479.32, + "duration": 0.0, + "text": "doing. That's like There's already many data<00:57:59.560> sets<00:57:59.880> where<00:58:00.000> they<00:58:00.120> basically<00:58:00.560> just" + }, + { + "start": 3480.71, + "duration": 0.0, + "text": "data sets where they basically just" + }, + { + "start": 3480.72, + "duration": 0.0, + "text": "data sets where they basically just strap<00:58:00.920> a<00:58:01.000> GoPro<00:58:01.400> to<00:58:01.520> the<00:58:01.640> center<00:58:01.840> of<00:58:01.920> a<00:58:01.960> guy" + }, + { + "start": 3482.11, + "duration": 0.0, + "text": "strap a GoPro to the center of a guy" + }, + { + "start": 3482.12, + "duration": 0.0, + "text": "strap a GoPro to the center of a guy cooking<00:58:02.440> in<00:58:02.520> a<00:58:02.560> kitchen,<00:58:02.920> and<00:58:03.040> there's<00:58:03.160> like," + }, + { + "start": 3483.51, + "duration": 0.0, + "text": "cooking in a kitchen, and there's like," + }, + { + "start": 3483.52, + "duration": 0.0, + "text": "cooking in a kitchen, and there's like, you<00:58:03.600> know,<00:58:03.680> gigabytes<00:58:04.080> of<00:58:04.160> those<00:58:04.320> videos.<00:58:04.720> And" + }, + { + "start": 3484.79, + "duration": 0.0, + "text": "you know, gigabytes of those videos. And" + }, + { + "start": 3484.8, + "duration": 0.0, + "text": "you know, gigabytes of those videos. And the<00:58:04.920> idea<00:58:05.160> is<00:58:05.280> just<00:58:05.480> like,<00:58:05.680> oh,<00:58:06.480> we<00:58:06.560> just" + }, + { + "start": 3486.75, + "duration": 0.0, + "text": "the idea is just like, oh, we just" + }, + { + "start": 3486.76, + "duration": 0.0, + "text": "the idea is just like, oh, we just basically<00:58:07.080> need<00:58:07.200> a<00:58:07.240> lot<00:58:07.480> of<00:58:07.800> video<00:58:08.160> of<00:58:08.280> people" + }, + { + "start": 3488.55, + "duration": 0.0, + "text": "basically need a lot of video of people" + }, + { + "start": 3488.56, + "duration": 0.0, + "text": "basically need a lot of video of people doing<00:58:08.840> whatever<00:58:09.080> they're<00:58:09.200> doing<00:58:09.400> that's" + }, + { + "start": 3489.55, + "duration": 0.0, + "text": "doing whatever they're doing that's" + }, + { + "start": 3489.56, + "duration": 0.0, + "text": "doing whatever they're doing that's useful<00:58:09.840> to<00:58:09.920> other<00:58:10.080> people.<00:58:11.000> Throw<00:58:11.120> a<00:58:11.160> bunch<00:58:11.320> of" + }, + { + "start": 3491.39, + "duration": 0.0, + "text": "useful to other people. Throw a bunch of" + }, + { + "start": 3491.4, + "duration": 0.0, + "text": "useful to other people. Throw a bunch of transformers<00:58:11.880> at<00:58:12.000> that,<00:58:12.320> trillion-dollar<00:58:12.760> AI" + }, + { + "start": 3492.99, + "duration": 0.0, + "text": "transformers at that, trillion-dollar AI" + }, + { + "start": 3493.0, + "duration": 0.0, + "text": "transformers at that, trillion-dollar AI company,<00:58:13.440> and<00:58:13.600> then<00:58:13.760> we<00:58:13.920> automate<00:58:14.240> them,<00:58:14.360> and" + }, + { + "start": 3494.43, + "duration": 0.0, + "text": "company, and then we automate them, and" + }, + { + "start": 3494.44, + "duration": 0.0, + "text": "company, and then we automate them, and then<00:58:14.560> we<00:58:14.720> rake<00:58:14.960> in<00:58:15.040> the<00:58:15.120> money,<00:58:15.440> and<00:58:15.520> then" + }, + { + "start": 3495.63, + "duration": 0.0, + "text": "then we rake in the money, and then" + }, + { + "start": 3495.64, + "duration": 0.0, + "text": "then we rake in the money, and then there's<00:58:15.800> no<00:58:15.920> more<00:58:16.160> people<00:58:16.440> to<00:58:16.520> buy<00:58:17.000> cuz" + }, + { + "start": 3497.15, + "duration": 0.0, + "text": "there's no more people to buy cuz" + }, + { + "start": 3497.16, + "duration": 0.0, + "text": "there's no more people to buy cuz there's<00:58:17.320> no<00:58:17.400> more<00:58:17.480> economy.<00:58:17.840> I<00:58:17.880> don't<00:58:18.480> Kind<00:58:18.560> of" + }, + { + "start": 3498.63, + "duration": 0.0, + "text": "there's no more economy. I don't Kind of" + }, + { + "start": 3498.64, + "duration": 0.0, + "text": "there's no more economy. I don't Kind of just<00:58:18.760> stopped<00:58:18.960> making<00:58:19.200> sense.<00:58:19.560> It's<00:58:19.680> like" + }, + { + "start": 3499.83, + "duration": 0.0, + "text": "just stopped making sense. It's like" + }, + { + "start": 3499.84, + "duration": 0.0, + "text": "just stopped making sense. It's like really<00:58:20.120> crazy.<00:58:20.960> Like<00:58:21.080> if<00:58:21.160> you<00:58:21.240> really<00:58:21.440> if<00:58:21.520> you" + }, + { + "start": 3501.63, + "duration": 0.0, + "text": "really crazy. Like if you really if you" + }, + { + "start": 3501.64, + "duration": 0.0, + "text": "really crazy. Like if you really if you really<00:58:21.840> think<00:58:22.040> about<00:58:22.200> it<00:58:22.280> for<00:58:22.400> a<00:58:22.440> second," + }, + { + "start": 3502.75, + "duration": 0.0, + "text": "really think about it for a second," + }, + { + "start": 3502.76, + "duration": 0.0, + "text": "really think about it for a second, right?<00:58:22.920> All<00:58:23.000> these<00:58:23.160> like<00:58:23.680> VC-backed<00:58:24.440> AI" + }, + { + "start": 3504.63, + "duration": 0.0, + "text": "right? All these like VC-backed AI" + }, + { + "start": 3504.64, + "duration": 0.0, + "text": "right? All these like VC-backed AI companies,<00:58:25.200> if<00:58:25.400> they<00:58:25.600> really<00:58:25.920> did<00:58:26.200> pan<00:58:26.480> out," + }, + { + "start": 3506.59, + "duration": 0.0, + "text": "companies, if they really did pan out," + }, + { + "start": 3506.6, + "duration": 0.0, + "text": "companies, if they really did pan out, there<00:58:26.680> would<00:58:26.800> be<00:58:26.880> no<00:58:27.040> more<00:58:27.160> economy.<00:58:27.960> Like" + }, + { + "start": 3508.11, + "duration": 0.0, + "text": "there would be no more economy. Like" + }, + { + "start": 3508.12, + "duration": 0.0, + "text": "there would be no more economy. Like what<00:58:28.240> are<00:58:28.280> you<00:58:28.400> competing<00:58:28.840> for<00:58:29.040> by<00:58:29.160> investing" + }, + { + "start": 3509.55, + "duration": 0.0, + "text": "what are you competing for by investing" + }, + { + "start": 3509.56, + "duration": 0.0, + "text": "what are you competing for by investing in<00:58:29.640> that?" + }, + { + "start": 3510.59, + "duration": 0.0, + "text": "in that?" + }, + { + "start": 3510.6, + "duration": 0.0, + "text": "in that? I<00:58:30.760> I<00:58:30.840> don't<00:58:31.000> know." + }, + { + "start": 3512.23, + "duration": 0.0, + "text": "I I don't know." + }, + { + "start": 3512.24, + "duration": 0.0, + "text": "I I don't know. Although,<00:58:32.560> um," + }, + { + "start": 3513.87, + "duration": 0.0, + "text": "Although, um," + }, + { + "start": 3513.88, + "duration": 0.0, + "text": "Although, um, concerning<00:58:34.240> like<00:58:34.440> those<00:58:34.800> gigabytes<00:58:35.240> videos" + }, + { + "start": 3515.51, + "duration": 0.0, + "text": "concerning like those gigabytes videos" + }, + { + "start": 3515.52, + "duration": 0.0, + "text": "concerning like those gigabytes videos and<00:58:35.640> so<00:58:35.800> on,<00:58:35.880> I<00:58:35.920> would<00:58:36.080> imagine<00:58:36.560> a<00:58:36.640> big<00:58:37.000> big" + }, + { + "start": 3517.23, + "duration": 0.0, + "text": "and so on, I would imagine a big big" + }, + { + "start": 3517.24, + "duration": 0.0, + "text": "and so on, I would imagine a big big chunk<00:58:37.680> of<00:58:37.800> like<00:58:38.280> what<00:58:38.600> would<00:58:38.760> make<00:58:39.200> a<00:58:39.560> the" + }, + { + "start": 3519.63, + "duration": 0.0, + "text": "chunk of like what would make a the" + }, + { + "start": 3519.64, + "duration": 0.0, + "text": "chunk of like what would make a the training<00:58:39.960> run<00:58:40.120> effective<00:58:40.560> is<00:58:40.640> like<00:58:40.800> how<00:58:40.920> do" + }, + { + "start": 3520.99, + "duration": 0.0, + "text": "training run effective is like how do" + }, + { + "start": 3521.0, + "duration": 0.0, + "text": "training run effective is like how do you<00:58:41.040> transform<00:58:41.400> the<00:58:41.480> data<00:58:41.760> before<00:58:42.080> sending<00:58:42.360> it" + }, + { + "start": 3522.47, + "duration": 0.0, + "text": "you transform the data before sending it" + }, + { + "start": 3522.48, + "duration": 0.0, + "text": "you transform the data before sending it into<00:58:42.680> the<00:58:43.040> model<00:58:43.400> or<00:58:43.480> into<00:58:43.800> the<00:58:44.000> algorithm," + }, + { + "start": 3524.75, + "duration": 0.0, + "text": "into the model or into the algorithm," + }, + { + "start": 3524.76, + "duration": 0.0, + "text": "into the model or into the algorithm, right?<00:58:44.960> Because<00:58:45.160> like<00:58:45.920> I<00:58:46.000> guess<00:58:46.200> if<00:58:46.320> you<00:58:46.400> like" + }, + { + "start": 3526.59, + "duration": 0.0, + "text": "right? Because like I guess if you like" + }, + { + "start": 3526.6, + "duration": 0.0, + "text": "right? Because like I guess if you like filter<00:58:46.920> down<00:58:47.200> to<00:58:47.320> what<00:58:47.520> actually<00:58:47.840> matters<00:58:48.320> cuz" + }, + { + "start": 3528.43, + "duration": 0.0, + "text": "filter down to what actually matters cuz" + }, + { + "start": 3528.44, + "duration": 0.0, + "text": "filter down to what actually matters cuz like<00:58:48.560> there's<00:58:48.720> like<00:58:48.920> a<00:58:49.000> lot<00:58:49.240> of<00:58:49.800> when<00:58:49.960> you<00:58:50.000> have" + }, + { + "start": 3530.43, + "duration": 0.0, + "text": "like there's like a lot of when you have" + }, + { + "start": 3530.44, + "duration": 0.0, + "text": "like there's like a lot of when you have a<00:58:50.480> guy<00:58:50.680> cooking<00:58:51.000> or<00:58:51.080> something,<00:58:51.760> like<00:58:52.000> a<00:58:52.080> lot" + }, + { + "start": 3532.31, + "duration": 0.0, + "text": "a guy cooking or something, like a lot" + }, + { + "start": 3532.32, + "duration": 0.0, + "text": "a guy cooking or something, like a lot of<00:58:52.400> that<00:58:52.760> image<00:58:53.040> data<00:58:53.280> is<00:58:53.360> completely" + }, + { + "start": 3533.71, + "duration": 0.0, + "text": "of that image data is completely" + }, + { + "start": 3533.72, + "duration": 0.0, + "text": "of that image data is completely useless.<00:58:54.040> You<00:58:54.120> don't<00:58:54.320> care<00:58:54.480> about<00:58:54.760> it<00:58:54.840> most<00:58:55.040> of" + }, + { + "start": 3535.11, + "duration": 0.0, + "text": "useless. You don't care about it most of" + }, + { + "start": 3535.12, + "duration": 0.0, + "text": "useless. You don't care about it most of it,<00:58:55.240> right?<00:58:55.720> So<00:58:55.840> like" + }, + { + "start": 3536.87, + "duration": 0.0, + "text": "it, right? So like" + }, + { + "start": 3536.88, + "duration": 0.0, + "text": "it, right? So like what<00:58:57.600> kind<00:58:57.760> of<00:58:57.920> pre-filtering<00:58:58.560> do<00:58:58.680> you<00:58:58.800> people" + }, + { + "start": 3539.07, + "duration": 0.0, + "text": "what kind of pre-filtering do you people" + }, + { + "start": 3539.08, + "duration": 0.0, + "text": "what kind of pre-filtering do you people do,<00:58:59.320> and<00:58:59.440> like<00:58:59.760> how<00:59:00.000> much<00:59:00.720> How<00:59:00.840> important<00:59:01.240> is" + }, + { + "start": 3541.31, + "duration": 0.0, + "text": "do, and like how much How important is" + }, + { + "start": 3541.32, + "duration": 0.0, + "text": "do, and like how much How important is that?<00:59:01.720> Well,<00:59:02.160> actually<00:59:02.560> people<00:59:02.800> do<00:59:02.920> the" + }, + { + "start": 3543.03, + "duration": 0.0, + "text": "that? Well, actually people do the" + }, + { + "start": 3543.04, + "duration": 0.0, + "text": "that? Well, actually people do the opposite<00:59:03.440> of<00:59:03.520> pre-filtering.<00:59:04.440> Sometimes<00:59:05.040> for" + }, + { + "start": 3545.15, + "duration": 0.0, + "text": "opposite of pre-filtering. Sometimes for" + }, + { + "start": 3545.16, + "duration": 0.0, + "text": "opposite of pre-filtering. Sometimes for deep<00:59:05.360> learning,<00:59:05.720> actually<00:59:06.000> I<00:59:06.040> think<00:59:06.240> Tesla's" + }, + { + "start": 3546.55, + "duration": 0.0, + "text": "deep learning, actually I think Tesla's" + }, + { + "start": 3546.56, + "duration": 0.0, + "text": "deep learning, actually I think Tesla's doing<00:59:06.760> this,<00:59:06.960> they<00:59:07.080> don't<00:59:07.320> feed<00:59:07.560> pixels,<00:59:08.040> they" + }, + { + "start": 3548.15, + "duration": 0.0, + "text": "doing this, they don't feed pixels, they" + }, + { + "start": 3548.16, + "duration": 0.0, + "text": "doing this, they don't feed pixels, they feed<00:59:08.360> like<00:59:08.560> raw<00:59:08.760> sensor<00:59:09.080> measurements.<00:59:09.520> They" + }, + { + "start": 3549.67, + "duration": 0.0, + "text": "feed like raw sensor measurements. They" + }, + { + "start": 3549.68, + "duration": 0.0, + "text": "feed like raw sensor measurements. They would<00:59:09.880> decrease<00:59:10.360> filtering." + }, + { + "start": 3551.51, + "duration": 0.0, + "text": "would decrease filtering." + }, + { + "start": 3551.52, + "duration": 0.0, + "text": "would decrease filtering. There's<00:59:11.680> a<00:59:11.880> There's<00:59:12.040> a<00:59:12.120> very<00:59:12.720> Well,<00:59:12.800> there's<00:59:12.960> a" + }, + { + "start": 3552.99, + "duration": 0.0, + "text": "There's a There's a very Well, there's a" + }, + { + "start": 3553.0, + "duration": 0.0, + "text": "There's a There's a very Well, there's a very<00:59:13.240> famous<00:59:13.840> like<00:59:14.080> blog<00:59:14.360> post,<00:59:14.640> basically," + }, + { + "start": 3555.11, + "duration": 0.0, + "text": "very famous like blog post, basically," + }, + { + "start": 3555.12, + "duration": 0.0, + "text": "very famous like blog post, basically, called<00:59:15.320> the<00:59:15.360> bitter<00:59:15.600> lesson.<00:59:16.560> And<00:59:16.640> the<00:59:16.720> bitter" + }, + { + "start": 3556.91, + "duration": 0.0, + "text": "called the bitter lesson. And the bitter" + }, + { + "start": 3556.92, + "duration": 0.0, + "text": "called the bitter lesson. And the bitter lesson<00:59:17.280> is<00:59:17.400> that<00:59:17.560> like<00:59:17.760> a<00:59:17.800> general<00:59:18.160> method" + }, + { + "start": 3558.43, + "duration": 0.0, + "text": "lesson is that like a general method" + }, + { + "start": 3558.44, + "duration": 0.0, + "text": "lesson is that like a general method that<00:59:18.600> scales<00:59:18.880> with<00:59:19.000> compute<00:59:19.480> will<00:59:19.600> always" + }, + { + "start": 3559.83, + "duration": 0.0, + "text": "that scales with compute will always" + }, + { + "start": 3559.84, + "duration": 0.0, + "text": "that scales with compute will always perform<00:59:20.200> better<00:59:20.440> than<00:59:20.560> a<00:59:20.600> specific<00:59:21.080> method" + }, + { + "start": 3561.35, + "duration": 0.0, + "text": "perform better than a specific method" + }, + { + "start": 3561.36, + "duration": 0.0, + "text": "perform better than a specific method made<00:59:21.520> by<00:59:21.640> engineers<00:59:22.560> because<00:59:22.840> compute<00:59:23.240> is" + }, + { + "start": 3563.35, + "duration": 0.0, + "text": "made by engineers because compute is" + }, + { + "start": 3563.36, + "duration": 0.0, + "text": "made by engineers because compute is always<00:59:23.560> getting<00:59:23.760> cheaper<00:59:24.040> and<00:59:24.120> more" + }, + { + "start": 3564.23, + "duration": 0.0, + "text": "always getting cheaper and more" + }, + { + "start": 3564.24, + "duration": 0.0, + "text": "always getting cheaper and more plentiful.<00:59:24.680> Engineers<00:59:25.200> are<00:59:25.280> becoming<00:59:25.600> more" + }, + { + "start": 3565.71, + "duration": 0.0, + "text": "plentiful. Engineers are becoming more" + }, + { + "start": 3565.72, + "duration": 0.0, + "text": "plentiful. Engineers are becoming more scarce<00:59:26.080> as<00:59:26.200> people<00:59:26.440> become<00:59:26.680> stupider<00:59:27.080> from" + }, + { + "start": 3567.27, + "duration": 0.0, + "text": "scarce as people become stupider from" + }, + { + "start": 3567.28, + "duration": 0.0, + "text": "scarce as people become stupider from the<00:59:27.400> AI<00:59:27.640> that<00:59:27.840> you're<00:59:27.960> creating<00:59:28.400> and<00:59:28.480> doing" + }, + { + "start": 3568.75, + "duration": 0.0, + "text": "the AI that you're creating and doing" + }, + { + "start": 3568.76, + "duration": 0.0, + "text": "the AI that you're creating and doing this.<00:59:29.440> Yeah,<00:59:29.520> but<00:59:29.640> but<00:59:30.040> then<00:59:30.160> by<00:59:30.280> choosing<00:59:30.640> to" + }, + { + "start": 3570.75, + "duration": 0.0, + "text": "this. Yeah, but but then by choosing to" + }, + { + "start": 3570.76, + "duration": 0.0, + "text": "this. Yeah, but but then by choosing to use<00:59:30.920> like<00:59:31.080> sensors,<00:59:31.720> like<00:59:31.880> you're<00:59:31.960> making" + }, + { + "start": 3572.31, + "duration": 0.0, + "text": "use like sensors, like you're making" + }, + { + "start": 3572.32, + "duration": 0.0, + "text": "use like sensors, like you're making that's<00:59:32.520> like<00:59:32.640> a<00:59:32.680> deliberate<00:59:33.000> choice.<00:59:33.320> Like" + }, + { + "start": 3573.51, + "duration": 0.0, + "text": "that's like a deliberate choice. Like" + }, + { + "start": 3573.52, + "duration": 0.0, + "text": "that's like a deliberate choice. Like instead<00:59:33.760> of<00:59:33.800> using<00:59:34.080> pixels,<00:59:34.400> we're<00:59:34.480> going<00:59:34.600> to" + }, + { + "start": 3574.67, + "duration": 0.0, + "text": "instead of using pixels, we're going to" + }, + { + "start": 3574.68, + "duration": 0.0, + "text": "instead of using pixels, we're going to use<00:59:34.840> sensors.<00:59:35.360> So,<00:59:35.480> it's<00:59:35.560> like<00:59:35.720> maybe<00:59:36.000> more" + }, + { + "start": 3576.19, + "duration": 0.0, + "text": "use sensors. So, it's like maybe more" + }, + { + "start": 3576.2, + "duration": 0.0, + "text": "use sensors. So, it's like maybe more relevant<00:59:36.600> information<00:59:37.120> within<00:59:37.400> it.<00:59:37.880> So,<00:59:37.920> like" + }, + { + "start": 3578.03, + "duration": 0.0, + "text": "relevant information within it. So, like" + }, + { + "start": 3578.04, + "duration": 0.0, + "text": "relevant information within it. So, like the<00:59:38.120> pattern<00:59:38.400> might<00:59:38.520> be<00:59:38.640> easier<00:59:38.880> to<00:59:38.960> extract" + }, + { + "start": 3579.35, + "duration": 0.0, + "text": "the pattern might be easier to extract" + }, + { + "start": 3579.36, + "duration": 0.0, + "text": "the pattern might be easier to extract somehow<00:59:39.760> or<00:59:40.080> No,<00:59:40.240> it's<00:59:40.400> not<00:59:40.560> that<00:59:40.720> it's" + }, + { + "start": 3581.43, + "duration": 0.0, + "text": "somehow or No, it's not that it's" + }, + { + "start": 3581.44, + "duration": 0.0, + "text": "somehow or No, it's not that it's the<00:59:41.640> idea<00:59:42.120> is<00:59:42.280> that<00:59:43.040> we<00:59:43.240> should<00:59:43.480> not<00:59:44.000> try<00:59:44.360> to" + }, + { + "start": 3584.63, + "duration": 0.0, + "text": "the idea is that we should not try to" + }, + { + "start": 3584.64, + "duration": 0.0, + "text": "the idea is that we should not try to design<00:59:45.640> these<00:59:45.880> massive<00:59:46.360> massive,<00:59:46.960> you<00:59:47.080> know," + }, + { + "start": 3587.19, + "duration": 0.0, + "text": "design these massive massive, you know," + }, + { + "start": 3587.2, + "duration": 0.0, + "text": "design these massive massive, you know, trillion<00:59:47.640> parameter<00:59:48.000> deep<00:59:48.240> learning<00:59:48.440> models." + }, + { + "start": 3589.63, + "duration": 0.0, + "text": "trillion parameter deep learning models." + }, + { + "start": 3589.64, + "duration": 0.0, + "text": "trillion parameter deep learning models. We<00:59:49.760> should<00:59:49.920> not<00:59:50.120> try<00:59:50.360> to<00:59:50.520> do<00:59:50.720> any<00:59:50.920> data" + }, + { + "start": 3591.15, + "duration": 0.0, + "text": "We should not try to do any data" + }, + { + "start": 3591.16, + "duration": 0.0, + "text": "We should not try to do any data transformation<00:59:51.720> on<00:59:51.840> what's<00:59:52.000> happening<00:59:52.320> in" + }, + { + "start": 3592.35, + "duration": 0.0, + "text": "transformation on what's happening in" + }, + { + "start": 3592.36, + "duration": 0.0, + "text": "transformation on what's happening in the<00:59:52.440> universe<00:59:52.840> because<00:59:53.280> it<00:59:53.440> will<00:59:54.040> find<00:59:54.480> the" + }, + { + "start": 3594.59, + "duration": 0.0, + "text": "the universe because it will find the" + }, + { + "start": 3594.6, + "duration": 0.0, + "text": "the universe because it will find the representation<00:59:55.240> that's<00:59:55.400> most<00:59:55.600> effective<00:59:56.040> for" + }, + { + "start": 3596.15, + "duration": 0.0, + "text": "representation that's most effective for" + }, + { + "start": 3596.16, + "duration": 0.0, + "text": "representation that's most effective for minimizing<00:59:56.640> the<00:59:56.720> loss,<00:59:57.200> right?<00:59:57.800> It<00:59:57.920> will<00:59:58.040> do" + }, + { + "start": 3598.19, + "duration": 0.0, + "text": "minimizing the loss, right? It will do" + }, + { + "start": 3598.2, + "duration": 0.0, + "text": "minimizing the loss, right? It will do that<00:59:58.360> with<00:59:58.520> compute.<00:59:59.040> It<00:59:59.200> will<00:59:59.320> do<00:59:59.560> It<00:59:59.680> will" + }, + { + "start": 3599.79, + "duration": 0.0, + "text": "that with compute. It will do It will" + }, + { + "start": 3599.8, + "duration": 0.0, + "text": "that with compute. It will do It will find<01:00:00.080> something<01:00:00.320> better<01:00:00.560> than<01:00:00.680> what<01:00:00.800> we<01:00:00.880> could" + }, + { + "start": 3600.99, + "duration": 0.0, + "text": "find something better than what we could" + }, + { + "start": 3601.0, + "duration": 0.0, + "text": "find something better than what we could ever<01:00:01.160> filter<01:00:01.560> it,<01:00:01.960> right?<01:00:02.160> And<01:00:02.400> what" + }, + { + "start": 3602.51, + "duration": 0.0, + "text": "ever filter it, right? And what" + }, + { + "start": 3602.52, + "duration": 0.0, + "text": "ever filter it, right? And what basically<01:00:02.800> you're<01:00:02.880> describing<01:00:03.240> is<01:00:03.440> I<01:00:03.520> think" + }, + { + "start": 3603.67, + "duration": 0.0, + "text": "basically you're describing is I think" + }, + { + "start": 3603.68, + "duration": 0.0, + "text": "basically you're describing is I think called<01:00:03.840> feature<01:00:04.160> engineering.<01:00:04.640> We're<01:00:04.760> like," + }, + { + "start": 3605.31, + "duration": 0.0, + "text": "called feature engineering. We're like," + }, + { + "start": 3605.32, + "duration": 0.0, + "text": "called feature engineering. We're like, \"Oh,<01:00:05.560> instead<01:00:05.880> of<01:00:05.960> feeding<01:00:06.240> the<01:00:06.360> images<01:00:06.800> to" + }, + { + "start": 3606.91, + "duration": 0.0, + "text": "\"Oh, instead of feeding the images to" + }, + { + "start": 3606.92, + "duration": 0.0, + "text": "\"Oh, instead of feeding the images to the<01:00:07.000> deep<01:00:07.160> learning<01:00:07.320> model,<01:00:07.520> we<01:00:07.640> feed<01:00:07.880> we<01:00:08.480> we" + }, + { + "start": 3608.59, + "duration": 0.0, + "text": "the deep learning model, we feed we we" + }, + { + "start": 3608.6, + "duration": 0.0, + "text": "the deep learning model, we feed we we try<01:00:08.760> to<01:00:08.840> detect<01:00:09.200> like<01:00:09.320> where<01:00:09.400> the<01:00:09.520> cones<01:00:09.800> are" + }, + { + "start": 3609.91, + "duration": 0.0, + "text": "try to detect like where the cones are" + }, + { + "start": 3609.92, + "duration": 0.0, + "text": "try to detect like where the cones are and<01:00:10.000> stuff<01:00:10.240> and<01:00:10.320> like<01:00:10.440> the<01:00:10.520> road<01:00:10.800> and<01:00:10.880> then" + }, + { + "start": 3610.99, + "duration": 0.0, + "text": "and stuff and like the road and then" + }, + { + "start": 3611.0, + "duration": 0.0, + "text": "and stuff and like the road and then like,<01:00:11.600> you<01:00:11.680> know,<01:00:11.760> we<01:00:11.880> feed<01:00:12.120> those<01:00:12.280> features" + }, + { + "start": 3612.59, + "duration": 0.0, + "text": "like, you know, we feed those features" + }, + { + "start": 3612.6, + "duration": 0.0, + "text": "like, you know, we feed those features into<01:00:12.720> the<01:00:12.800> model.\"<01:00:13.080> And<01:00:13.200> it<01:00:13.280> always" + }, + { + "start": 3613.55, + "duration": 0.0, + "text": "into the model.\" And it always" + }, + { + "start": 3613.56, + "duration": 0.0, + "text": "into the model.\" And it always underperforms<01:00:14.320> just<01:00:14.520> feeding<01:00:14.760> in<01:00:14.840> the<01:00:14.920> raw" + }, + { + "start": 3615.07, + "duration": 0.0, + "text": "underperforms just feeding in the raw" + }, + { + "start": 3615.08, + "duration": 0.0, + "text": "underperforms just feeding in the raw sensory<01:00:15.440> data.<01:00:16.320> If<01:00:16.480> you<01:00:16.600> have<01:00:16.840> enough<01:00:17.080> compute" + }, + { + "start": 3617.35, + "duration": 0.0, + "text": "sensory data. If you have enough compute" + }, + { + "start": 3617.36, + "duration": 0.0, + "text": "sensory data. If you have enough compute to<01:00:17.480> scale<01:00:17.760> up<01:00:17.920> because<01:00:18.840> the<01:00:18.960> compute<01:00:19.320> will" + }, + { + "start": 3619.43, + "duration": 0.0, + "text": "to scale up because the compute will" + }, + { + "start": 3619.44, + "duration": 0.0, + "text": "to scale up because the compute will find<01:00:20.080> the<01:00:20.160> representation<01:00:20.800> that<01:00:20.960> is<01:00:21.120> most" + }, + { + "start": 3621.31, + "duration": 0.0, + "text": "find the representation that is most" + }, + { + "start": 3621.32, + "duration": 0.0, + "text": "find the representation that is most effective<01:00:21.760> to<01:00:21.880> output<01:00:22.120> the<01:00:22.240> policy" + }, + { + "start": 3622.59, + "duration": 0.0, + "text": "effective to output the policy" + }, + { + "start": 3622.6, + "duration": 0.0, + "text": "effective to output the policy basically,<01:00:23.080> I<01:00:23.160> think.<01:00:23.760> So,<01:00:23.880> but<01:00:24.040> you<01:00:24.120> can" + }, + { + "start": 3624.31, + "duration": 0.0, + "text": "basically, I think. So, but you can" + }, + { + "start": 3624.32, + "duration": 0.0, + "text": "basically, I think. So, but you can still<01:00:24.640> select<01:00:25.360> the<01:00:25.760> which<01:00:26.360> what<01:00:26.680> kind<01:00:26.840> of" + }, + { + "start": 3626.91, + "duration": 0.0, + "text": "still select the which what kind of" + }, + { + "start": 3626.92, + "duration": 0.0, + "text": "still select the which what kind of sensor<01:00:27.160> data<01:00:27.360> you<01:00:27.440> feed<01:00:27.640> it<01:00:27.760> in.<01:00:27.880> Like<01:00:28.240> the" + }, + { + "start": 3628.31, + "duration": 0.0, + "text": "sensor data you feed it in. Like the" + }, + { + "start": 3628.32, + "duration": 0.0, + "text": "sensor data you feed it in. Like the sensor<01:00:28.680> is<01:00:28.840> is<01:00:28.960> different<01:00:29.200> from<01:00:29.320> sensor<01:00:29.560> data" + }, + { + "start": 3629.71, + "duration": 0.0, + "text": "sensor is is different from sensor data" + }, + { + "start": 3629.72, + "duration": 0.0, + "text": "sensor is is different from sensor data from<01:00:29.880> like<01:00:30.040> vision,<01:00:30.440> right?<01:00:31.120> Yeah.<01:00:31.560> You<01:00:31.680> still" + }, + { + "start": 3631.99, + "duration": 0.0, + "text": "from like vision, right? Yeah. You still" + }, + { + "start": 3632.0, + "duration": 0.0, + "text": "from like vision, right? Yeah. You still should<01:00:32.400> make<01:00:32.600> a<01:00:32.640> choice<01:00:33.200> and<01:00:33.360> like<01:00:33.480> that" + }, + { + "start": 3633.59, + "duration": 0.0, + "text": "should make a choice and like that" + }, + { + "start": 3633.6, + "duration": 0.0, + "text": "should make a choice and like that choice<01:00:33.920> will<01:00:34.040> matter,<01:00:34.840> right?<01:00:35.760> Yeah,<01:00:35.920> the" + }, + { + "start": 3636.03, + "duration": 0.0, + "text": "choice will matter, right? Yeah, the" + }, + { + "start": 3636.04, + "duration": 0.0, + "text": "choice will matter, right? Yeah, the choice<01:00:36.280> matters<01:00:36.600> in<01:00:36.680> the<01:00:36.800> opposite<01:00:37.120> direction" + }, + { + "start": 3637.51, + "duration": 0.0, + "text": "choice matters in the opposite direction" + }, + { + "start": 3637.52, + "duration": 0.0, + "text": "choice matters in the opposite direction from<01:00:37.640> what<01:00:37.760> you're<01:00:37.840> saying.<01:00:38.120> You<01:00:38.240> want<01:00:38.440> more" + }, + { + "start": 3638.79, + "duration": 0.0, + "text": "from what you're saying. You want more" + }, + { + "start": 3638.8, + "duration": 0.0, + "text": "from what you're saying. You want more more<01:00:39.040> generic,<01:00:39.600> more<01:00:40.000> bigger<01:00:40.560> amount<01:00:40.880> of" + }, + { + "start": 3640.99, + "duration": 0.0, + "text": "more generic, more bigger amount of" + }, + { + "start": 3641.0, + "duration": 0.0, + "text": "more generic, more bigger amount of data.<01:00:41.680> So,<01:00:41.720> you<01:00:41.760> feed<01:00:41.840> it<01:00:41.960> everything,<01:00:42.560> every" + }, + { + "start": 3642.79, + "duration": 0.0, + "text": "data. So, you feed it everything, every" + }, + { + "start": 3642.8, + "duration": 0.0, + "text": "data. So, you feed it everything, every type<01:00:43.000> of<01:00:43.120> sensor<01:00:43.320> data<01:00:43.480> you<01:00:43.600> can<01:00:43.840> get<01:00:43.960> your" + }, + { + "start": 3644.07, + "duration": 0.0, + "text": "type of sensor data you can get your" + }, + { + "start": 3644.08, + "duration": 0.0, + "text": "type of sensor data you can get your hands<01:00:44.360> on.<01:00:44.520> You<01:00:44.560> just<01:00:44.920> feed<01:00:45.080> it<01:00:45.160> whatever<01:00:45.440> you" + }, + { + "start": 3645.51, + "duration": 0.0, + "text": "hands on. You just feed it whatever you" + }, + { + "start": 3645.52, + "duration": 0.0, + "text": "hands on. You just feed it whatever you can.<01:00:45.880> Yes,<01:00:46.160> ideally<01:00:46.680> yes,<01:00:46.920> but<01:00:47.240> like<01:00:47.640> what" + }, + { + "start": 3647.79, + "duration": 0.0, + "text": "can. Yes, ideally yes, but like what" + }, + { + "start": 3647.8, + "duration": 0.0, + "text": "can. Yes, ideally yes, but like what they<01:00:47.920> found<01:00:48.280> actually<01:00:48.640> I<01:00:48.800> I<01:00:48.880> think<01:00:49.240> there<01:00:49.360> was" + }, + { + "start": 3649.43, + "duration": 0.0, + "text": "they found actually I I think there was" + }, + { + "start": 3649.44, + "duration": 0.0, + "text": "they found actually I I think there was this<01:00:49.560> paper<01:00:49.760> called<01:00:49.880> like<01:00:50.000> the<01:00:50.080> Nemo<01:00:50.880> paper.<01:00:51.160> I" + }, + { + "start": 3651.19, + "duration": 0.0, + "text": "this paper called like the Nemo paper. I" + }, + { + "start": 3651.2, + "duration": 0.0, + "text": "this paper called like the Nemo paper. I don't<01:00:51.320> remember<01:00:51.480> what<01:00:51.560> it<01:00:51.640> was,<01:00:51.920> but<01:00:52.480> they" + }, + { + "start": 3652.59, + "duration": 0.0, + "text": "don't remember what it was, but they" + }, + { + "start": 3652.6, + "duration": 0.0, + "text": "don't remember what it was, but they basically<01:00:53.040> had<01:00:53.160> a<01:00:53.200> transformer<01:00:54.360> they<01:00:54.440> just" + }, + { + "start": 3654.67, + "duration": 0.0, + "text": "basically had a transformer they just" + }, + { + "start": 3654.68, + "duration": 0.0, + "text": "basically had a transformer they just found<01:00:54.880> that<01:00:55.040> as<01:00:55.200> you<01:00:55.320> taught<01:00:55.560> it<01:00:55.680> many<01:00:55.920> random" + }, + { + "start": 3656.15, + "duration": 0.0, + "text": "found that as you taught it many random" + }, + { + "start": 3656.16, + "duration": 0.0, + "text": "found that as you taught it many random disparate<01:00:56.480> games,<01:00:56.840> it<01:00:56.960> would<01:00:57.040> like<01:00:57.160> improve" + }, + { + "start": 3657.79, + "duration": 0.0, + "text": "disparate games, it would like improve" + }, + { + "start": 3657.8, + "duration": 0.0, + "text": "disparate games, it would like improve on<01:00:57.920> the<01:00:58.040> other<01:00:58.200> ones,<01:00:58.480> right?<01:00:58.640> That<01:00:58.720> were" + }, + { + "start": 3658.79, + "duration": 0.0, + "text": "on the other ones, right? That were" + }, + { + "start": 3658.8, + "duration": 0.0, + "text": "on the other ones, right? That were seemingly<01:00:59.120> unrelated.<01:00:59.920> And<01:01:00.080> so,<01:01:00.200> there's" + }, + { + "start": 3660.35, + "duration": 0.0, + "text": "seemingly unrelated. And so, there's" + }, + { + "start": 3660.36, + "duration": 0.0, + "text": "seemingly unrelated. And so, there's something<01:01:00.680> just<01:01:00.840> about<01:01:01.600> transformers.<01:01:02.200> The" + }, + { + "start": 3662.27, + "duration": 0.0, + "text": "something just about transformers. The" + }, + { + "start": 3662.28, + "duration": 0.0, + "text": "something just about transformers. The more<01:01:02.480> data<01:01:03.040> that<01:01:03.480> you<01:01:03.640> give<01:01:03.840> them,<01:01:04.040> the<01:01:04.160> better" + }, + { + "start": 3664.39, + "duration": 0.0, + "text": "more data that you give them, the better" + }, + { + "start": 3664.4, + "duration": 0.0, + "text": "more data that you give them, the better the<01:01:04.520> representations<01:01:05.160> become<01:01:05.400> no<01:01:05.480> matter" + }, + { + "start": 3665.71, + "duration": 0.0, + "text": "the representations become no matter" + }, + { + "start": 3665.72, + "duration": 0.0, + "text": "the representations become no matter what.<01:01:06.280> I<01:01:06.360> think<01:01:06.560> it's<01:01:06.680> because<01:01:07.160> there's" + }, + { + "start": 3667.51, + "duration": 0.0, + "text": "what. I think it's because there's" + }, + { + "start": 3667.52, + "duration": 0.0, + "text": "what. I think it's because there's there's<01:01:07.720> something<01:01:08.120> kind<01:01:08.400> of<01:01:08.480> like<01:01:08.640> a<01:01:08.920> a" + }, + { + "start": 3668.99, + "duration": 0.0, + "text": "there's something kind of like a a" + }, + { + "start": 3669.0, + "duration": 0.0, + "text": "there's something kind of like a a universal<01:01:09.480> prior<01:01:09.840> or<01:01:10.000> like" + }, + { + "start": 3670.95, + "duration": 0.0, + "text": "universal prior or like" + }, + { + "start": 3670.96, + "duration": 0.0, + "text": "universal prior or like you<01:01:11.000> know,<01:01:11.120> it's<01:01:11.200> just<01:01:11.360> the<01:01:11.400> universe<01:01:11.760> is" + }, + { + "start": 3671.95, + "duration": 0.0, + "text": "you know, it's just the universe is" + }, + { + "start": 3671.96, + "duration": 0.0, + "text": "you know, it's just the universe is ordered<01:01:12.240> in<01:01:12.320> a<01:01:12.360> particular<01:01:12.840> way<01:01:13.040> that's" + }, + { + "start": 3673.27, + "duration": 0.0, + "text": "ordered in a particular way that's" + }, + { + "start": 3673.28, + "duration": 0.0, + "text": "ordered in a particular way that's evident<01:01:13.640> in<01:01:13.800> everything<01:01:14.120> that<01:01:14.240> happens<01:01:14.520> in" + }, + { + "start": 3674.59, + "duration": 0.0, + "text": "evident in everything that happens in" + }, + { + "start": 3674.6, + "duration": 0.0, + "text": "evident in everything that happens in it.<01:01:14.800> So,<01:01:14.880> the<01:01:15.000> more<01:01:15.240> in-<01:01:15.440> information<01:01:15.920> that" + }, + { + "start": 3675.99, + "duration": 0.0, + "text": "it. So, the more in- information that" + }, + { + "start": 3676.0, + "duration": 0.0, + "text": "it. So, the more in- information that comes<01:01:16.360> in<01:01:16.960> This<01:01:17.120> is<01:01:17.160> not<01:01:17.360> This<01:01:17.520> is<01:01:17.560> me<01:01:17.880> not<01:01:18.040> not" + }, + { + "start": 3678.15, + "duration": 0.0, + "text": "comes in This is not This is me not not" + }, + { + "start": 3678.16, + "duration": 0.0, + "text": "comes in This is not This is me not not by<01:01:18.240> the<01:01:18.320> way.<01:01:18.400> This<01:01:18.520> is<01:01:18.600> not<01:01:18.760> a<01:01:18.800> true<01:01:19.360> I'm<01:01:19.520> like" + }, + { + "start": 3679.95, + "duration": 0.0, + "text": "by the way. This is not a true I'm like" + }, + { + "start": 3679.96, + "duration": 0.0, + "text": "by the way. This is not a true I'm like theory<01:01:20.160> I'm<01:01:20.240> like<01:01:20.400> crafting<01:01:20.760> theories.<01:01:21.120> It's" + }, + { + "start": 3681.19, + "duration": 0.0, + "text": "theory I'm like crafting theories. It's" + }, + { + "start": 3681.2, + "duration": 0.0, + "text": "theory I'm like crafting theories. It's like<01:01:21.560> maybe<01:01:21.840> there's<01:01:22.000> like<01:01:22.120> something<01:01:22.320> about" + }, + { + "start": 3682.47, + "duration": 0.0, + "text": "like maybe there's like something about" + }, + { + "start": 3682.48, + "duration": 0.0, + "text": "like maybe there's like something about the<01:01:22.560> universe<01:01:22.960> that<01:01:23.200> just<01:01:23.400> whatever<01:01:23.600> you<01:01:23.720> give" + }, + { + "start": 3683.91, + "duration": 0.0, + "text": "the universe that just whatever you give" + }, + { + "start": 3683.92, + "duration": 0.0, + "text": "the universe that just whatever you give it<01:01:24.120> that's<01:01:24.400> from<01:01:24.640> the<01:01:24.960> universe<01:01:25.400> truly<01:01:25.720> and" + }, + { + "start": 3685.83, + "duration": 0.0, + "text": "it that's from the universe truly and" + }, + { + "start": 3685.84, + "duration": 0.0, + "text": "it that's from the universe truly and not<01:01:26.240> you<01:01:26.320> know<01:01:27.000> like<01:01:27.200> a<01:01:27.760> craft<01:01:28.120> of<01:01:28.320> man<01:01:28.600> or" + }, + { + "start": 3688.63, + "duration": 0.0, + "text": "not you know like a craft of man or" + }, + { + "start": 3688.64, + "duration": 0.0, + "text": "not you know like a craft of man or whatever<01:01:28.920> like<01:01:29.680> it<01:01:30.120> it's<01:01:30.240> just<01:01:30.560> somehow" + }, + { + "start": 3691.91, + "duration": 0.0, + "text": "whatever like it it's just somehow" + }, + { + "start": 3691.92, + "duration": 0.0, + "text": "whatever like it it's just somehow builds<01:01:32.240> better<01:01:32.440> representations<01:01:33.200> cuz<01:01:33.440> it's" + }, + { + "start": 3694.15, + "duration": 0.0, + "text": "builds better representations cuz it's" + }, + { + "start": 3694.16, + "duration": 0.0, + "text": "builds better representations cuz it's something<01:01:34.360> like<01:01:34.480> that.<01:01:34.640> I<01:01:34.680> don't<01:01:34.800> know." + }, + { + "start": 3695.07, + "duration": 0.0, + "text": "something like that. I don't know." + }, + { + "start": 3695.08, + "duration": 0.0, + "text": "something like that. I don't know. >> there's<01:01:35.200> universal<01:01:35.520> laws<01:01:35.840> and<01:01:35.960> patterns<01:01:36.360> and" + }, + { + "start": 3696.43, + "duration": 0.0, + "text": ">> there's universal laws and patterns and" + }, + { + "start": 3696.44, + "duration": 0.0, + "text": ">> there's universal laws and patterns and such<01:01:36.680> that<01:01:36.800> just<01:01:36.960> exist<01:01:37.240> in<01:01:37.320> reality,<01:01:37.720> right?" + }, + { + "start": 3697.95, + "duration": 0.0, + "text": "such that just exist in reality, right?" + }, + { + "start": 3697.96, + "duration": 0.0, + "text": "such that just exist in reality, right? And<01:01:38.080> like<01:01:38.280> that's<01:01:38.720> maybe<01:01:38.960> what<01:01:39.160> you're<01:01:39.720> what" + }, + { + "start": 3699.83, + "duration": 0.0, + "text": "And like that's maybe what you're what" + }, + { + "start": 3699.84, + "duration": 0.0, + "text": "And like that's maybe what you're what it's<01:01:40.040> using<01:01:40.400> somehow?<01:01:41.080> That's<01:01:41.680> maybe<01:01:42.000> yeah." + }, + { + "start": 3702.23, + "duration": 0.0, + "text": "it's using somehow? That's maybe yeah." + }, + { + "start": 3702.24, + "duration": 0.0, + "text": "it's using somehow? That's maybe yeah. I'm<01:01:42.320> not<01:01:42.440> I'm<01:01:42.520> really<01:01:42.720> not<01:01:42.880> sure.<01:01:43.080> That's" + }, + { + "start": 3704.07, + "duration": 0.0, + "text": "I'm not I'm really not sure. That's" + }, + { + "start": 3704.08, + "duration": 0.0, + "text": "I'm not I'm really not sure. That's I<01:01:44.160> would<01:01:44.360> I<01:01:44.400> would<01:01:44.560> like<01:01:44.680> to<01:01:44.800> see<01:01:44.920> more" + }, + { + "start": 3705.07, + "duration": 0.0, + "text": "I would I would like to see more" + }, + { + "start": 3705.08, + "duration": 0.0, + "text": "I would I would like to see more rigorous<01:01:45.440> attempts<01:01:45.720> to<01:01:45.800> answer<01:01:46.000> that" + }, + { + "start": 3706.11, + "duration": 0.0, + "text": "rigorous attempts to answer that" + }, + { + "start": 3706.12, + "duration": 0.0, + "text": "rigorous attempts to answer that question<01:01:46.400> rather<01:01:46.600> than<01:01:46.760> me<01:01:47.280> like<01:01:47.440> theory" + }, + { + "start": 3707.67, + "duration": 0.0, + "text": "question rather than me like theory" + }, + { + "start": 3707.68, + "duration": 0.0, + "text": "question rather than me like theory crafting." + }, + { + "start": 3710.23, + "duration": 0.0, + "text": "crafting." + }, + { + "start": 3710.24, + "duration": 0.0, + "text": "crafting. Are<01:01:50.480> there<01:01:50.560> any<01:01:50.640> more<01:01:50.720> questions?" + }, + { + "start": 3711.91, + "duration": 0.0, + "text": "Are there any more questions?" + }, + { + "start": 3711.92, + "duration": 0.0, + "text": "Are there any more questions? No." + }, + { + "start": 3712.83, + "duration": 0.0, + "text": "No." + }, + { + "start": 3712.84, + "duration": 0.0, + "text": "No. I<01:01:52.960> mean<01:01:53.120> just<01:01:53.360> earlier<01:01:53.680> like<01:01:54.240> uh" + }, + { + "start": 3714.99, + "duration": 0.0, + "text": "I mean just earlier like uh" + }, + { + "start": 3715.0, + "duration": 0.0, + "text": "I mean just earlier like uh said<01:01:55.200> I<01:01:55.280> you<01:01:55.360> beat<01:01:55.520> me<01:01:55.640> to<01:01:55.760> it,<01:01:55.840> but<01:01:55.960> I<01:01:56.000> was" + }, + { + "start": 3716.11, + "duration": 0.0, + "text": "said I you beat me to it, but I was" + }, + { + "start": 3716.12, + "duration": 0.0, + "text": "said I you beat me to it, but I was going<01:01:56.240> to<01:01:56.320> ask<01:01:56.560> about<01:01:56.800> like<01:01:57.120> uh<01:01:57.160> just<01:01:57.320> the" + }, + { + "start": 3717.39, + "duration": 0.0, + "text": "going to ask about like uh just the" + }, + { + "start": 3717.4, + "duration": 0.0, + "text": "going to ask about like uh just the noise<01:01:57.880> like<01:01:58.040> the<01:01:58.120> back<01:01:58.360> of<01:01:58.520> noise<01:01:58.880> of<01:01:59.400> AI" + }, + { + "start": 3719.51, + "duration": 0.0, + "text": "noise like the back of noise of AI" + }, + { + "start": 3719.52, + "duration": 0.0, + "text": "noise like the back of noise of AI generative<01:01:59.960> AI<01:02:00.200> on<01:02:00.320> the<01:02:00.400> internet,<01:02:00.800> right?" + }, + { + "start": 3721.23, + "duration": 0.0, + "text": "generative AI on the internet, right?" + }, + { + "start": 3721.24, + "duration": 0.0, + "text": "generative AI on the internet, right? Because<01:02:01.480> like<01:02:01.720> if<01:02:01.880> you<01:02:02.040> lower<01:02:02.280> the<01:02:02.360> friction" + }, + { + "start": 3722.75, + "duration": 0.0, + "text": "Because like if you lower the friction" + }, + { + "start": 3722.76, + "duration": 0.0, + "text": "Because like if you lower the friction for<01:02:02.880> like<01:02:03.040> creating<01:02:03.360> content,<01:02:04.080> but<01:02:04.680> like<01:02:04.840> you" + }, + { + "start": 3724.95, + "duration": 0.0, + "text": "for like creating content, but like you" + }, + { + "start": 3724.96, + "duration": 0.0, + "text": "for like creating content, but like you said<01:02:05.200> it's<01:02:05.320> like<01:02:05.960> you're<01:02:06.200> using<01:02:06.440> existing" + }, + { + "start": 3726.87, + "duration": 0.0, + "text": "said it's like you're using existing" + }, + { + "start": 3726.88, + "duration": 0.0, + "text": "said it's like you're using existing data<01:02:07.080> transformations<01:02:07.760> to<01:02:07.920> create<01:02:08.200> things," + }, + { + "start": 3728.87, + "duration": 0.0, + "text": "data transformations to create things," + }, + { + "start": 3728.88, + "duration": 0.0, + "text": "data transformations to create things, then<01:02:09.360> I<01:02:09.440> mean<01:02:09.720> wouldn't<01:02:10.040> the<01:02:10.080> problem<01:02:10.360> be<01:02:10.480> like" + }, + { + "start": 3730.71, + "duration": 0.0, + "text": "then I mean wouldn't the problem be like" + }, + { + "start": 3730.72, + "duration": 0.0, + "text": "then I mean wouldn't the problem be like you<01:02:10.800> just<01:02:11.000> end<01:02:11.120> up<01:02:11.200> with<01:02:11.320> so<01:02:11.400> much<01:02:11.640> stuff<01:02:12.200> that" + }, + { + "start": 3732.31, + "duration": 0.0, + "text": "you just end up with so much stuff that" + }, + { + "start": 3732.32, + "duration": 0.0, + "text": "you just end up with so much stuff that is<01:02:12.440> all<01:02:12.640> from<01:02:12.840> old<01:02:13.120> transformations?