Private
Public Access
0
0

conductor(platonic_intelligence_kumar): Phase 3 OCR - 62 frames OCR'd via winsdk in 3.7s

This commit is contained in:
2026-06-21 22:33:09 -04:00
parent f690b4dea4
commit f32e4fd268
3 changed files with 936 additions and 0 deletions
@@ -0,0 +1,932 @@
# OCR Results
## frame_00001.jpg
```
The World is not Random
```
## frame_00002.jpg
```
Towards a Platonic Intelligence with
Unified Factored Representations
sae
Good Adaptability
Unified Factored
Representation
Open—Ended
Search
Identical
output
behavior
Inputs
Solves Task
Function Space
4
Poor Adaptability
Fractured Entangled
Representation
Conventional
SGD
Akarsh Kumar
MIT CSAIL
November 4, 2025
1
```
## frame_00004.jpg
```
The World has Structure
Real World
```
## frame_00008.jpg
```
(no text extracted)
```
## frame_00016.jpg
```
Intelligent Agents must capture this Structure
```
## frame_00017.jpg
```
Capturing Structure with AI
```
## frame_00019.jpg
```
Capturing Structure with AI
• What about everything else?
• Example: lighting invariance
• How do you capture lighting invariance?
Architecture
Lighting
Invariance
6
```
## frame_00020.jpg
```
Capturing Structure with AI
• What about everything else?
• Example: lighting invariance
• How do you capture lighting invariance?
• We don't know
• Solution: train on lots of data with SGD
Architecture
Lighting
Invariance
```
## frame_00021.jpg
```
Does this Work?
ChatGPT
```
## frame_00022.jpg
```
Hypothesis: Fractured Entangled Representations
8
```
## frame_00023.jpg
```
Hypothesis: Fractured Entangled Representations
• Conventional SGD training finds
neural representations which are
fractured and entangled
output
behavior
Inputs
Solves Task *
Function S ace
Fractured Entangled
Representation
Conventional
SGD
8
```
## frame_00024.jpg
```
Hypothesis: Fractured Entangled Representations
• Conventional SGD training finds
neural representations which are
fractured and entangled
• Doesn't capture the underlying
regularities of the world
• Position: Open-Ended Search
may be the solution to learn
unified and factored neural
representations
Unified Factored
Representation
Open—Ended
Search
Identical
output
behavior
Inputs
Solves Task
Fractured Entangled
Representation
Conventional
SGD
Function S ace
8
```
## frame_00025.jpg
```
Hypothesis: Fractured Entangled Representations
• Conventional SGD training finds
neural representations which are
fractured and entangled
• Doesn't capture the underlying
regularities of the world
• Position: Open-Ended Search
may be the solution to learn
unified and factored neural
representations
• Internal representation affects
generalization, creativity, and
continual learning
Good Adaptability
Unified Factored
Representation
Open—Ended
Search
Identical
output
behavior
Inputs
Solves Task *
Function S ace
Poor Adaptability
Fractured Entangled
Representation
Conventional
SGD
8
```
## frame_00026.jpg
```
Compositional Pattern Producing Network
(CPPN)
```
## frame_00027.jpg
```
Compositional Pattern Producing Network
(CPPN)
Toy domain to study neural representations: implicitly represent an image
• Inspired by biological developmental process
```
## frame_00028.jpg
```
CPPNs are an Analogy
```
## frame_00029.jpg
```
Picbreeder!
restart TilJtate
SEjve
11
```
## frame_00030.jpg
```
Picbreeder!
• Evolve the underlying CPPNs
restart
11
```
## frame_00031.jpg
```
Picbreeder!
• Online website for humans to breed images to
their desire
• Evolve the underlying CPPNs
restart
11
```
## frame_00032.jpg
```
Picbreeder!
z. zzz
• Online website for humans to breed images to
their desire
• Evolve the underlying CPPNs
uzzzz
zzzzz
restart
11
```
## frame_00033.jpg
```
Picbreeder!
• Online website for humans to breed images to
their desire
• Evolve the underlying CPPNs
restart
11
```
## frame_00034.jpg
```
Picbreeder!
