Weights' relationships, not substrate, hold LLM intelligence, researcher argues
teortaxesTex · x · 2026-09-26
Researcher teortaxesTex argues people overindex on the substrate of neural networks: weights are mere numbers, but the relationships between weights are the real information, containing functions of unclear expressivity bounds.
Responding to the claim that LLMs are "gigantic memory and retrieval systems that can't invent solutions to new problems," he counters that von Neumann was likewise a memory-and-retrieval system—yet that's often the mark of creative genius in humans. His core intuition: LLMs are shapes, not code or mechanics; weights describe the shape of the model's representation space, and respecting that is the basic prerequisite for understanding LLMs.
Related event: Community Pushes Back on LeCun's "LLMs Are Just Retrieval" Claim(2 posts)→
More from AGI Musings
- The white pill on flag planting: 100k humans with AI agents will find good affordances — generativist · 2026-09-26
- Flag planting: how the AI space rushes to own the interface of every new model — generativist · 2026-09-26
- French Prix Goncourt drops novel accused of being 'almost entirely' AI-written — nordicinst · 2026-09-26
- Claim: China's AI progress is 'stunning', US has ~7 months to deploy open source plans — TinfoilTricorn · 2026-09-26
- Hype and doom are the same pitch: insiders see the bubble and want to exit first — gerardsans · 2026-09-26
- Ex-OpenAI policy lead: friends flood him with AI worries, and he can't tell them it's overblown — Miles_Brundage · 2026-09-26