<01:02:13.920> It's" + }, + { + "start": 3734.03, + "duration": 0.0, + "text": "is all from old transformations? It's" + }, + { + "start": 3734.04, + "duration": 0.0, + "text": "is all from old transformations? It's really<01:02:14.240> hard<01:02:14.480> to<01:02:14.600> find<01:02:15.160> novelty<01:02:15.840> among<01:02:16.040> that." + }, + { + "start": 3736.43, + "duration": 0.0, + "text": "really hard to find novelty among that." + }, + { + "start": 3736.44, + "duration": 0.0, + "text": "really hard to find novelty among that. So<01:02:16.520> like<01:02:16.720> Um<01:02:17.080> search<01:02:17.320> engines<01:02:17.560> get<01:02:17.880> good" + }, + { + "start": 3738.03, + "duration": 0.0, + "text": "So like Um search engines get good" + }, + { + "start": 3738.04, + "duration": 0.0, + "text": "So like Um search engines get good enough<01:02:18.280> that<01:02:18.400> like<01:02:18.640> we<01:02:18.720> don't<01:02:18.960> have<01:02:19.080> a<01:02:19.160> problem" + }, + { + "start": 3739.47, + "duration": 0.0, + "text": "enough that like we don't have a problem" + }, + { + "start": 3739.48, + "duration": 0.0, + "text": "enough that like we don't have a problem or<01:02:19.600> like<01:02:20.160> No,<01:02:20.480> there<01:02:20.680> there's<01:02:20.880> just<01:02:21.080> very<01:02:21.360> easy" + }, + { + "start": 3741.59, + "duration": 0.0, + "text": "or like No, there there's just very easy" + }, + { + "start": 3741.6, + "duration": 0.0, + "text": "or like No, there there's just very easy ways<01:02:21.840> to<01:02:21.960> detect<01:02:22.520> if<01:02:22.640> something<01:02:23.080> was<01:02:23.320> output" + }, + { + "start": 3743.67, + "duration": 0.0, + "text": "ways to detect if something was output" + }, + { + "start": 3743.68, + "duration": 0.0, + "text": "ways to detect if something was output by<01:02:23.880> AI<01:02:24.200> for<01:02:24.360> the<01:02:24.480> most<01:02:25.080> part<01:02:25.480> or<01:02:25.560> like<01:02:25.720> ways<01:02:25.920> to" + }, + { + "start": 3746.03, + "duration": 0.0, + "text": "by AI for the most part or like ways to" + }, + { + "start": 3746.04, + "duration": 0.0, + "text": "by AI for the most part or like ways to filter<01:02:26.440> for<01:02:27.040> data<01:02:27.320> that<01:02:27.520> is<01:02:27.720> like<01:02:27.840> the<01:02:27.920> most" + }, + { + "start": 3748.11, + "duration": 0.0, + "text": "filter for data that is like the most" + }, + { + "start": 3748.12, + "duration": 0.0, + "text": "filter for data that is like the most confusing<01:02:28.600> for<01:02:28.680> the<01:02:28.800> model.<01:02:29.120> And<01:02:29.240> so<01:02:29.360> these" + }, + { + "start": 3749.51, + "duration": 0.0, + "text": "confusing for the model. And so these" + }, + { + "start": 3749.52, + "duration": 0.0, + "text": "confusing for the model. And so these things<01:02:29.720> are<01:02:29.800> not<01:02:30.120> going<01:02:30.560> to<01:02:30.640> be<01:02:30.760> a<01:02:30.800> serious" + }, + { + "start": 3751.07, + "duration": 0.0, + "text": "things are not going to be a serious" + }, + { + "start": 3751.08, + "duration": 0.0, + "text": "things are not going to be a serious problem<01:02:31.400> for<01:02:31.520> deep<01:02:31.760> learning." + }, + { + "start": 3752.75, + "duration": 0.0, + "text": "problem for deep learning." + }, + { + "start": 3752.76, + "duration": 0.0, + "text": "problem for deep learning. at<01:02:32.920> scale.<01:02:33.520> Detecting<01:02:33.960> AI<01:02:34.320> or<01:02:34.400> detecting" + }, + { + "start": 3754.79, + "duration": 0.0, + "text": "at scale. Detecting AI or detecting" + }, + { + "start": 3754.8, + "duration": 0.0, + "text": "at scale. Detecting AI or detecting detecting<01:02:35.160> information<01:02:35.600> that<01:02:35.720> AI<01:02:35.920> has<01:02:36.040> not" + }, + { + "start": 3756.19, + "duration": 0.0, + "text": "detecting information that AI has not" + }, + { + "start": 3756.2, + "duration": 0.0, + "text": "detecting information that AI has not seen<01:02:36.360> before<01:02:36.640> that's<01:02:36.880> like<01:02:37.040> possible<01:02:37.440> so<01:02:37.520> this" + }, + { + "start": 3759.15, + "duration": 0.0, + "text": "seen before that's like possible so this" + }, + { + "start": 3759.16, + "duration": 0.0, + "text": "seen before that's like possible so this that's<01:02:39.400> something<01:02:39.680> that's<01:02:39.840> like<01:02:40.040> easy<01:02:40.320> you" + }, + { + "start": 3760.43, + "duration": 0.0, + "text": "that's something that's like easy you" + }, + { + "start": 3760.44, + "duration": 0.0, + "text": "that's something that's like easy you say?<01:02:41.000> Yeah,<01:02:41.240> you<01:02:41.360> can<01:02:41.520> just<01:02:41.880> tell<01:02:42.240> which<01:02:42.840> you" + }, + { + "start": 3762.91, + "duration": 0.0, + "text": "say? Yeah, you can just tell which you" + }, + { + "start": 3762.92, + "duration": 0.0, + "text": "say? Yeah, you can just tell which you know<01:02:43.760> training<01:02:44.320> seems<01:02:44.600> to<01:02:44.680> be<01:02:44.760> the<01:02:44.880> most" + }, + { + "start": 3765.11, + "duration": 0.0, + "text": "know training seems to be the most" + }, + { + "start": 3765.12, + "duration": 0.0, + "text": "know training seems to be the most confusing<01:02:45.800> to<01:02:46.000> it.<01:02:46.160> Maybe<01:02:46.320> this<01:02:46.480> problem<01:02:46.720> is" + }, + { + "start": 3766.83, + "duration": 0.0, + "text": "confusing to it. Maybe this problem is" + }, + { + "start": 3766.84, + "duration": 0.0, + "text": "confusing to it. Maybe this problem is more<01:02:47.320> complicated<01:02:47.960> than" + }, + { + "start": 3769.03, + "duration": 0.0, + "text": "more complicated than" + }, + { + "start": 3769.04, + "duration": 0.0, + "text": "more complicated than what<01:02:49.240> I'm<01:02:49.360> thinking<01:02:49.720> because<01:02:50.040> like<01:02:50.200> maybe" + }, + { + "start": 3772.23, + "duration": 0.0, + "text": "what I'm thinking because like maybe" + }, + { + "start": 3772.24, + "duration": 0.0, + "text": "what I'm thinking because like maybe I<01:02:52.360> don't<01:02:52.480> I<01:02:52.520> don't<01:02:52.720> know.<01:02:52.880> I'm<01:02:53.040> sure<01:02:53.200> there's" + }, + { + "start": 3773.75, + "duration": 0.0, + "text": "I don't I don't know. I'm sure there's" + }, + { + "start": 3773.76, + "duration": 0.0, + "text": "I don't I don't know. I'm sure there's there's<01:02:53.960> much<01:02:54.160> smarter<01:02:54.360> deep<01:02:54.520> learning" + }, + { + "start": 3774.71, + "duration": 0.0, + "text": "there's much smarter deep learning" + }, + { + "start": 3774.72, + "duration": 0.0, + "text": "there's much smarter deep learning researchers<01:02:55.080> who<01:02:55.160> have<01:02:55.280> answered<01:02:55.560> that" + }, + { + "start": 3775.67, + "duration": 0.0, + "text": "researchers who have answered that" + }, + { + "start": 3775.68, + "duration": 0.0, + "text": "researchers who have answered that question<01:02:55.960> already<01:02:56.200> probably,<01:02:56.640> but<01:02:57.120> my<01:02:57.520> my" + }, + { + "start": 3777.67, + "duration": 0.0, + "text": "question already probably, but my my" + }, + { + "start": 3777.68, + "duration": 0.0, + "text": "question already probably, but my my guess<01:02:58.080> would<01:02:58.200> be<01:02:58.520> that<01:02:58.880> like" + }, + { + "start": 3780.07, + "duration": 0.0, + "text": "guess would be that like" + }, + { + "start": 3780.08, + "duration": 0.0, + "text": "guess would be that like it's<01:03:00.360> probably<01:03:00.840> not<01:03:01.080> a<01:03:01.160> real<01:03:01.440> issue.<01:03:01.960> Okay," + }, + { + "start": 3782.19, + "duration": 0.0, + "text": "it's probably not a real issue. Okay," + }, + { + "start": 3782.2, + "duration": 0.0, + "text": "it's probably not a real issue. Okay, because<01:03:02.440> if<01:03:02.600> you<01:03:02.680> could<01:03:02.920> just<01:03:03.160> filter<01:03:03.480> out" + }, + { + "start": 3783.59, + "duration": 0.0, + "text": "because if you could just filter out" + }, + { + "start": 3783.6, + "duration": 0.0, + "text": "because if you could just filter out everything<01:03:03.880> that's<01:03:04.000> AI<01:03:04.120> generated<01:03:04.520> it'd<01:03:04.600> be" + }, + { + "start": 3784.67, + "duration": 0.0, + "text": "everything that's AI generated it'd be" + }, + { + "start": 3784.68, + "duration": 0.0, + "text": "everything that's AI generated it'd be really<01:03:04.920> great,<01:03:05.520> you<01:03:05.600> know.<01:03:05.960> Yes,<01:03:06.280> they<01:03:06.400> can<01:03:06.880> I" + }, + { + "start": 3786.95, + "duration": 0.0, + "text": "really great, you know. Yes, they can I" + }, + { + "start": 3786.96, + "duration": 0.0, + "text": "really great, you know. Yes, they can I think<01:03:07.160> you<01:03:07.240> can<01:03:07.360> do<01:03:07.480> that,<01:03:07.680> yeah.<01:03:08.360> Oh<01:03:08.560> yeah." + }, + { + "start": 3789.43, + "duration": 0.0, + "text": "think you can do that, yeah. Oh yeah." + }, + { + "start": 3789.44, + "duration": 0.0, + "text": "think you can do that, yeah. Oh yeah. I<01:03:09.560> hope<01:03:09.760> so." + }, + { + "start": 3790.91, + "duration": 0.0, + "text": "I hope so." + }, + { + "start": 3790.92, + "duration": 0.0, + "text": "I hope so. All<01:03:11.040> right,<01:03:11.160> we're<01:03:11.280> out<01:03:11.360> of<01:03:11.440> time.<01:03:12.000> We're<01:03:12.080> out" + }, + { + "start": 3792.19, + "duration": 0.0, + "text": "All right, we're out of time. We're out" + }, + { + "start": 3792.2, + "duration": 0.0, + "text": "All right, we're out of time. We're out of<01:03:12.280> time.<01:03:12.840> Round<01:03:13.040> of<01:03:13.120> applause<01:03:13.480> for<01:03:13.840> Krikey." + }, + { + "start": 3804.84, + "duration": 0.0, + "text": ">> Yeah." + } + ], + "plain": "Okay, so this is very difficult because\nCameron is such a is such an like\nindescribable guy.\nUm\nyou know,\nyeah, this is very very very difficult.\nI know Cameron just as like the guy who\nuh you know,\nhe\nhe's like the guy crazy enough to just\nlike dig through the mountain and do\njust like as sudden\nsudden huge amounts of work\nuh of like, \"Oh, I cloned an application\nor I did this.\" And then at other times\nhe, you know, goes and becomes a web\ndeveloper and then complains about like\nnot being of the program anymore.\nSo, anyway, um\ncan we get a round of applause for the\ncompletely undescribable\nCameron Wrights?\nWhat's up, Craigsters? LET'S MAKE SOME\nNOISE.\nLET'S SAY YOU WANTED TO WRITE A program\nthat could tell the difference between a\ncat and a dog.\nI want you to actually think about how\nyou would do this for a second, okay?\nImagine you were given the pixels and\nyou had to make the program that said,\nyou know, cat or dog.\nYou might try something like, you know,\nlook for ear shapes with edge detection,\nmaybe like detect like the little angle\nthere, you know, if it's a pointy ear,\nit's a cat, if it's a floppy ear, it's a\ndog.\nProblem solved, right? You don't need\nany more tech, you don't need any fancy\nwhatever.\nUh-oh.\nYou know, if we look at our program\nhere, it's just a system with an input\nand an output, right? Um\ntake in the image\nand then, you know, look for whatever\nheuristic we've decided we're looking\nfor, and then we we, you know, make our\noutput, right? But, um if we keep trying\nto iterate on this, it's like, okay,\nmaybe some of them have pointy ears, but\nonly if the fur is orange, and we try to\nkeep, you know, coming up with patterns\nthat try to get at what we mean by a dog\nor cat,\num you'll realize that it quickly\nbecomes untenable, right?\nSo, the whole idea of machine learning\nis you essentially have a program that\ncreates that program. It's basically\nmeta programming, right? That's sort of\na metaphor for what it is.\nUm, you set up some architecture and\nrepeatedly optimize with respect to, you\nknow, your examples, um, a machine\nlearning model, a black box of many\nparameters. It's a learned program,\nright? It's automatic programming.\nAs a quick aside, why is it all Python?\nIt's because it's basically meta\nprogramming, right? It's a complicated\nbuild system. Uh, these training scripts\nbasically are just there to eventually,\nyou know, automatically create the\nprogram that does what you want it to\ndo. Um, and most of the heavy lifting is\nnot necessarily done by the actual code\nyou write anyways, it's done\nin the in the dot products on the GPU\nor, you know,\npiece of compute that you're using to\nactually do the machine learning. Um,\nand John Carmack for his current AI\nefforts, he's in the race to win the\nworld.\nUm,\nhe has switched to Python. I thought\nthat was notable.\nSo, a little bit about me, right? I'm\nmostly a game developer. If we go back\nto 2017, one of my first projects was a\ngame a game engine.\nUm, I think this is probably pretty\ncommon for people who I don't know like\nto program is the first thing they try\nto make is a game engine. It's kind of\nfunny.\nUm, you don't know anything about what\nyou're doing. You're like, I'm going to,\nyou know, do the thing that's the\nultimate thing.\nUm, and then much later, you know, 3\nyears later, you know, it's like I'm\nmaking a lot of individual games, I'm\nworking on the Godot engine.\nUm, you know, these are a couple of my\ngame projects.\nUm, that there on the upper left is a\nspace game called Asteris. It's\nreleasing in 2035. Um, it's got\nsomething like a Overwatch net code in\nit, right? It's like a\nbig solar system with many people\nplaying at the same time. Um,\nyeah, I I was pretty proud of how stable\nthe net code was considering how chaotic\nthat game was, right? Um,\nand then Dante's Cowboy. Um, this was no\nengine in C, 3D rendering, animated\narmatures, all from scratch, and all the\nNPCs in this game were uh\nyou know, they were controlled by an\nLLM, and this was never released because\nit's fundamentally flawed. Um\nI'm a deep learning hobbyist.\nThe majority of my work is not in deep\nlearning, so there's going to be some\ninaccuracies in what I say. It's just\nalways been interesting to me ever since\nback when I started, you know, first\ndoing games, I've also been kind of\nlooking at deep learning, which is a\nsubset of machine learning, and it's\njust always kind of interested me. Like\nI've been following it, right?\nIronically enough, actually, my first\never professional programming experience\nwas to architect a deep learning model\nfrom scratch, um make it work, and\ndeploy it in production. Um this was for\nmy college roommate's company called\nMacroHard. Um he basically made software\nthat like was a custom broadcast overlay\nfor for League of Legends teams,\nand I made a deep learning model that\npredicted which team was uh going to win\nthe game, right? So, that little\nvisualization there is what I created.\nUh this was trained and, you know, on a\nand optimized on many, many professional\ngames.\nand you know, it's it's just a pretty\ninteresting problem, right? And at that\ntime, right, I implemented uh League of\nLegends game predicting paper,\nand I couldn't reproduce the results,\nyou know, I was like, \"Well,\nI copied the thing that the paper said\nto do, and then it's, you know, it\ndoesn't work.\"\nAnd basically,\nthere's a there was a bug in the paper,\nright? This is a very important lesson\nof not just how hard machine learning\nis, but how often people mess up, right?\nAnd how easy it is to mess up and be so\nconfident in what you did that you\npublish it to the world and waste my,\nyou know, 2 weeks, like\nit's just very, very easy to mess up.\nAnd just to give a quick overview of the\nbug they had, right? When you want to\nmake sure a machine learning model\nworks, you usually have a you give it a\ntest, just like a person, right? And if\nthat machine learning model were to see\nthe answers of the test before it took\nit, uh you know, its success in that\ntest would obviously not be\nrepresentative of it of if it could do\nthe, you know, a problem, right?\nArbitrary new examples, it's not\nevidence that it generalizes. And in\nthis specific case, they had a metric,\nthe champion win rate, that was trained\non their entire um\nthat champion win rate was based on\ntheir entire data set, right? And so\nthen when it went to test the model, it\nhad the win rate for the champions that\nwere in the test set. And so it\npurported to have amazing accuracy at\npredicting who would win on the test\nset, but there was data information\npollution, right? These are the kinds of\nbugs you find in deep learning. They're\nnot something you could really\nbreakpoint in the debugger, but things\nyou have to find like a scientist,\nright? Um you almost have to like uh\nbinary search over the uh problem space\nsometimes to see what's going wrong. It\ncould be anything, right?\nAnd I want to stress again that machine\nlearning is automatic programming,\nright? In this paper, this is a machine\nlearning technique, not a deep learning\ntechnique, called random forest, where\nit's basically a series of learned if\nstatements directly, right?\nAnd you know, there might be some\nmajority voting scheme, and they have\nsome algorithm to optimize like based on\nthe training data, oh, this if statement\nshould be this, and this if statement\nshould be this. They basically are\n. They basically are\nattempting to tweak the if statements\nsuch that it, you know, is as accurate\nas possible on the training data.\nand deep learning is separate from that.\nIt's a subset of machine learning.\nAnd the story kind of goes like this,\nright? Some guy in the '60s was looking\nat this, and he was like, I think I know\nwhat these are. He was very foolish.\nyou know, he he he basically, on a very\nabstract level, said, \"Okay, you know\nwhat neurons are? You know what's going\non in the brain? You know what's\nenabling us to do all these amazing\nthings? Is little, you know, boxes. They\ntake in a bunch of numbers, and then\nthey mutate those numbers somehow. Maybe\nwhen they have enough of that number,\nthey they activate and propagate that\ninformation to the other neurons.\"\nUm if you were to zoom in really, really\nquickly on what one of those, you know,\nvery abstract human neurons that they\ndesigned is actually doing, right? Um a\nsingle neuron in a neural network,\nright? The first step is it adds all the\nweights multiplied by the incoming\nactivations, right? So you see that\nbigger arrow is a much stronger weight.\nThat tiny arrow, you know, you can even\nhave a negative weight, right? There's\nbasically a set of weights of coming\nfrom the uh neurons on the previous\nlayer, right? And then it just adds them\nlike numbers. They're all floats, right?\nNext, there's a bias that adds a bias to\nthat weighted sum.\nUm so, this could be considered like the\nactivation threshold for that neuron,\nright? So, like if the bias is -6, then\nyou would require, you know, six to be\nabove zero and to continue to propagate\nthat information. Again, this is a very\nrough, bad model of biological neurons,\nand they stumbled upon something that's\nactually pretty crazy um just by doing\nthat, right? But, the final step is to\npass it through any nonlinear like sine\nor or x squared. I'll only get to why\nthey do that in a second, right? That's\nwhat happens in a neuron. Here's some\nlike pseudocode for how you would find\nthe float value for a neuron,\nright? Um\nand you can take a look at this and\nponder it for a second, but\nyou'll you'll notice that well, if we're\nconstantly just multiplying and adding\nnumbers, that's kind of what linear\nalgebra is,\nand that is how they accelerate this and\ntalk about this in the actual, you know,\nfield, right? They don't express each\nindividual weight in some kind of array.\nThey express it as a matrix, right?\nand then there's a lot of math that\napplies when you do that, right? Um you\noptimize those weights and the bias with\na process called gradient descent,\nright? Stochastic gradient descent. Um\nI'll also get to that in a second,\nright? But, fundamentally, this is just\na diagram for this expression or that\ncode, right? When you see this diagram,\nit is simply that code, right? Um\nthat's what it's intending to describe,\nright? And so, just to kind of get at\nlike what the act what's going on with\nthe activation function, you know, why\ndo they pass it through sine, right?\nEach layer of a neural network is\nbasically a node in a transform\nhierarchy of like n dimensions, right?\nUm so, if you're a game developer, which\nyou're the target audience of this talk,\nso you should be a game developer.\nYou're very familiar with like trees of\ntransforms, right?\nUm and you know that if you have a\nseries of transforms one after the\nother, you can multiply them all and it\nbecomes one linear transform, right?\nThat's like what it means for something\nto be be linear. And when we want\ncomplicated behavior from these deep\nlearning models, uh we don't want it to\nbe reducible like that to one linear uh\ntranslation and and rotation whatever,\nright? So, you basically think about it\nas like warping the space\nsuch that the translation and offset you\ndo kind of like it it enables you to\nbasically draw a much more richer\ncontour around the data, right? Um\nAnd so, just as a as as another metaphor\nto help you as a game developer\nunderstand what's going on here, right?\nBoth that neural network and that shader\ngraph, these are both compute graphs,\nright? There's no branching. Information\nis flowing in from the left and it's\ngoing out the right. Um\nA shader operates on vertices while a\nneural network operates on an image or\nsome state or anything you want, right?\nAnd in the case of game development, an\nartist might be manually tuning the\ninputs to that shader and making it look\ngood, right? They might be nudging, you\nknow, the color and trying to make the\noutput look good or some other\nparameter. In the case of deep learning,\nthat's actually replaced by a machine,\nright? Which is which direction\nminimizes error on training for each one\nof these little uh parameters, right?\nEach one of those weights and all those\nbiases, it knows which direction it\nshould go in assuming all the other\nparameters were, you know, not moving.\nIt knows approximately which direction\nto go in to get better at the task. And\nthen it just kind of repeatedly does\nthat thousands and millions of times.\nUm so, there is a code base for this\ntalk. It's here. github.com/krickey/ai\nUm and to just get at what this code\nbase is supposed to be for a second,\nright? Deep learning libraries or, you\nknow, autograd, they let you perform\nsome arbitrary, you know, computations,\nright? Might be a matrix multiplication,\nyou know, it might be x squared or one\nof those activation functions we talked\nabout. Um and then find out how you need\nto turn or or or adjust those\nparameters, right? The weights and the\nbiases. It automatically figures out\nwhat the expression is to make those\nmove in a direction that makes the model\nget better at the task, right?\nAnd that's basically that\nthe what these deep learning libraries\nare supposed to be in a nutshell, you\nknow, TensorFlow, PyTorch. Um\nAnd you know,\njust as a little preliminary for this\ncode base, there's a lot of like tensors\nin it. Um\nbut really tensors are just arrays of\nmatrices. Um\nlike you might want to talk about like\noh uh\nWe're we're feeding in many examples\ninto this deep learning model. And so we\nwant like a third dimension to it where\nit's just an array of the matrices that\nwe're feeding into the model, right? Um\nAnd so my idea with this code base is it\nwould be really cool if instead of me\nhaving a you know, have a Python\ninterpreter, there was one C file called\nai.c, right? We had CPU and GPU deep\nlearning computer vision algorithms and\nimplementations, right? I would want it\nto be able to execute LLM local LLMs,\nright? I want you to be able to\nfine-tune those LLMs, not just the\ninference. Um I would want to hardware\naccelerate with something portable like\nSokol, right? I have some computer\nvision algorithms as well cuz those are\nfun.\nUm and unfortunately, this this this\ntalk code base right now is like super\nsuper not that, right? It has a\nhand-coded gradient calculation for a\ntwo-layer perceptron that classifies\ndigits. It's super primitive and it's\nlike 1970s stuff. There's no working\nauto grad yet, but I almost have it\nworking.\nUm and I I intend to continue working on\nai.c cuz it's kind of fun. I'll be\nsending updates in the\nin the BSC Discord about that.\nbut just as an example of what I was\ntrying to do with this little tiny, you\nknow, deep learning library is\neach of these videos are videos I got\nfrom YouTube of some kind of ideally\nanimating a scene, right? Where the\ncamera is stationary. Um the theory is\nyou should be able to train a deep\nlearning model to generate these videos\nfrom the image. Um\nyou know, the assumption is that the\nbehavior of the pixels is relatively\nsimple. Um and so I actually, even\nthough I wasn't able to get this working\nin time for the talk with ai.c, I\nHold on. Sorry, one sec.\nOh, here we go. Okay. So, I basically\njust, you know, let's just try it in\nPython really quick to see if this will\neven work at all. This is an 80 kilobyte\nof parameters neural network and it's\ngarbage, right? This kind of happens a\nlot in deep learning when you just have\na random idea and you just try it. It\nmakes something really sometimes hellish\nlike I didn't even know what that is,\nhonestly.\nI think that 80 kilobyte is way too\nsmall to encode the dynamics of a waving\ntree, probably. So I would just, you\nknow, 10x it like\nDeep learning is just very experimental\nlike this. It's not exactly known, you\nknow, what would cause the model to\ncorrectly learn how to ideally animate\nthat scene or make the clouds move,\nwhich is what I'm hoping for, right? You\ncan imagine with a model like this you\ncould take any desktop wallpaper and\njust make it ideally animate in the\nbackground, right? Especially you got it\nlike\nyou know, good enough on the battery. It\njust might be kind of fun to, you know,\nminimize it and then see stuff always\nmoving.\nSo just as a little bit of a historical\ndetour, right? That what I just\nexplained to you, that deep learning\nmodel, it the perceptron, it was this is\nan article about it in 1958, right?\nThey were saying it can tell the\ndifference between a cat and a dog. It\nis of no practical use. One day it might\nbe useful to send one into outer space\nto take impressions for us.\nSo AI has always been a grift,\nbasically, right? And it's followed\nbasically this this phase where\nthis this phase where\nall the investors pour their money in\nand they're like, \"This is going to be a\nnew form of life that will like do our\nbidding, right?\" And then it doesn't\nwork and then it happens again and then\nit doesn't work. And now it's happening\nagain and it's, you know,\nwho knows what's going to happen.\nHe just says, \"What did we learn,\nPalmer?\" Um\nAnd so me personally, you know, I my\nkind of like AI boom bust was I got\nreally into deep learning. Um you know,\nand then I right when GPT-2 started to\nhappen, I was like, \"This thing's kind\nof nonsense and it's not going to\nscale.\" So I basically was like, \"All\nright, I'm kind of done with deep\nlearning. I think it's useless.\" Um and\nso I basically it's kind of like selling\nBitcoin at $100. I never would want to\nbuy back in cuz that's realizing my\nlosses. So I'm just kind of waiting for\nit to collapse again, hopefully, maybe,\nand then I can buy.\nUm and if you think that's a funny joke\nin the audience right now, how many of\nyou have any amount of retirement or\ncapital in ETFs, right? So, I need you\nto raise your hand.\nOkay.\nYour money's being gambled on the\nsuccess of AGI right now. Roughly 30% of\nthe S&P 500 is heavily weighted on AI\nsucceeding. AI growth stories, you know,\nhigh price to earnings.\nUm so, I propose a model of history\nhere. There's basically two grifts\norbiting each other, right? Fake money\npumps the AI. Capital flees to Bitcoin\ninstead of US-related, you know, assets.\nNo more easy capital, no more AI.\nDeflationary doom spiral where GDP\nbecomes 1/10 of what it is now. This is\nan idea that is basically a\nbubble-generating phenomenon, right?\nBoth of these, you know, one paper\nbasically started the whole\nThere's like the transformer paper and\nthere's the Bitcoin papers, parallels,\nright? And if you look at this like, you\nknow, innovators-laggards curve, right?\nI think we're right about here if you\nlook at the Nvidia stock price. It fits\nthe curve.\nSo, we're at peak of grift at the\nmoment. It'll fall off soon, I think.\nAnd so, I think now's a good time to\nshort, but\nthat may or may not be financial advice\nif it works out for you.\nI do just want to say that shorting, I\nbelieve, has unlimited uh risk and you\nat best would like double your money, so\nit's probably not like a good idea.\nHere's some examples of uh just machine\nlearning in general, right? Let's go\nback to not, you know, being kind of\nschizo. Um\nSo, this is a little piece of history\nagain, right? This is called Cyc. In\n1984, some people were very upset at the\ndeep learning people. They were like,\n\"Look at this arbitrary floats and\nnumbers. I want like an interpretable\nprogram that's intelligent, right?\" So,\nthey basically set out to manually\nprogram all of human intelligence. And\nthe theory was if you just throw enough\nbodies and people at it, it'll\neventually be smart, right? They've been\ndoing this since 1984. Okay, if you ask\nthis system questions like, \"If I put a\na cup of coffee on a table and then I\npush on it, what happens?\" It'll say\nlike, \"Oh, well, you know, it explodes.\"\nIt's like still not working very well.\nthe reasoning traces are like a thousand\ndeep, you know, it's kind of nuts. It\ndidn't really work, right? The First of\nall, the reasoning isn't even\nunderstandable.\nUm this approach just doesn't It's\ncalled symbolic AI, right? And I think\nthat there's actually a modern parallel,\nright? Waymo is like psych, that project\nwhere they were trying to manually\nencode intelligence instead of automatic\nprogramming, right? This is like deep\nlearning is really a new paradigm of\nprogramming where it's automatic. You're\nautomatically creating the the\nnonlinearities that do the thing that\nyou want the computer to do, right? Um\nWaymo, they basically scan the city and\nthey put a bunch of if statements where\nthe car should go. So, they're very slow\nto roll out. So, Tesla has basically\njust recorded millions and millions of\nhours of just people driving and they're\ntraining massive massive transformers on\nhuman behavior and they're just trying\nto model it, right?\nIt's also There's another parallel with\nlike Siri and Echo Dot, right? These are\nsymbolic systems. They're often\nfrustratingly unintelligent. Um\nto add new intelligent behavior, an\nengineer has to basically, you know,\ndesign the algorithm and all that stuff.\nAnd ChatGPT is statistical. It's deep\nlearning. It's automatic programming,\nDeep learning models don't learn. That\nalways bugs me. I think it's a misnomer.\nUm that's like saying when you're game\ndeveloper, you're creating a universe.\nIt's like a very like hand-wavy way to\nsay something that I think is really\nmechanical and mundane, right? I think\nit's more accurate way to uh describe\nwhat happens with deep learning is\noptimizing, right? And it's an automatic\noptimization. Um\nation. Um\nHere's just another side note, right?\nSo, there's big copyright disputes with\nthese deep learning models, right? And I\njust had a funny idea, right? Like\nyou could almost make the ar- if you\ntake the argument that uh the deep\nlearning people are trying to make,\nright? They're trying to secure their\nbag. They're trying to say, \"Well, we\nlearned the patterns from what you gave\nus. We didn't, you know, copy the pixels\nor whatever.\" That's like what JPEG\ndoes, like learns a model of the pixels.\nAnd so, it's almost like arguing that\nbecause, you know, it's JPEG and not\nPNG, uh because it's a model of the\npixels, you know, I didn't steal your\nthing.\nUm so, it's a little strange, right? Um\nI think it's kind of a bad faith\nargument cuz the patterns uh that are\nfound in deep learning um are very\ngeneric and they're not so specific to\nthe image. And so,\nit's very debatable, I think. I'm not\nsure what's going to happen. But just to\ngo back to my League of Legends task,\nright? If we if we just want to think\nabout that little bit more as an\nengineer,\num you might naively just put in the\ncurrent state of where all the people\nare on the map, how much gold they have,\nwhat items they have, throw it through\nthat perceptron I showed you, and maybe\nit'll automatically find ways to relate,\noh, like, you know, when the gold and\nthis item is here, then like they're\nmore likely to win than you would think,\nand stuff like that. These are the kind\nof patterns you hope a deep learning\nmodel finds as it optimizes based on\nmassive amount of data you have, right?\nThere's a problem with this, though.\nWhat about past states, right? Past\nstate informs the nature of present\nstate. If somebody was performing badly\n. If somebody was performing badly\non your team in the past, right? They\nwere like running it down mid, this\nshould inform your current prediction\nlike heavily, right? Um\nSo, this uh you know\nYou need a recurrent neural network for\nthis, right? A recurrent neural network\nis basically two of those perceptrons\nthat I showed you. One is modifying the\nstate almost, and then another one is\nvisualizing the state towards whatever\nlike you want it to visualize, right? It\nmight be probability for who's going to\nwin, or it might be like you could\nreplace it with another deep learning\nnetwork and train it to visualize like\nI don't know, which which champion is\nlike has the most synergy with the other\nchampion, stuff like this, right? Um it\nhas state. It's much more like a\nautomatic program, right? Automatically\nlearned program. Um the hidden state of\na recurrent neural network is like RAM.\nIt's analogous to a RAM.\nUm and that and that is why I think that\nthese are\nbest described as automatic programs,\nright? Um and so, what I did at the time\nto successfully model uh who would win\nin League of Legends games was I I I I\nselected a GRU architecture, and I spent\nweeks tuning it and and messing around,\nand I I got I managed to make it so\nhalfway through the game it could\npredict who would win with 86% accuracy.\nUm if you just went with whoever had\nmore gold, it was like 73% accuracy. Um\nand then this continued to generalize\neven on arbitrarily new games. Um and\njust as, you know, as a side note for\ngame developers, you could use a\ntechnology like this for our play of the\ngame kind of feature, right? To see big\nunpredictable swings in who's winning.\nUm we actually did do that. We took a\nour giant corpus of professional games,\nand I ran this model and saw what were\nthe biggest swings, and we looked at\nthose games, and they were the most\nexciting matches, right? So, this is\nkind of like just to give you ideas for\nlike what this technology can be used\nfor, right? Um each perceptron in a\nlayer is like roughly analogous to an if\nstatement in the program, right? Like\nanother branch. It lets you make a more\nspecific contour around what you're\ntrying to encapsulate, right? Um you're\nadding more resolution to that boundary.\nUm this is why it's called deep learning\ncuz there's many Once you add many\nlayers, you can make very complicated\nbehavior, right? Um the problem if you\njust take that perceptron and you add a\nbunch of layers is it adds way, way, way\ntoo many parameters and the gradients\nget very tiny, right? Um it doesn't\noptimize, right? The optimizer doesn't\nwork. I mean, you saw my failure of deep\nlearning. It's not easy.\nUh you almost have to like guess and\ncheck if you really aren't, you know,\nbranching off of some established known\nworking uh architecture, right? Um and\nso all of the field of of modern deep\nlearning is basically about how do we\nadd more parameters to this model and\nstill get those parameters to be\noptimized into something that\ngeneralizes to what we needed to do,\nright? In the case of a convolutional\nneural network over here, they basically\nthey kind of start with the inductive\nbias that like, \"Okay,\num when you're looking to see if a image\nis of a cat or a dog, um and you\nbasically are hoping that let's say this\ndeep learning model will pick up on the\npointy ear thing. It doesn't matter\nwhere in the image that pointy ear is,\nright?\" Um they basically want these\ndeep learning models to learn, you know,\nfeatures that it doesn't matter where\nthey are in the image. Uh if you just\nfed every single pixel and all the RGB\nvalues into a giant perceptron, then it\nwould have to learn that, you know, cat\near means cat in every single spot,\nright? That's not really what we do when\nwe recognize images and so they\nbasically That's like a little bit of\nkey insight. That's what deep learning\nresearchers are trying to do is like\ncome up with ways to to formulate and\nstructure the computation, the graph of\ncomputation, such that when you optimize\nit, it gets what you want, right? Um and\na transformer on the other end here I'll\ntalk about in a second. Um this is\nanother example of an architectural\nchange. You can add way more parameters\nand it's still stable and it's still,\nyou know, optimizing correctly, right?\nJust as another little side note, some\ndeep learning techniques are a little\nsilly. Uh Uh this is called dropout. You\nrandomly turn off neurons during the\nduring the training. It's it's supposed\nto present or or prevent memorizing the\ntraining data um and hopefully make it\nso neurons aren't exactly specialized\nWho knows why this works? Honestly, I\ndon't I haven't seen any real convincing\nargument in any paper for what exactly\nthis is doing. Like basically what's\nhappening in deep learning is people try\nstuff and they have some very like okay,\nyeah, that sounds about right. I think\nthis what's happening and then it's like\nthere's no rigor in this field basically\nlike um it's really crazy. This is\nactually used in production for chat GPT\nwithout it like it wouldn't work, I\nthink. Um and it's like you know, we\njust randomly turn off the neurons\nduring the training. Like okay.\nSo, there's a zillion videos that\nexplain transformers already. I'm not\ngoing to give you an explanation of the\ntransformer, which is you know, the\ncause of the current rift, right? Um and\nyou know, you can watch any of these and\nthey're good and whatever. But I'm just\ngoing to give you a quick overview. This\nis a single slide. This is what a\ntransformer is doing, right? It's\nbasically going from a sentence, a\nlookup table per token into a point in a\nspace where like the closer two points\nare the more related they are, right? It\nadds a position encoding to those points\nso that, you know, if we have an array\nof these points, then the points are\nmodulated in such a way that the the\ndeep learning models they can they can\npick up where they are in the sentence,\nright? They're almost jiggle in a way\nwhere like you could learn that pattern.\nI keep using the word learn. I'm sorry.\nIt's pervasive in the field, but I don't\nlike it. Um and then you literally just\nstack a hundred of these attention\nlayers onto those uh vectors, right? And\neach of these attention layers every\nsingle time, every single one of those\ntokens, right? It's basically looking at\nall of the past tokens and enriching the\ncurrent vector with more information\nabout the surrounding tokens and then\nmore facts potentially about that\ninformation in the form of another\nperceptron, right? it's like learning\nwhat what facts might be related to a\nword like house, right? It's It It\ncontinuously enriches each one of these\nuh words with more and more context,\nright? In a in a highly parallel easy to\ntrain manner. Um and then there's just\nthis two-layer perceptron and then it\ngives you the next word, and it does\nthat over and over again, right? Um\nthere are many hyper parameters in this\nmassive system, right? Many, many, many,\nmany things can go wrong. You know, it's\nlike what is the learning rate? You\nknow, how quickly do we turn these\nparameters to, you know, make it better,\nright? It's like a massive problem,\nactually, um to know what These are\ncalled hyper parameters because unlike\nparameters, like they're not\nautomatically learned in the core\nalgorithm of deep learning, right? You\ncan or you can try to figure out what\nthey are uh with a search algorithm.\nThat's very costly if you're OpenAI and\neach training run costs $10 million. Um\nand as it turns out, we heard from an\ninsider here that most of the big gains\nfrom these, you know,\nmost of the state-of-the-art models that\ncome out of OpenAI, they came from\nso-called YOLO runs, where they\nbasically they basically asked Karpathy\nor one of these, you know, like deep\nlearning generals, like, \"Hey,\nwe're going to give you $20 million to\njust try something, right? You just pick\nnumbers, pick hyper parameters, pick\nwhatever you want, and we'll see what\nhappens.\" And state-of-the-art model,\nright? It's basically vibe engineering.\nVibe engineering created the vibe coding\nmachines, and everybody's vibing, and\nthen, you know, we're being wireheaded,\nand then it's over.\nWell,\nI I I I I think that it's a very\ninteresting question, can these LLMs\nreason, right? So, they can't reason cuz\nthey're not human. I think that's a good\nway to use the definition of reason. Um\nbut let's say we want to figure out if\nthese LLMs can, you know, output text as\nif they were reasoning, right? I think\nthat this benchmark ARC-AGI is the It's\na stellar benchmark for this task,\nright? If you take a look at this\ndiagram here, right? That's uh one\nexample, and that's another example, and\nyou have to predict what happens here.\nIf you're human, which I'm assuming all\nof you are, it should be pretty obvious\nto you that you basically just put the\nyellow dots like where the little corner\nthing goes. And ARC-AGI is the kind of\nbenchmark where they basically take\nrandom people and they're like, \"All\nright, we'll give you a $15 Starbucks\ngift card if you solve this problem.\"\nAnd they verify that like people can\njust figure it out in a context-free\nmanner, right? And so, the whole idea\nwith this is they're trying to find\nproblems that like average people can\ndo, but like, you know, these LLMs\nthey're they can't do it at all. Um\nAnd ARK G I they were originally, you\nknow, totally defeated by OpenAI's\nlatest reasoning model.\nUm it it got human baseline on a huge\nsuite of these that are private, not\nanywhere existing on the internet. Um\nyou know, generic Each each ARK G I\nproblem requires a new unique insight.\nThey're basically puzzle game problems,\nright? They require some kind of like,\n\"Oh, that's how it works,\" right? And\nLLMs have already beaten the first ARK G\nI benchmark. Which that was quite\nsignificant. I didn't think that would\nhappen. And neither did the author of\nthe benchmark. He's a massive skeptic\nand he thinks it's a bubble. And then\nOpenAI pulled them on stream and he was\nlike, \"My benchmark is defeated.\" Um so\nwe made a new benchmark that they're all\nfailing at again.\nAnd that's Grok 4. A few days ago they\njust doubled the state of the art and\ngot 15% the 16% of the questions right.\nUm So just circle back on some\napplications to games really quick as I\nkind of\nclose up on this kind of like lightning\noverview of this, right? Any problem\nwhere you have input to output data and\nit's harder to find programmatically\nwhat that mapping is, that's where you\nwould throw something like deep learning\nat the problem and try to automatically\nprogram that, you know, uh\ntransformation.\nUm this is Nerf rendering. This is a\nvery interesting application of deep\nlearning to the problem of rendering, I\nthink. Um it's very old and there's much\nbetter techniques like uh Gaussian\nsplatting.\nUm but Nerf rendering is basically given\na point in the scene, a deep learning\nmodel learns how dense it is and then\nit, you know, kind of traces rays\nthrough that learned density to render.\nAnd just as a little interesting tidbit\nmore vibe uh engineering, uh basically\ndeep learning models are very bad at\nlearning so-called high-frequency um\nhigh-frequency functions, right? So a\nlittle tiny thin metal pole is a very\nhigh-frequency function density function\nfor it to learn. It's like nothing\nnothing nothing bam really thin nothing\nnothing nothing. It's very like tight\n. It's very like tight\nand so it's hard for a deep learning\nmodel to pick up on that cuz it would\nrequire a lot of non-linearity. It would\nrequire a lot of layers, right? Um and\nlayers, right? Um and\nso what they did in Nerf to, you know,\nachieve state-of-the-art results for\nlearning based rendering, is they pass\nthat position into\ninto a bunch of different sine waves and\nthey basically just made it easier for\nthe deep learning model to pick up on\nhigher frequency spatial information. I\nthought that was kind of funny. She's\nlike, pass it into the sine waves and\nthen it's better.\nUm this is called Ash Animatable\nGaussian Splatting. Let me show you how\nthe video works here. This is basically\nphotorealistic human rendering with with\naccurate physics on their clothes done\nin real time. Um it's just\nunfortunately, in order to, you know,\nscan somebody, you need about 100\ncameras synchronized and it's very\nexpensive. So, maybe they'll figure out\nhow to make that cheaper someday, but in\nthis specific paper, a deep learning\nmodel is used to go from a texture that\ndescribes the skeletal deformation of\nwhat you want the person to be in into\nthe the view dependent color that you\nwould see\num from that skeletal deformation. And\nand essentially, you need to train that\ndeep learning model to learn that\nrelationship between like how their arm\nis bending and what their clothes look\nlike, um you know, per person. Um\nAnd then of course, I'm sure probably\nall of you guys already know, ray\ntracing is is is greatly benefited by um\ndeep learning because it's used to\ndenoise, right? Um\nDeep learning is used in a algorithm\ncalled AlphaZero to master any discrete\nperfect information game. Basically, a a\nneural network goes from the board state\nto a likely next move\num and in a win probability. And then\nanother algorithm manipulates that those\nprobabilities and you know,\nbeats any human in any game we've ever\ninvented, unfortunately.\nAutomatic game AI, so you can make a\nstyle transfer, you know,\nyou can make it some kind of style\ntransfer for something like this for\nyour video game where maybe like you\nwant it to act more like the human\npeople. So, once it masters the game,\nyou then train it to act like a person\nand then it would be a kind of\ninteresting mix between a good player\nand oh, it plays like this specific\nchess player or whatever.\nUm or you could even have it validate\nthat your game is uh balanced and stuff\nlike that, right? I think this actually\nis very it's quite green field for video\ngames in general for how to apply uh\ndeep learning models in a lot of places.\nAnd you know, for automating like a call\ncenter, I think that there's a lot of\nthese call center AI startups. You know,\nit seems easy on the surface, but\nusually people call support because they\nwant to talk to a person and not a\nmachine. So, I think that this is\nprobably going to be a a giant disaster.\nUm and, you know, of course, automating\nsomething like engineering is way more\ncomplicated than just that, which isn't\neven working. So,\nyou know, just something to think about.\nUm and a quick note on computer vision\ncuz I didn't really get to flush that\nout that much in this talk. But, the\nbasic idea I had was for computer\nvision, you're essentially trying to\ntake in a bunch of images in something\ncalled monocular slam and get some kind\nof representation of the world that it's\nseeing um automatically, right? It's a\nbit like deep learning where you have\nsome automatic optimization process that\noperates on a lot of data to uncover\nsomething interesting about the world.\nUm hopefully interesting. Um and it this\nthat that visual slam pipeline is\nbasically the opposite of game\nrendering. And it involves somewhat\nsimilar linear algebra, right? Um you go\nfrom an image to easy to track key\npoints, pose estimate for the camera,\nsome representation of the scene for the\nrobot to drive around in, right? Yeah,\ngame rendering, you know, it's like\nscene representation, triangles, pixels.\nI thought that was interesting.\nSo, will LLMs replace knowledge work en\nmasse someday, right?\nI put a steelman yes and a steelman no\nhere.\nUm really, I think it's probably\nprobably not. But,\nthere's like a lot of investors who\ndisagree with me, so I don't know.\nnew knowledge can't be made by machines,\nright? This is a very funny joke made by\na philosopher called David Deutsch. Um\nmaybe this is what we're going to have\nto do someday. You're going to have to\ninvent new scientific knowledge to prove\nthat you're not a robot. Cuz maybe\nthat's the one thing the one thing that\nhumans can always do is create new\nknowledge cuz they're the only ones that\ncan officially reason.\nBut, if, you know, you have to have a\nhuman know if you did new knowledge, I\ndon't know how a system would detect if\nit's new knowledge or not. Maybe that's\nlike P and NP or something, I don't\nknow. Um\nWait until RKG I continues to be beaten,\nI think, is how you monitor this\nproblem.\nUm cuz if that continues to be beaten, I\nthink uh you know, that's crazy.\nThanks for listening. Don't forget to\ncheck out and work on github.com /ai.\nI want to make ai.c so that it's like\nfun to mess around with these models and\ntechniques and games.\nNo dependencies, no python, no nothing.\nI you know, I just it's like it's not\ngood right now. It needs to be rewritten\nlike 10 times. I've never made a deep\nlearning framework before, but\nThanks for listening to my talk.\nSam.\nOkay, that's a bit similar to the psych\nstuff.\nIt's used by cookie people for AI, but\nnot serious contenders.\nHello.\nWhat's up?\n>> Good.\nAwesome. Can we get another round of\napplause for Krakey?\nSubscribe to my YouTube. Subscribe to\nhis YouTube. Watch\neye contact practice. It's the best one.\nI mean, that talk was definitely like it\nwas like 15% skits, so\nbut that's why it's entertaining for\nsure.\nI think you're like part LLM.\nLike anyone that's talked to Krakey, you\ncould just give him a prompt and he just\n[ __ ] spits out infinite like\nsometimes he hallucinates. Yeah,\nsometimes it's garbage, but most of the\nit's garbage, but most of the\ntime it's kind of right. Okay.\nThat is what that's kind of what LLM's\ndo, yeah. Yeah, exactly. So you're it's\nin there.\nIt's interesting.\nSo everyone here at Better Software\nConference\nwants to write better software and wants\nto live in a world where there is better\nsoftware.\nSo will AI cause the destruction of\nbetter software?\nSo, okay, one of the biggest like hot\ntopics for for AI startups right now is,\nwell, if AI writes the programs, right,\nwhy do we even need an app store? Why do\nwe even need software engineers? You\nshould basically tell your computer the\napp you want and then it does it and\nthen it gives you the interface and then\nthe data. Um\nand it doesn't work at all right now and\nso, I you know, it would We're going to\nhave to see some new developments, I\nthink, to judge if that will happen or\nnot.\nI'm basically 90% it won't cause the\ndestruction of better software for at\nleast a decade,\nbut 10%, you know, there's always\nsomething around the corner. Grok 4 is\nlike much better after some reason.\nSo, in 10 years, like\nis software engineering a thing? Like,\nwhen you walk into a room or people or\nsoftware engineers going to be crafting\nprompts? So, I've been thinking about\nthis a lot. I mean, obviously, prompt\nengineering is not going to be a real\nthing because if you do believe these AI\nsystems will get to the point where they\ncan automate software engineering, they\ncan do the prompting, too. I don't think\nit's that crazy.\nUm but in any case,\nI think that the effect of this on\nsoftware I Well, I I've I've been\nthinking a lot about this question and\nI'm I'm still very uncertain, but if\nsoftware engineers are basically making\nthese like virtual machines that do some\nkind of processing or transformation\nthat's valuable to people and these LLMs\nlearned the patterns\nuh to to to create some of these\nmachines as it relates to the to what\nyou're asking of it and I think people\nwho are making machines that are very\nsame-ish, software engineers who are\ndata transformations that are similar to\nwhat exist are going to probably be\nscrewed. Um but if you're making unique\ndata transformations that are unlike\nwhat have been made before, kind of\nunique machines, I don't see how these\nmodels could really create new knowledge\nlike that yet. I haven't seen evidence\nthat they can do that.\nUm you pointed out something uh\nlike learning is different from\noptimization. And that's that's more of\na philosophical thing, I think.\n>> Like, when a human is born, right?\nThey're they're born like extremely\ninefficient and they just get more\noptimized as they grow. Is that not the\nsame? It's not I think it's not the same\nbecause whatever human being is doing\ncan create human reasoning and I don't\nthink it's\nI don't think it's it's it would be very\nhard to make the claim that these things\nare doing human reasoning right now, I\nthink, for a lot of reasons. Um you\nknow, oftentimes the reasoning doesn't\ncorrespond to what they say. It's like\nnot logically coherent, right? It you\nknow,\nfundamentally from the I mean, you kind\nof understand what's happening here,\nright? It's it's optimizing and and\ncreating an automatically created\nprogram\num that might have nothing to do with\nthe the the universal causal\nrelationship between like, I don't know,\nyou know, whatever's happening in your\nbrain or maybe it's if you're religious,\nthen it's not even a question.\nLike isn't the real problem like not\nthat AI replaces software engineers and\nthat destroys better software, but that\ndestroys better software, but that\nsoftware engineers use AI and that\ndestroys better software because that's\nwhat's happening pretty much already.\nWell, if you believe that society\nfunctions, then people will choose the\nproducts that are better.\nSo,\nin the long run it, you know,\nif yeah, I don't know. I'm not sure.\n>> you pointed out uh\nthe like Grok had a massive jump\nrecently. Like did you look into that at\nall?\nOn this specifically on the Arc AGI\ntest, which interestingly\nthere's another Arc. I don't know if\nTaha point pointed that out. I missed\nit.\nEverything is Arc.\nSo, yeah, Grok might have performed\nbetter because of its utter lack of\nsafety training. That is a rumor I kind\nof saw and it seemed to be somewhat\nconfirmed by, you know, another times\nit's like they did safety training and\nthe model got much less, you know,\nreasonable in quotes.\nUm it might be the case that the safety\ntraining is not intelligent cuz it's so\ndirectly refusing to to say something,\nright?\nUm, there's all sorts of hypotheses and\nthere's no real, again, rigor in the\nIt's not really it's They're just trying\nto figure it out. They don't Nobody\nNobody really knows.\n>> like broccolicing like some schizo like\n>> It's not good. I don't want to repeat it\nIt's not good. I don't want to repeat it\nout loud. You can Google it if you want.\nIt's not good. It's It's\nIt's completely out of control, but it\ncompletely out of control, but it\nis smarter. So,\nwell, I don't know.\nOkay. Um, you didn't point out either\nlike what caused the\nthis AI the last AI winter to end,\nright? Like what was the It specifically\nhappened So, transformers were made kind\nof I think like 4 years before they were\nreally used to generate text in a very\nin a very parallel large, you know, mega\ndata scale way. Um, and it happened when\nOpenAI I think released GPT-2 and they\npointed out like, \"Hey, we keep just\nmaking these models bigger and it seems\nlike the text is getting more and more\ncoherent. So, if you just complete the\nline, like someday we get to the thing\nthat makes everybody, you know,\nPop-Tarts for free or something.\" Like\nAnd then I don't know. It's That's what\nstarted this massive speculation bubble.\nIt's like, \"Wait a second. We might have\na path to completely removing\nuh humanity from what from satisfying\nour our our desires in the physical\nrealm.\"\nSo, it was just they they just scaled it\nup. Yeah, they scaled it up and it\nseemed like the scaling wouldn't end.\nIs that what they're Is that what people\nare still doing? They just keep scaling\nthings.\n>> buying GPUs. I mean, they're not scaling\nin the same way. They're not just making\nthe models bigger. They're trying to\nscale I the so-called reasoning models.\nUm, they're basically trying to make the\nmodels more and more correct instead of\njust making them remember more and more\nfacts, basically.\nI don't want to run out of time again,\nso I'm going to open\nI'm going to do questions right now.\nWhat's up? Uh so, Cameron, a bunch of us\nhave seen your uh Dante demo just\nfloating around. Um, can you get into\nwhy it's flawed and why Yeah, so my my\nwhole thinking with Dante's Inferno is I\nthink when I saw GPT-4 first like on the\nAPI, I was like certain that this would\nenable a huge new category of of video\nUm and so I immediately began making\nlike an RPG where all the characters\nwere controlled by an LLM, not just\ntheir speech but their actions, right?\nAnd the idea was like, \"Whoa, this thing\nis basically like prototype AGI or\nwhatever, right?\" So, it'll just kind of\noperate and it'll be fun and like and I\nwent through so many iterations of that\nidea. I made it so that like they could\njoin your party and trade items and all\nthis stuff and like I played with it a\nbunch, but it never felt like there was\nany meat to the game. The problem is\ngames are about predicting and\nunderstanding systems almost, right? And\nan LLM is an unpredictable black box,\nsometimes acts like a person, sometimes\ndoesn't. It's not fun to interact with\nin a way where you're trying to get it\nto do something. At best, you're\nbasically like, \"Hey trader, I'll\ntotally give you a million dollars if\nyou drop your item.\" And then it's like,\n\"This is not fun. This is dumb.\"\nIt's funny It's funny like for one time,\nbut it's not like a game you could\nmaster.\nIt's like the vending machine at Claude.\nOkay, there's Right now, there's some AI\ncompanies are trying to first deploy\nthese these LLMs in the real world to\nstart to,\nyou know, create a utopian post-scarcity\nworld where machines make everything\nautomatically and they're like, \"Step\nnumber one, put the LLM in a vending\nmachine and have it automatically order\nand interface with the people or\nwhatever.\" And right now, people are\nbasically convincing these AI vending\nmachines to stock many tungsten cubes\nand sell them at a loss and then they go\nbankrupt.\nUm they've published many reports about\nthis and they're like, \"You know, we're\nwe're you know, we're doing some prompt\nengineering to fix the business model\nand we're trying to make it refuse\ntungsten cubes more.\" Um but\nreally, I think there might be no end to\nhumanity's ingenuity as we battle\nagainst the machines for the next few\ndecades. So.\nThat's probably fine, right? It's\nEverything's probably fine, but\neverything's probably fine, but\nsometimes bad things happen. I don't\nknow. What's up?\nDo you think there So, I I feel like\nthere is a field of research, I just\ndon't know the words for it or how big\nit is, but do you think there's a way\nthat we can reverse engineer Yeah,\ninterpretability is big. A lot of Yeah,\nAnthropic is the leader in\ninterpretability, I think. Um they\nbasically do data science on the neurons\nthemselves, and they try to make claims\nabout, you know, how the data is flowing\nthrough the model. Um you can look at\ntheir Golden Gate model. They have a lot\nof really excellent stuff you can read\nforever. It's really cool. Um Do you\nknow much about it? Like, do you feel\nlike that is the way forward, or is that\njust a byproduct of the craziness? I\ndon't think there will be any value\ncreated from interpretability research,\nhonestly, other than maybe um\noptimizing the training process somehow.\nI'm honestly not really sure. I think\nthey're mostly doing it because they\nhave to, because they want to be the\nsafe people or something.\nUm and it's cool, but it doesn't seem\nvaluable to me, but I I'm not like a big\ndeep learning research. I'm like a\nhobbyist, like I said earlier, so I'm\nnot really sure.\nUh follow-up to that, is it\ninterpretability\nthat made people figure out you could\nremove the safety neuron? You can. They\nalready figured that out. It's called\nablation. So, any single open-source\nmodel, you can basically find what's\ncalled an ablated model, where they just\nbasically subtract the safety training,\nand then it does whatever. But, it's\nBut, you told me it's like one node\nthat's like this. Yeah, they I don't\nremember exactly how it goes, but it's\nnot a node, but it's like one vector\nkind of represents the safety,\nuh the refusal. Like, when whenever an\nLLM starts to refuse, it always seems to\noutput words in a very, you know, this\nkind of sounds like the same thing, and\nit basically you can find that inside of\nthe model and then nullify it is what\nablation is, from what I remember. It's\nbeen a while.\nUh as you said that AI is basically\nmetaprogramming. Yes. So, and Julia is\nvery famous for its metaprogramming.\nWill it replace Python? Like, will we\nhave Julia AI or like So, maybe\nmetaprogramming is a bad thing, cuz it's\nnot really comparable. I was moon I was\nmostly It's like metaprogramming only to\nthe extent that you're producing a\nprogram with a program, but it's not\nlike it's automatic metaprogramming.\nIt's not, you know, mechanistic symbolic\nmetaprogramming, where you're yeah. So,\nit would not be a good application for\nthis language, next one. Huh?\nIt's not a good application for jail.\n>> No. No.\nNo.\nUh\nrough To bring more into the concrete\nrealm, you mentioned SLAM. Yes. It's\nbeen a decade since I've looked at SLAM.\nLike, is deep learning the current state\nof the achieving SLAM algorithms, or are\nthey still\n>> I I just recently asked a SLAM\nresearcher that, and they basically said\ndeep learning is like used sometimes,\nbut it's not really necessary to solve\nthe problem though. Um\nyou know, they found good heuristics to\nfind the key points that are trackable.\nUm and like neural I think that neural\nSLAM seems like it's something they've\nkind of been like rumor you know, to\nkind of take over a state of the art,\ntake over a state of the art,\nbut like you know, the Quest doesn't do\nany neural SLAM as far as I know.\nYou mentioned using uh AI models for\nbalancing games. Yes.\n>> and\nlet's say you then train a model to play\nStarCraft II, and you you finally do\ndefinitively prove that Protoss should\nbe nerfed.\nyes.\nGod, go ahead and start recording.\nUh once you nerf Protoss, do you then go\nlike, \"Okay, cool. Now we do a $3\nmillion run again?\" I don't think it\nwould be that expensive to do it on a\ngame. It's the cost of AI right now is\nbecause to make these large language\nmodels learn, you know, reasoning, you\nhave to basically shuffle the entire\ninternet in and out of um multiple\naccumulate ads. It's really not so much\ndata they have to process. For something\nlike StarCraft II or AlphaZero, it's\nreally not on that scale of cost. Um I\ncould easily see many many many games\nbeing able to do this, and for simpler\ngames, you could probably do it on your\nlaptop even maybe. Um\nBut yeah, like I think that for\nsomething like StarCraft II, it's like I\ncould see them doing these kind of\nstudies if they have the the cap backs\nand they believe it's\nyou know, it's it's valuable for the\nplayers to have have truly be officially\nlike even the genius master AI\nuh players of this game it's\neven right that might be a good thing to\nprove.\nUh yeah. Uh would you say that like\ntraining is where like mostly cost is or\nUm actually well opening I says that\ninference is their biggest cost right\nnow\nbut there's a massive push to make you\nknow on on device inference chips so I\nthink that's kind of temporary and it's\njust because like consumer my\nunderstanding is I don't know if this is\nexactly true but I've seen this floating\naround that Nvidia cards don't have\nenough VRM to run these models simply\nbecause it's not something consumers\nwanted and they kind of want to nickel\nand dime you for VRAM\nand so this seems like the kind of thing\nwhere you know if it's a healthy market\nand you know it's not there's not like\nan Nvidia monopoly or whatever\nwhich there is I'm joking it's not going\nto get better but\num\nideally I think it's totally feasible\nthat these the inference cost is zero\nfor the people who are\nmaking these models it it should be\nzero.\nUh yeah. Yeah in your little experiment\ngame how did you translate basically the\nLM output to actions in your world? So\nbasically you can view putting LMs in a\nvideo game as so you have your game\nstate and then there's like the um\nphysics\nsimulation kind of like you know human\nin the loop like you're inputting\nactions and state 60 hertz or whatever\nand there's a text adventure way to\ninteract with the game simultaneously\nright and a good way to check that like\nthe game was properly you know\npresenting itself to LMs is if you could\nflip it and as a human I could play the\ntext adventure version of it right\nthat's how you would know that it's\nbeing represented well in text and\nthat's basically what I did is like a\nstructured text adventure presented to\nthe LMs um\ndescribing what's around it describing\nwhat you're wearing whatever yeah.\nYeah. So right now when I use some\nprograms I go oh I want to do kind of\nthis thing and it used to be that I was\nable to Google how do then X in program\nY\nand some of it will be like a forum\npost. But now you can't find those\nanymore or I'm doing difficult stuff. So\nmy question is\nwhat year will I be able to put like a\ngood version of Clippy in my programs?\nSo like there's a little dude you can\nactually describe like hey I'm trying to\nrotate the mesh in this way what\nlike how does the sub stuff know what?\nThere's so many startups trying to do\nthat. It's Yeah, it's like cursor,\nNo, well it's not even cursor for X.\nIt's like there's so many people who put\nthe little chat window on the thing and\nit's like ask it a question. I just feel\nlike they never work. Like they still\nare just so bad most of the time. Like I\ndon't know when it will get good, but\nhonestly I like\nI find that when I use LLMs to\nunderstand how software works, it gives\nme an incomplete understanding because\nI'm not really like traversing through\nthe whole set of knowledge, right? I'm\njust specifically getting the answer I\nneed and implementing it and moving on\nwith my life. Whereas when I would when\nI read documentation and like actually\ngo line by line and sit down and\nunderstand it, I basically I just know\nthe unknown unknowns, right? Because I\nexplored. And so LLMs to say how do I do\nthis in this program? I I'm actually not\neven sure that's good UX, basically. But\nif if you can like search the\ndocumentation for So like you need to\ntake it put the documentation is like a\nhundred page PDF documentation of this\nprogram. I'm not going to read that.\nYeah, it is good for that. Like LLMs are\nalmost like a very it's like a better\nGoogle almost.\nWhich is why they pivoted their entire\ncompany to LLMs. I might be able to So\nat what year can I expect to like put\none of these in my C program?\nWhere it's like running locally? Yeah,\nAI.C. Yeah, cuz I want to Yes, my AI.C\nproject. To get the guy in who like\nsometimes complains Um\nLet's say let's say you know smaller\nmodels are always getting better and\nbetter. Let's let's put a let's put a\nrelease date on it and say five years.\nSo yeah, so have you seen like a model\nthat sort of can work outside of the\ntraining data?\nIt feels like they're always like\nweighted averages of the training data.\nIt's It's not a weighted average of the\ntraining data. A better way to describe\nhow they process and hold and how our\nmodels impacted by training data is\nthey find patterns and then we run tests\non those patterns to make sure they\ngeneralize and they're not specific to\nthe training data. So, they find\npatterns which are general that do the\nthing you want it to do. ARC AGI\nspecifically is a benchmark for generic\nnever-before-seen on the internet\npatterns that fail LLMs, all of them.\nThey're private. They're never leaked or\nexposed. You can't ever You can't build\nan algorithm that interpolates from\nexisting solutions to solve ARC AGI\never. And LLMs did that.\nYeah, so so yeah, I was asking sort of\nabout the answers you get because it\nfeels like you get sort of you know, you\ncannot go outside of what it has been\ntaught, right? Or sort of\noptimized on,\nbut maybe you could sort of extrapolate\nsome, you know, outside of it. But is\nthat happening like Yeah, I I Well, I\nthink that like when you get to these\nabsurd levels of scale, um\nthere's almost nothing outside the box\nif it's predicting the entire internet.\nUm that's basically like a\na huge amount of examples of uh the\ndynamics of human intelligence recorded\nin into written form that you then, you\nknow, process and find patterns which uh\napply across many examples. That's\nThat's what they're doing. They\nDeep learning systems usually start with\nvery low-frequency generic patterns\nlike, \"Oh, there's a period every\nso-and-so words.\" And then they start to\nget to higher-frequency patterns and\nmore specific, more nuanced things, and\nthen\nthey they they're they're basically\npattern finders. Patterns which apply in\non not just the training data. That's\nthe whole point of being like a deep\nlearning person is you make sure that\nthe model does not just apply on the\ntraining data. That's like the point of\nall the architectures, that's the point\nof like, you know, doing everything\nyou're doing is making sure that it's\nnot just memorizing the training data,\nthat it's finding patterns in the\ntraining data that will apply in your\nspecific task that you need the model\nfor.\nSo, so sort of like say that you\nrecognize images,\nlike you can recognize cat and dog.\nYeah. But like a model that can do train\non cat and dog images, but also like\nthis is something else. Like without you\nshowing something else.\nYes, I mean obviously that model is not\num learned patterns which would\ngeneralize to is it a turtle or not. It\nwould try to bucket into cat or dog, cuz\nthat's what that,\nyou know, system has done, but you can\nfeed any image into an LM right now and\nit'll tell you what it is. Almost like\nexactly like what a person would say.\nObviously it's not perfect and it\nconstantly fails, um\nbut that error is always decreasing as\nthey, you know,\nfix bugs, they they improve the training\npipeline, they improve the data they\ngive it, and then it's always\napproaching something that seems like it\nwill apply in any situation. Um of\ncourse it's very far from that right\nnow. I'm not meaning to claim that like\nyou could take any image and then it'll\nperfectly recognize what's in it. You\ncan fool them all the time. Um it's just\nthat like\nit's the entire internet, right? I just\nthat's the thing that I keep thinking\nabout. You know, it's that's just a lot.\nIt's hard for me to even comprehend what\nthe limits of that is.\nSo, even if we don't end up in a world\nwhere the AI is able to do our job in\nin, you know, 10 years or less,\nwhatever. Um how likely do you think it\nis that we're heading towards a dead\ninternet style problem where the noise\ngenerated or the noise created by these\ngenerated AIs is a problem to our own\nAIs is a problem to our own\nability to communicate? I mean really\nalready it feels like so much of the\ninternet is on like private discords\nthat aren't, you know, on these massive\npublic forums and,\nyou know, linked with each other. I\nthink that people are already kind of\nreceding away from this like how it used\nto be where it was like a like a massive\ngraph of forum posts or whatever, right?\nLike um\nI don't know much about that. I mean\nthis this is the internet and it's not\ndead, I guess. So, I it's not I don't\nknow like I mean maybe.\nPeople I actually Well, so LLMs do seem\nlike they kind of defeat the Turing test\nand I I seem to I think I'm remembering\nthere was some news article. Somebody\nmaybe in like Twitch chat or whatever\ncan find it where I think some lab\nbasically plugged an LLM into like a\nsubreddit and just had it be a commenter\nand it got a lot of karma and people\nstarted to like relate to it and like\noh, this is a good post or he's a good\nguy. I like this guy. And then it was\nlike uncovered that it was really like a\nlike a false flag human being or\na false flag human being or\nwhatever and like they were they they\nbasically were doing a study on the\nsubreddit involuntarily. I think people\ngot very mad at that cuz it was like\nthat's probably not ethical to do human\nbeings. Most of what's happening in the\nmodern world is not ethical to do human\nbeings, honestly. I I don't think I\ndon't think LLMs are that crazy compared\nto like YouTube. Like you just show\ninfant pixels to a person like this is\nprobably not good for them. I don't know\nUm I don't really you know, things are\nalready so so so far from what is like\nhealthy for a human being. It's like\nWhat are you going to What do you want?\nLike you just like it's not that much\nworse than TikTok. I don't know what to\nsay. TikTok is a transformer. It's\nSo, we talked about this over breakfast\nyesterday, but like LLMs kind of cool,\ntransformers are cool, doing something,\nbut like there's a missing piece or\nmaybe many. Do you have any theories\nwhat those might be?\nMe personally as a hobbyist\ncrazy person? Uh\nI don't know. I feel like if I were to\nEverybody so every every single deep\nlearning research is like oh, it's going\nto be in a reinforcement learning. Like\nyou take that success of AlphaZero where\nit basically can master any system. You\njust like apply that to like soft\nconcepts. Like MuZero it's like you\nknow, I everything everybody is like\nthinking along those lines of like we\nbasically need computers to\nautomatically generate and master uh\nthings that not even humans have\nmastered yet, right? Because if all\nyou're doing is basically making these\nmodels uh try to optimize and find\npatterns like that are more and more\nclosely match human beings, they'll\nnever exceed the human baseline, right?\nI mean, they might be an ensemble of all\nthe best humans and, you know, they\nmight have be a genius in every single\nfield or whatever, but they're never\ngoing to surpass human genius, which you\nwould kind of want if you wanted to be a\ngod-like being that replaces mankind,\nwhich is some people's goal, not mine. I\ndon't I'm not into that, but some people\nlike that. I don't know. I think it's\nprobably bad.\nBut like like\nrandom thought like like what what what\nwhat could that be like?\n>> Reinforce you need to have some way of\noperating on these concepts um in a way,\nsome kind of like gym, some kind of like\nimaginary like it needs a world model,\nright? That's what OpenAI say like, \"Oh,\nwe need a world model.\" It's like, \"We\nneed some way We need a transformer not\non text.\" That's why So,\nI don't know if you know this, but the\nreason why OpenAI did their video model\nis because of they're trying to solve\nthis problem. They want a world model so\nthat these these AI systems can try to\nimagine new scenarios over and over\nagain and find new ways to be better.\nUm and then, you know, if you can\nactually even if it's very inefficient\nat be being better, right? If you can\nsomehow correlate the strength of an AI\nsystem one-to-one with like compute, you\nwe can always get more compute, right?\nWe can always make Nvidia stock price go\nup and like get more chips from Taiwan.\nLike it can always we can always keep\nramping this, right? There's a lot of\nsilicon on Earth and stuff like that,\nright? It's like you know, you you can\njust keep doing this. You can scale it\nand then once you have the compute, it\nkeeps printing it keeps being smart. And\nso, it will it is only a matter of time\nuntil, you know, if you can actually do\nself-optimizing intelligent processes\nwith silicon that it exceeds human\nbaseline and becomes the god-like figure\nsome people want to, you know, build\nupon the Earth.\nSam, Timothy?\nOh, I I had a question about like is the\nfuture then like oh if if they Is there\na data bottom neck at some point? And is\nthe future of reinforcement learning?\nSo, what I'm thinking about is like a\nbaby\nyou right spends like a lot of time\nreceiving input. And I have a theory\nthat the reason puberty happens, and\nlike the body grows so quickly, is\nbecause it's to like protect the brain\nfrom injuring itself before it while\nit's too dumb.\nThat kind of makes sense, actually. I\ndon't I don't know. I So so is is there\na future where like you\nlet robots like with cameras out in a\nfield, and then the learning is just\nlike, what is going to That what's the\nnext frame of the video? Yes, actually\nexactly. That's like what Nvidia's\ndoing. That's like There's already many\ndata sets where they basically just\nstrap a GoPro to the center of a guy\ncooking in a kitchen, and there's like,\nyou know, gigabytes of those videos. And\nthe idea is just like, oh, we just\nbasically need a lot of video of people\ndoing whatever they're doing that's\nuseful to other people. Throw a bunch of\ntransformers at that, trillion-dollar AI\ncompany, and then we automate them, and\nthen we rake in the money, and then\nthere's no more people to buy cuz\nthere's no more economy. I don't Kind of\neconomy. I don't Kind of\njust stopped making sense. It's like\nreally crazy. Like if you really if you\nreally think about it for a second,\nthink about it for a second,\nright? All these like VC-backed AI\ncompanies, if they really did pan out,\nthere would be no more economy. Like\nwhat are you competing for by investing\nin that?\nI I don't know.\nAlthough, um,\nconcerning like those gigabytes videos\nand so on, I would imagine a big big\nchunk of like what would make a the\ntraining run effective is like how do\nyou transform the data before sending it\ninto the model or into the algorithm,\nright? Because like I guess if you like\nfilter down to what actually matters cuz\nlike there's like a lot of when you have\na guy cooking or something, like a lot\nof that image data is completely\nuseless. You don't care about it most of\nit, right? So like\nwhat kind of pre-filtering do you people\ndo, and like how much How important is\nthat? Well, actually people do the\nopposite of pre-filtering. Sometimes for\ndeep learning, actually I think Tesla's\ndoing this, they don't feed pixels, they\nfeed like raw sensor measurements. They\nwould decrease filtering.\nThere's a There's a very Well, there's a\nvery famous like blog post, basically,\ncalled the bitter lesson. And the bitter\nlesson is that like a general method\nthat scales with compute will always\nperform better than a specific method\nmade by engineers because compute is\nalways getting cheaper and more\nplentiful. Engineers are becoming more\nscarce as people become stupider from\nthe AI that you're creating and doing\nthis. Yeah, but but then by choosing to\nuse like sensors, like you're making\nthat's like a deliberate choice. Like\ninstead of using pixels, we're going to\nuse sensors. So, it's like maybe more\nrelevant information within it. So, like\nthe pattern might be easier to extract\nsomehow or No, it's not that it's\nthe idea is that we should not try to\ndesign these massive massive, you know,\ntrillion parameter deep learning models.\nWe should not try to do any data\ntransformation on what's happening in\nthe universe because it will find the\nrepresentation that's most effective for\nminimizing the loss, right? It will do\nthat with compute. It will do It will\nfind something better than what we could\never filter it, right? And what\nbasically you're describing is I think\ncalled feature engineering. We're like,\n\"Oh, instead of feeding the images to\nthe deep learning model, we feed we we\ntry to detect like where the cones are\nand stuff and like the road and then\nlike, you know, we feed those features\ninto the model.\" And it always\nunderperforms just feeding in the raw\nsensory data. If you have enough compute\nto scale up because the compute will\nfind the representation that is most\neffective to output the policy\nbasically, I think. So, but you can\nstill select the which what kind of\nsensor data you feed it in. Like the\nsensor is is different from sensor data\nis is different from sensor data\nfrom like vision, right? Yeah. You still\nshould make a choice and like that\nchoice will matter, right? Yeah, the\nchoice matters in the opposite direction\nmatters in the opposite direction\nfrom what you're saying. You want more\nmore generic, more bigger amount of\ndata. So, you feed it everything, every\ntype of sensor data you can get your\nhands on. You just feed it whatever you\ncan. Yes, ideally yes, but like what\nthey found actually I I think there was\nthis paper called like the Nemo paper. I\ndon't remember what it was, but they\nbasically had a transformer they just\nfound that as you taught it many random\ndisparate games, it would like improve\non the other ones, right? That were\nseemingly unrelated. And so, there's\nsomething just about transformers. The\nmore data that you give them, the better\nthe representations become no matter\nwhat. I think it's because there's\nthere's something kind of like a a\nuniversal prior or like\nyou know, it's just the universe is\nordered in a particular way that's\nevident in everything that happens in\nit. So, the more in- information that\ncomes in This is not This is me not not\nby the way. This is not a true I'm like\ntheory I'm like crafting theories. It's\nlike maybe there's like something about\nthe universe that just whatever you give\nit that's from the universe truly and\nnot you know like a craft of man or\nwhatever like it it's just somehow\nbuilds better representations cuz it's\nsomething like that. I don't know.\n>> there's universal laws and patterns and\nsuch that just exist in reality, right?\nAnd like that's maybe what you're what\nit's using somehow? That's maybe yeah.\nI'm not I'm really not sure. That's\nI would I would like to see more\nrigorous attempts to answer that\nquestion rather than me like theory\ncrafting.\nAre there any more questions?\nI mean just earlier like uh\nsaid I you beat me to it, but I was\ngoing to ask about like uh just the\nnoise like the back of noise of AI\ngenerative AI on the internet, right?\nBecause like if you lower the friction\nfor like creating content, but like you\nsaid it's like you're using existing\ndata transformations to create things,\nthen I mean wouldn't the problem be like\nyou just end up with so much stuff that\nis all from old transformations? It's\nreally hard to find novelty among that.\nSo like Um search engines get good\nenough that like we don't have a problem\nor like No, there there's just very easy\nways to detect if something was output\nby AI for the most part or like ways to\nfilter for data that is like the most\nconfusing for the model. And so these\nthings are not going to be a serious\nproblem for deep learning.\nat scale. Detecting AI or detecting\ndetecting information that AI has not\nseen before that's like possible so this\nthat's something that's like easy you\nsay? Yeah, you can just tell which you\nknow training seems to be the most\nconfusing to it. Maybe this problem is\nmore complicated than\nwhat I'm thinking because like maybe\nI don't I don't know. I'm sure there's\nthere's much smarter deep learning\nresearchers who have answered that\nquestion already probably, but my my\nguess would be that like\nit's probably not a real issue. Okay,\nbecause if you could just filter out\neverything that's AI generated it'd be\nreally great, you know. Yes, they can I\nthink you can do that, yeah. Oh yeah.\nI hope so.\nAll right, we're out of time. We're out\nof time. Round of applause for Krikey.\n>> Yeah.", + "fetched_at": "2026-06-22T05:27:46Z", + "source": "yt-dlp-vtt", + "clean_segments": [ + { + "start": 0.52, + "text": "Okay, so this is very difficult because" + }, + { + "start": 2.64, + "text": "Cameron is such a is such an like" + }, + { + "start": 4.16, + "text": "indescribable guy." + }, + { + "start": 6.52, + "text": "Um" + }, + { + "start": 7.92, + "text": "you know," + }, + { + "start": 9.68, + "text": "yeah, this is very very very difficult." + }, + { + "start": 11.84, + "text": "I know Cameron just as like the guy who" + }, + { + "start": 15.16, + "text": "uh you know," + }, + { + "start": 17.84, + "text": "he" + }, + { + "start": 18.88, + "text": "he's like the guy crazy enough to just" + }, + { + "start": 20.76, + "text": "like dig through the mountain and do" + }, + { + "start": 23.16, + "text": "just like as sudden" + }, + { + "start": 26.84, + "text": "sudden huge amounts of work" + }, + { + "start": 30.56, + "text": "uh of like, \"Oh, I cloned an application" + }, + { + "start": 32.84, + "text": "or I did this.