• Online website for humans to breed images to
their desire
• Evolve the underlying CPPNs
• No end goal, do whatever you want!
oosæ
restart
11
```
## frame_00035.jpg
```
What Do You Expect to Find?
```
## frame_00036.jpg
```
What People Actually Found!
13
```
## frame_00037.jpg
```
Why Greatness Cannot be Planned
• Many insights on the nature of search
• Deception
• Serendipity
• Open-Endedness
• Case studies:
• Natural Evolution
• Scientific Innovation
Kenneth 0. Stanley • Joel Lehman
Why Greatness
Cannot Be Planned
The Myth of the Objective
@Springer
14
```
## frame_00038.jpg
```
Picbreeder has Intriguing Properties
Open-Ended
```
## frame_00039.jpg
```
Picbreeder has Intriguing Properties
Open-Ended
oäaua e
ana
16
```
## frame_00040.jpg
```
Picbreeder has Intriguing Properties
Serendipitous Exaptation
• When a trait evolved for one function but gets repurposed for another function
Warmh
Gliding
17
```
## frame_00041.jpg
```
Picbreeder has Intriguing Properties
Serendipitous Exaptation
Stepping stone to the Teapot
Stepping stone to the Skull
Stepping stone to Jupiter
Stepping stone to the Butterfly
Stepping stone to the Penguin
Stepping stone to the Lamp
18
```
## frame_00042.jpg
```
Picbreeder has Intriguing Properties
Emergence of Evolvability
• Natural evolution has developed adaptable genotypes:
• Canalization
• Regularity
Modularity
• Symmetry
• Certain axes of variation become more likely while
others become impossible
synurwtry
synvnetry
human
19
```
## frame_00043.jpg
```
Picbreeder has Intriguing Properties
Emergence of Evolvability
Most Picbreeder images
feature forms of canalization
Mutating single genes of these images
holistically affects distinct aspects of the image
mutation in
Nut.uon
Mutation in
ObJect Oni' Spatnght Only
These images have a structurally organized
(i.e. modular and hierarchical) genome
Global Lighting
genes
Shadow Only
Spotlight Only
Some Picbreeder images
show oor canalization
Mutating single genes of these images
affects none or many parts of the image
»ut.uøn Mutation in Nuuuon in
And
These images have little structural
organization in their genome
genes
Object Only
genes
genes
Shadow And
Object genes
No Effect
Global
Distortion
genes
Dolphin Eye
And Body
genes
These images have fit decendants
These images have unfit decendants
20
```
## frame_00045.jpg
```
Learning the Picbreeder Skull with SGD
22
```
## frame_00046.jpg
```
Learning the Picbreeder Skull with SGD
Picbreeder Skull
22
```
## frame_00047.jpg
```
Learning the Picbreeder Skull with SGD
• Let's train a conventional network to recreate the skull
• Perfect reconstruction!
Picbreeder Skull
SGD Trainin
SGD Skull
22
```
## frame_00048.jpg
```
Layerization
23
```
## frame_00049.jpg
```
Layerization
• Convert everything to a universal architecture space: MLP
sin(0)
do
cos(O)
bias
dx
lio
itie
-Afia
lin
ReLO •
tanh
23
```
## frame_00050.jpg
```
Layerization
• Convert everything to a universal architecture space: MLP
• Existence proof of Picbreeder solution MLP weight space
e
DLI
Neuron
24
```
## frame_00051.jpg
```
Picbreeder Skull
Unified Factored Representation
e
Neuron
ssgg
anon
25
```
## frame_00052.jpg
```
SGD Skull
Fractured Entangled Representation
Zigle
aaaa
Neuron
26
```
## frame_00053.jpg
```
Picbreeder Skull
Unified Factored Representation
Controls Mouth Opening
Controls Eye Winking
Controls Eye Width
Controls Jaw Width
Sweeping Weight Value
SGD Skull
Fractured Entangled Representation
Sweeping Weight Value
27
```
## frame_00054.jpg
```
Picbreeder Butterfly
Unified Factored Representation
Neuron
ooaa
28
```
## frame_00055.jpg
```
SGD Butterfly
Fractured Entangled Representation
Neuron
29
```
## frame_00056.jpg
```
Picbreeder Butterfly
Unified Factored Representation
Controls Wing Area
Controls Color
Converts Butterfly to Fly
Controls Vertical Shape
Sweeping Weight Value
SGD Butterfly
Fractured Entangled Representation
Aw= —1
Sweeping Weight Value
30
```
## frame_00057.jpg
```
Picbreeder Apple
Unified Factored Representation
Neuron
nannæg••
31
```
## frame_00058.jpg
```
SGD Apple
Fractured Entangled Representation
ooose•n
oec
Cloe
D •BRA
x yd 1
Neuron
32
```
## frame_00059.jpg
```
Picbreeder Apple
Unified Factored Representation
Controls Stem Angle
ooooo
Controls Apple Size
Cleans Background
ooooo
Removes Stem
AYI/ = +1
Sweeping Weight Value
SGD Apple
Fractured Entangled Representation
ecoe
Controls Apple Size
= Cleans Background
ooooe
Aw = +1
Sweeping Weight Value
33
```
## frame_00060.jpg
```
How does this Apply to LLMs?
