Adam Marblestone – AI is missing something fundamental about the brain

Adam Marblestone argues AI’s key missing ingredient is the brain’s complex, evolution-encoded reward functions — not architecture or learning algorithms.

• Evolution encoded thousands of bespoke reward/cost functions into subcortical brain regions, not the cortex — this is what makes biological learning so sample-efficient.
• The genome encodes reward functions compactly (like Python code), not full world models — explaining how ~3GB of DNA produces human-level generalization.
• Steering subsystem (hypothalamus, amygdala) has its own primitive sensory system; cortex learns to predict it, enabling abstract concepts like ‘Yan LeCun disapproves’ to trigger innate shame responses.
• Subcortical steering regions contain far more diverse, bespoke cell types than cortex — biological evidence that reward specification is where evolution invested complexity.
• Marblestone proposes cortex may do omnidirectional inference (predict any variable subset from any other), unlike LLMs which only predict next token — closer to Yann LeCun’s energy-based models.
• Formal math verification via Lean + RLVR is already producing billion-dollar companies (Harmonic, AlphaProof); Marblestone expects proof-search to largely automate mechanical mathematics.
• Convergent Research’s gap map identified ~hundreds of fundable infrastructure gaps across science, each roughly series-A scale — math proving infrastructure was a surprise entrant alongside bio/neuro tools.

2025-12-30 · Watch on YouTube