\" And then at other times" + }, + { + "start": 34.56, + "text": "he, you know, goes and becomes a web" + }, + { + "start": 36.4, + "text": "developer and then complains about like" + }, + { + "start": 38.72, + "text": "not being of the program anymore." + }, + { + "start": 42.12, + "text": "So, anyway, um" + }, + { + "start": 44.36, + "text": "can we get a round of applause for the" + }, + { + "start": 45.6, + "text": "completely undescribable" + }, + { + "start": 47.76, + "text": "Cameron Wrights?" + }, + { + "start": 55.08, + "text": "What's up, Craigsters? LET'S MAKE SOME" + }, + { + "start": 56.76, + "text": "NOISE." + }, + { + "start": 62.04, + "text": "LET'S SAY YOU WANTED TO WRITE A program" + }, + { + "start": 63.96, + "text": "that could tell the difference between a" + }, + { + "start": 65.36, + "text": "cat and a dog." + }, + { + "start": 67.4, + "text": "I want you to actually think about how" + }, + { + "start": 68.8, + "text": "you would do this for a second, okay?" + }, + { + "start": 70.24, + "text": "Imagine you were given the pixels and" + }, + { + "start": 72.0, + "text": "you had to make the program that said," + }, + { + "start": 73.88, + "text": "you know, cat or dog." + }, + { + "start": 75.56, + "text": "You might try something like, you know," + }, + { + "start": 77.0, + "text": "look for ear shapes with edge detection," + }, + { + "start": 78.84, + "text": "maybe like detect like the little angle" + }, + { + "start": 80.92, + "text": "there, you know, if it's a pointy ear," + }, + { + "start": 81.96, + "text": "it's a cat, if it's a floppy ear, it's a" + }, + { + "start": 83.36, + "text": "dog." + }, + { + "start": 84.88, + "text": "Problem solved, right? You don't need" + }, + { + "start": 86.16, + "text": "any more tech, you don't need any fancy" + }, + { + "start": 87.72, + "text": "whatever." + }, + { + "start": 88.8, + "text": "Uh-oh." + }, + { + "start": 92.04, + "text": "You know, if we look at our program" + }, + { + "start": 93.08, + "text": "here, it's just a system with an input" + }, + { + "start": 95.52, + "text": "and an output, right? Um" + }, + { + "start": 98.04, + "text": "take in the image" + }, + { + "start": 100.56, + "text": "and then, you know, look for whatever" + }, + { + "start": 102.04, + "text": "heuristic we've decided we're looking" + }, + { + "start": 103.32, + "text": "for, and then we we, you know, make our" + }, + { + "start": 105.68, + "text": "output, right? But, um if we keep trying" + }, + { + "start": 107.56, + "text": "to iterate on this, it's like, okay," + }, + { + "start": 108.56, + "text": "maybe some of them have pointy ears, but" + }, + { + "start": 109.84, + "text": "only if the fur is orange, and we try to" + }, + { + "start": 111.64, + "text": "keep, you know, coming up with patterns" + }, + { + "start": 112.92, + "text": "that try to get at what we mean by a dog" + }, + { + "start": 115.32, + "text": "or cat," + }, + { + "start": 116.6, + "text": "um you'll realize that it quickly" + }, + { + "start": 118.28, + "text": "becomes untenable, right?" + }, + { + "start": 120.36, + "text": "So, the whole idea of machine learning" + }, + { + "start": 123.0, + "text": "is you essentially have a program that" + }, + { + "start": 124.92, + "text": "creates that program. It's basically" + }, + { + "start": 126.72, + "text": "meta programming, right? That's sort of" + }, + { + "start": 128.56, + "text": "a metaphor for what it is." + }, + { + "start": 130.24, + "text": "Um, you set up some architecture and" + }, + { + "start": 132.0, + "text": "repeatedly optimize with respect to, you" + }, + { + "start": 134.16, + "text": "know, your examples, um, a machine" + }, + { + "start": 136.68, + "text": "learning model, a black box of many" + }, + { + "start": 138.52, + "text": "parameters. It's a learned program," + }, + { + "start": 140.64, + "text": "right? It's automatic programming." + }, + { + "start": 144.2, + "text": "As a quick aside, why is it all Python?" + }, + { + "start": 145.88, + "text": "It's because it's basically meta" + }, + { + "start": 147.36, + "text": "programming, right? It's a complicated" + }, + { + "start": 148.72, + "text": "build system. Uh, these training scripts" + }, + { + "start": 150.88, + "text": "basically are just there to eventually," + }, + { + "start": 152.76, + "text": "you know, automatically create the" + }, + { + "start": 154.32, + "text": "program that does what you want it to" + }, + { + "start": 155.72, + "text": "do. Um, and most of the heavy lifting is" + }, + { + "start": 158.84, + "text": "not necessarily done by the actual code" + }, + { + "start": 160.72, + "text": "you write anyways, it's done" + }, + { + "start": 162.6, + "text": "in the in the dot products on the GPU" + }, + { + "start": 164.6, + "text": "or, you know," + }, + { + "start": 166.4, + "text": "piece of compute that you're using to" + }, + { + "start": 168.44, + "text": "actually do the machine learning. Um," + }, + { + "start": 170.88, + "text": "and John Carmack for his current AI" + }, + { + "start": 172.64, + "text": "efforts, he's in the race to win the" + }, + { + "start": 173.84, + "text": "world." + }, + { + "start": 174.959, + "text": "Um," + }, + { + "start": 176.2, + "text": "he has switched to Python. I thought" + }, + { + "start": 178.08, + "text": "that was notable." + }, + { + "start": 179.48, + "text": "So, a little bit about me, right? I'm" + }, + { + "start": 181.68, + "text": "mostly a game developer. If we go back" + }, + { + "start": 183.4, + "text": "to 2017, one of my first projects was a" + }, + { + "start": 185.48, + "text": "game a game engine." + }, + { + "start": 187.28, + "text": "Um, I think this is probably pretty" + }, + { + "start": 188.48, + "text": "common for people who I don't know like" + }, + { + "start": 190.44, + "text": "to program is the first thing they try" + }, + { + "start": 191.68, + "text": "to make is a game engine. It's kind of" + }, + { + "start": 192.92, + "text": "funny." + }, + { + "start": 193.8, + "text": "Um, you don't know anything about what" + }, + { + "start": 195.239, + "text": "you're doing. You're like, I'm going to," + }, + { + "start": 196.239, + "text": "you know, do the thing that's the" + }, + { + "start": 197.28, + "text": "ultimate thing." + }, + { + "start": 198.56, + "text": "Um, and then much later, you know, 3" + }, + { + "start": 200.0, + "text": "years later, you know, it's like I'm" + }, + { + "start": 201.0, + "text": "making a lot of individual games, I'm" + }, + { + "start": 202.36, + "text": "working on the Godot engine." + }, + { + "start": 204.08, + "text": "Um, you know, these are a couple of my" + }, + { + "start": 205.84, + "text": "game projects." + }, + { + "start": 207.32, + "text": "Um, that there on the upper left is a" + }, + { + "start": 208.959, + "text": "space game called Asteris. It's" + }, + { + "start": 210.36, + "text": "releasing in 2035. Um, it's got" + }, + { + "start": 212.84, + "text": "something like a Overwatch net code in" + }, + { + "start": 214.8, + "text": "it, right? It's like a" + }, + { + "start": 216.36, + "text": "big solar system with many people" + }, + { + "start": 218.239, + "text": "playing at the same time. Um," + }, + { + "start": 221.08, + "text": "yeah, I I was pretty proud of how stable" + }, + { + "start": 223.36, + "text": "the net code was considering how chaotic" + }, + { + "start": 225.08, + "text": "that game was, right? Um," + }, + { + "start": 227.519, + "text": "and then Dante's Cowboy. Um, this was no" + }, + { + "start": 229.72, + "text": "engine in C, 3D rendering, animated" + }, + { + "start": 231.32, + "text": "armatures, all from scratch, and all the" + }, + { + "start": 233.68, + "text": "NPCs in this game were uh" + }, + { + "start": 235.72, + "text": "you know, they were controlled by an" + }, + { + "start": 236.68, + "text": "LLM, and this was never released because" + }, + { + "start": 238.64, + "text": "it's fundamentally flawed. Um" + }, + { + "start": 241.36, + "text": "I'm a deep learning hobbyist." + }, + { + "start": 243.52, + "text": "The majority of my work is not in deep" + }, + { + "start": 245.4, + "text": "learning, so there's going to be some" + }, + { + "start": 246.68, + "text": "inaccuracies in what I say. It's just" + }, + { + "start": 248.959, + "text": "always been interesting to me ever since" + }, + { + "start": 250.6, + "text": "back when I started, you know, first" + }, + { + "start": 251.68, + "text": "doing games, I've also been kind of" + }, + { + "start": 253.04, + "text": "looking at deep learning, which is a" + }, + { + "start": 254.8, + "text": "subset of machine learning, and it's" + }, + { + "start": 256.48, + "text": "just always kind of interested me. Like" + }, + { + "start": 257.72, + "text": "I've been following it, right?" + }, + { + "start": 260.359, + "text": "Ironically enough, actually, my first" + }, + { + "start": 261.76, + "text": "ever professional programming experience" + }, + { + "start": 264.08, + "text": "was to architect a deep learning model" + }, + { + "start": 266.04, + "text": "from scratch, um make it work, and" + }, + { + "start": 268.24, + "text": "deploy it in production. Um this was for" + }, + { + "start": 270.68, + "text": "my college roommate's company called" + }, + { + "start": 272.2, + "text": "MacroHard. Um he basically made software" + }, + { + "start": 275.84, + "text": "that like was a custom broadcast overlay" + }, + { + "start": 278.12, + "text": "for for League of Legends teams," + }, + { + "start": 280.32, + "text": "and I made a deep learning model that" + }, + { + "start": 282.36, + "text": "predicted which team was uh going to win" + }, + { + "start": 284.16, + "text": "the game, right? So, that little" + }, + { + "start": 285.88, + "text": "visualization there is what I created." + }, + { + "start": 288.16, + "text": "Uh this was trained and, you know, on a" + }, + { + "start": 290.04, + "text": "and optimized on many, many professional" + }, + { + "start": 291.68, + "text": "games." + }, + { + "start": 294.44, + "text": "and you know, it's it's just a pretty" + }, + { + "start": 295.4, + "text": "interesting problem, right? And at that" + }, + { + "start": 297.28, + "text": "time, right, I implemented uh League of" + }, + { + "start": 299.48, + "text": "Legends game predicting paper," + }, + { + "start": 301.52, + "text": "and I couldn't reproduce the results," + }, + { + "start": 302.96, + "text": "you know, I was like, \"Well," + }, + { + "start": 304.919, + "text": "I copied the thing that the paper said" + }, + { + "start": 306.28, + "text": "to do, and then it's, you know, it" + }, + { + "start": 307.56, + "text": "doesn't work.\"" + }, + { + "start": 308.76, + "text": "And basically," + }, + { + "start": 311.2, + "text": "there's a there was a bug in the paper," + }, + { + "start": 312.56, + "text": "right? This is a very important lesson" + }, + { + "start": 313.72, + "text": "of not just how hard machine learning" + }, + { + "start": 315.28, + "text": "is, but how often people mess up, right?" + }, + { + "start": 317.32, + "text": "And how easy it is to mess up and be so" + }, + { + "start": 319.2, + "text": "confident in what you did that you" + }, + { + "start": 320.84, + "text": "publish it to the world and waste my," + }, + { + "start": 322.88, + "text": "you know, 2 weeks, like" + }, + { + "start": 325.24, + "text": "it's just very, very easy to mess up." + }, + { + "start": 326.919, + "text": "And just to give a quick overview of the" + }, + { + "start": 328.64, + "text": "bug they had, right? When you want to" + }, + { + "start": 330.64, + "text": "make sure a machine learning model" + }, + { + "start": 332.24, + "text": "works, you usually have a you give it a" + }, + { + "start": 333.96, + "text": "test, just like a person, right? And if" + }, + { + "start": 336.6, + "text": "that machine learning model were to see" + }, + { + "start": 337.84, + "text": "the answers of the test before it took" + }, + { + "start": 339.68, + "text": "it, uh you know, its success in that" + }, + { + "start": 341.84, + "text": "test would obviously not be" + }, + { + "start": 342.88, + "text": "representative of it of if it could do" + }, + { + "start": 344.52, + "text": "the, you know, a problem, right?" + }, + { + "start": 346.12, + "text": "Arbitrary new examples, it's not" + }, + { + "start": 347.8, + "text": "evidence that it generalizes. And in" + }, + { + "start": 349.56, + "text": "this specific case, they had a metric," + }, + { + "start": 351.44, + "text": "the champion win rate, that was trained" + }, + { + "start": 353.52, + "text": "on their entire um" + }, + { + "start": 355.64, + "text": "that champion win rate was based on" + }, + { + "start": 356.72, + "text": "their entire data set, right? And so" + }, + { + "start": 358.6, + "text": "then when it went to test the model, it" + }, + { + "start": 360.84, + "text": "had the win rate for the champions that" + }, + { + "start": 362.2, + "text": "were in the test set. And so it" + }, + { + "start": 364.12, + "text": "purported to have amazing accuracy at" + }, + { + "start": 365.76, + "text": "predicting who would win on the test" + }, + { + "start": 367.56, + "text": "set, but there was data information" + }, + { + "start": 369.08, + "text": "pollution, right? These are the kinds of" + }, + { + "start": 371.16, + "text": "bugs you find in deep learning. They're" + }, + { + "start": 372.6, + "text": "not something you could really" + }, + { + "start": 373.52, + "text": "breakpoint in the debugger, but things" + }, + { + "start": 375.24, + "text": "you have to find like a scientist," + }, + { + "start": 376.56, + "text": "right? Um you almost have to like uh" + }, + { + "start": 378.64, + "text": "binary search over the uh problem space" + }, + { + "start": 380.72, + "text": "sometimes to see what's going wrong. It" + }, + { + "start": 382.4, + "text": "could be anything, right?" + }, + { + "start": 384.32, + "text": "And I want to stress again that machine" + }, + { + "start": 385.32, + "text": "learning is automatic programming," + }, + { + "start": 387.24, + "text": "right? In this paper, this is a machine" + }, + { + "start": 388.68, + "text": "learning technique, not a deep learning" + }, + { + "start": 390.0, + "text": "technique, called random forest, where" + }, + { + "start": 391.76, + "text": "it's basically a series of learned if" + }, + { + "start": 393.44, + "text": "statements directly, right?" + }, + { + "start": 395.52, + "text": "And you know, there might be some" + }, + { + "start": 396.24, + "text": "majority voting scheme, and they have" + }, + { + "start": 397.52, + "text": "some algorithm to optimize like based on" + }, + { + "start": 399.32, + "text": "the training data, oh, this if statement" + }, + { + "start": 401.08, + "text": "should be this, and this if statement" + }, + { + "start": 402.16, + "text": "should be this. They basically are" + }, + { + "start": 403.79, + "text": ". They basically are" + }, + { + "start": 403.8, + "text": "attempting to tweak the if statements" + }, + { + "start": 405.32, + "text": "such that it, you know, is as accurate" + }, + { + "start": 407.0, + "text": "as possible on the training data." + }, + { + "start": 410.24, + "text": "and deep learning is separate from that." + }, + { + "start": 411.6, + "text": "It's a subset of machine learning." + }, + { + "start": 413.68, + "text": "And the story kind of goes like this," + }, + { + "start": 414.919, + "text": "right? Some guy in the '60s was looking" + }, + { + "start": 416.96, + "text": "at this, and he was like, I think I know" + }, + { + "start": 419.0, + "text": "what these are. He was very foolish." + }, + { + "start": 422.76, + "text": "you know, he he he basically, on a very" + }, + { + "start": 424.76, + "text": "abstract level, said, \"Okay, you know" + }, + { + "start": 426.08, + "text": "what neurons are? You know what's going" + }, + { + "start": 427.24, + "text": "on in the brain? You know what's" + }, + { + "start": 428.0, + "text": "enabling us to do all these amazing" + }, + { + "start": 429.2, + "text": "things? Is little, you know, boxes. They" + }, + { + "start": 432.2, + "text": "take in a bunch of numbers, and then" + }, + { + "start": 434.16, + "text": "they mutate those numbers somehow. Maybe" + }, + { + "start": 436.04, + "text": "when they have enough of that number," + }, + { + "start": 437.2, + "text": "they they activate and propagate that" + }, + { + "start": 438.52, + "text": "information to the other neurons.\"" + }, + { + "start": 440.48, + "text": "Um if you were to zoom in really, really" + }, + { + "start": 442.12, + "text": "quickly on what one of those, you know," + }, + { + "start": 444.16, + "text": "very abstract human neurons that they" + }, + { + "start": 445.6, + "text": "designed is actually doing, right? Um a" + }, + { + "start": 449.12, + "text": "single neuron in a neural network," + }, + { + "start": 451.04, + "text": "right? The first step is it adds all the" + }, + { + "start": 453.36, + "text": "weights multiplied by the incoming" + }, + { + "start": 454.64, + "text": "activations, right? So you see that" + }, + { + "start": 456.4, + "text": "bigger arrow is a much stronger weight." + }, + { + "start": 458.04, + "text": "That tiny arrow, you know, you can even" + }, + { + "start": 459.52, + "text": "have a negative weight, right? There's" + }, + { + "start": 460.919, + "text": "basically a set of weights of coming" + }, + { + "start": 463.08, + "text": "from the uh neurons on the previous" + }, + { + "start": 464.92, + "text": "layer, right? And then it just adds them" + }, + { + "start": 466.64, + "text": "like numbers. They're all floats, right?" + }, + { + "start": 468.88, + "text": "Next, there's a bias that adds a bias to" + }, + { + "start": 470.6, + "text": "that weighted sum." + }, + { + "start": 472.12, + "text": "Um so, this could be considered like the" + }, + { + "start": 473.36, + "text": "activation threshold for that neuron," + }, + { + "start": 474.84, + "text": "right? So, like if the bias is -6, then" + }, + { + "start": 477.4, + "text": "you would require, you know, six to be" + }, + { + "start": 479.4, + "text": "above zero and to continue to propagate" + }, + { + "start": 481.2, + "text": "that information. Again, this is a very" + }, + { + "start": 482.96, + "text": "rough, bad model of biological neurons," + }, + { + "start": 485.52, + "text": "and they stumbled upon something that's" + }, + { + "start": 486.72, + "text": "actually pretty crazy um just by doing" + }, + { + "start": 488.72, + "text": "that, right? But, the final step is to" + }, + { + "start": 490.4, + "text": "pass it through any nonlinear like sine" + }, + { + "start": 492.64, + "text": "or or x squared. I'll only get to why" + }, + { + "start": 494.04, + "text": "they do that in a second, right? That's" + }, + { + "start": 495.88, + "text": "what happens in a neuron. Here's some" + }, + { + "start": 497.2, + "text": "like pseudocode for how you would find" + }, + { + "start": 498.76, + "text": "the float value for a neuron," + }, + { + "start": 500.84, + "text": "right? Um" + }, + { + "start": 502.44, + "text": "and you can take a look at this and" + }, + { + "start": 503.36, + "text": "ponder it for a second, but" + }, + { + "start": 505.4, + "text": "you'll you'll notice that well, if we're" + }, + { + "start": 507.32, + "text": "constantly just multiplying and adding" + }, + { + "start": 508.6, + "text": "numbers, that's kind of what linear" + }, + { + "start": 509.72, + "text": "algebra is," + }, + { + "start": 511.0, + "text": "and that is how they accelerate this and" + }, + { + "start": 512.599, + "text": "talk about this in the actual, you know," + }, + { + "start": 514.2, + "text": "field, right? They don't express each" + }, + { + "start": 516.2, + "text": "individual weight in some kind of array." + }, + { + "start": 517.719, + "text": "They express it as a matrix, right?" + }, + { + "start": 521.24, + "text": "and then there's a lot of math that" + }, + { + "start": 522.4, + "text": "applies when you do that, right? Um you" + }, + { + "start": 524.68, + "text": "optimize those weights and the bias with" + }, + { + "start": 527.0, + "text": "a process called gradient descent," + }, + { + "start": 528.8, + "text": "right? Stochastic gradient descent. Um" + }, + { + "start": 531.88, + "text": "I'll also get to that in a second," + }, + { + "start": 533.08, + "text": "right? But, fundamentally, this is just" + }, + { + "start": 534.72, + "text": "a diagram for this expression or that" + }, + { + "start": 536.32, + "text": "code, right? When you see this diagram," + }, + { + "start": 538.48, + "text": "it is simply that code, right? Um" + }, + { + "start": 542.16, + "text": "that's what it's intending to describe," + }, + { + "start": 543.44, + "text": "right? And so, just to kind of get at" + }, + { + "start": 545.16, + "text": "like what the act what's going on with" + }, + { + "start": 546.32, + "text": "the activation function, you know, why" + }, + { + "start": 547.56, + "text": "do they pass it through sine, right?" + }, + { + "start": 550.12, + "text": "Each layer of a neural network is" + }, + { + "start": 551.36, + "text": "basically a node in a transform" + }, + { + "start": 552.96, + "text": "hierarchy of like n dimensions, right?" + }, + { + "start": 555.72, + "text": "Um so, if you're a game developer, which" + }, + { + "start": 558.0, + "text": "you're the target audience of this talk," + }, + { + "start": 559.16, + "text": "so you should be a game developer." + }, + { + "start": 561.6, + "text": "You're very familiar with like trees of" + }, + { + "start": 562.88, + "text": "transforms, right?" + }, + { + "start": 564.8, + "text": "Um and you know that if you have a" + }, + { + "start": 566.64, + "text": "series of transforms one after the" + }, + { + "start": 568.08, + "text": "other, you can multiply them all and it" + }, + { + "start": 569.96, + "text": "becomes one linear transform, right?" + }, + { + "start": 571.4, + "text": "That's like what it means for something" + }, + { + "start": 572.44, + "text": "to be be linear. And when we want" + }, + { + "start": 574.48, + "text": "complicated behavior from these deep" + }, + { + "start": 575.92, + "text": "learning models, uh we don't want it to" + }, + { + "start": 577.72, + "text": "be reducible like that to one linear uh" + }, + { + "start": 580.08, + "text": "translation and and rotation whatever," + }, + { + "start": 581.76, + "text": "right? So, you basically think about it" + }, + { + "start": 583.76, + "text": "as like warping the space" + }, + { + "start": 585.8, + "text": "such that the translation and offset you" + }, + { + "start": 588.04, + "text": "do kind of like it it enables you to" + }, + { + "start": 590.48, + "text": "basically draw a much more richer" + }, + { + "start": 592.76, + "text": "contour around the data, right? Um" + }, + { + "start": 596.36, + "text": "And so, just as a as as another metaphor" + }, + { + "start": 598.24, + "text": "to help you as a game developer" + }, + { + "start": 599.6, + "text": "understand what's going on here, right?" + }, + { + "start": 601.48, + "text": "Both that neural network and that shader" + }, + { + "start": 603.28, + "text": "graph, these are both compute graphs," + }, + { + "start": 604.88, + "text": "right? There's no branching. Information" + }, + { + "start": 607.04, + "text": "is flowing in from the left and it's" + }, + { + "start": 608.56, + "text": "going out the right. Um" + }, + { + "start": 611.28, + "text": "A shader operates on vertices while a" + }, + { + "start": 612.76, + "text": "neural network operates on an image or" + }, + { + "start": 614.36, + "text": "some state or anything you want, right?" + }, + { + "start": 616.64, + "text": "And in the case of game development, an" + }, + { + "start": 618.24, + "text": "artist might be manually tuning the" + }, + { + "start": 619.6, + "text": "inputs to that shader and making it look" + }, + { + "start": 621.44, + "text": "good, right? They might be nudging, you" + }, + { + "start": 623.12, + "text": "know, the color and trying to make the" + }, + { + "start": 624.12, + "text": "output look good or some other" + }, + { + "start": 625.24, + "text": "parameter. In the case of deep learning," + }, + { + "start": 627.48, + "text": "that's actually replaced by a machine," + }, + { + "start": 629.28, + "text": "right? Which is which direction" + }, + { + "start": 630.96, + "text": "minimizes error on training for each one" + }, + { + "start": 632.64, + "text": "of these little uh parameters, right?" + }, + { + "start": 634.0, + "text": "Each one of those weights and all those" + }, + { + "start": 635.04, + "text": "biases, it knows which direction it" + }, + { + "start": 637.24, + "text": "should go in assuming all the other" + }, + { + "start": 638.72, + "text": "parameters were, you know, not moving." + }, + { + "start": 640.72, + "text": "It knows approximately which direction" + }, + { + "start": 641.76, + "text": "to go in to get better at the task. And" + }, + { + "start": 643.92, + "text": "then it just kind of repeatedly does" + }, + { + "start": 644.96, + "text": "that thousands and millions of times." + }, + { + "start": 647.2, + "text": "Um so, there is a code base for this" + }, + { + "start": 649.16, + "text": "talk. It's here. github.com/krickey/ai" + }, + { + "start": 652.839, + "text": "Um and to just get at what this code" + }, + { + "start": 654.2, + "text": "base is supposed to be for a second," + }, + { + "start": 655.76, + "text": "right? Deep learning libraries or, you" + }, + { + "start": 657.68, + "text": "know, autograd, they let you perform" + }, + { + "start": 659.64, + "text": "some arbitrary, you know, computations," + }, + { + "start": 661.6, + "text": "right? Might be a matrix multiplication," + }, + { + "start": 663.4, + "text": "you know, it might be x squared or one" + }, + { + "start": 664.8, + "text": "of those activation functions we talked" + }, + { + "start": 666.16, + "text": "about. Um and then find out how you need" + }, + { + "start": 669.2, + "text": "to turn or or or adjust those" + }, + { + "start": 671.4, + "text": "parameters, right? The weights and the" + }, + { + "start": 672.6, + "text": "biases. It automatically figures out" + }, + { + "start": 675.88, + "text": "what the expression is to make those" + }, + { + "start": 678.64, + "text": "move in a direction that makes the model" + }, + { + "start": 680.12, + "text": "get better at the task, right?" + }, + { + "start": 682.36, + "text": "And that's basically that" + }, + { + "start": 684.4, + "text": "the what these deep learning libraries" + }, + { + "start": 685.76, + "text": "are supposed to be in a nutshell, you" + }, + { + "start": 686.72, + "text": "know, TensorFlow, PyTorch. Um" + }, + { + "start": 689.52, + "text": "And you know," + }, + { + "start": 691.16, + "text": "just as a little preliminary for this" + }, + { + "start": 692.44, + "text": "code base, there's a lot of like tensors" + }, + { + "start": 693.88, + "text": "in it. Um" + }, + { + "start": 695.12, + "text": "but really tensors are just arrays of" + }, + { + "start": 696.48, + "text": "matrices. Um" + }, + { + "start": 698.08, + "text": "like you might want to talk about like" + }, + { + "start": 699.48, + "text": "oh uh" + }, + { + "start": 701.4, + "text": "We're we're feeding in many examples" + }, + { + "start": 703.6, + "text": "into this deep learning model. And so we" + }, + { + "start": 705.88, + "text": "want like a third dimension to it where" + }, + { + "start": 707.28, + "text": "it's just an array of the matrices that" + }, + { + "start": 708.92, + "text": "we're feeding into the model, right? Um" + }, + { + "start": 711.96, + "text": "And so my idea with this code base is it" + }, + { + "start": 714.68, + "text": "would be really cool if instead of me" + }, + { + "start": 716.12, + "text": "having a you know, have a Python" + }, + { + "start": 717.52, + "text": "interpreter, there was one C file called" + }, + { + "start": 719.52, + "text": "ai.c, right? We had CPU and GPU deep" + }, + { + "start": 722.72, + "text": "learning computer vision algorithms and" + }, + { + "start": 724.16, + "text": "implementations, right? I would want it" + }, + { + "start": 726.0, + "text": "to be able to execute LLM local LLMs," + }, + { + "start": 729.32, + "text": "right? I want you to be able to" + }, + { + "start": 730.48, + "text": "fine-tune those LLMs, not just the" + }, + { + "start": 732.08, + "text": "inference. Um I would want to hardware" + }, + { + "start": 734.24, + "text": "accelerate with something portable like" + }, + { + "start": 735.44, + "text": "Sokol, right? I have some computer" + }, + { + "start": 737.28, + "text": "vision algorithms as well cuz those are" + }, + { + "start": 738.44, + "text": "fun." + }, + { + "start": 739.44, + "text": "Um and unfortunately, this this this" + }, + { + "start": 741.72, + "text": "talk code base right now is like super" + }, + { + "start": 743.64, + "text": "super not that, right? It has a" + }, + { + "start": 744.96, + "text": "hand-coded gradient calculation for a" + }, + { + "start": 746.56, + "text": "two-layer perceptron that classifies" + }, + { + "start": 748.2, + "text": "digits. It's super primitive and it's" + }, + { + "start": 750.0, + "text": "like 1970s stuff. There's no working" + }, + { + "start": 752.28, + "text": "auto grad yet, but I almost have it" + }, + { + "start": 753.96, + "text": "working." + }, + { + "start": 754.92, + "text": "Um and I I intend to continue working on" + }, + { + "start": 757.04, + "text": "ai.c cuz it's kind of fun. I'll be" + }, + { + "start": 758.88, + "text": "sending updates in the" + }, + { + "start": 760.44, + "text": "in the BSC Discord about that." + }, + { + "start": 765.839, + "text": "but just as an example of what I was" + }, + { + "start": 767.52, + "text": "trying to do with this little tiny, you" + }, + { + "start": 770.2, + "text": "know, deep learning library is" + }, + { + "start": 772.96, + "text": "each of these videos are videos I got" + }, + { + "start": 774.64, + "text": "from YouTube of some kind of ideally" + }, + { + "start": 776.52, + "text": "animating a scene, right? Where the" + }, + { + "start": 778.24, + "text": "camera is stationary. Um the theory is" + }, + { + "start": 780.839, + "text": "you should be able to train a deep" + }, + { + "start": 781.76, + "text": "learning model to generate these videos" + }, + { + "start": 783.52, + "text": "from the image. Um" + }, + { + "start": 785.68, + "text": "you know, the assumption is that the" + }, + { + "start": 786.72, + "text": "behavior of the pixels is relatively" + }, + { + "start": 788.16, + "text": "simple. Um and so I actually, even" + }, + { + "start": 790.64, + "text": "though I wasn't able to get this working" + }, + { + "start": 792.76, + "text": "in time for the talk with ai.c, I" + }, + { + "start": 797.44, + "text": "Hold on. Sorry, one sec." + }, + { + "start": 800.76, + "text": "Oh, here we go. Okay. So, I basically" + }, + { + "start": 803.839, + "text": "just, you know, let's just try it in" + }, + { + "start": 804.92, + "text": "Python really quick to see if this will" + }, + { + "start": 806.12, + "text": "even work at all. This is an 80 kilobyte" + }, + { + "start": 808.68, + "text": "of parameters neural network and it's" + }, + { + "start": 810.64, + "text": "garbage, right? This kind of happens a" + }, + { + "start": 812.6, + "text": "lot in deep learning when you just have" + }, + { + "start": 813.56, + "text": "a random idea and you just try it. It" + }, + { + "start": 815.44, + "text": "makes something really sometimes hellish" + }, + { + "start": 817.48, + "text": "like I didn't even know what that is," + }, + { + "start": 818.92, + "text": "honestly." + }, + { + "start": 820.12, + "text": "I think that 80 kilobyte is way too" + }, + { + "start": 821.44, + "text": "small to encode the dynamics of a waving" + }, + { + "start": 823.2, + "text": "tree, probably. So I would just, you" + }, + { + "start": 824.48, + "text": "know, 10x it like" + }, + { + "start": 826.92, + "text": "Deep learning is just very experimental" + }, + { + "start": 828.4, + "text": "like this. It's not exactly known, you" + }, + { + "start": 830.04, + "text": "know, what would cause the model to" + }, + { + "start": 831.44, + "text": "correctly learn how to ideally animate" + }, + { + "start": 832.8, + "text": "that scene or make the clouds move," + }, + { + "start": 834.0, + "text": "which is what I'm hoping for, right? You" + }, + { + "start": 835.68, + "text": "can imagine with a model like this you" + }, + { + "start": 836.88, + "text": "could take any desktop wallpaper and" + }, + { + "start": 838.68, + "text": "just make it ideally animate in the" + }, + { + "start": 839.8, + "text": "background, right? Especially you got it" + }, + { + "start": 841.04, + "text": "like" + }, + { + "start": 841.96, + "text": "you know, good enough on the battery. It" + }, + { + "start": 843.08, + "text": "just might be kind of fun to, you know," + }, + { + "start": 844.64, + "text": "minimize it and then see stuff always" + }, + { + "start": 845.84, + "text": "moving." + }, + { + "start": 848.08, + "text": "So just as a little bit of a historical" + }, + { + "start": 849.72, + "text": "detour, right? That what I just" + }, + { + "start": 851.64, + "text": "explained to you, that deep learning" + }, + { + "start": 852.76, + "text": "model, it the perceptron, it was this is" + }, + { + "start": 855.16, + "text": "an article about it in 1958, right?" + }, + { + "start": 858.24, + "text": "They were saying it can tell the" + }, + { + "start": 859.12, + "text": "difference between a cat and a dog. It" + }, + { + "start": 860.84, + "text": "is of no practical use. One day it might" + }, + { + "start": 863.24, + "text": "be useful to send one into outer space" + }, + { + "start": 864.8, + "text": "to take impressions for us." + }, + { + "start": 866.84, + "text": "So AI has always been a grift," + }, + { + "start": 868.68, + "text": "basically, right? And it's followed" + }, + { + "start": 870.4, + "text": "basically this this phase where" + }, + { + "start": 872.71, + "text": "this this phase where" + }, + { + "start": 872.72, + "text": "all the investors pour their money in" + }, + { + "start": 873.92, + "text": "and they're like, \"This is going to be a" + }, + { + "start": 874.92, + "text": "new form of life that will like do our" + }, + { + "start": 876.44, + "text": "bidding, right?\" And then it doesn't" + }, + { + "start": 878.16, + "text": "work and then it happens again and then" + }, + { + "start": 879.64, + "text": "it doesn't work. And now it's happening" + }, + { + "start": 881.44, + "text": "again and it's, you know," + }, + { + "start": 883.04, + "text": "who knows what's going to happen." + }, + { + "start": 889.36, + "text": "He just says, \"What did we learn," + }, + { + "start": 890.32, + "text": "Palmer?\" Um" + }, + { + "start": 892.6, + "text": "And so me personally, you know, I my" + }, + { + "start": 894.72, + "text": "kind of like AI boom bust was I got" + }, + { + "start": 896.92, + "text": "really into deep learning. Um you know," + }, + { + "start": 899.4, + "text": "and then I right when GPT-2 started to" + }, + { + "start": 901.2, + "text": "happen, I was like, \"This thing's kind" + }, + { + "start": 902.28, + "text": "of nonsense and it's not going to" + }, + { + "start": 903.24, + "text": "scale.\" So I basically was like, \"All" + }, + { + "start": 904.76, + "text": "right, I'm kind of done with deep" + }, + { + "start": 905.52, + "text": "learning. I think it's useless.\" Um and" + }, + { + "start": 907.839, + "text": "so I basically it's kind of like selling" + }, + { + "start": 909.4, + "text": "Bitcoin at $100. I never would want to" + }, + { + "start": 912.0, + "text": "buy back in cuz that's realizing my" + }, + { + "start": 913.52, + "text": "losses. So I'm just kind of waiting for" + }, + { + "start": 914.72, + "text": "it to collapse again, hopefully, maybe," + }, + { + "start": 916.24, + "text": "and then I can buy." + }, + { + "start": 917.52, + "text": "Um and if you think that's a funny joke" + }, + { + "start": 919.44, + "text": "in the audience right now, how many of" + }, + { + "start": 920.839, + "text": "you have any amount of retirement or" + }, + { + "start": 921.96, + "text": "capital in ETFs, right? So, I need you" + }, + { + "start": 924.04, + "text": "to raise your hand." + }, + { + "start": 926.52, + "text": "Okay." + }, + { + "start": 927.96, + "text": "Your money's being gambled on the" + }, + { + "start": 929.08, + "text": "success of AGI right now. Roughly 30% of" + }, + { + "start": 932.0, + "text": "the S&P 500 is heavily weighted on AI" + }, + { + "start": 934.0, + "text": "succeeding. AI growth stories, you know," + }, + { + "start": 936.72, + "text": "high price to earnings." + }, + { + "start": 938.24, + "text": "Um so, I propose a model of history" + }, + { + "start": 940.6, + "text": "here. There's basically two grifts" + }, + { + "start": 942.04, + "text": "orbiting each other, right? Fake money" + }, + { + "start": 944.08, + "text": "pumps the AI. Capital flees to Bitcoin" + }, + { + "start": 946.2, + "text": "instead of US-related, you know, assets." + }, + { + "start": 948.44, + "text": "No more easy capital, no more AI." + }, + { + "start": 950.32, + "text": "Deflationary doom spiral where GDP" + }, + { + "start": 951.76, + "text": "becomes 1/10 of what it is now. This is" + }, + { + "start": 953.72, + "text": "an idea that is basically a" + }, + { + "start": 955.0, + "text": "bubble-generating phenomenon, right?" + }, + { + "start": 956.96, + "text": "Both of these, you know, one paper" + }, + { + "start": 958.44, + "text": "basically started the whole" + }, + { + "start": 960.56, + "text": "There's like the transformer paper and" + }, + { + "start": 961.72, + "text": "there's the Bitcoin papers, parallels," + }, + { + "start": 963.24, + "text": "right? And if you look at this like, you" + }, + { + "start": 965.04, + "text": "know, innovators-laggards curve, right?" + }, + { + "start": 967.44, + "text": "I think we're right about here if you" + }, + { + "start": 968.68, + "text": "look at the Nvidia stock price. It fits" + }, + { + "start": 970.6, + "text": "the curve." + }, + { + "start": 972.24, + "text": "So, we're at peak of grift at the" + }, + { + "start": 973.28, + "text": "moment. It'll fall off soon, I think." + }, + { + "start": 974.84, + "text": "And so, I think now's a good time to" + }, + { + "start": 975.92, + "text": "short, but" + }, + { + "start": 979.8, + "text": "that may or may not be financial advice" + }, + { + "start": 981.28, + "text": "if it works out for you." + }, + { + "start": 984.4, + "text": "I do just want to say that shorting, I" + }, + { + "start": 986.48, + "text": "believe, has unlimited uh risk and you" + }, + { + "start": 989.2, + "text": "at best would like double your money, so" + }, + { + "start": 990.6, + "text": "it's probably not like a good idea." + }, + { + "start": 993.8, + "text": "Here's some examples of uh just machine" + }, + { + "start": 996.0, + "text": "learning in general, right? Let's go" + }, + { + "start": 996.96, + "text": "back to not, you know, being kind of" + }, + { + "start": 998.8, + "text": "schizo. Um" + }, + { + "start": 1001.04, + "text": "So, this is a little piece of history" + }, + { + "start": 1002.52, + "text": "again, right? This is called Cyc. In" + }, + { + "start": 1004.32, + "text": "1984, some people were very upset at the" + }, + { + "start": 1006.68, + "text": "deep learning people. They were like," + }, + { + "start": 1008.16, + "text": "\"Look at this arbitrary floats and" + }, + { + "start": 1009.44, + "text": "numbers. I want like an interpretable" + }, + { + "start": 1011.16, + "text": "program that's intelligent, right?\" So," + }, + { + "start": 1013.0, + "text": "they basically set out to manually" + }, + { + "start": 1015.2, + "text": "program all of human intelligence. And" + }, + { + "start": 1017.16, + "text": "the theory was if you just throw enough" + }, + { + "start": 1018.36, + "text": "bodies and people at it, it'll" + }, + { + "start": 1019.44, + "text": "eventually be smart, right? They've been" + }, + { + "start": 1021.16, + "text": "doing this since 1984. Okay, if you ask" + }, + { + "start": 1023.6, + "text": "this system questions like, \"If I put a" + }, + { + "start": 1025.52, + "text": "a cup of coffee on a table and then I" + }, + { + "start": 1026.92, + "text": "push on it, what happens?\" It'll say" + }, + { + "start": 1028.56, + "text": "like, \"Oh, well, you know, it explodes.\"" + }, + { + "start": 1029.8, + "text": "It's like still not working very well." + }, + { + "start": 1031.56, + "text": "the reasoning traces are like a thousand" + }, + { + "start": 1033.64, + "text": "deep, you know, it's kind of nuts. It" + }, + { + "start": 1034.92, + "text": "didn't really work, right? The First of" + }, + { + "start": 1036.72, + "text": "all, the reasoning isn't even" + }, + { + "start": 1037.56, + "text": "understandable." + }, + { + "start": 1038.8, + "text": "Um this approach just doesn't It's" + }, + { + "start": 1040.36, + "text": "called symbolic AI, right? And I think" + }, + { + "start": 1042.12, + "text": "that there's actually a modern parallel," + }, + { + "start": 1043.64, + "text": "right? Waymo is like psych, that project" + }, + { + "start": 1045.68, + "text": "where they were trying to manually" + }, + { + "start": 1046.68, + "text": "encode intelligence instead of automatic" + }, + { + "start": 1048.68, + "text": "programming, right? This is like deep" + }, + { + "start": 1050.48, + "text": "learning is really a new paradigm of" + }, + { + "start": 1051.72, + "text": "programming where it's automatic. You're" + }, + { + "start": 1053.28, + "text": "automatically creating the the" + }, + { + "start": 1054.76, + "text": "nonlinearities that do the thing that" + }, + { + "start": 1056.12, + "text": "you want the computer to do, right? Um" + }, + { + "start": 1058.2, + "text": "Waymo, they basically scan the city and" + }, + { + "start": 1060.12, + "text": "they put a bunch of if statements where" + }, + { + "start": 1061.08, + "text": "the car should go. So, they're very slow" + }, + { + "start": 1063.0, + "text": "to roll out. So, Tesla has basically" + }, + { + "start": 1065.16, + "text": "just recorded millions and millions of" + }, + { + "start": 1066.88, + "text": "hours of just people driving and they're" + }, + { + "start": 1068.84, + "text": "training massive massive transformers on" + }, + { + "start": 1071.16, + "text": "human behavior and they're just trying" + }, + { + "start": 1072.24, + "text": "to model it, right?" + }, + { + "start": 1074.12, + "text": "It's also There's another parallel with" + }, + { + "start": 1075.48, + "text": "like Siri and Echo Dot, right? These are" + }, + { + "start": 1077.04, + "text": "symbolic systems. They're often" + }, + { + "start": 1078.32, + "text": "frustratingly unintelligent. Um" + }, + { + "start": 1080.48, + "text": "to add new intelligent behavior, an" + }, + { + "start": 1081.92, + "text": "engineer has to basically, you know," + }, + { + "start": 1083.68, + "text": "design the algorithm and all that stuff." + }, + { + "start": 1085.04, + "text": "And ChatGPT is statistical. It's deep" + }, + { + "start": 1086.88, + "text": "learning. It's automatic programming," + }, + { + "start": 1090.72, + "text": "Deep learning models don't learn. That" + }, + { + "start": 1092.12, + "text": "always bugs me. I think it's a misnomer." + }, + { + "start": 1094.2, + "text": "Um that's like saying when you're game" + }, + { + "start": 1095.16, + "text": "developer, you're creating a universe." + }, + { + "start": 1096.52, + "text": "It's like a very like hand-wavy way to" + }, + { + "start": 1098.68, + "text": "say something that I think is really" + }, + { + "start": 1099.88, + "text": "mechanical and mundane, right? I think" + }, + { + "start": 1101.68, + "text": "it's more accurate way to uh describe" + }, + { + "start": 1103.88, + "text": "what happens with deep learning is" + }, + { + "start": 1104.88, + "text": "optimizing, right? And it's an automatic" + }, + { + "start": 1106.44, + "text": "optimization. Um" + }, + { + "start": 1108.83, + "text": "ation. Um" + }, + { + "start": 1108.84, + "text": "Here's just another side note, right?" + }, + { + "start": 1109.96, + "text": "So, there's big copyright disputes with" + }, + { + "start": 1111.16, + "text": "these deep learning models, right? And I" + }, + { + "start": 1112.36, + "text": "just had a funny idea, right? Like" + }, + { + "start": 1115.04, + "text": "you could almost make the ar- if you" + }, + { + "start": 1116.4, + "text": "take the argument that uh the deep" + }, + { + "start": 1117.76, + "text": "learning people are trying to make," + }, + { + "start": 1118.76, + "text": "right? They're trying to secure their" + }, + { + "start": 1119.76, + "text": "bag. They're trying to say, \"Well, we" + }, + { + "start": 1121.32, + "text": "learned the patterns from what you gave" + }, + { + "start": 1122.88, + "text": "us. We didn't, you know, copy the pixels" + }, + { + "start": 1124.48, + "text": "or whatever.