34
```
## frame_00061.jpg
```
FER In LLMs
Evidence in GPT-3
Example 1:
Me: I have 3 pencils, 2 pens, and 4 erasers. How many things do I
have?
GPT-3: You have 9 things. [always correct]
Example 2:
Me: I have 3 chickens, 2 ducks, and 4 geese. How many things do I
have?
GPT-3: You have 10 animals total. [always incorrect]
35
```
## frame_00062.jpg
```
GSM-Symbolic: Understanding the Limitations of
Mathematical Reasoning in Large Language Models
Iman Mirzadeht Keivan Alizadeh Hooman Shahrokhi*
Samy Bengio Mehrdad Farajtabart
Oncel Tuzel
Apple
Abstract
Recent advancements in Large Language Models (LLMs) have sparked interest in their formal
reasoning capabilities, particularly in mathematics. The GSM8K benchmark is widely used
to assess the mathematical reasoning of models on grade-school-level questions. While the
performance of LLMs on GSM8K has significantly improved in recent years, it remains unclear
whether their mathematical reasoning capabilities have genuinely advanced, raising questions
about the reliability of the reported metrics. To address these concerns, we conduct a large-
scale study on several state-of-the-art open and closed models. To overcome the limitations of
existing evaluations, we introduce GSM-Symbolic, an improved benchmark created from symbolic
templates that allow for the generation of a diverse set of questions. GSM-SymboIic enables
more controllable evaluations, providing key insights and more reliable metrics for measuring the
reasoning capabilities of models.Our findings reveal that LLMs exhibit noticeable variance when
responding to different instantiations of the same question. Specifically, the performance of all
models declines when only the numerical values in the question are altered in the GSM-Symbolic
benchmark. Furthermore, we investigate the fragility of mathematical reasoning in these models
and demonstrate that their performance significantly deteriorates as the number of clauses in
a question increases. We hypothesize that this decline is due to the fact that current LLMs
are not capable of genuine logical reasoning; instead, they attempt to replicate the reasoning
steps observed in their training data. When we add a single clause that appears relevant to the
question, we observe significant performance drops (up to 65%) across all state-of-the-art models,
even though the added clause does not contribute to the reasoning chain needed to reach the
final answer. Overall, our work provides a more nuanced understanding of LLMs' capabilities
and limitations in mathematical reasoning.
36
```
## frame_00063.jpg
```
Reasoning or Reciting? Exploring the Capabilities and Limitations of
Language Models Through Counterfactual Tasks
Zhaofeng Wue Linlu Qiue Alexis Rosse EkinAkyürekC Boyuan Chene
Bailin Wange Najoung Kima Jacob Andrease Yoon Kim
e MIT O Boston University
zfw@csail.mit.edu
G PT-4
Performance
Default
Counterfactual
Spatial
COO Of
Arithmetic
Drawing
Draw a
Code Exec.
pytim
L"bo". "nb"l
"bo"l
Chord Fingering
play C
a guitar
Code Gen.
Sort list by the
second c
in
Note in Melody
The 4th note of
Twinkle Twinkle
in C major
in Ai
Basic Syntax
Find the main
subject and verb
think 'MS
(they. think)
(they, 'hink)
IS the move
TX are Y, Y are Z.
Are X Z?
SET Game
Rule:
or
•hap..
