Karpathy: Future LLMs Will Shrink to 'Cognitive Core' by Shedding Knowledge

ZeroStateReflex · x · 2026-08-21

Andrej Karpathy discusses the 'cognitive core' of LLMs—small models that sacrifice encyclopedic knowledge for capability. He argues that limiting or stripping away memory improves generalization, preventing over-reliance on memorized data. Models must first scale up to help refactor training data into synthetic formats before shrinking. Human inability to memorize easily is framed as a feature (regularization) rather than a bug in this context.

Related event: Karpathy Proposes "Cognitive Core" LLMs: Small Models That Strip Encyclopedic Knowledge(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →