Adam Marblestone – AI is missing something fundamental about the brain

Core Hypothesis: The Learning vs. Steering Subsystem Framework

• The brain's capability gap with LLMs stems not from data volume or architecture alone, but from the specificity of its loss (cost) functions and reward signals, which evolution has encoded over millions of years.
• Steve Burrows' Theory: The brain operates via a dual-subsystem architecture:
  • Learning Subsystem (Cortex): A general-purpose, omni-directional inference engine capable of modeling any subset of variables given any other subset, rather than predicting a single next token.
  • Steering Subsystem (Subcortical/Amygdala/Hypothalamus): An innate, pre-wired set of "thought assessor" cells and reflexes that generate reward signals and biological drives (e.g., fear, social status, hunger).
• Loss Function Complexity: Evolution likely encoded complex, multi-stage learning curriculums directly into the brain's loss functions, allowing for high sample efficiency and generalization that current simple "next-token prediction" models lack.
• Amortized Inference vs. Biological Constraints:
  • Digital minds can "amortize" inference (distill complex reasoning into weights) because models can be copied and edited; biological minds cannot easily copy their internal states, so evolution must "hard-wire" critical behaviors via the steering subsystem.
  • The brain likely uses a mix of amortized forward passes for perception and non-amortized sampling (probabilistic/energy-based models) for complex decision-making, though the exact balance remains unknown.

Neuroscientific Evidence and Biological Architecture

• Cellular Heterogeneity: The steering subsystem contains a "gazillion" diverse cell types (thousands of bespoke types) dedicated to specific innate responses (e.g., a specific cell type for "flinching at a skittering spider"), whereas the learning subsystem (cortex) uses a more repeating, general-purpose architecture.
• Genomic Efficiency: The human genome (approx. 3GB) contains a tiny fraction of code relevant to brain wiring; this is explained by the fact that evolution encodes compact "Python-like" reward functions and cell-type specifications rather than detailed neural connection maps.
• Predicting the Steering System: The cortex learns to predict the state of the steering subsystem (e.g., "Am I about to flinch?") using abstract inputs (like the word "spider"), allowing generalization to novel situations (hearing about a spider on the back) without direct sensory exposure to the threat.
• Multi-Modal and Omni-Directional Inference: The cortex may function as a multimodal foundation model where any area can predict any other area's variables (e.g., predicting vision from audition), unlike LLMs which are unidirectional (input -> next token).
• Hardware Co-Design: Biological hardware constraints (20 Watts, 200Hz, stochastic neurons) force the brain to co-design algorithms with physics, potentially utilizing cellular machinery (e.g., cerebellar time-delay storage) rather than just synaptic weights.

Implications for AI and Research Timelines

• Multi-Agent Scaling Experiment:
  • An experiment comparing single-agent AlphaZero training vs. splitting compute across 16 diverse agents showed that diversity yields higher performance; the best agent in the 16-agent population (using 1/16th the compute) outperformed the single-agent model trained on the full budget.
  • This suggests future AI scaling may benefit from co-evolutionary leagues and diverse strategies rather than massive single-model scaling.
• The "Connectome" Investment Case:
  • Cost Reduction: Projects like E11 Bio aim to reduce mouse connectome mapping costs from billions to tens of millions via optical microscopy (molecularly annotated connectomes), moving toward a "Human Connectome Project" priced at hundreds of millions to low billions.
  • Scientific Value: A connectome provides "hard constraints" to distinguish between competing AI theories (e.g., backprop vs. energy-based models) and map the wiring of the steering subsystem.
  • Timeline: The speaker estimates a 10-year timeline for a transformative paradigm shift driven by understanding these biological principles, rather than the 3-year timelines often cited for current LLM scaling.
• Brain-Augmented Training (Behavior Cloning on Steroids):
  • A proposed method to train AI involves adding an auxiliary loss function where the model predicts its own neural activity patterns when processing data, not just the labels.
  • This forces the AI to learn representations consistent with human brain geometry, potentially improving generalization and robustness against adversarial examples.

Mathematics, Formal Verification, and AGI

• Lean and Formal Proofs:
  • Mathematical proofs formalized in Lean (a verifiable programming language) provide a "perfect RLVR" (Reinforcement Learning with Verifiable Rewards) signal.
  • AI can search for proofs (like AlphaGo for Go), automating the mechanical validation of theorems, though generating new conjectures (creativity) remains a challenge.
• Provable Software Security:
  • Formal verification via Lean can create "unhackable" software by mathematically proving security properties, a field currently hindered by the "specification problem" (defining formal specs for complex systems).
• Symbolic vs. Neural Representations:
  • The speaker hypothesizes that the brain does not use a symbolic language for its world model but rather a "huge mess" of learned cost functions and architectural constraints that approximate symbolic operations.
  • Future AGI civilization might rely on provable symbolic interactions for safe collaboration, filtering LLM "cleverness" through formal verification to ensure alignment and prevent social engineering attacks between agents.

Specific Research Gaps and Infrastructure Needs

• The "Gap Map": Convergent Research identified ~300 "fundamental capabilities" (infrastructure gaps) in science that, if solved, would unlock massive progress (analogous to the Hubble Space Telescope for astronomy).
  • Surprising Gaps: Includes formal math proving infrastructure and scalable data collection tools, not just biology.
  • Funding: Most require focused engineering teams and billions in investment, which traditional academic grants cannot easily provide.
• Neuroscience Technology Scaling: Just as the Human Genome Project drove down sequencing costs, a concerted effort to map connectomes must parallelize data collection to become scientifically tractable.
• Interpretability Limits: Fully understanding a neural network's internal logic (finding "Gold Gate Bridge circuits") may be impossible; the goal is to describe the system via architecture, learning rules, initialization, and loss functions rather than tracing individual neuron pathways.