Adam Marblestone – AI is missing something fundamental about the brain

Marblestone argues that evolution has built extensive complexity into loss functions rather than architecture, encoding specific curricula for different brain regions. Unlike ML's simple loss functions (next token prediction), the brain may use many specialized cost functions activated at different developmental stages, essentially 'Python code' for learning curricula that evolution has optimized over millions of iterations.

• •Machine learning uses mathematically simple loss functions while evolution may encode complex, stage-specific loss functions
• •The cortex might implement omnidirectional prediction - predicting any subset of variables from any other subset, not just forward prediction
• •This is similar to energy-based models and probabilistic AI approaches that Jan LeCun advocates
• •The brain predicts not just sensory data but also innate behaviors and physiological responses from the 'lizard brain'

Addresses Ilya Sutskever's question about how genomes encode high-level desires. Marblestone explains Steve Byrnes' theory: the brain has a 'steering subsystem' (innate responses) and 'learning subsystem' (cortex). Evolution wires learned features to innate rewards by having the cortex predict steering subsystem responses, allowing abstract concepts like 'embarrassing a famous scientist' to trigger innate shame responses despite evolution never encountering podcasts.

• •The steering subsystem has primitive sensory systems (like superior colliculus for face detection) with innate heuristics
• •Parts of cortex/amygdala learn to predict these innate responses, creating 'thought assessors'
• •Neurons that predict innate responses become the substrate for abstract concepts to trigger appropriate rewards
• •This solves the generalization problem: word 'spider' can activate spider-flinch reflex through learned prediction, not just visual spiders

Explores how neural networks 'amortize' Bayesian inference - instead of sampling possible causes for observations, they directly map observations to likely causes. Discusses whether brains do true probabilistic sampling or amortized inference, and how test-time compute in LLMs relates to this distinction. Digital minds can copy amortized solutions, potentially changing what's worth building in versus computing at runtime.

• •Amortization means pre-computing inference patterns into network weights rather than sampling at test time
• •Test-time compute (like chain-of-thought) may be doing the sampling that amortization skips
• •Capabilities requiring inference-time compute can be distilled back into models through training
• •Digital minds' copyability creates different trade-offs than biological evolution faced

Explains how evolution achieves sample-efficient learning with minimal genomic information (3GB genome, small fraction for brain). The key is that reward functions are compact 'Python code' that exploit the learning subsystem's generalization. Single-cell atlases reveal the steering subsystem has vastly more diverse, bespoke cell types than cortex, suggesting most genomic complexity goes into specifying innate behaviors and reward circuits, not the learning algorithm itself.

• •Steering subsystem has thousands of specialized cell types vs cortex's relatively uniform architecture
• •Each reward function may need distinct cell types for innate wiring without learning
• •Cortex is mostly replicated architecture (like adding transformer layers), requiring little genetic code
• •This explains rapid hominid brain expansion: scaling up existing cortical architecture is genetically cheap

Current LLMs use primitive RL (upweighting entire successful trajectories) compared to decade-old techniques like Q-learning with value functions. The brain likely implements both: basal ganglia doing simple model-free RL with finite action spaces, and cortex building world models that include predictions of rewards. Some theories suggest 'RL as inference' where you sample plans conditional on high reward.

• •Basal ganglia may implement simple model-free RL for motor and cognitive actions
• •Dopamine provides reward prediction error signals (temporal difference learning), not just raw reward
• •Cortex can model when rewards occur, enabling more sophisticated planning
• •RL as inference: clamp reward as high and sample backwards to infer plans that would achieve it

Biological brains face energy constraints (20W, 200Hz) but may have advantages in co-locating memory and compute, unstructured sparsity, and natural stochasticity for sampling. The inability to copy or directly access brain states is a major disadvantage. Future AI hardware may adopt brain-like features (low voltage, co-located memory) while keeping digital advantages (copyability, random access).

• •Brain's inability to be copied or randomly accessed is fundamental limitation
• •Neurons' natural stochasticity may be optimized for probabilistic sampling without explicit random number generation
• •Co-located memory and compute, low voltage operation are likely advantages worth copying
• •Probabilistic AI researchers argue test-time compute is inherently necessary, not just a current limitation

Most cellular machinery in neurons is implementation detail for executing learning rules (like weight normalization or memory consolidation) rather than fundamentally new algorithms. However, some examples like cerebellar timing circuits suggest cells can store computational primitives (time constants) that would require complex wiring in pure connectionist models.

• •Weight normalization in digital networks is trivial; in neurons requires complex cellular machinery
• •Memory consolidation and synaptic plasticity involve extensive cellular changes without adding algorithmic complexity
• •Cerebellum stores time delays in cell bodies rather than requiring longer synaptic chains
• •Main algorithmic story is still synaptic weight changes, not radical cellular computation

E11 Bio is developing optical connectomics to reduce mouse brain mapping costs from billions to tens of millions of dollars. Unlike electron microscopy, optical approaches enable 'molecularly annotated connectomes' showing synapse types and cell properties. The strategy mirrors the Human Genome Project's shift from expensive techniques to massively parallel sequencing, aiming for million-fold cost reductions through technology development before large-scale data collection.

• •Current estimates: several billion for first mouse connectome with EM, target is $10-30M with optical methods
• •Optical connectomics allows molecular annotation (synapse types, cell markers) impossible with EM
• •Human brain is 1000x larger than mouse; may not need every neuron, focus on steering subsystem
• •Parallel approach: entire mouse brain + human steering subsystem + multiple mammal species for comparison

Proposes using brain activity patterns as auxiliary training signals beyond simple labels. Instead of just 'cat/dog' labels, predict the full neural activity pattern humans exhibit when seeing cats. This could sculpt networks to represent information more like human brains, potentially improving generalization. Discusses practical challenges of collecting brain data at scale and whether this approach could create commercial value.

• •Auxiliary loss: predict human neural patterns in addition to task labels
• •Could align AI representations with human conceptual geometry and generalization patterns
• •Requires solving brain recording at scale - much harder than generating labeled images
• •Potential for 'human data companies' if brain recording becomes cheap enough and shows clear AI benefits

Marblestone's research assumes AI timelines longer than 5 years, as comprehensive neuroscience won't impact 2027 AGI scenarios. Estimates low billions needed for comprehensive brain mapping across species. Discusses funding mix: current work is philanthropy-based, but NSF tech labs, billionaire-backed moonshots, and potentially AI lab investment could accelerate progress. Key risk: moonshot companies might pursue flashy goals without doing fundamental science.

• •Comprehensive neuroscience research needs hundreds of millions to low billions over ~10 years
• •Current LLM paradigm is 'weirdly different' from brain architecture, creating uncertainty about relevance
• •If model-based RL had achieved GPT-4 capabilities, confidence in neuroscience approach would be higher
• •Potential for AI labs to fund at 1/100th of their 2030 training budgets if neuroscience insights prove valuable