Adam Marblestone – AI is missing something fundamental about the brain

In this podcast episode, Adam Marblestone delves into the profound question of how the brain achieves its remarkable capabilities and why current large language models (LLMs), despite being trained on massive datasets, exhibit only a fraction of human intelligence. The conversation explores neuroscience, artificial intelligence, reward functions, learning algorithms, and the technological challenges facing the decoding of the brain's architecture. Adam also reflects on the impact of recent advancements in AI and mathematics, discussing the future of formal verification, brain-inspired AI, and the extensive infrastructure needed to unlock the secrets of intelligence.

The Brain's Learning Algorithm and Cost Functions

Adam opens by outlining the challenge of understanding how the brain manages to learn and adapt so efficiently compared to current AI models. He frames this through the lens of machine learning components: architecture, learning algorithm, initialization, and cost or loss functions. While AI researchers tend to favor simple, mathematically convenient loss functions like next-token prediction in language models, Adam suggests evolution likely crafted a far more complex set of intertwined loss functions and reward signals. These cost functions might be orchestrated as an elaborate curriculum, turning different brain regions on and off at various developmental stages, enabling sample-efficient learning. This complexity, embedded deeply into the brain's innate structures, likely gives humans their "special sauce."

Omnidirectional Inference and the Cortex as a General Prediction Engine

A key speculative idea Adam discusses is that the cerebral cortex might fundamentally operate as a general-purpose prediction engine capable of omnidirectional inference—predicting any subset of variables from any other complementary subset within its input. This jumps beyond LLMs, which specialize in one-directional next-token prediction. He draws parallels with Yann LeCun's vision of energy-based models that represent joint probability distributions, allowing sampling conditioned on any subset of variables. The cortical architecture, composed of six physical layers connected across regions, might function as a mechanism approximating backpropagation but applied generally to arbitrary prediction tasks, forming a flexible foundation for complex cognition.

Steering Subsystem, Innate Reward Functions, and Learning Subsystem Dynamics

Adam expounds Steve Byrnes' theory differentiating the brain into the Learning Subsystem (largely cortical) and the Steering Subsystem (subcortical regions such as the amygdala, hypothalamus, and brainstem). The Steering Subsystem provides genetically ingrained reward signals and cost functions encoding primitive drives and survival instincts, like reactions to spiders or pain. The Learning Subsystem, in turn, develops rich world models and predictors that anticipate the Steering Subsystem's responses—essentially modeling what the innate drives will signal in a given situation. This dynamic wiring enables learned experiences, such as understanding the abstract concept of "spider" beyond direct sensory input to tap into innate fear circuits.

Evolutionary Encoding of Cost Functions and Cell-Type Diversity

A fundamental question raised is how evolution, which cannot directly know about futuristic concepts (like Yann LeCun or podcasts), encodes flexible reward systems adaptive enough to deal with such novel stimuli. The answer seems to lie in highly specialized cell types and genetically programmed wiring rules, particularly within the Steering Subsystem, which is rich in bespoke neuron types. This ensures distinct, innate circuits compute different basic reward-relevant variables—like "I'm about to flinch" or "this is social status." Meanwhile, the Learning Subsystem leverages a more uniform architecture with plastic synapses to develop generalized representations, learning from those innate reward signals. Adam relates this distribution of complexity to the limited information capacity of the genome and explains why such compact, reusable "Python-like" code for loss functions could suffice to create the full intelligence.

Probabilistic AI, Amortized Inference, and Energy-Based Models

The podcast touches on the fundamental tension between probabilistic inference—often formulated via costly Monte Carlo sampling or energy-based models—and the more feedforward amortized inference done by neural networks. Neural nets learn to approximate posterior distributions directly, effectively "compiling" inference into a single forward pass for efficiency. Adam discusses how biological brains may blend both, with neurons exhibiting stochasticity and sampling-like behavior, yet producing fast perception and response. The long-term vision involves whether AI systems will embed ever more of their inference at train time or continue to rely on expensive test-time computation, and how evolution's "design" of the brain reflects tradeoffs between architectural constraints, speed, and energy efficiency.

RL and Model-Free vs Model-Based Learning in the Brain

The difference between simplistic reinforcement learning (RL) used in current AI and more nuanced biological RL is highlighted. Existing LLMs use crude forms of RL without value functions, particularly lacking temporal difference (TD) learning mechanisms important in biological systems like the basal ganglia and striatum. Dopamine signals in the brain act as reward prediction errors, supporting more sophisticated model-free and model-based RL. The cortex builds richer world models incorporating actions, rewards, and outcomes in a model-based framework, while the subcortical areas handle simpler, more finite-action control sequences. This layered RL, from evolutionary dynamics down to culture and individual brain subsystems, creates a complex reinforcement architecture underpinning learning and decision-making.

Advantages and Disadvantages of Biological Hardware Versus Digital AI

Adam weighs the pros and cons of brain-inspired biological hardware versus modern digital computers. The brain is extremely energy efficient, operates asynchronously with low-speed electrical pulses, and integrates computation and memory in the same structures, allowing what might be called "cognitive dexterity." On the other hand, digital computers offer rapid, random-access memory and copy-ability, allowing for exact replicability and parameter tuning, something biological brains lack. He conjectures that future AI hardware and algorithms may evolve to incorporate co-design principles that capture the brain's advantageous features like co-located memory-compute, stochasticity, and low voltage operation, while overcoming its limitations.