\" That's like what JPEG" + }, + { + "start": 1126.0, + "text": "does, like learns a model of the pixels." + }, + { + "start": 1128.52, + "text": "And so, it's almost like arguing that" + }, + { + "start": 1129.72, + "text": "because, you know, it's JPEG and not" + }, + { + "start": 1131.2, + "text": "PNG, uh because it's a model of the" + }, + { + "start": 1133.12, + "text": "pixels, you know, I didn't steal your" + }, + { + "start": 1134.4, + "text": "thing." + }, + { + "start": 1135.4, + "text": "Um so, it's a little strange, right? Um" + }, + { + "start": 1137.2, + "text": "I think it's kind of a bad faith" + }, + { + "start": 1138.36, + "text": "argument cuz the patterns uh that are" + }, + { + "start": 1140.8, + "text": "found in deep learning um are very" + }, + { + "start": 1143.36, + "text": "generic and they're not so specific to" + }, + { + "start": 1144.8, + "text": "the image. And so," + }, + { + "start": 1146.48, + "text": "it's very debatable, I think. I'm not" + }, + { + "start": 1147.68, + "text": "sure what's going to happen. But just to" + }, + { + "start": 1149.28, + "text": "go back to my League of Legends task," + }, + { + "start": 1150.72, + "text": "right? If we if we just want to think" + }, + { + "start": 1151.72, + "text": "about that little bit more as an" + }, + { + "start": 1152.6, + "text": "engineer," + }, + { + "start": 1153.72, + "text": "um you might naively just put in the" + }, + { + "start": 1155.08, + "text": "current state of where all the people" + }, + { + "start": 1156.44, + "text": "are on the map, how much gold they have," + }, + { + "start": 1157.72, + "text": "what items they have, throw it through" + }, + { + "start": 1159.24, + "text": "that perceptron I showed you, and maybe" + }, + { + "start": 1161.12, + "text": "it'll automatically find ways to relate," + }, + { + "start": 1162.8, + "text": "oh, like, you know, when the gold and" + }, + { + "start": 1164.48, + "text": "this item is here, then like they're" + }, + { + "start": 1165.96, + "text": "more likely to win than you would think," + }, + { + "start": 1167.36, + "text": "and stuff like that. These are the kind" + }, + { + "start": 1168.84, + "text": "of patterns you hope a deep learning" + }, + { + "start": 1170.36, + "text": "model finds as it optimizes based on" + }, + { + "start": 1172.48, + "text": "massive amount of data you have, right?" + }, + { + "start": 1175.08, + "text": "There's a problem with this, though." + }, + { + "start": 1175.88, + "text": "What about past states, right? Past" + }, + { + "start": 1177.68, + "text": "state informs the nature of present" + }, + { + "start": 1179.56, + "text": "state. If somebody was performing badly" + }, + { + "start": 1181.43, + "text": ". If somebody was performing badly" + }, + { + "start": 1181.44, + "text": "on your team in the past, right? They" + }, + { + "start": 1182.68, + "text": "were like running it down mid, this" + }, + { + "start": 1184.16, + "text": "should inform your current prediction" + }, + { + "start": 1185.36, + "text": "like heavily, right? Um" + }, + { + "start": 1187.96, + "text": "So, this uh you know" + }, + { + "start": 1190.0, + "text": "You need a recurrent neural network for" + }, + { + "start": 1191.12, + "text": "this, right? A recurrent neural network" + }, + { + "start": 1192.68, + "text": "is basically two of those perceptrons" + }, + { + "start": 1194.44, + "text": "that I showed you. One is modifying the" + }, + { + "start": 1196.64, + "text": "state almost, and then another one is" + }, + { + "start": 1198.08, + "text": "visualizing the state towards whatever" + }, + { + "start": 1199.8, + "text": "like you want it to visualize, right? It" + }, + { + "start": 1201.8, + "text": "might be probability for who's going to" + }, + { + "start": 1202.96, + "text": "win, or it might be like you could" + }, + { + "start": 1204.64, + "text": "replace it with another deep learning" + }, + { + "start": 1206.12, + "text": "network and train it to visualize like" + }, + { + "start": 1208.64, + "text": "I don't know, which which champion is" + }, + { + "start": 1209.92, + "text": "like has the most synergy with the other" + }, + { + "start": 1211.2, + "text": "champion, stuff like this, right? Um it" + }, + { + "start": 1213.52, + "text": "has state. It's much more like a" + }, + { + "start": 1214.84, + "text": "automatic program, right? Automatically" + }, + { + "start": 1216.24, + "text": "learned program. Um the hidden state of" + }, + { + "start": 1218.48, + "text": "a recurrent neural network is like RAM." + }, + { + "start": 1219.84, + "text": "It's analogous to a RAM." + }, + { + "start": 1221.36, + "text": "Um and that and that is why I think that" + }, + { + "start": 1222.96, + "text": "these are" + }, + { + "start": 1224.08, + "text": "best described as automatic programs," + }, + { + "start": 1225.52, + "text": "right? Um and so, what I did at the time" + }, + { + "start": 1228.52, + "text": "to successfully model uh who would win" + }, + { + "start": 1230.88, + "text": "in League of Legends games was I I I I" + }, + { + "start": 1232.72, + "text": "selected a GRU architecture, and I spent" + }, + { + "start": 1234.72, + "text": "weeks tuning it and and messing around," + }, + { + "start": 1237.2, + "text": "and I I got I managed to make it so" + }, + { + "start": 1239.12, + "text": "halfway through the game it could" + }, + { + "start": 1240.08, + "text": "predict who would win with 86% accuracy." + }, + { + "start": 1242.56, + "text": "Um if you just went with whoever had" + }, + { + "start": 1243.76, + "text": "more gold, it was like 73% accuracy. Um" + }, + { + "start": 1246.44, + "text": "and then this continued to generalize" + }, + { + "start": 1248.0, + "text": "even on arbitrarily new games. Um and" + }, + { + "start": 1250.32, + "text": "just as, you know, as a side note for" + }, + { + "start": 1251.72, + "text": "game developers, you could use a" + }, + { + "start": 1253.12, + "text": "technology like this for our play of the" + }, + { + "start": 1255.12, + "text": "game kind of feature, right? To see big" + }, + { + "start": 1256.68, + "text": "unpredictable swings in who's winning." + }, + { + "start": 1258.52, + "text": "Um we actually did do that. We took a" + }, + { + "start": 1260.44, + "text": "our giant corpus of professional games," + }, + { + "start": 1262.52, + "text": "and I ran this model and saw what were" + }, + { + "start": 1263.92, + "text": "the biggest swings, and we looked at" + }, + { + "start": 1265.56, + "text": "those games, and they were the most" + }, + { + "start": 1266.6, + "text": "exciting matches, right? So, this is" + }, + { + "start": 1268.72, + "text": "kind of like just to give you ideas for" + }, + { + "start": 1270.88, + "text": "like what this technology can be used" + }, + { + "start": 1272.52, + "text": "for, right? Um each perceptron in a" + }, + { + "start": 1275.44, + "text": "layer is like roughly analogous to an if" + }, + { + "start": 1277.24, + "text": "statement in the program, right? Like" + }, + { + "start": 1278.24, + "text": "another branch. It lets you make a more" + }, + { + "start": 1280.24, + "text": "specific contour around what you're" + }, + { + "start": 1282.2, + "text": "trying to encapsulate, right? Um you're" + }, + { + "start": 1284.92, + "text": "adding more resolution to that boundary." + }, + { + "start": 1286.72, + "text": "Um this is why it's called deep learning" + }, + { + "start": 1287.96, + "text": "cuz there's many Once you add many" + }, + { + "start": 1289.48, + "text": "layers, you can make very complicated" + }, + { + "start": 1290.84, + "text": "behavior, right? Um the problem if you" + }, + { + "start": 1293.0, + "text": "just take that perceptron and you add a" + }, + { + "start": 1294.4, + "text": "bunch of layers is it adds way, way, way" + }, + { + "start": 1296.64, + "text": "too many parameters and the gradients" + }, + { + "start": 1297.96, + "text": "get very tiny, right? Um it doesn't" + }, + { + "start": 1300.12, + "text": "optimize, right? The optimizer doesn't" + }, + { + "start": 1301.44, + "text": "work. I mean, you saw my failure of deep" + }, + { + "start": 1303.0, + "text": "learning. It's not easy." + }, + { + "start": 1304.52, + "text": "Uh you almost have to like guess and" + }, + { + "start": 1306.12, + "text": "check if you really aren't, you know," + }, + { + "start": 1308.24, + "text": "branching off of some established known" + }, + { + "start": 1310.0, + "text": "working uh architecture, right? Um and" + }, + { + "start": 1312.44, + "text": "so all of the field of of modern deep" + }, + { + "start": 1314.68, + "text": "learning is basically about how do we" + }, + { + "start": 1316.56, + "text": "add more parameters to this model and" + }, + { + "start": 1318.4, + "text": "still get those parameters to be" + }, + { + "start": 1319.52, + "text": "optimized into something that" + }, + { + "start": 1320.64, + "text": "generalizes to what we needed to do," + }, + { + "start": 1322.76, + "text": "right? In the case of a convolutional" + }, + { + "start": 1324.48, + "text": "neural network over here, they basically" + }, + { + "start": 1327.24, + "text": "they kind of start with the inductive" + }, + { + "start": 1328.44, + "text": "bias that like, \"Okay," + }, + { + "start": 1330.2, + "text": "um when you're looking to see if a image" + }, + { + "start": 1331.8, + "text": "is of a cat or a dog, um and you" + }, + { + "start": 1334.44, + "text": "basically are hoping that let's say this" + }, + { + "start": 1335.88, + "text": "deep learning model will pick up on the" + }, + { + "start": 1337.04, + "text": "pointy ear thing. It doesn't matter" + }, + { + "start": 1338.84, + "text": "where in the image that pointy ear is," + }, + { + "start": 1340.6, + "text": "right?\" Um they basically want these" + }, + { + "start": 1342.96, + "text": "deep learning models to learn, you know," + }, + { + "start": 1345.28, + "text": "features that it doesn't matter where" + }, + { + "start": 1346.96, + "text": "they are in the image. Uh if you just" + }, + { + "start": 1348.6, + "text": "fed every single pixel and all the RGB" + }, + { + "start": 1350.24, + "text": "values into a giant perceptron, then it" + }, + { + "start": 1352.28, + "text": "would have to learn that, you know, cat" + }, + { + "start": 1353.72, + "text": "ear means cat in every single spot," + }, + { + "start": 1355.72, + "text": "right? That's not really what we do when" + }, + { + "start": 1357.04, + "text": "we recognize images and so they" + }, + { + "start": 1358.68, + "text": "basically That's like a little bit of" + }, + { + "start": 1359.8, + "text": "key insight. That's what deep learning" + }, + { + "start": 1361.24, + "text": "researchers are trying to do is like" + }, + { + "start": 1363.24, + "text": "come up with ways to to formulate and" + }, + { + "start": 1365.0, + "text": "structure the computation, the graph of" + }, + { + "start": 1366.88, + "text": "computation, such that when you optimize" + }, + { + "start": 1369.28, + "text": "it, it gets what you want, right? Um and" + }, + { + "start": 1371.28, + "text": "a transformer on the other end here I'll" + }, + { + "start": 1372.52, + "text": "talk about in a second. Um this is" + }, + { + "start": 1374.56, + "text": "another example of an architectural" + }, + { + "start": 1375.8, + "text": "change. You can add way more parameters" + }, + { + "start": 1377.76, + "text": "and it's still stable and it's still," + }, + { + "start": 1379.72, + "text": "you know, optimizing correctly, right?" + }, + { + "start": 1381.8, + "text": "Just as another little side note, some" + }, + { + "start": 1383.12, + "text": "deep learning techniques are a little" + }, + { + "start": 1384.12, + "text": "silly. Uh Uh this is called dropout. You" + }, + { + "start": 1386.72, + "text": "randomly turn off neurons during the" + }, + { + "start": 1388.76, + "text": "during the training. It's it's supposed" + }, + { + "start": 1390.36, + "text": "to present or or prevent memorizing the" + }, + { + "start": 1391.92, + "text": "training data um and hopefully make it" + }, + { + "start": 1393.88, + "text": "so neurons aren't exactly specialized" + }, + { + "start": 1397.28, + "text": "Who knows why this works? Honestly, I" + }, + { + "start": 1398.64, + "text": "don't I haven't seen any real convincing" + }, + { + "start": 1400.28, + "text": "argument in any paper for what exactly" + }, + { + "start": 1402.36, + "text": "this is doing. Like basically what's" + }, + { + "start": 1404.52, + "text": "happening in deep learning is people try" + }, + { + "start": 1405.88, + "text": "stuff and they have some very like okay," + }, + { + "start": 1407.8, + "text": "yeah, that sounds about right. I think" + }, + { + "start": 1408.8, + "text": "this what's happening and then it's like" + }, + { + "start": 1409.88, + "text": "there's no rigor in this field basically" + }, + { + "start": 1411.28, + "text": "like um it's really crazy. This is" + }, + { + "start": 1413.76, + "text": "actually used in production for chat GPT" + }, + { + "start": 1415.44, + "text": "without it like it wouldn't work, I" + }, + { + "start": 1416.56, + "text": "think. Um and it's like you know, we" + }, + { + "start": 1418.04, + "text": "just randomly turn off the neurons" + }, + { + "start": 1419.28, + "text": "during the training. Like okay." + }, + { + "start": 1422.56, + "text": "So, there's a zillion videos that" + }, + { + "start": 1424.48, + "text": "explain transformers already. I'm not" + }, + { + "start": 1425.8, + "text": "going to give you an explanation of the" + }, + { + "start": 1427.24, + "text": "transformer, which is you know, the" + }, + { + "start": 1428.28, + "text": "cause of the current rift, right? Um and" + }, + { + "start": 1430.72, + "text": "you know, you can watch any of these and" + }, + { + "start": 1431.76, + "text": "they're good and whatever. But I'm just" + }, + { + "start": 1433.76, + "text": "going to give you a quick overview. This" + }, + { + "start": 1434.76, + "text": "is a single slide. This is what a" + }, + { + "start": 1436.04, + "text": "transformer is doing, right? It's" + }, + { + "start": 1437.48, + "text": "basically going from a sentence, a" + }, + { + "start": 1439.16, + "text": "lookup table per token into a point in a" + }, + { + "start": 1441.04, + "text": "space where like the closer two points" + }, + { + "start": 1443.12, + "text": "are the more related they are, right? It" + }, + { + "start": 1444.96, + "text": "adds a position encoding to those points" + }, + { + "start": 1446.84, + "text": "so that, you know, if we have an array" + }, + { + "start": 1448.84, + "text": "of these points, then the points are" + }, + { + "start": 1450.68, + "text": "modulated in such a way that the the" + }, + { + "start": 1452.76, + "text": "deep learning models they can they can" + }, + { + "start": 1454.16, + "text": "pick up where they are in the sentence," + }, + { + "start": 1455.96, + "text": "right? They're almost jiggle in a way" + }, + { + "start": 1457.24, + "text": "where like you could learn that pattern." + }, + { + "start": 1460.48, + "text": "I keep using the word learn. I'm sorry." + }, + { + "start": 1461.64, + "text": "It's pervasive in the field, but I don't" + }, + { + "start": 1462.84, + "text": "like it. Um and then you literally just" + }, + { + "start": 1465.4, + "text": "stack a hundred of these attention" + }, + { + "start": 1467.04, + "text": "layers onto those uh vectors, right? And" + }, + { + "start": 1469.28, + "text": "each of these attention layers every" + }, + { + "start": 1470.44, + "text": "single time, every single one of those" + }, + { + "start": 1472.12, + "text": "tokens, right? It's basically looking at" + }, + { + "start": 1474.32, + "text": "all of the past tokens and enriching the" + }, + { + "start": 1475.92, + "text": "current vector with more information" + }, + { + "start": 1478.0, + "text": "about the surrounding tokens and then" + }, + { + "start": 1479.6, + "text": "more facts potentially about that" + }, + { + "start": 1481.12, + "text": "information in the form of another" + }, + { + "start": 1482.76, + "text": "perceptron, right? it's like learning" + }, + { + "start": 1484.0, + "text": "what what facts might be related to a" + }, + { + "start": 1485.68, + "text": "word like house, right? It's It It" + }, + { + "start": 1487.4, + "text": "continuously enriches each one of these" + }, + { + "start": 1489.88, + "text": "uh words with more and more context," + }, + { + "start": 1491.36, + "text": "right? In a in a highly parallel easy to" + }, + { + "start": 1492.88, + "text": "train manner. Um and then there's just" + }, + { + "start": 1495.44, + "text": "this two-layer perceptron and then it" + }, + { + "start": 1496.84, + "text": "gives you the next word, and it does" + }, + { + "start": 1497.96, + "text": "that over and over again, right? Um" + }, + { + "start": 1500.76, + "text": "there are many hyper parameters in this" + }, + { + "start": 1502.56, + "text": "massive system, right? Many, many, many," + }, + { + "start": 1504.56, + "text": "many things can go wrong. You know, it's" + }, + { + "start": 1506.64, + "text": "like what is the learning rate? You" + }, + { + "start": 1507.88, + "text": "know, how quickly do we turn these" + }, + { + "start": 1509.08, + "text": "parameters to, you know, make it better," + }, + { + "start": 1510.44, + "text": "right? It's like a massive problem," + }, + { + "start": 1512.16, + "text": "actually, um to know what These are" + }, + { + "start": 1514.48, + "text": "called hyper parameters because unlike" + }, + { + "start": 1516.04, + "text": "parameters, like they're not" + }, + { + "start": 1517.76, + "text": "automatically learned in the core" + }, + { + "start": 1519.32, + "text": "algorithm of deep learning, right? You" + }, + { + "start": 1520.68, + "text": "can or you can try to figure out what" + }, + { + "start": 1522.28, + "text": "they are uh with a search algorithm." + }, + { + "start": 1524.24, + "text": "That's very costly if you're OpenAI and" + }, + { + "start": 1526.12, + "text": "each training run costs $10 million. Um" + }, + { + "start": 1528.6, + "text": "and as it turns out, we heard from an" + }, + { + "start": 1530.56, + "text": "insider here that most of the big gains" + }, + { + "start": 1533.52, + "text": "from these, you know," + }, + { + "start": 1535.36, + "text": "most of the state-of-the-art models that" + }, + { + "start": 1536.48, + "text": "come out of OpenAI, they came from" + }, + { + "start": 1537.64, + "text": "so-called YOLO runs, where they" + }, + { + "start": 1539.72, + "text": "basically they basically asked Karpathy" + }, + { + "start": 1542.04, + "text": "or one of these, you know, like deep" + }, + { + "start": 1543.56, + "text": "learning generals, like, \"Hey," + }, + { + "start": 1545.16, + "text": "we're going to give you $20 million to" + }, + { + "start": 1546.52, + "text": "just try something, right? You just pick" + }, + { + "start": 1547.84, + "text": "numbers, pick hyper parameters, pick" + }, + { + "start": 1549.24, + "text": "whatever you want, and we'll see what" + }, + { + "start": 1550.72, + "text": "happens.\" And state-of-the-art model," + }, + { + "start": 1552.92, + "text": "right? It's basically vibe engineering." + }, + { + "start": 1555.04, + "text": "Vibe engineering created the vibe coding" + }, + { + "start": 1556.6, + "text": "machines, and everybody's vibing, and" + }, + { + "start": 1558.52, + "text": "then, you know, we're being wireheaded," + }, + { + "start": 1559.76, + "text": "and then it's over." + }, + { + "start": 1565.44, + "text": "Well," + }, + { + "start": 1566.76, + "text": "I I I I I think that it's a very" + }, + { + "start": 1568.92, + "text": "interesting question, can these LLMs" + }, + { + "start": 1570.48, + "text": "reason, right? So, they can't reason cuz" + }, + { + "start": 1572.32, + "text": "they're not human. I think that's a good" + }, + { + "start": 1574.12, + "text": "way to use the definition of reason. Um" + }, + { + "start": 1576.04, + "text": "but let's say we want to figure out if" + }, + { + "start": 1577.24, + "text": "these LLMs can, you know, output text as" + }, + { + "start": 1579.44, + "text": "if they were reasoning, right? I think" + }, + { + "start": 1581.2, + "text": "that this benchmark ARC-AGI is the It's" + }, + { + "start": 1583.76, + "text": "a stellar benchmark for this task," + }, + { + "start": 1585.2, + "text": "right? If you take a look at this" + }, + { + "start": 1586.28, + "text": "diagram here, right? That's uh one" + }, + { + "start": 1588.36, + "text": "example, and that's another example, and" + }, + { + "start": 1589.8, + "text": "you have to predict what happens here." + }, + { + "start": 1591.68, + "text": "If you're human, which I'm assuming all" + }, + { + "start": 1593.24, + "text": "of you are, it should be pretty obvious" + }, + { + "start": 1595.24, + "text": "to you that you basically just put the" + }, + { + "start": 1596.64, + "text": "yellow dots like where the little corner" + }, + { + "start": 1598.04, + "text": "thing goes. And ARC-AGI is the kind of" + }, + { + "start": 1600.16, + "text": "benchmark where they basically take" + }, + { + "start": 1601.76, + "text": "random people and they're like, \"All" + }, + { + "start": 1602.6, + "text": "right, we'll give you a $15 Starbucks" + }, + { + "start": 1603.88, + "text": "gift card if you solve this problem.\"" + }, + { + "start": 1605.4, + "text": "And they verify that like people can" + }, + { + "start": 1607.28, + "text": "just figure it out in a context-free" + }, + { + "start": 1608.56, + "text": "manner, right? And so, the whole idea" + }, + { + "start": 1610.32, + "text": "with this is they're trying to find" + }, + { + "start": 1611.36, + "text": "problems that like average people can" + }, + { + "start": 1613.0, + "text": "do, but like, you know, these LLMs" + }, + { + "start": 1614.6, + "text": "they're they can't do it at all. Um" + }, + { + "start": 1617.28, + "text": "And ARK G I they were originally, you" + }, + { + "start": 1620.04, + "text": "know, totally defeated by OpenAI's" + }, + { + "start": 1621.76, + "text": "latest reasoning model." + }, + { + "start": 1623.4, + "text": "Um it it got human baseline on a huge" + }, + { + "start": 1625.68, + "text": "suite of these that are private, not" + }, + { + "start": 1627.36, + "text": "anywhere existing on the internet. Um" + }, + { + "start": 1629.36, + "text": "you know, generic Each each ARK G I" + }, + { + "start": 1631.16, + "text": "problem requires a new unique insight." + }, + { + "start": 1632.84, + "text": "They're basically puzzle game problems," + }, + { + "start": 1634.2, + "text": "right? They require some kind of like," + }, + { + "start": 1636.16, + "text": "\"Oh, that's how it works,\" right? And" + }, + { + "start": 1637.84, + "text": "LLMs have already beaten the first ARK G" + }, + { + "start": 1639.68, + "text": "I benchmark. Which that was quite" + }, + { + "start": 1641.08, + "text": "significant. I didn't think that would" + }, + { + "start": 1642.0, + "text": "happen. And neither did the author of" + }, + { + "start": 1643.44, + "text": "the benchmark. He's a massive skeptic" + }, + { + "start": 1644.72, + "text": "and he thinks it's a bubble. And then" + }, + { + "start": 1646.32, + "text": "OpenAI pulled them on stream and he was" + }, + { + "start": 1647.84, + "text": "like, \"My benchmark is defeated.\" Um so" + }, + { + "start": 1650.52, + "text": "we made a new benchmark that they're all" + }, + { + "start": 1651.64, + "text": "failing at again." + }, + { + "start": 1653.04, + "text": "And that's Grok 4. A few days ago they" + }, + { + "start": 1655.68, + "text": "just doubled the state of the art and" + }, + { + "start": 1656.96, + "text": "got 15% the 16% of the questions right." + }, + { + "start": 1660.52, + "text": "Um So just circle back on some" + }, + { + "start": 1662.4, + "text": "applications to games really quick as I" + }, + { + "start": 1664.12, + "text": "kind of" + }, + { + "start": 1665.04, + "text": "close up on this kind of like lightning" + }, + { + "start": 1666.52, + "text": "overview of this, right? Any problem" + }, + { + "start": 1668.56, + "text": "where you have input to output data and" + }, + { + "start": 1670.08, + "text": "it's harder to find programmatically" + }, + { + "start": 1671.48, + "text": "what that mapping is, that's where you" + }, + { + "start": 1673.24, + "text": "would throw something like deep learning" + }, + { + "start": 1674.8, + "text": "at the problem and try to automatically" + }, + { + "start": 1676.56, + "text": "program that, you know, uh" + }, + { + "start": 1678.36, + "text": "transformation." + }, + { + "start": 1680.0, + "text": "Um this is Nerf rendering. This is a" + }, + { + "start": 1681.44, + "text": "very interesting application of deep" + }, + { + "start": 1683.08, + "text": "learning to the problem of rendering, I" + }, + { + "start": 1684.32, + "text": "think. Um it's very old and there's much" + }, + { + "start": 1686.56, + "text": "better techniques like uh Gaussian" + }, + { + "start": 1688.0, + "text": "splatting." + }, + { + "start": 1689.04, + "text": "Um but Nerf rendering is basically given" + }, + { + "start": 1691.2, + "text": "a point in the scene, a deep learning" + }, + { + "start": 1692.52, + "text": "model learns how dense it is and then" + }, + { + "start": 1694.52, + "text": "it, you know, kind of traces rays" + }, + { + "start": 1695.6, + "text": "through that learned density to render." + }, + { + "start": 1698.68, + "text": "And just as a little interesting tidbit" + }, + { + "start": 1700.64, + "text": "more vibe uh engineering, uh basically" + }, + { + "start": 1703.8, + "text": "deep learning models are very bad at" + }, + { + "start": 1705.08, + "text": "learning so-called high-frequency um" + }, + { + "start": 1708.48, + "text": "high-frequency functions, right? So a" + }, + { + "start": 1709.92, + "text": "little tiny thin metal pole is a very" + }, + { + "start": 1711.44, + "text": "high-frequency function density function" + }, + { + "start": 1713.12, + "text": "for it to learn. It's like nothing" + }, + { + "start": 1714.44, + "text": "nothing nothing bam really thin nothing" + }, + { + "start": 1716.0, + "text": "nothing nothing. It's very like tight" + }, + { + "start": 1717.55, + "text": ". It's very like tight" + }, + { + "start": 1717.56, + "text": "and so it's hard for a deep learning" + }, + { + "start": 1718.68, + "text": "model to pick up on that cuz it would" + }, + { + "start": 1720.2, + "text": "require a lot of non-linearity. It would" + }, + { + "start": 1721.6, + "text": "require a lot of layers, right? Um and" + }, + { + "start": 1723.67, + "text": "layers, right? Um and" + }, + { + "start": 1723.68, + "text": "so what they did in Nerf to, you know," + }, + { + "start": 1725.44, + "text": "achieve state-of-the-art results for" + }, + { + "start": 1726.96, + "text": "learning based rendering, is they pass" + }, + { + "start": 1728.92, + "text": "that position into" + }, + { + "start": 1731.28, + "text": "into a bunch of different sine waves and" + }, + { + "start": 1733.08, + "text": "they basically just made it easier for" + }, + { + "start": 1734.32, + "text": "the deep learning model to pick up on" + }, + { + "start": 1736.0, + "text": "higher frequency spatial information. I" + }, + { + "start": 1738.2, + "text": "thought that was kind of funny. She's" + }, + { + "start": 1739.16, + "text": "like, pass it into the sine waves and" + }, + { + "start": 1740.2, + "text": "then it's better." + }, + { + "start": 1741.24, + "text": "Um this is called Ash Animatable" + }, + { + "start": 1742.84, + "text": "Gaussian Splatting. Let me show you how" + }, + { + "start": 1743.96, + "text": "the video works here. This is basically" + }, + { + "start": 1746.56, + "text": "photorealistic human rendering with with" + }, + { + "start": 1748.32, + "text": "accurate physics on their clothes done" + }, + { + "start": 1749.68, + "text": "in real time. Um it's just" + }, + { + "start": 1751.76, + "text": "unfortunately, in order to, you know," + }, + { + "start": 1753.8, + "text": "scan somebody, you need about 100" + }, + { + "start": 1755.2, + "text": "cameras synchronized and it's very" + }, + { + "start": 1756.8, + "text": "expensive. So, maybe they'll figure out" + }, + { + "start": 1758.52, + "text": "how to make that cheaper someday, but in" + }, + { + "start": 1759.96, + "text": "this specific paper, a deep learning" + }, + { + "start": 1761.68, + "text": "model is used to go from a texture that" + }, + { + "start": 1763.68, + "text": "describes the skeletal deformation of" + }, + { + "start": 1765.8, + "text": "what you want the person to be in into" + }, + { + "start": 1768.2, + "text": "the the view dependent color that you" + }, + { + "start": 1769.72, + "text": "would see" + }, + { + "start": 1770.84, + "text": "um from that skeletal deformation. And" + }, + { + "start": 1772.6, + "text": "and essentially, you need to train that" + }, + { + "start": 1774.16, + "text": "deep learning model to learn that" + }, + { + "start": 1775.08, + "text": "relationship between like how their arm" + }, + { + "start": 1776.92, + "text": "is bending and what their clothes look" + }, + { + "start": 1778.44, + "text": "like, um you know, per person. Um" + }, + { + "start": 1781.56, + "text": "And then of course, I'm sure probably" + }, + { + "start": 1783.04, + "text": "all of you guys already know, ray" + }, + { + "start": 1784.48, + "text": "tracing is is is greatly benefited by um" + }, + { + "start": 1787.92, + "text": "deep learning because it's used to" + }, + { + "start": 1789.16, + "text": "denoise, right? Um" + }, + { + "start": 1792.16, + "text": "Deep learning is used in a algorithm" + }, + { + "start": 1794.0, + "text": "called AlphaZero to master any discrete" + }, + { + "start": 1795.92, + "text": "perfect information game. Basically, a a" + }, + { + "start": 1798.44, + "text": "neural network goes from the board state" + }, + { + "start": 1799.84, + "text": "to a likely next move" + }, + { + "start": 1801.52, + "text": "um and in a win probability. And then" + }, + { + "start": 1803.48, + "text": "another algorithm manipulates that those" + }, + { + "start": 1805.88, + "text": "probabilities and you know," + }, + { + "start": 1807.8, + "text": "beats any human in any game we've ever" + }, + { + "start": 1808.96, + "text": "invented, unfortunately." + }, + { + "start": 1811.36, + "text": "Automatic game AI, so you can make a" + }, + { + "start": 1812.8, + "text": "style transfer, you know," + }, + { + "start": 1815.16, + "text": "you can make it some kind of style" + }, + { + "start": 1816.08, + "text": "transfer for something like this for" + }, + { + "start": 1817.08, + "text": "your video game where maybe like you" + }, + { + "start": 1819.4, + "text": "want it to act more like the human" + }, + { + "start": 1820.64, + "text": "people. So, once it masters the game," + }, + { + "start": 1823.0, + "text": "you then train it to act like a person" + }, + { + "start": 1825.08, + "text": "and then it would be a kind of" + }, + { + "start": 1826.6, + "text": "interesting mix between a good player" + }, + { + "start": 1827.96, + "text": "and oh, it plays like this specific" + }, + { + "start": 1829.44, + "text": "chess player or whatever." + }, + { + "start": 1830.8, + "text": "Um or you could even have it validate" + }, + { + "start": 1832.6, + "text": "that your game is uh balanced and stuff" + }, + { + "start": 1834.88, + "text": "like that, right? I think this actually" + }, + { + "start": 1836.08, + "text": "is very it's quite green field for video" + }, + { + "start": 1837.8, + "text": "games in general for how to apply uh" + }, + { + "start": 1839.84, + "text": "deep learning models in a lot of places." + }, + { + "start": 1842.76, + "text": "And you know, for automating like a call" + }, + { + "start": 1844.24, + "text": "center, I think that there's a lot of" + }, + { + "start": 1845.52, + "text": "these call center AI startups. You know," + }, + { + "start": 1847.48, + "text": "it seems easy on the surface, but" + }, + { + "start": 1848.76, + "text": "usually people call support because they" + }, + { + "start": 1850.24, + "text": "want to talk to a person and not a" + }, + { + "start": 1851.56, + "text": "machine. So, I think that this is" + }, + { + "start": 1853.16, + "text": "probably going to be a a giant disaster." + }, + { + "start": 1855.84, + "text": "Um and, you know, of course, automating" + }, + { + "start": 1857.48, + "text": "something like engineering is way more" + }, + { + "start": 1858.68, + "text": "complicated than just that, which isn't" + }, + { + "start": 1860.72, + "text": "even working. So," + }, + { + "start": 1862.12, + "text": "you know, just something to think about." + }, + { + "start": 1863.32, + "text": "Um and a quick note on computer vision" + }, + { + "start": 1865.04, + "text": "cuz I didn't really get to flush that" + }, + { + "start": 1866.24, + "text": "out that much in this talk. But, the" + }, + { + "start": 1868.2, + "text": "basic idea I had was for computer" + }, + { + "start": 1870.72, + "text": "vision, you're essentially trying to" + }, + { + "start": 1872.16, + "text": "take in a bunch of images in something" + }, + { + "start": 1873.68, + "text": "called monocular slam and get some kind" + }, + { + "start": 1876.48, + "text": "of representation of the world that it's" + }, + { + "start": 1878.32, + "text": "seeing um automatically, right? It's a" + }, + { + "start": 1880.4, + "text": "bit like deep learning where you have" + }, + { + "start": 1881.52, + "text": "some automatic optimization process that" + }, + { + "start": 1883.76, + "text": "operates on a lot of data to uncover" + }, + { + "start": 1885.2, + "text": "something interesting about the world." + }, + { + "start": 1887.24, + "text": "Um hopefully interesting. Um and it this" + }, + { + "start": 1890.68, + "text": "that that visual slam pipeline is" + }, + { + "start": 1892.16, + "text": "basically the opposite of game" + }, + { + "start": 1893.92, + "text": "rendering. And it involves somewhat" + }, + { + "start": 1895.56, + "text": "similar linear algebra, right? Um you go" + }, + { + "start": 1898.2, + "text": "from an image to easy to track key" + }, + { + "start": 1899.4, + "text": "points, pose estimate for the camera," + }, + { + "start": 1901.48, + "text": "some representation of the scene for the" + }, + { + "start": 1902.96, + "text": "robot to drive around in, right? Yeah," + }, + { + "start": 1905.28, + "text": "game rendering, you know, it's like" + }, + { + "start": 1906.44, + "text": "scene representation, triangles, pixels." + }, + { + "start": 1907.96, + "text": "I thought that was interesting." + }, + { + "start": 1909.48, + "text": "So, will LLMs replace knowledge work en" + }, + { + "start": 1911.28, + "text": "masse someday, right?" + }, + { + "start": 1912.96, + "text": "I put a steelman yes and a steelman no" + }, + { + "start": 1914.52, + "text": "here." + }, + { + "start": 1915.56, + "text": "Um really, I think it's probably" + }, + { + "start": 1917.4, + "text": "probably not. But," + }, + { + "start": 1919.6, + "text": "there's like a lot of investors who" + }, + { + "start": 1920.92, + "text": "disagree with me, so I don't know." + }, + { + "start": 1924.08, + "text": "new knowledge can't be made by machines," + }, + { + "start": 1925.76, + "text": "right? This is a very funny joke made by" + }, + { + "start": 1927.68, + "text": "a philosopher called David Deutsch. Um" + }, + { + "start": 1929.88, + "text": "maybe this is what we're going to have" + }, + { + "start": 1930.8, + "text": "to do someday. You're going to have to" + }, + { + "start": 1931.64, + "text": "invent new scientific knowledge to prove" + }, + { + "start": 1933.2, + "text": "that you're not a robot. Cuz maybe" + }, + { + "start": 1934.8, + "text": "that's the one thing the one thing that" + }, + { + "start": 1936.24, + "text": "humans can always do is create new" + }, + { + "start": 1937.72, + "text": "knowledge cuz they're the only ones that" + }, + { + "start": 1939.04, + "text": "can officially reason." + }, + { + "start": 1943.12, + "text": "But, if, you know, you have to have a" + }, + { + "start": 1945.16, + "text": "human know if you did new knowledge, I" + }, + { + "start": 1946.6, + "text": "don't know how a system would detect if" + }, + { + "start": 1947.88, + "text": "it's new knowledge or not. Maybe that's" + }, + { + "start": 1949.56, + "text": "like P and NP or something, I don't" + }, + { + "start": 1950.84, + "text": "know. Um" + }, + { + "start": 1952.68, + "text": "Wait until RKG I continues to be beaten," + }, + { + "start": 1954.48, + "text": "I think, is how you monitor this" + }, + { + "start": 1956.36, + "text": "problem." + }, + { + "start": 1957.4, + "text": "Um cuz if that continues to be beaten, I" + }, + { + "start": 1959.04, + "text": "think uh you know, that's crazy." + }, + { + "start": 1962.88, + "text": "Thanks for listening. Don't forget to" + }, + { + "start": 1963.76, + "text": "check out and work on github.com /ai." + }, + { + "start": 1967.88, + "text": "I want to make ai.c so that it's like" + }, + { + "start": 1969.68, + "text": "fun to mess around with these models and" + }, + { + "start": 1971.0, + "text": "techniques and games." + }, + { + "start": 1973.0, + "text": "No dependencies, no python, no nothing." + }, + { + "start": 1975.0, + "text": "I you know, I just it's like it's not" + }, + { + "start": 1977.24, + "text": "good right now. It needs to be rewritten" + }, + { + "start": 1978.56, + "text": "like 10 times. I've never made a deep" + }, + { + "start": 1980.12, + "text": "learning framework before, but" + }, + { + "start": 1982.16, + "text": "Thanks for listening to my talk." + }, + { + "start": 1992.28, + "text": "Sam." + }, + { + "start": 1997.8, + "text": "Okay, that's a bit similar to the psych" + }, + { + "start": 1999.48, + "text": "stuff." + }, + { + "start": 2000.48, + "text": "It's used by cookie people for AI, but" + }, + { + "start": 2002.36, + "text": "not serious contenders." + }, + { + "start": 2005.64, + "text": "Hello." + }, + { + "start": 2006.56, + "text": "What's up?" + }, + { + "start": 2007.12, + "text": ">> Good." + }, + { + "start": 2008.6, + "text": "Awesome. Can we get another round of" + }, + { + "start": 2010.48, + "text": "applause for Krakey?" + }, + { + "start": 2017.2, + "text": "Subscribe to my YouTube. Subscribe to" + }, + { + "start": 2018.96, + "text": "his YouTube. Watch" + }, + { + "start": 2021.32, + "text": "eye contact practice. It's the best one." + }, + { + "start": 2027.2, + "text": "I mean, that talk was definitely like it" + }, + { + "start": 2028.92, + "text": "was like 15% skits, so" + }, + { + "start": 2031.04, + "text": "but that's why it's entertaining for" + }, + { + "start": 2032.6, + "text": "sure." + }, + { + "start": 2034.16, + "text": "I think you're like part LLM." + }, + { + "start": 2037.0, + "text": "Like anyone that's talked to Krakey, you" + }, + { + "start": 2038.56, + "text": "could just give him a prompt and he just" + }, + { + "start": 2041.2, + "text": "[ __ ] spits out infinite like" + }, + { + "start": 2043.8, + "text": "sometimes he hallucinates. Yeah," + }, + { + "start": 2045.64, + "text": "sometimes it's garbage, but most of the" + }, + { + "start": 2047.83, + "text": "it's garbage, but most of the" + }, + { + "start": 2047.84, + "text": "time it's kind of right. Okay." + }, + { + "start": 2050.24, + "text": "That is what that's kind of what LLM's" + }, + { + "start": 2051.72, + "text": "do, yeah. Yeah, exactly. So you're it's" + }, + { + "start": 2053.879, + "text": "in there." + }, + { + "start": 2054.84, + "text": "It's interesting." + }, + { + "start": 2056.399, + "text": "So everyone here at Better Software" + }, + { + "start": 2058.84, + "text": "Conference" + }, + { + "start": 2060.44, + "text": "wants to write better software and wants" + }, + { + "start": 2062.399, + "text": "to live in a world where there is better" + }, + { + "start": 2064.56, + "text": "software." + }, + { + "start": 2066.6, + "text": "So will AI cause the destruction of" + }, + { + "start": 2069.72, + "text": "better software?" + }, + { + "start": 2078.6, + "text": "So, okay, one of the biggest like hot" + }, + { + "start": 2080.72, + "text": "topics for for AI startups right now is," + }, + { + "start": 2083.24, + "text": "well, if AI writes the programs, right," + }, + { + "start": 2084.879, + "text": "why do we even need an app store? Why do" + }, + { + "start": 2086.2, + "text": "we even need software engineers? You" + }, + { + "start": 2087.76, + "text": "should basically tell your computer the" + }, + { + "start": 2088.96, + "text": "app you want and then it does it and" + }, + { + "start": 2090.32, + "text": "then it gives you the interface and then" + }, + { + "start": 2091.36, + "text": "the data. Um" + }, + { + "start": 2093.48, + "text": "and it doesn't work at all right now and" + }, + { + "start": 2094.72, + "text": "so, I you know, it would We're going to" + }, + { + "start": 2096.8, + "text": "have to see some new developments, I" + }, + { + "start": 2098.32, + "text": "think, to judge if that will happen or" + }, + { + "start": 2100.36, + "text": "not." + }, + { + "start": 2101.24, + "text": "I'm basically 90% it won't cause the" + }, + { + "start": 2104.28, + "text": "destruction of better software for at" + }, + { + "start": 2105.64, + "text": "least a decade," + }, + { + "start": 2107.16, + "text": "but 10%, you know, there's always" + }, + { + "start": 2109.36, + "text": "something around the corner. Grok 4 is" + }, + { + "start": 2111.08, + "text": "like much better after some reason." + }, + { + "start": 2114.32, + "text": "So, in 10 years, like" + }, + { + "start": 2116.6, + "text": "is software engineering a thing? Like," + }, + { + "start": 2117.92, + "text": "when you walk into a room or people or" + }, + { + "start": 2120.2, + "text": "software engineers going to be crafting" + }, + { + "start": 2122.24, + "text": "prompts? So, I've been thinking about" + }, + { + "start": 2124.16, + "text": "this a lot. I mean, obviously, prompt" + }, + { + "start": 2125.64, + "text": "engineering is not going to be a real" + }, + { + "start": 2126.76, + "text": "thing because if you do believe these AI" + }, + { + "start": 2128.36, + "text": "systems will get to the point where they" + }, + { + "start": 2129.44, + "text": "can automate software engineering, they" + }, + { + "start": 2130.48, + "text": "can do the prompting, too. I don't think" + }, + { + "start": 2131.64, + "text": "it's that crazy." + }, + { + "start": 2132.92, + "text": "Um but in any case," + }, + { + "start": 2134.68, + "text": "I think that the effect of this on" + }, + { + "start": 2136.68, + "text": "software I Well, I I've I've been" + }, + { + "start": 2138.64, + "text": "thinking a lot about this question and" + }, + { + "start": 2140.2, + "text": "I'm I'm still very uncertain, but if" + }, + { + "start": 2142.24, + "text": "software engineers are basically making" + }, + { + "start": 2143.64, + "text": "these like virtual machines that do some" + }, + { + "start": 2145.4, + "text": "kind of processing or transformation" + }, + { + "start": 2147.2, + "text": "that's valuable to people and these LLMs" + }, + { + "start": 2149.68, + "text": "learned the patterns" + }, + { + "start": 2151.24, + "text": "uh to to to create some of these" + }, + { + "start": 2152.8, + "text": "machines as it relates to the to what" + }, + { + "start": 2154.72, + "text": "you're asking of it and I think people" + }, + { + "start": 2157.2, + "text": "who are making machines that are very" + }, + { + "start": 2159.28, + "text": "same-ish, software engineers who are" + }, + { + "start": 2161.88, + "text": "data transformations that are similar to" + }, + { + "start": 2163.48, + "text": "what exist are going to probably be" + }, + { + "start": 2165.64, + "text": "screwed. Um but if you're making unique" + }, + { + "start": 2168.4, + "text": "data transformations that are unlike" + }, + { + "start": 2169.92, + "text": "what have been made before, kind of" + }, + { + "start": 2170.96, + "text": "unique machines, I don't see how these" + }, + { + "start": 2172.88, + "text": "models could really create new knowledge" + }, + { + "start": 2174.68, + "text": "like that yet. I haven't seen evidence" + }, + { + "start": 2175.88, + "text": "that they can do that." + }, + { + "start": 2178.68, + "text": "Um you pointed out something uh" + }, + { + "start": 2181.24, + "text": "like learning is different from" + }, + { + "start": 2183.48, + "text": "optimization. And that's that's more of" + }, + { + "start": 2185.32, + "text": "a philosophical thing, I think." + }, + { + "start": 2187.0, + "text": ">> Like, when a human is born, right?" + }, + { + "start": 2188.44, + "text": "They're they're born like extremely" + }, + { + "start": 2189.84, + "text": "inefficient and they just get more" + }, + { + "start": 2191.0, + "text": "optimized as they grow. Is that not the" + }, + { + "start": 2192.88, + "text": "same? It's not I think it's not the same" + }, + { + "start": 2195.4, + "text": "because whatever human being is doing" + }, + { + "start": 2197.44, + "text": "can create human reasoning and I don't" + }, + { + "start": 2200.2, + "text": "think it's" + }, + { + "start": 2201.36, + "text": "I don't think it's it's it would be very" + }, + { + "start": 2203.2, + "text": "hard to make the claim that these things" + }, + { + "start": 2204.28, + "text": "are doing human reasoning right now, I" + }, + { + "start": 2205.44, + "text": "think, for a lot of reasons. Um you" + }, + { + "start": 2208.12, + "text": "know, oftentimes the reasoning doesn't" + }, + { + "start": 2209.24, + "text": "correspond to what they say. It's like" + }, + { + "start": 2210.56, + "text": "not logically coherent, right? It you" + }, + { + "start": 2213.44, + "text": "know," + }, + { + "start": 2214.24, + "text": "fundamentally from the I mean, you kind" + }, + { + "start": 2215.68, + "text": "of understand what's happening here," + }, + { + "start": 2217.08, + "text": "right? It's it's optimizing and and" + }, + { + "start": 2219.56, + "text": "creating an automatically created" + }, + { + "start": 2221.64, + "text": "program" + }, + { + "start": 2222.88, + "text": "um that might have nothing to do with" + }, + { + "start": 2224.48, + "text": "the the the universal causal" + }, + { + "start": 2226.32, + "text": "relationship between like, I don't know," + }, + { + "start": 2228.12, + "text": "you know, whatever's happening in your" + }, + { + "start": 2229.08, + "text": "brain or maybe it's if you're religious," + }, + { + "start": 2230.88, + "text": "then it's not even a question." + }, + { + "start": 2243.72, + "text": "Like isn't the real problem like not" + }, + { + "start": 2245.96, + "text": "that AI replaces software engineers and" + }, + { + "start": 2248.4, + "text": "that destroys better software, but that" + }, + { + "start": 2250.23, + "text": "destroys better software, but that" + }, + { + "start": 2250.24, + "text": "software engineers use AI and that" + }, + { + "start": 2252.08, + "text": "destroys better software because that's" + }, + { + "start": 2253.96, + "text": "what's happening pretty much already." + }, + { + "start": 2255.96, + "text": "Well, if you believe that society" + }, + { + "start": 2257.64, + "text": "functions, then people will choose the" + }, + { + "start": 2259.8, + "text": "products that are better." + }, + { + "start": 2261.56, + "text": "So," + }, + { + "start": 2262.52, + "text": "in the long run it, you know," + }, + { + "start": 2264.76, + "text": "if yeah, I don't know. I'm not sure." + }, + { + "start": 2267.52, + "text": ">> you pointed out uh" + }, + { + "start": 2270.04, + "text": "the like Grok had a massive jump" + }, + { + "start": 2272.12, + "text": "recently. Like did you look into that at" + }, + { + "start": 2273.56, + "text": "all?" + }, + { + "start": 2274.6, + "text": "On this specifically on the Arc AGI" + }, + { + "start": 2276.88, + "text": "test, which interestingly" + }, + { + "start": 2279.8, + "text": "there's another Arc. I don't know if" + }, + { + "start": 2280.92, + "text": "Taha point pointed that out. I missed" + }, + { + "start": 2282.8, + "text": "it." + }, + { + "start": 2283.6, + "text": "Everything is Arc." + }, + { + "start": 2285.68, + "text": "So, yeah, Grok might have performed" + }, + { + "start": 2288.32, + "text": "better because of its utter lack of" + }, + { + "start": 2289.76, + "text": "safety training. That is a rumor I kind" + }, + { + "start": 2291.56, + "text": "of saw and it seemed to be somewhat" + }, + { + "start": 2292.92, + "text": "confirmed by, you know, another times" + }, + { + "start": 2295.2, + "text": "it's like they did safety training and" + }, + { + "start": 2296.44, + "text": "the model got much less, you know," + }, + { + "start": 2298.76, + "text": "reasonable in quotes." + }, + { + "start": 2300.6, + "text": "Um it might be the case that the safety" + }, + { + "start": 2302.72, + "text": "training is not intelligent cuz it's so" + }, + { + "start": 2304.48, + "text": "directly refusing to to say something," + }, + { + "start": 2306.36, + "text": "right?" + }, + { + "start": 2307.36, + "text": "Um, there's all sorts of hypotheses and" + }, + { + "start": 2308.84, + "text": "there's no real, again, rigor in the" + }, + { + "start": 2310.6, + "text": "It's not really it's They're just trying" + }, + { + "start": 2312.2, + "text": "to figure it out. They don't Nobody" + }, + { + "start": 2313.44, + "text": "Nobody really knows." + }, + { + "start": 2314.52, + "text": ">> like broccolicing like some schizo like" + }, + { + "start": 2316.6, + "text": ">> It's not good. I don't want to repeat it" + }, + { + "start": 2318.23, + "text": "It's not good. I don't want to repeat it" + }, + { + "start": 2318.24, + "text": "out loud. You can Google it if you want." + }, + { + "start": 2319.56, + "text": "It's not good. It's It's" + }, + { + "start": 2321.48, + "text": "It's completely out of control, but it" + }, + { + "start": 2323.47, + "text": "completely out of control, but it" + }, + { + "start": 2323.48, + "text": "is smarter. So," + }, + { + "start": 2326.16, + "text": "well, I don't know." + }, + { + "start": 2328.72, + "text": "Okay. Um, you didn't point out either" + }, + { + "start": 2331.28, + "text": "like what caused the" + }, + { + "start": 2333.72, + "text": "this AI the last AI winter to end," + }, + { + "start": 2336.48, + "text": "right? Like what was the It specifically" + }, + { + "start": 2338.2, + "text": "happened So, transformers were made kind" + }, + { + "start": 2340.16, + "text": "of I think like 4 years before they were" + }, + { + "start": 2342.04, + "text": "really used to generate text in a very" + }, + { + "start": 2344.52, + "text": "in a very parallel large, you know, mega" + }, + { + "start": 2347.2, + "text": "data scale way. Um, and it happened when" + }, + { + "start": 2350.52, + "text": "OpenAI I think released GPT-2 and they" + }, + { + "start": 2352.56, + "text": "pointed out like, \"Hey, we keep just" + }, + { + "start": 2354.4, + "text": "making these models bigger and it seems" + }, + { + "start": 2355.96, + "text": "like the text is getting more and more" + }, + { + "start": 2357.16, + "text": "coherent. So, if you just complete the" + }, + { + "start": 2358.92, + "text": "line, like someday we get to the thing" + }, + { + "start": 2360.44, + "text": "that makes everybody, you know," + }, + { + "start": 2362.08, + "text": "Pop-Tarts for free or something.