Figure 1: GPT-4's performance on the default version of various tasks (blue) and counterfactual counterparts
The shown results use O-shot chain-of-thought prompting (54; GPT-4 consistently
orange
and substantially underperforms on counterfactual variants compared to default task instantiations.
37
```
## frame_00068.jpg
```
On the Biology of a Large Language Model
We investigate the internal mechanisms used by Claude 3.5 Haiku — Anthropic's lightweight
production model — in a variety of contexts, using our circuit tracing methodology.
add —57
95
sum = _5
add _9
Inputs near 30 make this
early feature fire
-30
calc :
36
calc :
Hover to see
visualizations!
Example
low precision
features
_6
36
36+59=
sum —92
-40 + -50
5-
-59
sum =
-36 + -60
59
59
Exam mod O
eat ures
_9
Ull
Sum Features
The model has finally computed information
about the sum: its value mod 10, mod 100,
and its approximate magnitude.
Lookup Table Features
The model has stored information about
particular pairs of input properties. They take
input from the original addends (via attention)
and the Add Function features. Operand plots
ssibly with repetition (modular) or
are points.
smearing (tw•preciston)
Add Function Features
The model separately determines the ones
digit Of the number to be added and its
approximate magnitude. Operand plots show
vertical or horizontal stripes,
Input Features
The model has features specific to the ones
digit and Io the approximate magnitude, at
various scales.
Most computation
takes place on the
•zu token
38
```
## frame_00069.jpg
```
Scaling helps... but in what way?
Scaling Laws for Neural Language Models
2024
10—9
L = (Cmin/2.3 •
10-7
10-5
10-3
10—1
Compute
PF-days, non-embedding
4.2
3.9
3.6
3.3
3.0
2.7
101
L = (0/5.4
108
Dataset Size
tokens
Kaplan et al. (2020)
5.6
4.8
3.2
2.4
109
105
L = (N18.8 •
107
Parameters
non-embedding
109
39
```
## frame_00070.jpg
```
Scaling helps... but in what way?
Platonic Representation Hypothesis
The Platonic Representation Hypothesis
Neural networks, trained with different objectives
on different data and modalities, are converging to a
shared statistical model of reality in their representa-
tion spaces.
z
x
Img
A red sphere next to
ft ext
Huh et al. (2020)
40
```
## frame_00071.jpg
```
Scaling helps... but in what way?
41
```
## frame_00072.jpg
```
What could be better?
42
```
## frame_00073.jpg
```
What could be better?
• Complexification (ex: morphogenesis, etc.)
• Builds regularities on top of other regularities (bottom up)
• Emergence
• Adaptability
• Pressures the learned regularities to be robust to environmental changes
• Representation must capture axes of variation which "carve nature at its joints"
• Serendipity (order matters for learning!)
• Much higher chance of finding a useful learning curriculum
• What learning paradigm captures all of these? Open-Endedness!
Function Space
FER
olution Space
of Skull
UFR
: Open—Ended
Search
Conventional
SGD
42
```
## frame_00074.jpg
```
Is this a Platonic Intelligence?
Space of Forms
Real World
Intelligent Agents
Unified Factored
Representation
43
```
## frame_00075.jpg
```
Is this a Platonic Intelligence?
Aspirational Ideal
Unified Factored
Representation
Instantiation
Fractured Entangled
Representation
44
```
## frame_00076.jpg
```
Collaborators
Jeff Clune
UBC
Vector Institute
Joel Lehman
University of Oxford
Kenneth Stanley
Lila Sciences
UBC
45
```
## frame_00077.jpg
```
Thank You!
46
```
## frame_00081.jpg
```
Thanh You!
```
## frame_00082.jpg
```
(no text extracted)
```
@@ -0,0 +1,2 @@
Phase 2 Keyframes for C:\projects\manual_slop\conductor\tracks\video_analysis_platonic_intelligence_kumar_20260621\artifacts\video.mp4
OK: kept 62 frames
@@ -0,0 +1,2 @@
Phase 3 OCR for C:\projects\manual_slop\conductor\tracks\video_analysis_platonic_intelligence_kumar_20260621\artifacts\frames (winsdk)
OK: OCR'd 62 frames in 3.7s