Integrating Neuroscience and AI: Research Programs and Connectomics

Marblestone reflects on the state and future of neuroscience, particularly the need for large-scale, systematic technologies like connectomics—the high resolution mapping of every neuron and synapse in a brain. He stresses that understanding brain algorithms requires not just observing neuron activity but decoding the architecture, cell types, synaptic wiring rules, and the molecular details that govern connectivity and function. Technologies like E11 Bio and efforts to move beyond electron microscopy to molecularly annotated optical connectomes aim to reduce costs dramatically and accelerate progress. These data-driven approaches can constrain AI-inspired models and ground speculative theories of brain function into testable frameworks.

Bridging Brain Research and AI Safety/Alignment

A major motivation underlying this research is AI safety and alignment, particularly understanding how innate drives and higher-order desires emerge and integrate within an intelligent system. Human social instincts, ethical norms, and status-seeking behaviors are deeply tied to the Steering Subsystem and associated reward functions. Understanding the wiring between the learned cortex and these innate systems could inform how to design AI systems that "care" about human values or social status in robust ways. The complexity and modularity of these systems underscore why alignment is a deeply challenging problem and why neuroscience insights are crucial.

Advances in Formal Mathematical Proof and AI-Assisted Verification

Adam discusses progress in formalizing mathematics and software verification with tools like Lean, a proof assistant that enables mathematicians to encode proofs machine-checkably. This paradigm opens the door to AI systems that can "reinforcement-learn via verifiable reward" (RLVR) by producing provably correct theorems and software. While automating the creative aspects of mathematics, such as conjecture formulation and concept generation, remains challenging, the mechanical process of verifying and searching for proofs is accelerating. This shift has wide-ranging implications from mathematical research acceleration to provably secure software and cybersecurity, potentially transforming how we build and trust complex systems.

The Future of AI, Symbolic Representations, and World Models

The episode touches on speculation about whether the future of AI might reintegrate symbolic methods and explicit world models because their interpretability and verifiability could support safer collaboration among AI systems. Adam considers that if machines could generate and prove vast networks of theorems or represent their world models in symbolic, equation-based forms, this might mitigate risks inherent in opaque neural networks. While it's uncertain whether this vision will unfold imminently, it suggests a possible path where symbolic and neural approaches converge, facilitating both power and safety.

Representation in the Brain: Symbolic or Messy and Distributed?

One philosophical and scientific question Adam wrestles with is the nature of world representations in the brain, whether they resemble symbolic registers with variable binding or are more like high-dimensional distributed activation patterns within neural circuits. Although there is evidence of functionally specialized neurons—for instance, face-selective cells or place cells in the hippocampus—Adam suspects the overall picture will be complex and messy rather than clean symbolic encoding. Understanding these representations will require integrating architectural, algorithmic, and loss function perspectives.

Continual Learning and Memory Consolidation

Addressing continual learning, Adam acknowledges it remains an enigmatic area in neuroscience. The interplay between the hippocampus and cortex—where the hippocampus may store short-term memories and replay them to train the cortex—is a leading hypothesis. Multiple timescales of plasticity and forms of synaptic adaptation might underlie the brain's ability to learn continually without catastrophic forgetting. However, the precise biological mechanisms remain incompletely understood.

Attention and Fast Weights Analogues in Biological Systems

Adam reflects on whether the brain has mechanisms analogous to fast weights or key-value attention seen in transformer architectures. Attention in the brain is multifaceted—ranging from cortical region selection to thalamic gating and dynamic inter-area communication. While similar principles of matching and constraint satisfaction might be at play, whether this maps cleanly onto transformer-style attention is uncertain. These corticothalamic circuits are critical yet not well-understood substrates for selective information routing in biological intelligence.

The Gap Map and Infrastructure for Scientific Progress

Marblestone explains the Gap Map initiative, an effort to identify fundamental infrastructure "gaps" or needs across scientific fields that limit rapid advancement. Inspired by examples like how the Hubble Space Telescope transformed astronomy, these "mini Hubble" projects encompass new technologies, data platforms, and engineered tools necessary for breakthroughs. The Gap Map spans hundreds of capabilities, suggesting that focused investment on infrastructure rather than just hypothesis-driven science could accelerate discoveries even in mature disciplines like mathematics and neuroscience.

The Role of Outsiders and Democratization of Science

Adam highlights how advances in AI-driven formal verification and neuroscience could empower "outsider" researchers who are not traditionally credentialed in a given field. By lowering barriers to rigorous mathematical or scientific contributions, tools like Lean may facilitate broader participation and cross-disciplinary innovation. This democratization could spark novel breakthroughs in fields ranging from string theory to neuroscience, changing the landscape of scientific creativity.

Funding and Timelines for Neuroscience-AI Integration

Finally, Adam candidly discusses the scale and funding necessary to achieve breakthroughs in brain mapping, AI-aligned neuroscience, and understanding intelligence. Mouse brain connectomics projects, transitioning from over a billion dollars to potentially tens of millions with new technology, offer a benchmark. Achieving comprehensive, multi-species brain wiring maps—including humans—remains expensive but feasible within a multi-billion-dollar coordinated campaign. He contrasts this with the rapid pace of AI development and emphasizes that these efforts are long-term, foundational investments crucial for shaping safe and capable future AI.