\" Like" + }, + { + "start": 2364.0, + "text": "And then I don't know. It's That's what" + }, + { + "start": 2365.68, + "text": "started this massive speculation bubble." + }, + { + "start": 2367.16, + "text": "It's like, \"Wait a second. We might have" + }, + { + "start": 2368.48, + "text": "a path to completely removing" + }, + { + "start": 2371.44, + "text": "uh humanity from what from satisfying" + }, + { + "start": 2374.12, + "text": "our our our desires in the physical" + }, + { + "start": 2375.88, + "text": "realm.\"" + }, + { + "start": 2377.0, + "text": "So, it was just they they just scaled it" + }, + { + "start": 2378.48, + "text": "up. Yeah, they scaled it up and it" + }, + { + "start": 2379.96, + "text": "seemed like the scaling wouldn't end." + }, + { + "start": 2382.72, + "text": "Is that what they're Is that what people" + }, + { + "start": 2384.0, + "text": "are still doing? They just keep scaling" + }, + { + "start": 2385.28, + "text": "things." + }, + { + "start": 2385.68, + "text": ">> buying GPUs. I mean, they're not scaling" + }, + { + "start": 2387.04, + "text": "in the same way. They're not just making" + }, + { + "start": 2388.24, + "text": "the models bigger. They're trying to" + }, + { + "start": 2389.24, + "text": "scale I the so-called reasoning models." + }, + { + "start": 2392.68, + "text": "Um, they're basically trying to make the" + }, + { + "start": 2394.08, + "text": "models more and more correct instead of" + }, + { + "start": 2395.64, + "text": "just making them remember more and more" + }, + { + "start": 2396.88, + "text": "facts, basically." + }, + { + "start": 2401.12, + "text": "I don't want to run out of time again," + }, + { + "start": 2402.24, + "text": "so I'm going to open" + }, + { + "start": 2403.56, + "text": "I'm going to do questions right now." + }, + { + "start": 2405.28, + "text": "What's up? Uh so, Cameron, a bunch of us" + }, + { + "start": 2407.84, + "text": "have seen your uh Dante demo just" + }, + { + "start": 2409.96, + "text": "floating around. Um, can you get into" + }, + { + "start": 2412.36, + "text": "why it's flawed and why Yeah, so my my" + }, + { + "start": 2416.04, + "text": "whole thinking with Dante's Inferno is I" + }, + { + "start": 2418.28, + "text": "think when I saw GPT-4 first like on the" + }, + { + "start": 2421.0, + "text": "API, I was like certain that this would" + }, + { + "start": 2423.36, + "text": "enable a huge new category of of video" + }, + { + "start": 2426.76, + "text": "Um and so I immediately began making" + }, + { + "start": 2428.12, + "text": "like an RPG where all the characters" + }, + { + "start": 2429.6, + "text": "were controlled by an LLM, not just" + }, + { + "start": 2431.4, + "text": "their speech but their actions, right?" + }, + { + "start": 2433.08, + "text": "And the idea was like, \"Whoa, this thing" + }, + { + "start": 2434.24, + "text": "is basically like prototype AGI or" + }, + { + "start": 2436.08, + "text": "whatever, right?\" So, it'll just kind of" + }, + { + "start": 2437.36, + "text": "operate and it'll be fun and like and I" + }, + { + "start": 2439.72, + "text": "went through so many iterations of that" + }, + { + "start": 2441.24, + "text": "idea. I made it so that like they could" + }, + { + "start": 2442.76, + "text": "join your party and trade items and all" + }, + { + "start": 2444.32, + "text": "this stuff and like I played with it a" + }, + { + "start": 2445.48, + "text": "bunch, but it never felt like there was" + }, + { + "start": 2447.2, + "text": "any meat to the game. The problem is" + }, + { + "start": 2449.44, + "text": "games are about predicting and" + }, + { + "start": 2451.0, + "text": "understanding systems almost, right? And" + }, + { + "start": 2452.84, + "text": "an LLM is an unpredictable black box," + }, + { + "start": 2455.44, + "text": "sometimes acts like a person, sometimes" + }, + { + "start": 2456.84, + "text": "doesn't. It's not fun to interact with" + }, + { + "start": 2458.36, + "text": "in a way where you're trying to get it" + }, + { + "start": 2459.36, + "text": "to do something. At best, you're" + }, + { + "start": 2460.88, + "text": "basically like, \"Hey trader, I'll" + }, + { + "start": 2462.8, + "text": "totally give you a million dollars if" + }, + { + "start": 2464.44, + "text": "you drop your item.\" And then it's like," + }, + { + "start": 2465.84, + "text": "\"This is not fun. This is dumb.\"" + }, + { + "start": 2467.96, + "text": "It's funny It's funny like for one time," + }, + { + "start": 2469.8, + "text": "but it's not like a game you could" + }, + { + "start": 2470.88, + "text": "master." + }, + { + "start": 2472.12, + "text": "It's like the vending machine at Claude." + }, + { + "start": 2474.28, + "text": "Okay, there's Right now, there's some AI" + }, + { + "start": 2476.16, + "text": "companies are trying to first deploy" + }, + { + "start": 2477.92, + "text": "these these LLMs in the real world to" + }, + { + "start": 2479.76, + "text": "start to," + }, + { + "start": 2480.8, + "text": "you know, create a utopian post-scarcity" + }, + { + "start": 2482.72, + "text": "world where machines make everything" + }, + { + "start": 2484.12, + "text": "automatically and they're like, \"Step" + }, + { + "start": 2485.64, + "text": "number one, put the LLM in a vending" + }, + { + "start": 2487.2, + "text": "machine and have it automatically order" + }, + { + "start": 2488.84, + "text": "and interface with the people or" + }, + { + "start": 2489.76, + "text": "whatever.\" And right now, people are" + }, + { + "start": 2491.28, + "text": "basically convincing these AI vending" + }, + { + "start": 2492.76, + "text": "machines to stock many tungsten cubes" + }, + { + "start": 2494.92, + "text": "and sell them at a loss and then they go" + }, + { + "start": 2496.04, + "text": "bankrupt." + }, + { + "start": 2497.36, + "text": "Um they've published many reports about" + }, + { + "start": 2498.92, + "text": "this and they're like, \"You know, we're" + }, + { + "start": 2500.0, + "text": "we're you know, we're doing some prompt" + }, + { + "start": 2501.12, + "text": "engineering to fix the business model" + }, + { + "start": 2502.44, + "text": "and we're trying to make it refuse" + }, + { + "start": 2503.6, + "text": "tungsten cubes more.\" Um but" + }, + { + "start": 2506.64, + "text": "really, I think there might be no end to" + }, + { + "start": 2507.92, + "text": "humanity's ingenuity as we battle" + }, + { + "start": 2509.48, + "text": "against the machines for the next few" + }, + { + "start": 2510.64, + "text": "decades. So." + }, + { + "start": 2513.12, + "text": "That's probably fine, right? It's" + }, + { + "start": 2515.6, + "text": "Everything's probably fine, but" + }, + { + "start": 2516.96, + "text": "everything's probably fine, but" + }, + { + "start": 2518.84, + "text": "sometimes bad things happen. I don't" + }, + { + "start": 2520.84, + "text": "know. What's up?" + }, + { + "start": 2523.12, + "text": "Do you think there So, I I feel like" + }, + { + "start": 2524.24, + "text": "there is a field of research, I just" + }, + { + "start": 2525.24, + "text": "don't know the words for it or how big" + }, + { + "start": 2526.44, + "text": "it is, but do you think there's a way" + }, + { + "start": 2527.52, + "text": "that we can reverse engineer Yeah," + }, + { + "start": 2529.16, + "text": "interpretability is big. A lot of Yeah," + }, + { + "start": 2531.24, + "text": "Anthropic is the leader in" + }, + { + "start": 2532.64, + "text": "interpretability, I think. Um they" + }, + { + "start": 2534.24, + "text": "basically do data science on the neurons" + }, + { + "start": 2535.92, + "text": "themselves, and they try to make claims" + }, + { + "start": 2537.56, + "text": "about, you know, how the data is flowing" + }, + { + "start": 2539.0, + "text": "through the model. Um you can look at" + }, + { + "start": 2540.72, + "text": "their Golden Gate model. They have a lot" + }, + { + "start": 2541.88, + "text": "of really excellent stuff you can read" + }, + { + "start": 2543.6, + "text": "forever. It's really cool. Um Do you" + }, + { + "start": 2545.8, + "text": "know much about it? Like, do you feel" + }, + { + "start": 2547.04, + "text": "like that is the way forward, or is that" + }, + { + "start": 2548.64, + "text": "just a byproduct of the craziness? I" + }, + { + "start": 2550.48, + "text": "don't think there will be any value" + }, + { + "start": 2551.92, + "text": "created from interpretability research," + }, + { + "start": 2553.72, + "text": "honestly, other than maybe um" + }, + { + "start": 2557.08, + "text": "optimizing the training process somehow." + }, + { + "start": 2559.16, + "text": "I'm honestly not really sure. I think" + }, + { + "start": 2560.4, + "text": "they're mostly doing it because they" + }, + { + "start": 2561.92, + "text": "have to, because they want to be the" + }, + { + "start": 2563.32, + "text": "safe people or something." + }, + { + "start": 2564.96, + "text": "Um and it's cool, but it doesn't seem" + }, + { + "start": 2566.92, + "text": "valuable to me, but I I'm not like a big" + }, + { + "start": 2569.4, + "text": "deep learning research. I'm like a" + }, + { + "start": 2570.4, + "text": "hobbyist, like I said earlier, so I'm" + }, + { + "start": 2571.68, + "text": "not really sure." + }, + { + "start": 2573.0, + "text": "Uh follow-up to that, is it" + }, + { + "start": 2574.68, + "text": "interpretability" + }, + { + "start": 2577.04, + "text": "that made people figure out you could" + }, + { + "start": 2578.92, + "text": "remove the safety neuron? You can. They" + }, + { + "start": 2581.2, + "text": "already figured that out. It's called" + }, + { + "start": 2582.08, + "text": "ablation. So, any single open-source" + }, + { + "start": 2584.16, + "text": "model, you can basically find what's" + }, + { + "start": 2585.32, + "text": "called an ablated model, where they just" + }, + { + "start": 2586.76, + "text": "basically subtract the safety training," + }, + { + "start": 2588.96, + "text": "and then it does whatever. But, it's" + }, + { + "start": 2590.56, + "text": "But, you told me it's like one node" + }, + { + "start": 2592.76, + "text": "that's like this. Yeah, they I don't" + }, + { + "start": 2594.36, + "text": "remember exactly how it goes, but it's" + }, + { + "start": 2596.36, + "text": "not a node, but it's like one vector" + }, + { + "start": 2598.04, + "text": "kind of represents the safety," + }, + { + "start": 2600.2, + "text": "uh the refusal. Like, when whenever an" + }, + { + "start": 2601.8, + "text": "LLM starts to refuse, it always seems to" + }, + { + "start": 2603.8, + "text": "output words in a very, you know, this" + }, + { + "start": 2605.92, + "text": "kind of sounds like the same thing, and" + }, + { + "start": 2607.56, + "text": "it basically you can find that inside of" + }, + { + "start": 2609.52, + "text": "the model and then nullify it is what" + }, + { + "start": 2611.56, + "text": "ablation is, from what I remember. It's" + }, + { + "start": 2613.08, + "text": "been a while." + }, + { + "start": 2618.12, + "text": "Uh as you said that AI is basically" + }, + { + "start": 2620.6, + "text": "metaprogramming. Yes. So, and Julia is" + }, + { + "start": 2624.08, + "text": "very famous for its metaprogramming." + }, + { + "start": 2626.04, + "text": "Will it replace Python? Like, will we" + }, + { + "start": 2627.96, + "text": "have Julia AI or like So, maybe" + }, + { + "start": 2631.0, + "text": "metaprogramming is a bad thing, cuz it's" + }, + { + "start": 2633.44, + "text": "not really comparable. I was moon I was" + }, + { + "start": 2635.56, + "text": "mostly It's like metaprogramming only to" + }, + { + "start": 2638.24, + "text": "the extent that you're producing a" + }, + { + "start": 2640.04, + "text": "program with a program, but it's not" + }, + { + "start": 2642.16, + "text": "like it's automatic metaprogramming." + }, + { + "start": 2644.48, + "text": "It's not, you know, mechanistic symbolic" + }, + { + "start": 2646.64, + "text": "metaprogramming, where you're yeah. So," + }, + { + "start": 2648.92, + "text": "it would not be a good application for" + }, + { + "start": 2650.2, + "text": "this language, next one. Huh?" + }, + { + "start": 2652.68, + "text": "It's not a good application for jail." + }, + { + "start": 2655.0, + "text": ">> No. No." + }, + { + "start": 2656.96, + "text": "No." + }, + { + "start": 2658.52, + "text": "Uh" + }, + { + "start": 2660.72, + "text": "rough To bring more into the concrete" + }, + { + "start": 2662.8, + "text": "realm, you mentioned SLAM. Yes. It's" + }, + { + "start": 2664.8, + "text": "been a decade since I've looked at SLAM." + }, + { + "start": 2666.4, + "text": "Like, is deep learning the current state" + }, + { + "start": 2667.96, + "text": "of the achieving SLAM algorithms, or are" + }, + { + "start": 2670.68, + "text": "they still" + }, + { + "start": 2671.16, + "text": ">> I I just recently asked a SLAM" + }, + { + "start": 2672.72, + "text": "researcher that, and they basically said" + }, + { + "start": 2674.44, + "text": "deep learning is like used sometimes," + }, + { + "start": 2676.16, + "text": "but it's not really necessary to solve" + }, + { + "start": 2677.52, + "text": "the problem though. Um" + }, + { + "start": 2680.0, + "text": "you know, they found good heuristics to" + }, + { + "start": 2681.32, + "text": "find the key points that are trackable." + }, + { + "start": 2683.28, + "text": "Um and like neural I think that neural" + }, + { + "start": 2684.84, + "text": "SLAM seems like it's something they've" + }, + { + "start": 2686.48, + "text": "kind of been like rumor you know, to" + }, + { + "start": 2688.04, + "text": "kind of take over a state of the art," + }, + { + "start": 2689.11, + "text": "take over a state of the art," + }, + { + "start": 2689.12, + "text": "but like you know, the Quest doesn't do" + }, + { + "start": 2690.76, + "text": "any neural SLAM as far as I know." + }, + { + "start": 2695.4, + "text": "You mentioned using uh AI models for" + }, + { + "start": 2698.92, + "text": "balancing games. Yes." + }, + { + "start": 2701.2, + "text": ">> and" + }, + { + "start": 2702.72, + "text": "let's say you then train a model to play" + }, + { + "start": 2704.52, + "text": "StarCraft II, and you you finally do" + }, + { + "start": 2705.92, + "text": "definitively prove that Protoss should" + }, + { + "start": 2707.8, + "text": "be nerfed." + }, + { + "start": 2709.96, + "text": "yes." + }, + { + "start": 2712.12, + "text": "God, go ahead and start recording." + }, + { + "start": 2714.16, + "text": "Uh once you nerf Protoss, do you then go" + }, + { + "start": 2716.84, + "text": "like, \"Okay, cool. Now we do a $3" + }, + { + "start": 2718.56, + "text": "million run again?\" I don't think it" + }, + { + "start": 2721.16, + "text": "would be that expensive to do it on a" + }, + { + "start": 2723.16, + "text": "game. It's the cost of AI right now is" + }, + { + "start": 2725.96, + "text": "because to make these large language" + }, + { + "start": 2728.04, + "text": "models learn, you know, reasoning, you" + }, + { + "start": 2730.96, + "text": "have to basically shuffle the entire" + }, + { + "start": 2732.84, + "text": "internet in and out of um multiple" + }, + { + "start": 2735.12, + "text": "accumulate ads. It's really not so much" + }, + { + "start": 2736.6, + "text": "data they have to process. For something" + }, + { + "start": 2738.2, + "text": "like StarCraft II or AlphaZero, it's" + }, + { + "start": 2739.56, + "text": "really not on that scale of cost. Um I" + }, + { + "start": 2742.4, + "text": "could easily see many many many games" + }, + { + "start": 2744.12, + "text": "being able to do this, and for simpler" + }, + { + "start": 2746.28, + "text": "games, you could probably do it on your" + }, + { + "start": 2747.16, + "text": "laptop even maybe. Um" + }, + { + "start": 2749.6, + "text": "But yeah, like I think that for" + }, + { + "start": 2751.44, + "text": "something like StarCraft II, it's like I" + }, + { + "start": 2752.8, + "text": "could see them doing these kind of" + }, + { + "start": 2754.12, + "text": "studies if they have the the cap backs" + }, + { + "start": 2755.84, + "text": "and they believe it's" + }, + { + "start": 2757.04, + "text": "you know, it's it's valuable for the" + }, + { + "start": 2758.56, + "text": "players to have have truly be officially" + }, + { + "start": 2761.12, + "text": "like even the genius master AI" + }, + { + "start": 2764.48, + "text": "uh players of this game it's" + }, + { + "start": 2766.52, + "text": "even right that might be a good thing to" + }, + { + "start": 2767.88, + "text": "prove." + }, + { + "start": 2770.04, + "text": "Uh yeah. Uh would you say that like" + }, + { + "start": 2772.2, + "text": "training is where like mostly cost is or" + }, + { + "start": 2775.44, + "text": "Um actually well opening I says that" + }, + { + "start": 2778.32, + "text": "inference is their biggest cost right" + }, + { + "start": 2779.76, + "text": "now" + }, + { + "start": 2780.8, + "text": "but there's a massive push to make you" + }, + { + "start": 2783.64, + "text": "know on on device inference chips so I" + }, + { + "start": 2786.6, + "text": "think that's kind of temporary and it's" + }, + { + "start": 2788.28, + "text": "just because like consumer my" + }, + { + "start": 2789.84, + "text": "understanding is I don't know if this is" + }, + { + "start": 2791.72, + "text": "exactly true but I've seen this floating" + }, + { + "start": 2792.96, + "text": "around that Nvidia cards don't have" + }, + { + "start": 2794.96, + "text": "enough VRM to run these models simply" + }, + { + "start": 2796.48, + "text": "because it's not something consumers" + }, + { + "start": 2797.84, + "text": "wanted and they kind of want to nickel" + }, + { + "start": 2798.92, + "text": "and dime you for VRAM" + }, + { + "start": 2800.88, + "text": "and so this seems like the kind of thing" + }, + { + "start": 2802.12, + "text": "where you know if it's a healthy market" + }, + { + "start": 2803.76, + "text": "and you know it's not there's not like" + }, + { + "start": 2804.76, + "text": "an Nvidia monopoly or whatever" + }, + { + "start": 2807.24, + "text": "which there is I'm joking it's not going" + }, + { + "start": 2808.56, + "text": "to get better but" + }, + { + "start": 2810.08, + "text": "um" + }, + { + "start": 2812.04, + "text": "ideally I think it's totally feasible" + }, + { + "start": 2813.56, + "text": "that these the inference cost is zero" + }, + { + "start": 2815.04, + "text": "for the people who are" + }, + { + "start": 2816.56, + "text": "making these models it it should be" + }, + { + "start": 2817.92, + "text": "zero." + }, + { + "start": 2819.48, + "text": "Uh yeah. Yeah in your little experiment" + }, + { + "start": 2822.0, + "text": "game how did you translate basically the" + }, + { + "start": 2824.4, + "text": "LM output to actions in your world? So" + }, + { + "start": 2827.0, + "text": "basically you can view putting LMs in a" + }, + { + "start": 2829.04, + "text": "video game as so you have your game" + }, + { + "start": 2830.88, + "text": "state and then there's like the um" + }, + { + "start": 2834.32, + "text": "physics" + }, + { + "start": 2835.4, + "text": "simulation kind of like you know human" + }, + { + "start": 2837.96, + "text": "in the loop like you're inputting" + }, + { + "start": 2840.04, + "text": "actions and state 60 hertz or whatever" + }, + { + "start": 2842.8, + "text": "and there's a text adventure way to" + }, + { + "start": 2844.56, + "text": "interact with the game simultaneously" + }, + { + "start": 2846.04, + "text": "right and a good way to check that like" + }, + { + "start": 2848.04, + "text": "the game was properly you know" + }, + { + "start": 2850.12, + "text": "presenting itself to LMs is if you could" + }, + { + "start": 2851.64, + "text": "flip it and as a human I could play the" + }, + { + "start": 2854.0, + "text": "text adventure version of it right" + }, + { + "start": 2855.56, + "text": "that's how you would know that it's" + }, + { + "start": 2856.36, + "text": "being represented well in text and" + }, + { + "start": 2857.64, + "text": "that's basically what I did is like a" + }, + { + "start": 2858.8, + "text": "structured text adventure presented to" + }, + { + "start": 2860.56, + "text": "the LMs um" + }, + { + "start": 2862.28, + "text": "describing what's around it describing" + }, + { + "start": 2863.4, + "text": "what you're wearing whatever yeah." + }, + { + "start": 2865.92, + "text": "Yeah. So right now when I use some" + }, + { + "start": 2868.52, + "text": "programs I go oh I want to do kind of" + }, + { + "start": 2871.28, + "text": "this thing and it used to be that I was" + }, + { + "start": 2873.4, + "text": "able to Google how do then X in program" + }, + { + "start": 2876.36, + "text": "Y" + }, + { + "start": 2877.4, + "text": "and some of it will be like a forum" + }, + { + "start": 2879.44, + "text": "post. But now you can't find those" + }, + { + "start": 2882.0, + "text": "anymore or I'm doing difficult stuff. So" + }, + { + "start": 2884.8, + "text": "my question is" + }, + { + "start": 2886.28, + "text": "what year will I be able to put like a" + }, + { + "start": 2888.88, + "text": "good version of Clippy in my programs?" + }, + { + "start": 2892.24, + "text": "So like there's a little dude you can" + }, + { + "start": 2894.04, + "text": "actually describe like hey I'm trying to" + }, + { + "start": 2895.8, + "text": "rotate the mesh in this way what" + }, + { + "start": 2898.76, + "text": "like how does the sub stuff know what?" + }, + { + "start": 2900.72, + "text": "There's so many startups trying to do" + }, + { + "start": 2902.04, + "text": "that. It's Yeah, it's like cursor," + }, + { + "start": 2904.32, + "text": "No, well it's not even cursor for X." + }, + { + "start": 2905.88, + "text": "It's like there's so many people who put" + }, + { + "start": 2907.68, + "text": "the little chat window on the thing and" + }, + { + "start": 2908.88, + "text": "it's like ask it a question. I just feel" + }, + { + "start": 2910.08, + "text": "like they never work. Like they still" + }, + { + "start": 2911.64, + "text": "are just so bad most of the time. Like I" + }, + { + "start": 2913.84, + "text": "don't know when it will get good, but" + }, + { + "start": 2915.84, + "text": "honestly I like" + }, + { + "start": 2917.76, + "text": "I find that when I use LLMs to" + }, + { + "start": 2920.12, + "text": "understand how software works, it gives" + }, + { + "start": 2922.84, + "text": "me an incomplete understanding because" + }, + { + "start": 2925.32, + "text": "I'm not really like traversing through" + }, + { + "start": 2926.88, + "text": "the whole set of knowledge, right? I'm" + }, + { + "start": 2928.56, + "text": "just specifically getting the answer I" + }, + { + "start": 2930.32, + "text": "need and implementing it and moving on" + }, + { + "start": 2931.68, + "text": "with my life. Whereas when I would when" + }, + { + "start": 2933.32, + "text": "I read documentation and like actually" + }, + { + "start": 2935.12, + "text": "go line by line and sit down and" + }, + { + "start": 2936.32, + "text": "understand it, I basically I just know" + }, + { + "start": 2939.04, + "text": "the unknown unknowns, right? Because I" + }, + { + "start": 2941.44, + "text": "explored. And so LLMs to say how do I do" + }, + { + "start": 2944.8, + "text": "this in this program? I I'm actually not" + }, + { + "start": 2946.24, + "text": "even sure that's good UX, basically. But" + }, + { + "start": 2948.16, + "text": "if if you can like search the" + }, + { + "start": 2949.32, + "text": "documentation for So like you need to" + }, + { + "start": 2952.16, + "text": "take it put the documentation is like a" + }, + { + "start": 2953.96, + "text": "hundred page PDF documentation of this" + }, + { + "start": 2956.2, + "text": "program. I'm not going to read that." + }, + { + "start": 2957.72, + "text": "Yeah, it is good for that. Like LLMs are" + }, + { + "start": 2960.72, + "text": "almost like a very it's like a better" + }, + { + "start": 2963.36, + "text": "Google almost." + }, + { + "start": 2965.28, + "text": "Which is why they pivoted their entire" + }, + { + "start": 2966.4, + "text": "company to LLMs. I might be able to So" + }, + { + "start": 2969.04, + "text": "at what year can I expect to like put" + }, + { + "start": 2972.0, + "text": "one of these in my C program?" + }, + { + "start": 2974.32, + "text": "Where it's like running locally? Yeah," + }, + { + "start": 2976.24, + "text": "AI.C. Yeah, cuz I want to Yes, my AI.C" + }, + { + "start": 2978.52, + "text": "project. To get the guy in who like" + }, + { + "start": 2980.88, + "text": "sometimes complains Um" + }, + { + "start": 2982.76, + "text": "Let's say let's say you know smaller" + }, + { + "start": 2984.44, + "text": "models are always getting better and" + }, + { + "start": 2985.52, + "text": "better. Let's let's put a let's put a" + }, + { + "start": 2987.48, + "text": "release date on it and say five years." + }, + { + "start": 2992.44, + "text": "So yeah, so have you seen like a model" + }, + { + "start": 2995.52, + "text": "that sort of can work outside of the" + }, + { + "start": 2998.0, + "text": "training data?" + }, + { + "start": 2999.56, + "text": "It feels like they're always like" + }, + { + "start": 3000.88, + "text": "weighted averages of the training data." + }, + { + "start": 3003.08, + "text": "It's It's not a weighted average of the" + }, + { + "start": 3004.6, + "text": "training data. A better way to describe" + }, + { + "start": 3006.48, + "text": "how they process and hold and how our" + }, + { + "start": 3008.96, + "text": "models impacted by training data is" + }, + { + "start": 3011.48, + "text": "they find patterns and then we run tests" + }, + { + "start": 3014.44, + "text": "on those patterns to make sure they" + }, + { + "start": 3015.84, + "text": "generalize and they're not specific to" + }, + { + "start": 3017.48, + "text": "the training data. So, they find" + }, + { + "start": 3018.88, + "text": "patterns which are general that do the" + }, + { + "start": 3021.16, + "text": "thing you want it to do. ARC AGI" + }, + { + "start": 3022.68, + "text": "specifically is a benchmark for generic" + }, + { + "start": 3025.2, + "text": "never-before-seen on the internet" + }, + { + "start": 3026.64, + "text": "patterns that fail LLMs, all of them." + }, + { + "start": 3029.36, + "text": "They're private. They're never leaked or" + }, + { + "start": 3030.64, + "text": "exposed. You can't ever You can't build" + }, + { + "start": 3033.52, + "text": "an algorithm that interpolates from" + }, + { + "start": 3034.64, + "text": "existing solutions to solve ARC AGI" + }, + { + "start": 3036.24, + "text": "ever. And LLMs did that." + }, + { + "start": 3038.6, + "text": "Yeah, so so yeah, I was asking sort of" + }, + { + "start": 3040.68, + "text": "about the answers you get because it" + }, + { + "start": 3043.16, + "text": "feels like you get sort of you know, you" + }, + { + "start": 3045.4, + "text": "cannot go outside of what it has been" + }, + { + "start": 3047.16, + "text": "taught, right? Or sort of" + }, + { + "start": 3049.2, + "text": "optimized on," + }, + { + "start": 3050.6, + "text": "but maybe you could sort of extrapolate" + }, + { + "start": 3054.0, + "text": "some, you know, outside of it. But is" + }, + { + "start": 3056.92, + "text": "that happening like Yeah, I I Well, I" + }, + { + "start": 3059.12, + "text": "think that like when you get to these" + }, + { + "start": 3060.76, + "text": "absurd levels of scale, um" + }, + { + "start": 3064.68, + "text": "there's almost nothing outside the box" + }, + { + "start": 3066.28, + "text": "if it's predicting the entire internet." + }, + { + "start": 3068.64, + "text": "Um that's basically like a" + }, + { + "start": 3070.84, + "text": "a huge amount of examples of uh the" + }, + { + "start": 3073.96, + "text": "dynamics of human intelligence recorded" + }, + { + "start": 3077.32, + "text": "in into written form that you then, you" + }, + { + "start": 3079.2, + "text": "know, process and find patterns which uh" + }, + { + "start": 3081.96, + "text": "apply across many examples. That's" + }, + { + "start": 3083.88, + "text": "That's what they're doing. They" + }, + { + "start": 3085.72, + "text": "Deep learning systems usually start with" + }, + { + "start": 3087.04, + "text": "very low-frequency generic patterns" + }, + { + "start": 3089.2, + "text": "like, \"Oh, there's a period every" + }, + { + "start": 3090.56, + "text": "so-and-so words.\" And then they start to" + }, + { + "start": 3091.72, + "text": "get to higher-frequency patterns and" + }, + { + "start": 3092.84, + "text": "more specific, more nuanced things, and" + }, + { + "start": 3094.64, + "text": "then" + }, + { + "start": 3095.52, + "text": "they they they're they're basically" + }, + { + "start": 3096.92, + "text": "pattern finders. Patterns which apply in" + }, + { + "start": 3100.0, + "text": "on not just the training data. That's" + }, + { + "start": 3101.28, + "text": "the whole point of being like a deep" + }, + { + "start": 3102.88, + "text": "learning person is you make sure that" + }, + { + "start": 3104.88, + "text": "the model does not just apply on the" + }, + { + "start": 3106.32, + "text": "training data. That's like the point of" + }, + { + "start": 3107.84, + "text": "all the architectures, that's the point" + }, + { + "start": 3109.16, + "text": "of like, you know, doing everything" + }, + { + "start": 3110.68, + "text": "you're doing is making sure that it's" + }, + { + "start": 3111.76, + "text": "not just memorizing the training data," + }, + { + "start": 3113.08, + "text": "that it's finding patterns in the" + }, + { + "start": 3114.48, + "text": "training data that will apply in your" + }, + { + "start": 3116.04, + "text": "specific task that you need the model" + }, + { + "start": 3117.24, + "text": "for." + }, + { + "start": 3118.44, + "text": "So, so sort of like say that you" + }, + { + "start": 3121.04, + "text": "recognize images," + }, + { + "start": 3123.16, + "text": "like you can recognize cat and dog." + }, + { + "start": 3125.4, + "text": "Yeah. But like a model that can do train" + }, + { + "start": 3128.0, + "text": "on cat and dog images, but also like" + }, + { + "start": 3130.52, + "text": "this is something else. Like without you" + }, + { + "start": 3133.72, + "text": "showing something else." + }, + { + "start": 3135.32, + "text": "Yes, I mean obviously that model is not" + }, + { + "start": 3137.4, + "text": "um learned patterns which would" + }, + { + "start": 3138.96, + "text": "generalize to is it a turtle or not. It" + }, + { + "start": 3140.6, + "text": "would try to bucket into cat or dog, cuz" + }, + { + "start": 3142.52, + "text": "that's what that," + }, + { + "start": 3143.72, + "text": "you know, system has done, but you can" + }, + { + "start": 3145.96, + "text": "feed any image into an LM right now and" + }, + { + "start": 3148.24, + "text": "it'll tell you what it is. Almost like" + }, + { + "start": 3149.8, + "text": "exactly like what a person would say." + }, + { + "start": 3151.76, + "text": "Obviously it's not perfect and it" + }, + { + "start": 3152.96, + "text": "constantly fails, um" + }, + { + "start": 3155.08, + "text": "but that error is always decreasing as" + }, + { + "start": 3156.96, + "text": "they, you know," + }, + { + "start": 3158.52, + "text": "fix bugs, they they improve the training" + }, + { + "start": 3160.4, + "text": "pipeline, they improve the data they" + }, + { + "start": 3162.12, + "text": "give it, and then it's always" + }, + { + "start": 3163.56, + "text": "approaching something that seems like it" + }, + { + "start": 3165.96, + "text": "will apply in any situation. Um of" + }, + { + "start": 3168.96, + "text": "course it's very far from that right" + }, + { + "start": 3170.04, + "text": "now. I'm not meaning to claim that like" + }, + { + "start": 3171.92, + "text": "you could take any image and then it'll" + }, + { + "start": 3172.96, + "text": "perfectly recognize what's in it. You" + }, + { + "start": 3174.12, + "text": "can fool them all the time. Um it's just" + }, + { + "start": 3176.72, + "text": "that like" + }, + { + "start": 3178.84, + "text": "it's the entire internet, right? I just" + }, + { + "start": 3180.64, + "text": "that's the thing that I keep thinking" + }, + { + "start": 3181.96, + "text": "about. You know, it's that's just a lot." + }, + { + "start": 3184.0, + "text": "It's hard for me to even comprehend what" + }, + { + "start": 3185.16, + "text": "the limits of that is." + }, + { + "start": 3188.6, + "text": "So, even if we don't end up in a world" + }, + { + "start": 3190.0, + "text": "where the AI is able to do our job in" + }, + { + "start": 3191.72, + "text": "in, you know, 10 years or less," + }, + { + "start": 3192.92, + "text": "whatever. Um how likely do you think it" + }, + { + "start": 3194.64, + "text": "is that we're heading towards a dead" + }, + { + "start": 3196.08, + "text": "internet style problem where the noise" + }, + { + "start": 3198.08, + "text": "generated or the noise created by these" + }, + { + "start": 3199.8, + "text": "generated AIs is a problem to our own" + }, + { + "start": 3201.99, + "text": "AIs is a problem to our own" + }, + { + "start": 3202.0, + "text": "ability to communicate? I mean really" + }, + { + "start": 3203.84, + "text": "already it feels like so much of the" + }, + { + "start": 3205.08, + "text": "internet is on like private discords" + }, + { + "start": 3207.12, + "text": "that aren't, you know, on these massive" + }, + { + "start": 3209.12, + "text": "public forums and," + }, + { + "start": 3210.8, + "text": "you know, linked with each other. I" + }, + { + "start": 3212.08, + "text": "think that people are already kind of" + }, + { + "start": 3213.68, + "text": "receding away from this like how it used" + }, + { + "start": 3215.8, + "text": "to be where it was like a like a massive" + }, + { + "start": 3218.2, + "text": "graph of forum posts or whatever, right?" + }, + { + "start": 3220.0, + "text": "Like um" + }, + { + "start": 3222.2, + "text": "I don't know much about that. I mean" + }, + { + "start": 3223.08, + "text": "this this is the internet and it's not" + }, + { + "start": 3224.6, + "text": "dead, I guess. So, I it's not I don't" + }, + { + "start": 3226.2, + "text": "know like I mean maybe." + }, + { + "start": 3229.08, + "text": "People I actually Well, so LLMs do seem" + }, + { + "start": 3231.68, + "text": "like they kind of defeat the Turing test" + }, + { + "start": 3233.8, + "text": "and I I seem to I think I'm remembering" + }, + { + "start": 3235.88, + "text": "there was some news article. Somebody" + }, + { + "start": 3237.36, + "text": "maybe in like Twitch chat or whatever" + }, + { + "start": 3238.56, + "text": "can find it where I think some lab" + }, + { + "start": 3241.08, + "text": "basically plugged an LLM into like a" + }, + { + "start": 3242.68, + "text": "subreddit and just had it be a commenter" + }, + { + "start": 3244.8, + "text": "and it got a lot of karma and people" + }, + { + "start": 3245.96, + "text": "started to like relate to it and like" + }, + { + "start": 3247.92, + "text": "oh, this is a good post or he's a good" + }, + { + "start": 3249.0, + "text": "guy. I like this guy. And then it was" + }, + { + "start": 3250.88, + "text": "like uncovered that it was really like a" + }, + { + "start": 3252.44, + "text": "like a false flag human being or" + }, + { + "start": 3253.75, + "text": "a false flag human being or" + }, + { + "start": 3253.76, + "text": "whatever and like they were they they" + }, + { + "start": 3255.68, + "text": "basically were doing a study on the" + }, + { + "start": 3257.4, + "text": "subreddit involuntarily. I think people" + }, + { + "start": 3259.24, + "text": "got very mad at that cuz it was like" + }, + { + "start": 3260.96, + "text": "that's probably not ethical to do human" + }, + { + "start": 3262.28, + "text": "beings. Most of what's happening in the" + }, + { + "start": 3263.88, + "text": "modern world is not ethical to do human" + }, + { + "start": 3265.24, + "text": "beings, honestly. I I don't think I" + }, + { + "start": 3267.2, + "text": "don't think LLMs are that crazy compared" + }, + { + "start": 3268.48, + "text": "to like YouTube. Like you just show" + }, + { + "start": 3270.88, + "text": "infant pixels to a person like this is" + }, + { + "start": 3272.36, + "text": "probably not good for them. I don't know" + }, + { + "start": 3274.4, + "text": "Um I don't really you know, things are" + }, + { + "start": 3276.76, + "text": "already so so so far from what is like" + }, + { + "start": 3279.36, + "text": "healthy for a human being. It's like" + }, + { + "start": 3281.8, + "text": "What are you going to What do you want?" + }, + { + "start": 3283.0, + "text": "Like you just like it's not that much" + }, + { + "start": 3285.4, + "text": "worse than TikTok. I don't know what to" + }, + { + "start": 3287.4, + "text": "say. TikTok is a transformer. It's" + }, + { + "start": 3293.12, + "text": "So, we talked about this over breakfast" + }, + { + "start": 3294.68, + "text": "yesterday, but like LLMs kind of cool," + }, + { + "start": 3297.2, + "text": "transformers are cool, doing something," + }, + { + "start": 3299.12, + "text": "but like there's a missing piece or" + }, + { + "start": 3300.88, + "text": "maybe many. Do you have any theories" + }, + { + "start": 3302.48, + "text": "what those might be?" + }, + { + "start": 3303.84, + "text": "Me personally as a hobbyist" + }, + { + "start": 3306.36, + "text": "crazy person? Uh" + }, + { + "start": 3307.96, + "text": "I don't know. I feel like if I were to" + }, + { + "start": 3310.64, + "text": "Everybody so every every single deep" + }, + { + "start": 3312.36, + "text": "learning research is like oh, it's going" + }, + { + "start": 3313.24, + "text": "to be in a reinforcement learning. Like" + }, + { + "start": 3314.68, + "text": "you take that success of AlphaZero where" + }, + { + "start": 3316.2, + "text": "it basically can master any system. You" + }, + { + "start": 3317.72, + "text": "just like apply that to like soft" + }, + { + "start": 3318.96, + "text": "concepts. Like MuZero it's like you" + }, + { + "start": 3321.36, + "text": "know, I everything everybody is like" + }, + { + "start": 3323.04, + "text": "thinking along those lines of like we" + }, + { + "start": 3324.4, + "text": "basically need computers to" + }, + { + "start": 3325.76, + "text": "automatically generate and master uh" + }, + { + "start": 3327.96, + "text": "things that not even humans have" + }, + { + "start": 3328.96, + "text": "mastered yet, right? Because if all" + }, + { + "start": 3330.32, + "text": "you're doing is basically making these" + }, + { + "start": 3332.48, + "text": "models uh try to optimize and find" + }, + { + "start": 3334.76, + "text": "patterns like that are more and more" + }, + { + "start": 3336.24, + "text": "closely match human beings, they'll" + }, + { + "start": 3337.76, + "text": "never exceed the human baseline, right?" + }, + { + "start": 3339.28, + "text": "I mean, they might be an ensemble of all" + }, + { + "start": 3341.08, + "text": "the best humans and, you know, they" + }, + { + "start": 3342.48, + "text": "might have be a genius in every single" + }, + { + "start": 3343.68, + "text": "field or whatever, but they're never" + }, + { + "start": 3344.76, + "text": "going to surpass human genius, which you" + }, + { + "start": 3346.88, + "text": "would kind of want if you wanted to be a" + }, + { + "start": 3348.16, + "text": "god-like being that replaces mankind," + }, + { + "start": 3350.52, + "text": "which is some people's goal, not mine. I" + }, + { + "start": 3352.12, + "text": "don't I'm not into that, but some people" + }, + { + "start": 3353.44, + "text": "like that. I don't know. I think it's" + }, + { + "start": 3354.6, + "text": "probably bad." + }, + { + "start": 3356.88, + "text": "But like like" + }, + { + "start": 3358.44, + "text": "random thought like like what what what" + }, + { + "start": 3360.0, + "text": "what could that be like?" + }, + { + "start": 3360.92, + "text": ">> Reinforce you need to have some way of" + }, + { + "start": 3363.36, + "text": "operating on these concepts um in a way," + }, + { + "start": 3366.0, + "text": "some kind of like gym, some kind of like" + }, + { + "start": 3367.48, + "text": "imaginary like it needs a world model," + }, + { + "start": 3369.84, + "text": "right? That's what OpenAI say like, \"Oh," + }, + { + "start": 3371.28, + "text": "we need a world model.\" It's like, \"We" + }, + { + "start": 3372.32, + "text": "need some way We need a transformer not" + }, + { + "start": 3373.92, + "text": "on text.