conductor(platonic_intelligence_kumar): Phase 4 Synthesis - report.md (1564 lines, 104KB) + summary.md (384 words)

conductor/tracks/video_analysis_platonic_intelligence_kumar_20260621/summary.md

@@ -0,0 +1,25 @@
+# Summary: Towards a Platonic Intelligence (Kumar)
+
+**Source:** https://youtu.be/1mXUFweWOug
+**Author:** Akarsh Kumar (MIT CSAIL)
+**Track:** Child #5 of `video_analysis_campaign_20260621`
+**Cluster:** B (Platonic / geometric AI representations)
+**Pass:** 1 of 3 (research-only deep-dive)
+
+---
+
+## One-paragraph synthesis
+
+Kumar presents a **position paper** arguing that conventional SGD finds **Fractured Entangled Representations (FER)** — neural networks producing the right output behavior but with weights that do not correspond to human-interpretable factors — and proposes **open-ended search** as a paradigm that finds **Unified Factored Representations (UFR)**. Evidence from Picbreeder (Stanley & Lehman, 2008): when humans evolve CPPNs by selecting images they like with no fixed objective, the resulting CPPNs layerize to MLPs where each weight controls one semantic axis (mouth opening, eye width). SGD-trained MLPs reproducing the same images have entangled weights. The FER diagnosis predicts three LLM failure modes: GPT-3's chicken/duck/pencil counting failure (Mirzadeh et al. GSM-Symbolic), GPT-4's counterfactual-task degradation (Wu et al. "Reasoning or Reciting"), and Claude 3.5 Haiku's magnitude-heuristic arithmetic (Anthropic circuit tracing) — bag-of-heuristics rather than unified algorithms. Kumar critiques the **Platonic Representation Hypothesis** (Huh et al. 2024) as a **statistical** observation that doesn't address factorization. The proposed solution is **open-endedness** with four properties: complexification (building regularities on top of regularities, bottom-up), emergence, adaptability (training for the ability to adapt, not for any specific task — the most important driver per the author), and serendipity. **Backward connections:** score_dynamics_giorgini (alternative route via score matching), entropy_epiplexity (algorithmic-information perspective), cs229_building_llms (the SGD paradigm critiqued). **Forward connections:** cs336_architectures (scaling doesn't fix FER), creikey_dl_cv (DDPM as alternative), free_lunches_levin (open-endedness + information theory).
+
+---
+
+## Three key takeaways
+
+1. **SGD finds FER, open-ended search finds UFR** — same loss, same architecture, completely different internal organization. Picbreeder CPPNs layerize to MLPs where each weight controls one semantic axis; SGD MLPs have entangled weights.
+2. **FER predicts LLM jagged intelligence** — GPT-3 fails at chicken/duck counting (surface pattern), GPT-4 fails at counterfactual tasks, Claude uses magnitude heuristics not digit-by-digit arithmetic. Correct answers, wrong internal mechanisms.
+3. **Pressure to adapt creates UFR** — the most actionable conjecture. Train for adaptability to changing environments (not any specific task), and representations will capture underlying regularities.
+
+---
+
+*Pass 2 (de-obfuscation via user's mathematical encoding) to follow.*