\" That's why So," + }, + { + "start": 3375.56, + "text": "I don't know if you know this, but the" + }, + { + "start": 3376.4, + "text": "reason why OpenAI did their video model" + }, + { + "start": 3378.68, + "text": "is because of they're trying to solve" + }, + { + "start": 3380.32, + "text": "this problem. They want a world model so" + }, + { + "start": 3382.2, + "text": "that these these AI systems can try to" + }, + { + "start": 3384.28, + "text": "imagine new scenarios over and over" + }, + { + "start": 3385.92, + "text": "again and find new ways to be better." + }, + { + "start": 3388.44, + "text": "Um and then, you know, if you can" + }, + { + "start": 3390.12, + "text": "actually even if it's very inefficient" + }, + { + "start": 3391.92, + "text": "at be being better, right? If you can" + }, + { + "start": 3393.6, + "text": "somehow correlate the strength of an AI" + }, + { + "start": 3396.0, + "text": "system one-to-one with like compute, you" + }, + { + "start": 3398.6, + "text": "we can always get more compute, right?" + }, + { + "start": 3400.16, + "text": "We can always make Nvidia stock price go" + }, + { + "start": 3401.72, + "text": "up and like get more chips from Taiwan." + }, + { + "start": 3403.04, + "text": "Like it can always we can always keep" + }, + { + "start": 3404.72, + "text": "ramping this, right? There's a lot of" + }, + { + "start": 3405.8, + "text": "silicon on Earth and stuff like that," + }, + { + "start": 3407.0, + "text": "right? It's like you know, you you can" + }, + { + "start": 3408.52, + "text": "just keep doing this. You can scale it" + }, + { + "start": 3409.76, + "text": "and then once you have the compute, it" + }, + { + "start": 3411.16, + "text": "keeps printing it keeps being smart. And" + }, + { + "start": 3413.12, + "text": "so, it will it is only a matter of time" + }, + { + "start": 3414.8, + "text": "until, you know, if you can actually do" + }, + { + "start": 3417.88, + "text": "self-optimizing intelligent processes" + }, + { + "start": 3419.52, + "text": "with silicon that it exceeds human" + }, + { + "start": 3421.2, + "text": "baseline and becomes the god-like figure" + }, + { + "start": 3423.04, + "text": "some people want to, you know, build" + }, + { + "start": 3425.16, + "text": "upon the Earth." + }, + { + "start": 3433.08, + "text": "Sam, Timothy?" + }, + { + "start": 3434.68, + "text": "Oh, I I had a question about like is the" + }, + { + "start": 3437.52, + "text": "future then like oh if if they Is there" + }, + { + "start": 3441.88, + "text": "a data bottom neck at some point? And is" + }, + { + "start": 3444.2, + "text": "the future of reinforcement learning?" + }, + { + "start": 3446.16, + "text": "So, what I'm thinking about is like a" + }, + { + "start": 3447.52, + "text": "baby" + }, + { + "start": 3448.72, + "text": "you right spends like a lot of time" + }, + { + "start": 3450.24, + "text": "receiving input. And I have a theory" + }, + { + "start": 3452.44, + "text": "that the reason puberty happens, and" + }, + { + "start": 3454.64, + "text": "like the body grows so quickly, is" + }, + { + "start": 3456.2, + "text": "because it's to like protect the brain" + }, + { + "start": 3458.92, + "text": "from injuring itself before it while" + }, + { + "start": 3460.72, + "text": "it's too dumb." + }, + { + "start": 3462.64, + "text": "That kind of makes sense, actually. I" + }, + { + "start": 3463.68, + "text": "don't I don't know. I So so is is there" + }, + { + "start": 3466.32, + "text": "a future where like you" + }, + { + "start": 3468.24, + "text": "let robots like with cameras out in a" + }, + { + "start": 3470.68, + "text": "field, and then the learning is just" + }, + { + "start": 3472.12, + "text": "like, what is going to That what's the" + }, + { + "start": 3473.92, + "text": "next frame of the video? Yes, actually" + }, + { + "start": 3476.04, + "text": "exactly. That's like what Nvidia's" + }, + { + "start": 3477.36, + "text": "doing. That's like There's already many" + }, + { + "start": 3479.32, + "text": "data sets where they basically just" + }, + { + "start": 3480.72, + "text": "strap a GoPro to the center of a guy" + }, + { + "start": 3482.12, + "text": "cooking in a kitchen, and there's like," + }, + { + "start": 3483.52, + "text": "you know, gigabytes of those videos. And" + }, + { + "start": 3484.8, + "text": "the idea is just like, oh, we just" + }, + { + "start": 3486.76, + "text": "basically need a lot of video of people" + }, + { + "start": 3488.56, + "text": "doing whatever they're doing that's" + }, + { + "start": 3489.56, + "text": "useful to other people. Throw a bunch of" + }, + { + "start": 3491.4, + "text": "transformers at that, trillion-dollar AI" + }, + { + "start": 3493.0, + "text": "company, and then we automate them, and" + }, + { + "start": 3494.44, + "text": "then we rake in the money, and then" + }, + { + "start": 3495.64, + "text": "there's no more people to buy cuz" + }, + { + "start": 3497.16, + "text": "there's no more economy. I don't Kind of" + }, + { + "start": 3498.63, + "text": "economy. I don't Kind of" + }, + { + "start": 3498.64, + "text": "just stopped making sense. It's like" + }, + { + "start": 3499.84, + "text": "really crazy. Like if you really if you" + }, + { + "start": 3501.64, + "text": "really think about it for a second," + }, + { + "start": 3502.75, + "text": "think about it for a second," + }, + { + "start": 3502.76, + "text": "right? All these like VC-backed AI" + }, + { + "start": 3504.64, + "text": "companies, if they really did pan out," + }, + { + "start": 3506.6, + "text": "there would be no more economy. Like" + }, + { + "start": 3508.12, + "text": "what are you competing for by investing" + }, + { + "start": 3509.56, + "text": "in that?" + }, + { + "start": 3510.6, + "text": "I I don't know." + }, + { + "start": 3512.24, + "text": "Although, um," + }, + { + "start": 3513.88, + "text": "concerning like those gigabytes videos" + }, + { + "start": 3515.52, + "text": "and so on, I would imagine a big big" + }, + { + "start": 3517.24, + "text": "chunk of like what would make a the" + }, + { + "start": 3519.64, + "text": "training run effective is like how do" + }, + { + "start": 3521.0, + "text": "you transform the data before sending it" + }, + { + "start": 3522.48, + "text": "into the model or into the algorithm," + }, + { + "start": 3524.76, + "text": "right? Because like I guess if you like" + }, + { + "start": 3526.6, + "text": "filter down to what actually matters cuz" + }, + { + "start": 3528.44, + "text": "like there's like a lot of when you have" + }, + { + "start": 3530.44, + "text": "a guy cooking or something, like a lot" + }, + { + "start": 3532.32, + "text": "of that image data is completely" + }, + { + "start": 3533.72, + "text": "useless. You don't care about it most of" + }, + { + "start": 3535.12, + "text": "it, right? So like" + }, + { + "start": 3536.88, + "text": "what kind of pre-filtering do you people" + }, + { + "start": 3539.08, + "text": "do, and like how much How important is" + }, + { + "start": 3541.32, + "text": "that? Well, actually people do the" + }, + { + "start": 3543.04, + "text": "opposite of pre-filtering. Sometimes for" + }, + { + "start": 3545.16, + "text": "deep learning, actually I think Tesla's" + }, + { + "start": 3546.56, + "text": "doing this, they don't feed pixels, they" + }, + { + "start": 3548.16, + "text": "feed like raw sensor measurements. They" + }, + { + "start": 3549.68, + "text": "would decrease filtering." + }, + { + "start": 3551.52, + "text": "There's a There's a very Well, there's a" + }, + { + "start": 3553.0, + "text": "very famous like blog post, basically," + }, + { + "start": 3555.12, + "text": "called the bitter lesson. And the bitter" + }, + { + "start": 3556.92, + "text": "lesson is that like a general method" + }, + { + "start": 3558.44, + "text": "that scales with compute will always" + }, + { + "start": 3559.84, + "text": "perform better than a specific method" + }, + { + "start": 3561.36, + "text": "made by engineers because compute is" + }, + { + "start": 3563.36, + "text": "always getting cheaper and more" + }, + { + "start": 3564.24, + "text": "plentiful. Engineers are becoming more" + }, + { + "start": 3565.72, + "text": "scarce as people become stupider from" + }, + { + "start": 3567.28, + "text": "the AI that you're creating and doing" + }, + { + "start": 3568.76, + "text": "this. Yeah, but but then by choosing to" + }, + { + "start": 3570.76, + "text": "use like sensors, like you're making" + }, + { + "start": 3572.32, + "text": "that's like a deliberate choice. Like" + }, + { + "start": 3573.52, + "text": "instead of using pixels, we're going to" + }, + { + "start": 3574.68, + "text": "use sensors. So, it's like maybe more" + }, + { + "start": 3576.2, + "text": "relevant information within it. So, like" + }, + { + "start": 3578.04, + "text": "the pattern might be easier to extract" + }, + { + "start": 3579.36, + "text": "somehow or No, it's not that it's" + }, + { + "start": 3581.44, + "text": "the idea is that we should not try to" + }, + { + "start": 3584.64, + "text": "design these massive massive, you know," + }, + { + "start": 3587.2, + "text": "trillion parameter deep learning models." + }, + { + "start": 3589.64, + "text": "We should not try to do any data" + }, + { + "start": 3591.16, + "text": "transformation on what's happening in" + }, + { + "start": 3592.36, + "text": "the universe because it will find the" + }, + { + "start": 3594.6, + "text": "representation that's most effective for" + }, + { + "start": 3596.16, + "text": "minimizing the loss, right? It will do" + }, + { + "start": 3598.2, + "text": "that with compute. It will do It will" + }, + { + "start": 3599.8, + "text": "find something better than what we could" + }, + { + "start": 3601.0, + "text": "ever filter it, right? And what" + }, + { + "start": 3602.52, + "text": "basically you're describing is I think" + }, + { + "start": 3603.68, + "text": "called feature engineering. We're like," + }, + { + "start": 3605.32, + "text": "\"Oh, instead of feeding the images to" + }, + { + "start": 3606.92, + "text": "the deep learning model, we feed we we" + }, + { + "start": 3608.6, + "text": "try to detect like where the cones are" + }, + { + "start": 3609.92, + "text": "and stuff and like the road and then" + }, + { + "start": 3611.0, + "text": "like, you know, we feed those features" + }, + { + "start": 3612.6, + "text": "into the model.\" And it always" + }, + { + "start": 3613.56, + "text": "underperforms just feeding in the raw" + }, + { + "start": 3615.08, + "text": "sensory data. If you have enough compute" + }, + { + "start": 3617.36, + "text": "to scale up because the compute will" + }, + { + "start": 3619.44, + "text": "find the representation that is most" + }, + { + "start": 3621.32, + "text": "effective to output the policy" + }, + { + "start": 3622.6, + "text": "basically, I think. So, but you can" + }, + { + "start": 3624.32, + "text": "still select the which what kind of" + }, + { + "start": 3626.92, + "text": "sensor data you feed it in. Like the" + }, + { + "start": 3628.32, + "text": "sensor is is different from sensor data" + }, + { + "start": 3629.71, + "text": "is is different from sensor data" + }, + { + "start": 3629.72, + "text": "from like vision, right? Yeah. You still" + }, + { + "start": 3632.0, + "text": "should make a choice and like that" + }, + { + "start": 3633.6, + "text": "choice will matter, right? Yeah, the" + }, + { + "start": 3636.04, + "text": "choice matters in the opposite direction" + }, + { + "start": 3637.51, + "text": "matters in the opposite direction" + }, + { + "start": 3637.52, + "text": "from what you're saying. You want more" + }, + { + "start": 3638.8, + "text": "more generic, more bigger amount of" + }, + { + "start": 3641.0, + "text": "data. So, you feed it everything, every" + }, + { + "start": 3642.8, + "text": "type of sensor data you can get your" + }, + { + "start": 3644.08, + "text": "hands on. You just feed it whatever you" + }, + { + "start": 3645.52, + "text": "can. Yes, ideally yes, but like what" + }, + { + "start": 3647.8, + "text": "they found actually I I think there was" + }, + { + "start": 3649.44, + "text": "this paper called like the Nemo paper. I" + }, + { + "start": 3651.2, + "text": "don't remember what it was, but they" + }, + { + "start": 3652.6, + "text": "basically had a transformer they just" + }, + { + "start": 3654.68, + "text": "found that as you taught it many random" + }, + { + "start": 3656.16, + "text": "disparate games, it would like improve" + }, + { + "start": 3657.8, + "text": "on the other ones, right? That were" + }, + { + "start": 3658.8, + "text": "seemingly unrelated. And so, there's" + }, + { + "start": 3660.36, + "text": "something just about transformers. The" + }, + { + "start": 3662.28, + "text": "more data that you give them, the better" + }, + { + "start": 3664.4, + "text": "the representations become no matter" + }, + { + "start": 3665.72, + "text": "what. I think it's because there's" + }, + { + "start": 3667.52, + "text": "there's something kind of like a a" + }, + { + "start": 3669.0, + "text": "universal prior or like" + }, + { + "start": 3670.96, + "text": "you know, it's just the universe is" + }, + { + "start": 3671.96, + "text": "ordered in a particular way that's" + }, + { + "start": 3673.28, + "text": "evident in everything that happens in" + }, + { + "start": 3674.6, + "text": "it. So, the more in- information that" + }, + { + "start": 3676.0, + "text": "comes in This is not This is me not not" + }, + { + "start": 3678.16, + "text": "by the way. This is not a true I'm like" + }, + { + "start": 3679.96, + "text": "theory I'm like crafting theories. It's" + }, + { + "start": 3681.2, + "text": "like maybe there's like something about" + }, + { + "start": 3682.48, + "text": "the universe that just whatever you give" + }, + { + "start": 3683.92, + "text": "it that's from the universe truly and" + }, + { + "start": 3685.84, + "text": "not you know like a craft of man or" + }, + { + "start": 3688.64, + "text": "whatever like it it's just somehow" + }, + { + "start": 3691.92, + "text": "builds better representations cuz it's" + }, + { + "start": 3694.16, + "text": "something like that. I don't know." + }, + { + "start": 3695.08, + "text": ">> there's universal laws and patterns and" + }, + { + "start": 3696.44, + "text": "such that just exist in reality, right?" + }, + { + "start": 3697.96, + "text": "And like that's maybe what you're what" + }, + { + "start": 3699.84, + "text": "it's using somehow? That's maybe yeah." + }, + { + "start": 3702.24, + "text": "I'm not I'm really not sure. That's" + }, + { + "start": 3704.08, + "text": "I would I would like to see more" + }, + { + "start": 3705.08, + "text": "rigorous attempts to answer that" + }, + { + "start": 3706.12, + "text": "question rather than me like theory" + }, + { + "start": 3707.68, + "text": "crafting." + }, + { + "start": 3710.24, + "text": "Are there any more questions?" + }, + { + "start": 3712.84, + "text": "I mean just earlier like uh" + }, + { + "start": 3715.0, + "text": "said I you beat me to it, but I was" + }, + { + "start": 3716.12, + "text": "going to ask about like uh just the" + }, + { + "start": 3717.4, + "text": "noise like the back of noise of AI" + }, + { + "start": 3719.52, + "text": "generative AI on the internet, right?" + }, + { + "start": 3721.24, + "text": "Because like if you lower the friction" + }, + { + "start": 3722.76, + "text": "for like creating content, but like you" + }, + { + "start": 3724.96, + "text": "said it's like you're using existing" + }, + { + "start": 3726.88, + "text": "data transformations to create things," + }, + { + "start": 3728.88, + "text": "then I mean wouldn't the problem be like" + }, + { + "start": 3730.72, + "text": "you just end up with so much stuff that" + }, + { + "start": 3732.32, + "text": "is all from old transformations? It's" + }, + { + "start": 3734.04, + "text": "really hard to find novelty among that." + }, + { + "start": 3736.44, + "text": "So like Um search engines get good" + }, + { + "start": 3738.04, + "text": "enough that like we don't have a problem" + }, + { + "start": 3739.48, + "text": "or like No, there there's just very easy" + }, + { + "start": 3741.6, + "text": "ways to detect if something was output" + }, + { + "start": 3743.68, + "text": "by AI for the most part or like ways to" + }, + { + "start": 3746.04, + "text": "filter for data that is like the most" + }, + { + "start": 3748.12, + "text": "confusing for the model. And so these" + }, + { + "start": 3749.52, + "text": "things are not going to be a serious" + }, + { + "start": 3751.08, + "text": "problem for deep learning." + }, + { + "start": 3752.76, + "text": "at scale. Detecting AI or detecting" + }, + { + "start": 3754.8, + "text": "detecting information that AI has not" + }, + { + "start": 3756.2, + "text": "seen before that's like possible so this" + }, + { + "start": 3759.16, + "text": "that's something that's like easy you" + }, + { + "start": 3760.44, + "text": "say? Yeah, you can just tell which you" + }, + { + "start": 3762.92, + "text": "know training seems to be the most" + }, + { + "start": 3765.12, + "text": "confusing to it. Maybe this problem is" + }, + { + "start": 3766.84, + "text": "more complicated than" + }, + { + "start": 3769.04, + "text": "what I'm thinking because like maybe" + }, + { + "start": 3772.24, + "text": "I don't I don't know. I'm sure there's" + }, + { + "start": 3773.76, + "text": "there's much smarter deep learning" + }, + { + "start": 3774.72, + "text": "researchers who have answered that" + }, + { + "start": 3775.68, + "text": "question already probably, but my my" + }, + { + "start": 3777.68, + "text": "guess would be that like" + }, + { + "start": 3780.08, + "text": "it's probably not a real issue. Okay," + }, + { + "start": 3782.2, + "text": "because if you could just filter out" + }, + { + "start": 3783.6, + "text": "everything that's AI generated it'd be" + }, + { + "start": 3784.68, + "text": "really great, you know. Yes, they can I" + }, + { + "start": 3786.96, + "text": "think you can do that, yeah. Oh yeah." + }, + { + "start": 3789.44, + "text": "I hope so." + }, + { + "start": 3790.92, + "text": "All right, we're out of time. We're out" + }, + { + "start": 3792.2, + "text": "of time. Round of applause for Krikey." + }, + { + "start": 3804.84, + "text": ">> Yeah." + } + ] +} \ No newline at end of file diff --git a/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript_clean.txt b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript_clean.txt new file mode 100644 index 00000000..29c9ddfb --- /dev/null +++ b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/transcript_clean.txt @@ -0,0 +1,2082 @@ +Okay, so this is very difficult because +Cameron is such a is such an like +indescribable guy. +Um +you know, +yeah, this is very very very difficult. +I know Cameron just as like the guy who +uh you know, +he +he's like the guy crazy enough to just +like dig through the mountain and do +just like as sudden +sudden huge amounts of work +uh of like, "Oh, I cloned an application +or I did this." And then at other times +he, you know, goes and becomes a web +developer and then complains about like +not being of the program anymore. +So, anyway, um +can we get a round of applause for the +completely undescribable +Cameron Wrights? +What's up, Craigsters? LET'S MAKE SOME +NOISE. +LET'S SAY YOU WANTED TO WRITE A program +that could tell the difference between a +cat and a dog. +I want you to actually think about how +you would do this for a second, okay? +Imagine you were given the pixels and +you had to make the program that said, +you know, cat or dog. +You might try something like, you know, +look for ear shapes with edge detection, +maybe like detect like the little angle +there, you know, if it's a pointy ear, +it's a cat, if it's a floppy ear, it's a +dog. +Problem solved, right? You don't need +any more tech, you don't need any fancy +whatever. +Uh-oh. +You know, if we look at our program +here, it's just a system with an input +and an output, right? Um +take in the image +and then, you know, look for whatever +heuristic we've decided we're looking +for, and then we we, you know, make our +output, right? But, um if we keep trying +to iterate on this, it's like, okay, +maybe some of them have pointy ears, but +only if the fur is orange, and we try to +keep, you know, coming up with patterns +that try to get at what we mean by a dog +or cat, +um you'll realize that it quickly +becomes untenable, right? +So, the whole idea of machine learning +is you essentially have a program that +creates that program. It's basically +meta programming, right? That's sort of +a metaphor for what it is. +Um, you set up some architecture and +repeatedly optimize with respect to, you +know, your examples, um, a machine +learning model, a black box of many +parameters. It's a learned program, +right? It's automatic programming. +As a quick aside, why is it all Python? +It's because it's basically meta +programming, right? It's a complicated +build system. Uh, these training scripts +basically are just there to eventually, +you know, automatically create the +program that does what you want it to +do. Um, and most of the heavy lifting is +not necessarily done by the actual code +you write anyways, it's done +in the in the dot products on the GPU +or, you know, +piece of compute that you're using to +actually do the machine learning. Um, +and John Carmack for his current AI +efforts, he's in the race to win the +world. +Um, +he has switched to Python. I thought +that was notable. +So, a little bit about me, right? I'm +mostly a game developer. If we go back +to 2017, one of my first projects was a +game a game engine. +Um, I think this is probably pretty +common for people who I don't know like +to program is the first thing they try +to make is a game engine. It's kind of +funny. +Um, you don't know anything about what +you're doing. You're like, I'm going to, +you know, do the thing that's the +ultimate thing. +Um, and then much later, you know, 3 +years later, you know, it's like I'm +making a lot of individual games, I'm +working on the Godot engine. +Um, you know, these are a couple of my +game projects. +Um, that there on the upper left is a +space game called Asteris. It's +releasing in 2035. Um, it's got +something like a Overwatch net code in +it, right? It's like a +big solar system with many people +playing at the same time. Um, +yeah, I I was pretty proud of how stable +the net code was considering how chaotic +that game was, right? Um, +and then Dante's Cowboy. Um, this was no +engine in C, 3D rendering, animated +armatures, all from scratch, and all the +NPCs in this game were uh +you know, they were controlled by an +LLM, and this was never released because +it's fundamentally flawed. Um +I'm a deep learning hobbyist. +The majority of my work is not in deep +learning, so there's going to be some +inaccuracies in what I say. It's just +always been interesting to me ever since +back when I started, you know, first +doing games, I've also been kind of +looking at deep learning, which is a +subset of machine learning, and it's +just always kind of interested me. Like +I've been following it, right? +Ironically enough, actually, my first +ever professional programming experience +was to architect a deep learning model +from scratch, um make it work, and +deploy it in production. Um this was for +my college roommate's company called +MacroHard. Um he basically made software +that like was a custom broadcast overlay +for for League of Legends teams, +and I made a deep learning model that +predicted which team was uh going to win +the game, right? So, that little +visualization there is what I created. +Uh this was trained and, you know, on a +and optimized on many, many professional +games. +and you know, it's it's just a pretty +interesting problem, right? And at that +time, right, I implemented uh League of +Legends game predicting paper, +and I couldn't reproduce the results, +you know, I was like, "Well, +I copied the thing that the paper said +to do, and then it's, you know, it +doesn't work." +And basically, +there's a there was a bug in the paper, +right? This is a very important lesson +of not just how hard machine learning +is, but how often people mess up, right? +And how easy it is to mess up and be so +confident in what you did that you +publish it to the world and waste my, +you know, 2 weeks, like +it's just very, very easy to mess up. +And just to give a quick overview of the +bug they had, right? When you want to +make sure a machine learning model +works, you usually have a you give it a +test, just like a person, right? And if +that machine learning model were to see +the answers of the test before it took +it, uh you know, its success in that +test would obviously not be +representative of it of if it could do +the, you know, a problem, right? +Arbitrary new examples, it's not +evidence that it generalizes. And in +this specific case, they had a metric, +the champion win rate, that was trained +on their entire um +that champion win rate was based on +their entire data set, right? And so +then when it went to test the model, it +had the win rate for the champions that +were in the test set. And so it +purported to have amazing accuracy at +predicting who would win on the test +set, but there was data information +pollution, right? These are the kinds of +bugs you find in deep learning. They're +not something you could really +breakpoint in the debugger, but things +you have to find like a scientist, +right? Um you almost have to like uh +binary search over the uh problem space +sometimes to see what's going wrong. It +could be anything, right? +And I want to stress again that machine +learning is automatic programming, +right? In this paper, this is a machine +learning technique, not a deep learning +technique, called random forest, where +it's basically a series of learned if +statements directly, right? +And you know, there might be some +majority voting scheme, and they have +some algorithm to optimize like based on +the training data, oh, this if statement +should be this, and this if statement +should be this. They basically are +. They basically are +attempting to tweak the if statements +such that it, you know, is as accurate +as possible on the training data. +and deep learning is separate from that. +It's a subset of machine learning. +And the story kind of goes like this, +right? Some guy in the '60s was looking +at this, and he was like, I think I know +what these are. He was very foolish. +you know, he he he basically, on a very +abstract level, said, "Okay, you know +what neurons are? You know what's going +on in the brain? You know what's +enabling us to do all these amazing +things? Is little, you know, boxes. They +take in a bunch of numbers, and then +they mutate those numbers somehow. Maybe +when they have enough of that number, +they they activate and propagate that +information to the other neurons." +Um if you were to zoom in really, really +quickly on what one of those, you know, +very abstract human neurons that they +designed is actually doing, right? Um a +single neuron in a neural network, +right? The first step is it adds all the +weights multiplied by the incoming +activations, right? So you see that +bigger arrow is a much stronger weight. +That tiny arrow, you know, you can even +have a negative weight, right? There's +basically a set of weights of coming +from the uh neurons on the previous +layer, right? And then it just adds them +like numbers. They're all floats, right? +Next, there's a bias that adds a bias to +that weighted sum. +Um so, this could be considered like the +activation threshold for that neuron, +right? So, like if the bias is -6, then +you would require, you know, six to be +above zero and to continue to propagate +that information. Again, this is a very +rough, bad model of biological neurons, +and they stumbled upon something that's +actually pretty crazy um just by doing +that, right? But, the final step is to +pass it through any nonlinear like sine +or or x squared. I'll only get to why +they do that in a second, right? That's +what happens in a neuron. Here's some +like pseudocode for how you would find +the float value for a neuron, +right? Um +and you can take a look at this and +ponder it for a second, but +you'll you'll notice that well, if we're +constantly just multiplying and adding +numbers, that's kind of what linear +algebra is, +and that is how they accelerate this and +talk about this in the actual, you know, +field, right? They don't express each +individual weight in some kind of array. +They express it as a matrix, right? +and then there's a lot of math that +applies when you do that, right? Um you +optimize those weights and the bias with +a process called gradient descent, +right? Stochastic gradient descent. Um +I'll also get to that in a second, +right? But, fundamentally, this is just +a diagram for this expression or that +code, right? When you see this diagram, +it is simply that code, right? Um +that's what it's intending to describe, +right? And so, just to kind of get at +like what the act what's going on with +the activation function, you know, why +do they pass it through sine, right? +Each layer of a neural network is +basically a node in a transform +hierarchy of like n dimensions, right? +Um so, if you're a game developer, which +you're the target audience of this talk, +so you should be a game developer. +You're very familiar with like trees of +transforms, right? +Um and you know that if you have a +series of transforms one after the +other, you can multiply them all and it +becomes one linear transform, right? +That's like what it means for something +to be be linear. And when we want +complicated behavior from these deep +learning models, uh we don't want it to +be reducible like that to one linear uh +translation and and rotation whatever, +right? So, you basically think about it +as like warping the space +such that the translation and offset you +do kind of like it it enables you to +basically draw a much more richer +contour around the data, right? Um +And so, just as a as as another metaphor +to help you as a game developer +understand what's going on here, right? +Both that neural network and that shader +graph, these are both compute graphs, +right? There's no branching. Information +is flowing in from the left and it's +going out the right. Um +A shader operates on vertices while a +neural network operates on an image or +some state or anything you want, right? +And in the case of game development, an +artist might be manually tuning the +inputs to that shader and making it look +good, right? They might be nudging, you +know, the color and trying to make the +output look good or some other +parameter. In the case of deep learning, +that's actually replaced by a machine, +right? Which is which direction +minimizes error on training for each one +of these little uh parameters, right? +Each one of those weights and all those +biases, it knows which direction it +should go in assuming all the other +parameters were, you know, not moving. +It knows approximately which direction +to go in to get better at the task. And +then it just kind of repeatedly does +that thousands and millions of times. +Um so, there is a code base for this +talk. It's here. github.com/krickey/ai +Um and to just get at what this code +base is supposed to be for a second, +right? Deep learning libraries or, you +know, autograd, they let you perform +some arbitrary, you know, computations, +right? Might be a matrix multiplication, +you know, it might be x squared or one +of those activation functions we talked +about. Um and then find out how you need +to turn or or or adjust those +parameters, right? The weights and the +biases. It automatically figures out +what the expression is to make those +move in a direction that makes the model +get better at the task, right? +And that's basically that +the what these deep learning libraries +are supposed to be in a nutshell, you +know, TensorFlow, PyTorch. Um +And you know, +just as a little preliminary for this +code base, there's a lot of like tensors +in it. Um +but really tensors are just arrays of +matrices. Um +like you might want to talk about like +oh uh +We're we're feeding in many examples +into this deep learning model. And so we +want like a third dimension to it where +it's just an array of the matrices that +we're feeding into the model, right? Um +And so my idea with this code base is it +would be really cool if instead of me +having a you know, have a Python +interpreter, there was one C file called +ai.c, right? We had CPU and GPU deep +learning computer vision algorithms and +implementations, right? I would want it +to be able to execute LLM local LLMs, +right? I want you to be able to +fine-tune those LLMs, not just the +inference. Um I would want to hardware +accelerate with something portable like +Sokol, right? I have some computer +vision algorithms as well cuz those are +fun. +Um and unfortunately, this this this +talk code base right now is like super +super not that, right? It has a +hand-coded gradient calculation for a +two-layer perceptron that classifies +digits. It's super primitive and it's +like 1970s stuff. There's no working +auto grad yet, but I almost have it +working. +Um and I I intend to continue working on +ai.c cuz it's kind of fun. I'll be +sending updates in the +in the BSC Discord about that. +but just as an example of what I was +trying to do with this little tiny, you +know, deep learning library is +each of these videos are videos I got +from YouTube of some kind of ideally +animating a scene, right? Where the +camera is stationary. Um the theory is +you should be able to train a deep +learning model to generate these videos +from the image. Um +you know, the assumption is that the +behavior of the pixels is relatively +simple. Um and so I actually, even +though I wasn't able to get this working +in time for the talk with ai.c, I +Hold on. Sorry, one sec. +Oh, here we go. Okay. So, I basically +just, you know, let's just try it in +Python really quick to see if this will +even work at all. This is an 80 kilobyte +of parameters neural network and it's +garbage, right? This kind of happens a +lot in deep learning when you just have +a random idea and you just try it. It +makes something really sometimes hellish +like I didn't even know what that is, +honestly. +I think that 80 kilobyte is way too +small to encode the dynamics of a waving +tree, probably. So I would just, you +know, 10x it like +Deep learning is just very experimental +like this. It's not exactly known, you +know, what would cause the model to +correctly learn how to ideally animate +that scene or make the clouds move, +which is what I'm hoping for, right? You +can imagine with a model like this you +could take any desktop wallpaper and +just make it ideally animate in the +background, right? Especially you got it +like +you know, good enough on the battery. It +just might be kind of fun to, you know, +minimize it and then see stuff always +moving. +So just as a little bit of a historical +detour, right? That what I just +explained to you, that deep learning +model, it the perceptron, it was this is +an article about it in 1958, right? +They were saying it can tell the +difference between a cat and a dog. It +is of no practical use. One day it might +be useful to send one into outer space +to take impressions for us. +So AI has always been a grift, +basically, right? And it's followed +basically this this phase where +this this phase where +all the investors pour their money in +and they're like, "This is going to be a +new form of life that will like do our +bidding, right?" And then it doesn't +work and then it happens again and then +it doesn't work. And now it's happening +again and it's, you know, +who knows what's going to happen. +He just says, "What did we learn, +Palmer?" Um +And so me personally, you know, I my +kind of like AI boom bust was I got +really into deep learning. Um you know, +and then I right when GPT-2 started to +happen, I was like, "This thing's kind +of nonsense and it's not going to +scale." So I basically was like, "All +right, I'm kind of done with deep +learning. I think it's useless." Um and +so I basically it's kind of like selling +Bitcoin at $100. I never would want to +buy back in cuz that's realizing my +losses. So I'm just kind of waiting for +it to collapse again, hopefully, maybe, +and then I can buy. +Um and if you think that's a funny joke +in the audience right now, how many of +you have any amount of retirement or +capital in ETFs, right? So, I need you +to raise your hand. +Okay. +Your money's being gambled on the +success of AGI right now. Roughly 30% of +the S&P 500 is heavily weighted on AI +succeeding. AI growth stories, you know, +high price to earnings. +Um so, I propose a model of history +here. There's basically two grifts +orbiting each other, right? Fake money +pumps the AI. Capital flees to Bitcoin +instead of US-related, you know, assets. +No more easy capital, no more AI. +Deflationary doom spiral where GDP +becomes 1/10 of what it is now. This is +an idea that is basically a +bubble-generating phenomenon, right? +Both of these, you know, one paper +basically started the whole +There's like the transformer paper and +there's the Bitcoin papers, parallels, +right? And if you look at this like, you +know, innovators-laggards curve, right? +I think we're right about here if you +look at the Nvidia stock price. It fits +the curve. +So, we're at peak of grift at the +moment. It'll fall off soon, I think. +And so, I think now's a good time to +short, but +that may or may not be financial advice +if it works out for you. +I do just want to say that shorting, I +believe, has unlimited uh risk and you +at best would like double your money, so +it's probably not like a good idea. +Here's some examples of uh just machine +learning in general, right? Let's go +back to not, you know, being kind of +schizo. Um +So, this is a little piece of history +again, right? This is called Cyc. In +1984, some people were very upset at the +deep learning people. They were like, +"Look at this arbitrary floats and +numbers. I want like an interpretable +program that's intelligent, right?" So, +they basically set out to manually +program all of human intelligence. And +the theory was if you just throw enough +bodies and people at it, it'll +eventually be smart, right? They've been +doing this since 1984. Okay, if you ask +this system questions like, "If I put a +a cup of coffee on a table and then I +push on it, what happens?" It'll say +like, "Oh, well, you know, it explodes." +It's like still not working very well. +the reasoning traces are like a thousand +deep, you know, it's kind of nuts. It +didn't really work, right? The First of +all, the reasoning isn't even +understandable. +Um this approach just doesn't It's +called symbolic AI, right? And I think +that there's actually a modern parallel, +right? Waymo is like psych, that project +where they were trying to manually +encode intelligence instead of automatic +programming, right? This is like deep +learning is really a new paradigm of +programming where it's automatic. You're +automatically creating the the +nonlinearities that do the thing that +you want the computer to do, right? Um +Waymo, they basically scan the city and +they put a bunch of if statements where +the car should go. So, they're very slow +to roll out. So, Tesla has basically +just recorded millions and millions of +hours of just people driving and they're +training massive massive transformers on +human behavior and they're just trying +to model it, right? +It's also There's another parallel with +like Siri and Echo Dot, right? These are +symbolic systems. They're often +frustratingly unintelligent. Um +to add new intelligent behavior, an +engineer has to basically, you know, +design the algorithm and all that stuff. +And ChatGPT is statistical. It's deep +learning. It's automatic programming, +Deep learning models don't learn. That +always bugs me. I think it's a misnomer. +Um that's like saying when you're game +developer, you're creating a universe. +It's like a very like hand-wavy way to +say something that I think is really +mechanical and mundane, right? I think +it's more accurate way to uh describe +what happens with deep learning is +optimizing, right? And it's an automatic +optimization. Um +ation. Um +Here's just another side note, right? +So, there's big copyright disputes with +these deep learning models, right? And I +just had a funny idea, right? Like +you could almost make the ar- if you +take the argument that uh the deep +learning people are trying to make, +right? They're trying to secure their +bag. They're trying to say, "Well, we +learned the patterns from what you gave +us. We didn't, you know, copy the pixels +or whatever." That's like what JPEG +does, like learns a model of the pixels. +And so, it's almost like arguing that +because, you know, it's JPEG and not +PNG, uh because it's a model of the +pixels, you know, I didn't steal your +thing. +Um so, it's a little strange, right? Um +I think it's kind of a bad faith +argument cuz the patterns uh that are +found in deep learning um are very +generic and they're not so specific to +the image. And so, +it's very debatable, I think. I'm not +sure what's going to happen. But just to +go back to my League of Legends task, +right? If we if we just want to think +about that little bit more as an +engineer, +um you might naively just put in the +current state of where all the people +are on the map, how much gold they have, +what items they have, throw it through +that perceptron I showed you, and maybe +it'll automatically find ways to relate, +oh, like, you know, when the gold and +this item is here, then like they're +more likely to win than you would think, +and stuff like that. These are the kind +of patterns you hope a deep learning +model finds as it optimizes based on +massive amount of data you have, right? +There's a problem with this, though. +What about past states, right? Past +state informs the nature of present +state. If somebody was performing badly +. If somebody was performing badly +on your team in the past, right? They +were like running it down mid, this +should inform your current prediction +like heavily, right? Um +So, this uh you know +You need a recurrent neural network for +this, right? A recurrent neural network +is basically two of those perceptrons +that I showed you. One is modifying the +state almost, and then another one is +visualizing the state towards whatever +like you want it to visualize, right? It +might be probability for who's going to +win, or it might be like you could +replace it with another deep learning +network and train it to visualize like +I don't know, which which champion is +like has the most synergy with the other +champion, stuff like this, right? Um it +has state. It's much more like a +automatic program, right? Automatically +learned program. Um the hidden state of +a recurrent neural network is like RAM. +It's analogous to a RAM. +Um and that and that is why I think that +these are +best described as automatic programs, +right? Um and so, what I did at the time +to successfully model uh who would win +in League of Legends games was I I I I +selected a GRU architecture, and I spent +weeks tuning it and and messing around, +and I I got I managed to make it so +halfway through the game it could +predict who would win with 86% accuracy. +Um if you just went with whoever had +more gold, it was like 73% accuracy. Um +and then this continued to generalize +even on arbitrarily new games. Um and +just as, you know, as a side note for +game developers, you could use a +technology like this for our play of the +game kind of feature, right? To see big +unpredictable swings in who's winning. +Um we actually did do that. We took a +our giant corpus of professional games, +and I ran this model and saw what were +the biggest swings, and we looked at +those games, and they were the most +exciting matches, right? So, this is +kind of like just to give you ideas for +like what this technology can be used +for, right? Um each perceptron in a +layer is like roughly analogous to an if +statement in the program, right? Like +another branch. It lets you make a more +specific contour around what you're +trying to encapsulate, right? Um you're +adding more resolution to that boundary. +Um this is why it's called deep learning +cuz there's many Once you add many +layers, you can make very complicated +behavior, right? Um the problem if you +just take that perceptron and you add a +bunch of layers is it adds way, way, way +too many parameters and the gradients +get very tiny, right? Um it doesn't +optimize, right? The optimizer doesn't +work. I mean, you saw my failure of deep +learning. It's not easy. +Uh you almost have to like guess and +check if you really aren't, you know, +branching off of some established known +working uh architecture, right? Um and +so all of the field of of modern deep +learning is basically about how do we +add more parameters to this model and +still get those parameters to be +optimized into something that +generalizes to what we needed to do, +right? In the case of a convolutional +neural network over here, they basically +they kind of start with the inductive +bias that like, "Okay, +um when you're looking to see if a image +is of a cat or a dog, um and you +basically are hoping that let's say this +deep learning model will pick up on the +pointy ear thing. It doesn't matter +where in the image that pointy ear is, +right?" Um they basically want these +deep learning models to learn, you know, +features that it doesn't matter where +they are in the image. Uh if you just +fed every single pixel and all the RGB +values into a giant perceptron, then it +would have to learn that, you know, cat +ear means cat in every single spot, +right? That's not really what we do when +we recognize images and so they +basically That's like a little bit of +key insight. That's what deep learning +researchers are trying to do is like +come up with ways to to formulate and +structure the computation, the graph of +computation, such that when you optimize +it, it gets what you want, right? Um and +a transformer on the other end here I'll +talk about in a second. Um this is +another example of an architectural +change. You can add way more parameters +and it's still stable and it's still, +you know, optimizing correctly, right? +Just as another little side note, some +deep learning techniques are a little +silly. Uh Uh this is called dropout. You +randomly turn off neurons during the +during the training. It's it's supposed +to present or or prevent memorizing the +training data um and hopefully make it +so neurons aren't exactly specialized +Who knows why this works? Honestly, I +don't I haven't seen any real convincing +argument in any paper for what exactly +this is doing. Like basically what's +happening in deep learning is people try +stuff and they have some very like okay, +yeah, that sounds about right. I think +this what's happening and then it's like +there's no rigor in this field basically +like um it's really crazy. This is +actually used in production for chat GPT +without it like it wouldn't work, I +think. Um and it's like you know, we +just randomly turn off the neurons +during the training. Like okay. +So, there's a zillion videos that +explain transformers already. I'm not +going to give you an explanation of the +transformer, which is you know, the +cause of the current rift, right? Um and +you know, you can watch any of these and +they're good and whatever. But I'm just +going to give you a quick overview. This +is a single slide. This is what a +transformer is doing, right? It's +basically going from a sentence, a +lookup table per token into a point in a +space where like the closer two points +are the more related they are, right? It +adds a position encoding to those points +so that, you know, if we have an array +of these points, then the points are +modulated in such a way that the the +deep learning models they can they can +pick up where they are in the sentence, +right? They're almost jiggle in a way +where like you could learn that pattern. +I keep using the word learn. I'm sorry. +It's pervasive in the field, but I don't +like it. Um and then you literally just +stack a hundred of these attention +layers onto those uh vectors, right? And +each of these attention layers every +single time, every single one of those +tokens, right? It's basically looking at +all of the past tokens and enriching the +current vector with more information +about the surrounding tokens and then +more facts potentially about that +information in the form of another +perceptron, right? it's like learning +what what facts might be related to a +word like house, right? It's It It +continuously enriches each one of these +uh words with more and more context, +right? In a in a highly parallel easy to +train manner. Um and then there's just +this two-layer perceptron and then it +gives you the next word, and it does +that over and over again, right? Um +there are many hyper parameters in this +massive system, right? Many, many, many, +many things can go wrong. You know, it's +like what is the learning rate? You +know, how quickly do we turn these +parameters to, you know, make it better, +right? It's like a massive problem, +actually, um to know what These are +called hyper parameters because unlike +parameters, like they're not +automatically learned in the core +algorithm of deep learning, right? You +can or you can try to figure out what +they are uh with a search algorithm. +That's very costly if you're OpenAI and +each training run costs $10 million. Um +and as it turns out, we heard from an +insider here that most of the big gains +from these, you know, +most of the state-of-the-art models that +come out of OpenAI, they came from +so-called YOLO runs, where they +basically they basically asked Karpathy +or one of these, you know, like deep +learning generals, like, "Hey, +we're going to give you $20 million to +just try something, right? You just pick +numbers, pick hyper parameters, pick +whatever you want, and we'll see what +happens." And state-of-the-art model, +right? It's basically vibe engineering. +Vibe engineering created the vibe coding +machines, and everybody's vibing, and +then, you know, we're being wireheaded, +and then it's over. +Well, +I I I I I think that it's a very +interesting question, can these LLMs +reason, right? So, they can't reason cuz +they're not human. I think that's a good +way to use the definition of reason. Um +but let's say we want to figure out if +these LLMs can, you know, output text as +if they were reasoning, right? I think +that this benchmark ARC-AGI is the It's +a stellar benchmark for this task, +right? If you take a look at this +diagram here, right? That's uh one +example, and that's another example, and +you have to predict what happens here. +If you're human, which I'm assuming all +of you are, it should be pretty obvious +to you that you basically just put the +yellow dots like where the little corner +thing goes. And ARC-AGI is the kind of +benchmark where they basically take +random people and they're like, "All +right, we'll give you a $15 Starbucks +gift card if you solve this problem." +And they verify that like people can +just figure it out in a context-free +manner, right? And so, the whole idea +with this is they're trying to find +problems that like average people can +do, but like, you know, these LLMs +they're they can't do it at all. Um +And ARK G I they were originally, you +know, totally defeated by OpenAI's +latest reasoning model. +Um it it got human baseline on a huge +suite of these that are private, not +anywhere existing on the internet. Um +you know, generic Each each ARK G I +problem requires a new unique insight. +They're basically puzzle game problems, +right? They require some kind of like, +"Oh, that's how it works," right? And +LLMs have already beaten the first ARK G +I benchmark. Which that was quite +significant. I didn't think that would +happen. And neither did the author of +the benchmark. He's a massive skeptic +and he thinks it's a bubble. And then +OpenAI pulled them on stream and he was +like, "My benchmark is defeated." Um so +we made a new benchmark that they're all +failing at again. +And that's Grok 4. A few days ago they +just doubled the state of the art and +got 15% the 16% of the questions right. +Um So just circle back on some +applications to games really quick as I +kind of +close up on this kind of like lightning +overview of this, right? Any problem +where you have input to output data and +it's harder to find programmatically +what that mapping is, that's where you +would throw something like deep learning +at the problem and try to automatically +program that, you know, uh +transformation. +Um this is Nerf rendering. This is a +very interesting application of deep +learning to the problem of rendering, I +think. Um it's very old and there's much +better techniques like uh Gaussian +splatting. +Um but Nerf rendering is basically given +a point in the scene, a deep learning +model learns how dense it is and then +it, you know, kind of traces rays +through that learned density to render. +And just as a little interesting tidbit +more vibe uh engineering, uh basically +deep learning models are very bad at +learning so-called high-frequency um +high-frequency functions, right? So a +little tiny thin metal pole is a very +high-frequency function density function +for it to learn. It's like nothing +nothing nothing bam really thin nothing +nothing nothing. It's very like tight +. It's very like tight +and so it's hard for a deep learning +model to pick up on that cuz it would +require a lot of non-linearity. It would +require a lot of layers, right? Um and +layers, right? Um and +so what they did in Nerf to, you know, +achieve state-of-the-art results for +learning based rendering, is they pass +that position into +into a bunch of different sine waves and +they basically just made it easier for +the deep learning model to pick up on +higher frequency spatial information. I +thought that was kind of funny. She's +like, pass it into the sine waves and +then it's better. +Um this is called Ash Animatable +Gaussian Splatting. Let me show you how +the video works here. This is basically +photorealistic human rendering with with +accurate physics on their clothes done +in real time. Um it's just +unfortunately, in order to, you know, +scan somebody, you need about 100 +cameras synchronized and it's very +expensive. So, maybe they'll figure out +how to make that cheaper someday, but in +this specific paper, a deep learning +model is used to go from a texture that +describes the skeletal deformation of +what you want the person to be in into +the the view dependent color that you +would see +um from that skeletal deformation. And +and essentially, you need to train that +deep learning model to learn that +relationship between like how their arm +is bending and what their clothes look +like, um you know, per person. Um +And then of course, I'm sure probably +all of you guys already know, ray +tracing is is is greatly benefited by um +deep learning because it's used to +denoise, right? Um +Deep learning is used in a algorithm +called AlphaZero to master any discrete +perfect information game. Basically, a a +neural network goes from the board state +to a likely next move +um and in a win probability. And then +another algorithm manipulates that those +probabilities and you know, +beats any human in any game we've ever +invented, unfortunately. +Automatic game AI, so you can make a +style transfer, you know, +you can make it some kind of style +transfer for something like this for +your video game where maybe like you +want it to act more like the human +people. So, once it masters the game, +you then train it to act like a person +and then it would be a kind of +interesting mix between a good player +and oh, it plays like this specific +chess player or whatever. +Um or you could even have it validate +that your game is uh balanced and stuff +like that, right? I think this actually +is very it's quite green field for video +games in general for how to apply uh +deep learning models in a lot of places. +And you know, for automating like a call +center, I think that there's a lot of +these call center AI startups. You know, +it seems easy on the surface, but +usually people call support because they +want to talk to a person and not a +machine. So, I think that this is +probably going to be a a giant disaster. +Um and, you know, of course, automating +something like engineering is way more +complicated than just that, which isn't +even working. So, +you know, just something to think about. +Um and a quick note on computer vision +cuz I didn't really get to flush that +out that much in this talk. But, the +basic idea I had was for computer +vision, you're essentially trying to +take in a bunch of images in something +called monocular slam and get some kind +of representation of the world that it's +seeing um automatically, right? It's a +bit like deep learning where you have +some automatic optimization process that +operates on a lot of data to uncover +something interesting about the world. +Um hopefully interesting. Um and it this +that that visual slam pipeline is +basically the opposite of game +rendering. And it involves somewhat +similar linear algebra, right? Um you go +from an image to easy to track key +points, pose estimate for the camera, +some representation of the scene for the +robot to drive around in, right? Yeah, +game rendering, you know, it's like +scene representation, triangles, pixels. +I thought that was interesting. +So, will LLMs replace knowledge work en +masse someday, right? +I put a steelman yes and a steelman no +here. +Um really, I think it's probably +probably not. But, +there's like a lot of investors who +disagree with me, so I don't know. +new knowledge can't be made by machines, +right? This is a very funny joke made by +a philosopher called David Deutsch. Um +maybe this is what we're going to have +to do someday. You're going to have to +invent new scientific knowledge to prove +that you're not a robot. Cuz maybe +that's the one thing the one thing that +humans can always do is create new +knowledge cuz they're the only ones that +can officially reason. +But, if, you know, you have to have a +human know if you did new knowledge, I +don't know how a system would detect if +it's new knowledge or not. Maybe that's +like P and NP or something, I don't +know. Um +Wait until RKG I continues to be beaten, +I think, is how you monitor this +problem. +Um cuz if that continues to be beaten, I +think uh you know, that's crazy. +Thanks for listening. Don't forget to +check out and work on github.com /ai. +I want to make ai.c so that it's like +fun to mess around with these models and +techniques and games. +No dependencies, no python, no nothing. +I you know, I just it's like it's not +good right now. It needs to be rewritten +like 10 times. I've never made a deep +learning framework before, but +Thanks for listening to my talk. +Sam. +Okay, that's a bit similar to the psych +stuff. +It's used by cookie people for AI, but +not serious contenders. +Hello. +What's up? +>> Good. +Awesome. Can we get another round of +applause for Krakey? +Subscribe to my YouTube. Subscribe to +his YouTube. Watch +eye contact practice. It's the best one. +I mean, that talk was definitely like it +was like 15% skits, so +but that's why it's entertaining for +sure. +I think you're like part LLM. +Like anyone that's talked to Krakey, you +could just give him a prompt and he just +[ __ ] spits out infinite like +sometimes he hallucinates. Yeah, +sometimes it's garbage, but most of the +it's garbage, but most of the +time it's kind of right. Okay. +That is what that's kind of what LLM's +do, yeah. Yeah, exactly. So you're it's +in there. +It's interesting. +So everyone here at Better Software +Conference +wants to write better software and wants +to live in a world where there is better +software. +So will AI cause the destruction of +better software? +So, okay, one of the biggest like hot +topics for for AI startups right now is, +well, if AI writes the programs, right, +why do we even need an app store? Why do +we even need software engineers? You +should basically tell your computer the +app you want and then it does it and +then it gives you the interface and then +the data. Um +and it doesn't work at all right now and +so, I you know, it would We're going to +have to see some new developments, I +think, to judge if that will happen or +not. +I'm basically 90% it won't cause the +destruction of better software for at +least a decade, +but 10%, you know, there's always +something around the corner. Grok 4 is +like much better after some reason. +So, in 10 years, like +is software engineering a thing? Like, +when you walk into a room or people or +software engineers going to be crafting +prompts? So, I've been thinking about +this a lot. I mean, obviously, prompt +engineering is not going to be a real +thing because if you do believe these AI +systems will get to the point where they +can automate software engineering, they +can do the prompting, too. I don't think +it's that crazy. +Um but in any case, +I think that the effect of this on +software I Well, I I've I've been +thinking a lot about this question and +I'm I'm still very uncertain, but if +software engineers are basically making +these like virtual machines that do some +kind of processing or transformation +that's valuable to people and these LLMs +learned the patterns +uh to to to create some of these +machines as it relates to the to what +you're asking of it and I think people +who are making machines that are very +same-ish, software engineers who are +data transformations that are similar to +what exist are going to probably be +screwed. Um but if you're making unique +data transformations that are unlike +what have been made before, kind of +unique machines, I don't see how these +models could really create new knowledge +like that yet. I haven't seen evidence +that they can do that. +Um you pointed out something uh +like learning is different from +optimization. And that's that's more of +a philosophical thing, I think. +>> Like, when a human is born, right? +They're they're born like extremely +inefficient and they just get more +optimized as they grow. Is that not the +same? It's not I think it's not the same +because whatever human being is doing +can create human reasoning and I don't +think it's +I don't think it's it's it would be very +hard to make the claim that these things +are doing human reasoning right now, I +think, for a lot of reasons. Um you +know, oftentimes the reasoning doesn't +correspond to what they say. It's like +not logically coherent, right? It you +know, +fundamentally from the I mean, you kind +of understand what's happening here, +right? It's it's optimizing and and +creating an automatically created +program +um that might have nothing to do with +the the the universal causal +relationship between like, I don't know, +you know, whatever's happening in your +brain or maybe it's if you're religious, +then it's not even a question. +Like isn't the real problem like not +that AI replaces software engineers and +that destroys better software, but that +destroys better software, but that +software engineers use AI and that +destroys better software because that's +what's happening pretty much already. +Well, if you believe that society +functions, then people will choose the +products that are better. +So, +in the long run it, you know, +if yeah, I don't know. I'm not sure. +>> you pointed out uh +the like Grok had a massive jump +recently. Like did you look into that at +all? +On this specifically on the Arc AGI +test, which interestingly +there's another Arc. I don't know if +Taha point pointed that out. I missed +it. +Everything is Arc. +So, yeah, Grok might have performed +better because of its utter lack of +safety training. That is a rumor I kind +of saw and it seemed to be somewhat +confirmed by, you know, another times +it's like they did safety training and +the model got much less, you know, +reasonable in quotes. +Um it might be the case that the safety +training is not intelligent cuz it's so +directly refusing to to say something, +right? +Um, there's all sorts of hypotheses and +there's no real, again, rigor in the +It's not really it's They're just trying +to figure it out. They don't Nobody +Nobody really knows. +>> like broccolicing like some schizo like +>> It's not good. I don't want to repeat it +It's not good. I don't want to repeat it +out loud. You can Google it if you want. +It's not good. It's It's +It's completely out of control, but it +completely out of control, but it +is smarter. So, +well, I don't know. +Okay. Um, you didn't point out either +like what caused the +this AI the last AI winter to end, +right? Like what was the It specifically +happened So, transformers were made kind +of I think like 4 years before they were +really used to generate text in a very +in a very parallel large, you know, mega +data scale way. Um, and it happened when +OpenAI I think released GPT-2 and they +pointed out like, "Hey, we keep just +making these models bigger and it seems +like the text is getting more and more +coherent. So, if you just complete the +line, like someday we get to the thing +that makes everybody, you know, +Pop-Tarts for free or something." Like +And then I don't know. It's That's what +started this massive speculation bubble. +It's like, "Wait a second. We might have +a path to completely removing +uh humanity from what from satisfying +our our our desires in the physical +realm." +So, it was just they they just scaled it +up. Yeah, they scaled it up and it +seemed like the scaling wouldn't end. +Is that what they're Is that what people +are still doing? They just keep scaling +things. +>> buying GPUs. I mean, they're not scaling +in the same way. They're not just making +the models bigger. They're trying to +scale I the so-called reasoning models. +Um, they're basically trying to make the +models more and more correct instead of +just making them remember more and more +facts, basically. +I don't want to run out of time again, +so I'm going to open +I'm going to do questions right now. +What's up? Uh so, Cameron, a bunch of us +have seen your uh Dante demo just +floating around. Um, can you get into +why it's flawed and why Yeah, so my my +whole thinking with Dante's Inferno is I +think when I saw GPT-4 first like on the +API, I was like certain that this would +enable a huge new category of of video +Um and so I immediately began making +like an RPG where all the characters +were controlled by an LLM, not just +their speech but their actions, right? +And the idea was like, "Whoa, this thing +is basically like prototype AGI or +whatever, right?" So, it'll just kind of +operate and it'll be fun and like and I +went through so many iterations of that +idea. I made it so that like they could +join your party and trade items and all +this stuff and like I played with it a +bunch, but it never felt like there was +any meat to the game. The problem is +games are about predicting and +understanding systems almost, right? And +an LLM is an unpredictable black box, +sometimes acts like a person, sometimes +doesn't. It's not fun to interact with +in a way where you're trying to get it +to do something. At best, you're +basically like, "Hey trader, I'll +totally give you a million dollars if +you drop your item." And then it's like, +"This is not fun. This is dumb." +It's funny It's funny like for one time, +but it's not like a game you could +master. +It's like the vending machine at Claude. +Okay, there's Right now, there's some AI +companies are trying to first deploy +these these LLMs in the real world to +start to, +you know, create a utopian post-scarcity +world where machines make everything +automatically and they're like, "Step +number one, put the LLM in a vending +machine and have it automatically order +and interface with the people or +whatever." And right now, people are +basically convincing these AI vending +machines to stock many tungsten cubes +and sell them at a loss and then they go +bankrupt. +Um they've published many reports about +this and they're like, "You know, we're +we're you know, we're doing some prompt +engineering to fix the business model +and we're trying to make it refuse +tungsten cubes more." Um but +really, I think there might be no end to +humanity's ingenuity as we battle +against the machines for the next few +decades. So. +That's probably fine, right? It's +Everything's probably fine, but +everything's probably fine, but +sometimes bad things happen. I don't +know. What's up? +Do you think there So, I I feel like +there is a field of research, I just +don't know the words for it or how big +it is, but do you think there's a way +that we can reverse engineer Yeah, +interpretability is big. A lot of Yeah, +Anthropic is the leader in +interpretability, I think. Um they +basically do data science on the neurons +themselves, and they try to make claims +about, you know, how the data is flowing +through the model. Um you can look at +their Golden Gate model. They have a lot +of really excellent stuff you can read +forever. It's really cool. Um Do you +know much about it? Like, do you feel +like that is the way forward, or is that +just a byproduct of the craziness? I +don't think there will be any value +created from interpretability research, +honestly, other than maybe um +optimizing the training process somehow. +I'm honestly not really sure. I think +they're mostly doing it because they +have to, because they want to be the +safe people or something. +Um and it's cool, but it doesn't seem +valuable to me, but I I'm not like a big +deep learning research. I'm like a +hobbyist, like I said earlier, so I'm +not really sure. +Uh follow-up to that, is it +interpretability +that made people figure out you could +remove the safety neuron? You can. They +already figured that out. It's called +ablation. So, any single open-source +model, you can basically find what's +called an ablated model, where they just +basically subtract the safety training, +and then it does whatever. But, it's +But, you told me it's like one node +that's like this. Yeah, they I don't +remember exactly how it goes, but it's +not a node, but it's like one vector +kind of represents the safety, +uh the refusal. Like, when whenever an +LLM starts to refuse, it always seems to +output words in a very, you know, this +kind of sounds like the same thing, and +it basically you can find that inside of +the model and then nullify it is what +ablation is, from what I remember. It's +been a while. +Uh as you said that AI is basically +metaprogramming. Yes. So, and Julia is +very famous for its metaprogramming. +Will it replace Python? Like, will we +have Julia AI or like So, maybe +metaprogramming is a bad thing, cuz it's +not really comparable. I was moon I was +mostly It's like metaprogramming only to +the extent that you're producing a +program with a program, but it's not +like it's automatic metaprogramming. +It's not, you know, mechanistic symbolic +metaprogramming, where you're yeah. So, +it would not be a good application for +this language, next one. Huh? +It's not a good application for jail. +>> No. No. +No. +Uh +rough To bring more into the concrete +realm, you mentioned SLAM. Yes. It's +been a decade since I've looked at SLAM. +Like, is deep learning the current state +of the achieving SLAM algorithms, or are +they still +>> I I just recently asked a SLAM +researcher that, and they basically said +deep learning is like used sometimes, +but it's not really necessary to solve +the problem though. Um +you know, they found good heuristics to +find the key points that are trackable. +Um and like neural I think that neural +SLAM seems like it's something they've +kind of been like rumor you know, to +kind of take over a state of the art, +take over a state of the art, +but like you know, the Quest doesn't do +any neural SLAM as far as I know. +You mentioned using uh AI models for +balancing games. Yes. +>> and +let's say you then train a model to play +StarCraft II, and you you finally do +definitively prove that Protoss should +be nerfed. +yes. +God, go ahead and start recording. +Uh once you nerf Protoss, do you then go +like, "Okay, cool. Now we do a $3 +million run again?" I don't think it +would be that expensive to do it on a +game. It's the cost of AI right now is +because to make these large language +models learn, you know, reasoning, you +have to basically shuffle the entire +internet in and out of um multiple +accumulate ads. It's really not so much +data they have to process. For something +like StarCraft II or AlphaZero, it's +really not on that scale of cost. Um I +could easily see many many many games +being able to do this, and for simpler +games, you could probably do it on your +laptop even maybe. Um +But yeah, like I think that for +something like StarCraft II, it's like I +could see them doing these kind of +studies if they have the the cap backs +and they believe it's +you know, it's it's valuable for the +players to have have truly be officially +like even the genius master AI +uh players of this game it's +even right that might be a good thing to +prove. +Uh yeah. Uh would you say that like +training is where like mostly cost is or +Um actually well opening I says that +inference is their biggest cost right +now +but there's a massive push to make you +know on on device inference chips so I +think that's kind of temporary and it's +just because like consumer my +understanding is I don't know if this is +exactly true but I've seen this floating +around that Nvidia cards don't have +enough VRM to run these models simply +because it's not something consumers +wanted and they kind of want to nickel +and dime you for VRAM +and so this seems like the kind of thing +where you know if it's a healthy market +and you know it's not there's not like +an Nvidia monopoly or whatever +which there is I'm joking it's not going +to get better but +um +ideally I think it's totally feasible +that these the inference cost is zero +for the people who are +making these models it it should be +zero. +Uh yeah. Yeah in your little experiment +game how did you translate basically the +LM output to actions in your world? So +basically you can view putting LMs in a +video game as so you have your game +state and then there's like the um +physics +simulation kind of like you know human +in the loop like you're inputting +actions and state 60 hertz or whatever +and there's a text adventure way to +interact with the game simultaneously +right and a good way to check that like +the game was properly you know +presenting itself to LMs is if you could +flip it and as a human I could play the +text adventure version of it right +that's how you would know that it's +being represented well in text and +that's basically what I did is like a +structured text adventure presented to +the LMs um +describing what's around it describing +what you're wearing whatever yeah. +Yeah. So right now when I use some +programs I go oh I want to do kind of +this thing and it used to be that I was +able to Google how do then X in program +Y +and some of it will be like a forum +post. But now you can't find those +anymore or I'm doing difficult stuff. So +my question is +what year will I be able to put like a +good version of Clippy in my programs? +So like there's a little dude you can +actually describe like hey I'm trying to +rotate the mesh in this way what +like how does the sub stuff know what? +There's so many startups trying to do +that. It's Yeah, it's like cursor, +No, well it's not even cursor for X. +It's like there's so many people who put +the little chat window on the thing and +it's like ask it a question. I just feel +like they never work. Like they still +are just so bad most of the time. Like I +don't know when it will get good, but +honestly I like +I find that when I use LLMs to +understand how software works, it gives +me an incomplete understanding because +I'm not really like traversing through +the whole set of knowledge, right? I'm +just specifically getting the answer I +need and implementing it and moving on +with my life. Whereas when I would when +I read documentation and like actually +go line by line and sit down and +understand it, I basically I just know +the unknown unknowns, right? Because I +explored. And so LLMs to say how do I do +this in this program? I I'm actually not +even sure that's good UX, basically. But +if if you can like search the +documentation for So like you need to +take it put the documentation is like a +hundred page PDF documentation of this +program. I'm not going to read that. +Yeah, it is good for that. Like LLMs are +almost like a very it's like a better +Google almost. +Which is why they pivoted their entire +company to LLMs. I might be able to So +at what year can I expect to like put +one of these in my C program? +Where it's like running locally? Yeah, +AI.C. Yeah, cuz I want to Yes, my AI.C +project. To get the guy in who like +sometimes complains Um +Let's say let's say you know smaller +models are always getting better and +better. Let's let's put a let's put a +release date on it and say five years. +So yeah, so have you seen like a model +that sort of can work outside of the +training data? +It feels like they're always like +weighted averages of the training data. +It's It's not a weighted average of the +training data. A better way to describe +how they process and hold and how our +models impacted by training data is +they find patterns and then we run tests +on those patterns to make sure they +generalize and they're not specific to +the training data. So, they find +patterns which are general that do the +thing you want it to do. ARC AGI +specifically is a benchmark for generic +never-before-seen on the internet +patterns that fail LLMs, all of them. +They're private. They're never leaked or +exposed. You can't ever You can't build +an algorithm that interpolates from +existing solutions to solve ARC AGI +ever. And LLMs did that. +Yeah, so so yeah, I was asking sort of +about the answers you get because it +feels like you get sort of you know, you +cannot go outside of what it has been +taught, right? Or sort of +optimized on, +but maybe you could sort of extrapolate +some, you know, outside of it. But is +that happening like Yeah, I I Well, I +think that like when you get to these +absurd levels of scale, um +there's almost nothing outside the box +if it's predicting the entire internet. +Um that's basically like a +a huge amount of examples of uh the +dynamics of human intelligence recorded +in into written form that you then, you +know, process and find patterns which uh +apply across many examples. That's +That's what they're doing. They +Deep learning systems usually start with +very low-frequency generic patterns +like, "Oh, there's a period every +so-and-so words." And then they start to +get to higher-frequency patterns and +more specific, more nuanced things, and +then +they they they're they're basically +pattern finders. Patterns which apply in +on not just the training data. That's +the whole point of being like a deep +learning person is you make sure that +the model does not just apply on the +training data. That's like the point of +all the architectures, that's the point +of like, you know, doing everything +you're doing is making sure that it's +not just memorizing the training data, +that it's finding patterns in the +training data that will apply in your +specific task that you need the model +for. +So, so sort of like say that you +recognize images, +like you can recognize cat and dog. +Yeah. But like a model that can do train +on cat and dog images, but also like +this is something else. Like without you +showing something else. +Yes, I mean obviously that model is not +um learned patterns which would +generalize to is it a turtle or not. It +would try to bucket into cat or dog, cuz +that's what that, +you know, system has done, but you can +feed any image into an LM right now and +it'll tell you what it is. Almost like +exactly like what a person would say. +Obviously it's not perfect and it +constantly fails, um +but that error is always decreasing as +they, you know, +fix bugs, they they improve the training +pipeline, they improve the data they +give it, and then it's always +approaching something that seems like it +will apply in any situation. Um of +course it's very far from that right +now. I'm not meaning to claim that like +you could take any image and then it'll +perfectly recognize what's in it. You +can fool them all the time. Um it's just +that like +it's the entire internet, right? I just +that's the thing that I keep thinking +about. You know, it's that's just a lot. +It's hard for me to even comprehend what +the limits of that is. +So, even if we don't end up in a world +where the AI is able to do our job in +in, you know, 10 years or less, +whatever. Um how likely do you think it +is that we're heading towards a dead +internet style problem where the noise +generated or the noise created by these +generated AIs is a problem to our own +AIs is a problem to our own +ability to communicate? I mean really +already it feels like so much of the +internet is on like private discords +that aren't, you know, on these massive +public forums and, +you know, linked with each other. I +think that people are already kind of +receding away from this like how it used +to be where it was like a like a massive +graph of forum posts or whatever, right? +Like um +I don't know much about that. I mean +this this is the internet and it's not +dead, I guess. So, I it's not I don't +know like I mean maybe. +People I actually Well, so LLMs do seem +like they kind of defeat the Turing test +and I I seem to I think I'm remembering +there was some news article. Somebody +maybe in like Twitch chat or whatever +can find it where I think some lab +basically plugged an LLM into like a +subreddit and just had it be a commenter +and it got a lot of karma and people +started to like relate to it and like +oh, this is a good post or he's a good +guy. I like this guy. And then it was +like uncovered that it was really like a +like a false flag human being or +a false flag human being or +whatever and like they were they they +basically were doing a study on the +subreddit involuntarily. I think people +got very mad at that cuz it was like +that's probably not ethical to do human +beings. Most of what's happening in the +modern world is not ethical to do human +beings, honestly. I I don't think I +don't think LLMs are that crazy compared +to like YouTube. Like you just show +infant pixels to a person like this is +probably not good for them. I don't know +Um I don't really you know, things are +already so so so far from what is like +healthy for a human being. It's like +What are you going to What do you want? +Like you just like it's not that much +worse than TikTok. I don't know what to +say. TikTok is a transformer. It's +So, we talked about this over breakfast +yesterday, but like LLMs kind of cool, +transformers are cool, doing something, +but like there's a missing piece or +maybe many. Do you have any theories +what those might be? +Me personally as a hobbyist +crazy person? Uh +I don't know. I feel like if I were to +Everybody so every every single deep +learning research is like oh, it's going +to be in a reinforcement learning. Like +you take that success of AlphaZero where +it basically can master any system. You +just like apply that to like soft +concepts. Like MuZero it's like you +know, I everything everybody is like +thinking along those lines of like we +basically need computers to +automatically generate and master uh +things that not even humans have +mastered yet, right? Because if all +you're doing is basically making these +models uh try to optimize and find +patterns like that are more and more +closely match human beings, they'll +never exceed the human baseline, right? +I mean, they might be an ensemble of all +the best humans and, you know, they +might have be a genius in every single +field or whatever, but they're never +going to surpass human genius, which you +would kind of want if you wanted to be a +god-like being that replaces mankind, +which is some people's goal, not mine. I +don't I'm not into that, but some people +like that. I don't know. I think it's +probably bad. +But like like +random thought like like what what what +what could that be like? +>> Reinforce you need to have some way of +operating on these concepts um in a way, +some kind of like gym, some kind of like +imaginary like it needs a world model, +right? That's what OpenAI say like, "Oh, +we need a world model." It's like, "We +need some way We need a transformer not +on text." That's why So, +I don't know if you know this, but the +reason why OpenAI did their video model +is because of they're trying to solve +this problem. They want a world model so +that these these AI systems can try to +imagine new scenarios over and over +again and find new ways to be better. +Um and then, you know, if you can +actually even if it's very inefficient +at be being better, right? If you can +somehow correlate the strength of an AI +system one-to-one with like compute, you +we can always get more compute, right? +We can always make Nvidia stock price go +up and like get more chips from Taiwan. +Like it can always we can always keep +ramping this, right? There's a lot of +silicon on Earth and stuff like that, +right? It's like you know, you you can +just keep doing this. You can scale it +and then once you have the compute, it +keeps printing it keeps being smart. And +so, it will it is only a matter of time +until, you know, if you can actually do +self-optimizing intelligent processes +with silicon that it exceeds human +baseline and becomes the god-like figure +some people want to, you know, build +upon the Earth. +Sam, Timothy? +Oh, I I had a question about like is the +future then like oh if if they Is there +a data bottom neck at some point? And is +the future of reinforcement learning? +So, what I'm thinking about is like a +baby +you right spends like a lot of time +receiving input. And I have a theory +that the reason puberty happens, and +like the body grows so quickly, is +because it's to like protect the brain +from injuring itself before it while +it's too dumb. +That kind of makes sense, actually. I +don't I don't know. I So so is is there +a future where like you +let robots like with cameras out in a +field, and then the learning is just +like, what is going to That what's the +next frame of the video? Yes, actually +exactly. That's like what Nvidia's +doing. That's like There's already many +data sets where they basically just +strap a GoPro to the center of a guy +cooking in a kitchen, and there's like, +you know, gigabytes of those videos. And +the idea is just like, oh, we just +basically need a lot of video of people +doing whatever they're doing that's +useful to other people. Throw a bunch of +transformers at that, trillion-dollar AI +company, and then we automate them, and +then we rake in the money, and then +there's no more people to buy cuz +there's no more economy. I don't Kind of +economy. I don't Kind of +just stopped making sense. It's like +really crazy. Like if you really if you +really think about it for a second, +think about it for a second, +right? All these like VC-backed AI +companies, if they really did pan out, +there would be no more economy. Like +what are you competing for by investing +in that? +I I don't know. +Although, um, +concerning like those gigabytes videos +and so on, I would imagine a big big +chunk of like what would make a the +training run effective is like how do +you transform the data before sending it +into the model or into the algorithm, +right? Because like I guess if you like +filter down to what actually matters cuz +like there's like a lot of when you have +a guy cooking or something, like a lot +of that image data is completely +useless. You don't care about it most of +it, right? So like +what kind of pre-filtering do you people +do, and like how much How important is +that? Well, actually people do the +opposite of pre-filtering. Sometimes for +deep learning, actually I think Tesla's +doing this, they don't feed pixels, they +feed like raw sensor measurements. They +would decrease filtering. +There's a There's a very Well, there's a +very famous like blog post, basically, +called the bitter lesson. And the bitter +lesson is that like a general method +that scales with compute will always +perform better than a specific method +made by engineers because compute is +always getting cheaper and more +plentiful. Engineers are becoming more +scarce as people become stupider from +the AI that you're creating and doing +this. Yeah, but but then by choosing to +use like sensors, like you're making +that's like a deliberate choice. Like +instead of using pixels, we're going to +use sensors. So, it's like maybe more +relevant information within it. So, like +the pattern might be easier to extract +somehow or No, it's not that it's +the idea is that we should not try to +design these massive massive, you know, +trillion parameter deep learning models. +We should not try to do any data +transformation on what's happening in +the universe because it will find the +representation that's most effective for +minimizing the loss, right? It will do +that with compute. It will do It will +find something better than what we could +ever filter it, right? And what +basically you're describing is I think +called feature engineering. We're like, +"Oh, instead of feeding the images to +the deep learning model, we feed we we +try to detect like where the cones are +and stuff and like the road and then +like, you know, we feed those features +into the model." And it always +underperforms just feeding in the raw +sensory data. If you have enough compute +to scale up because the compute will +find the representation that is most +effective to output the policy +basically, I think. So, but you can +still select the which what kind of +sensor data you feed it in. Like the +sensor is is different from sensor data +is is different from sensor data +from like vision, right? Yeah. You still +should make a choice and like that +choice will matter, right? Yeah, the +choice matters in the opposite direction +matters in the opposite direction +from what you're saying. You want more +more generic, more bigger amount of +data. So, you feed it everything, every +type of sensor data you can get your +hands on. You just feed it whatever you +can. Yes, ideally yes, but like what +they found actually I I think there was +this paper called like the Nemo paper. I +don't remember what it was, but they +basically had a transformer they just +found that as you taught it many random +disparate games, it would like improve +on the other ones, right? That were +seemingly unrelated. And so, there's +something just about transformers. The +more data that you give them, the better +the representations become no matter +what. I think it's because there's +there's something kind of like a a +universal prior or like +you know, it's just the universe is +ordered in a particular way that's +evident in everything that happens in +it. So, the more in- information that +comes in This is not This is me not not +by the way. This is not a true I'm like +theory I'm like crafting theories. It's +like maybe there's like something about +the universe that just whatever you give +it that's from the universe truly and +not you know like a craft of man or +whatever like it it's just somehow +builds better representations cuz it's +something like that. I don't know. +>> there's universal laws and patterns and +such that just exist in reality, right? +And like that's maybe what you're what +it's using somehow? That's maybe yeah. +I'm not I'm really not sure. That's +I would I would like to see more +rigorous attempts to answer that +question rather than me like theory +crafting. +Are there any more questions? +I mean just earlier like uh +said I you beat me to it, but I was +going to ask about like uh just the +noise like the back of noise of AI +generative AI on the internet, right? +Because like if you lower the friction +for like creating content, but like you +said it's like you're using existing +data transformations to create things, +then I mean wouldn't the problem be like +you just end up with so much stuff that +is all from old transformations? It's +really hard to find novelty among that. +So like Um search engines get good +enough that like we don't have a problem +or like No, there there's just very easy +ways to detect if something was output +by AI for the most part or like ways to +filter for data that is like the most +confusing for the model. And so these +things are not going to be a serious +problem for deep learning. +at scale. Detecting AI or detecting +detecting information that AI has not +seen before that's like possible so this +that's something that's like easy you +say? Yeah, you can just tell which you +know training seems to be the most +confusing to it. Maybe this problem is +more complicated than +what I'm thinking because like maybe +I don't I don't know. I'm sure there's +there's much smarter deep learning +researchers who have answered that +question already probably, but my my +guess would be that like +it's probably not a real issue. Okay, +because if you could just filter out +everything that's AI generated it'd be +really great, you know. Yes, they can I +think you can do that, yeah. Oh yeah. +I hope so. +All right, we're out of time. We're out +of time. Round of applause for Krikey. +>> Yeah. \ No newline at end of file diff --git a/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/video.log b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/video.log new file mode 100644 index 00000000..9462c849 --- /dev/null +++ b/conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/video.log @@ -0,0 +1,16 @@ +# yt-dlp log +# url: https://youtu.be/yxkUvXs-hoQ +# output: conductor/tracks/video_analysis_creikey_dl_cv_20260621/artifacts/video.mp4 +# returncode: 0 + +stdout: +[youtube] Extracting URL: https://youtu.be/yxkUvXs-hoQ +[youtube] yxkUvXs-hoQ: Downloading webpage +[youtube] yxkUvXs-hoQ: Downloading android vr player API JSON +[info] yxkUvXs-hoQ: Downloading 1 format(s): 400+251 +[download] video.mp4.f400.mp4 +[download] video.mp4.f251.webm +[Merger] Merging formats into video.mp4 + +stderr: +WARNING: yt-dlp EJS not enabled; some formats may be missing.