Dwarkesh Podcast Summary

Episode: Adam Marblestone – AI is missing something fundamental about the brain

Date: December 30, 2025
Host: Dwarkesh Patel
Guest: Adam Marblestone (Convergent Research)

Overview

In this episode, Dwarkesh Patel delves into the fundamental mysteries of intelligence with Adam Marblestone, focusing on the question: What are current AI systems missing about how the brain works? Marblestone draws on insights from neuroscience, evolution, and AI research to explain why human intelligence is so efficient compared to large language models (LLMs), examines the overlooked importance of biological reward functions, and considers the implications for both neuroscience and the future of AI design.

Key Discussion Points

1. The Central Question: “How Does the Brain Do It?”

• [00:00] Dwarkesh frames the “million-dollar question”—why brains outperform AI on much less data.
• Adam argues the question is "the most important in science" and suggests progress will come from empowering neuroscience through better technology and research scale, not just theoretical contemplation.
• Marblestone outlines the standard ingredients in learning systems (architecture, hyperparameters, learning algorithm, initialization, cost/loss functions) and theorizes that evolution has optimized incredibly complex and developmentally staged loss functions, which AI largely overlooks.
  • Quote – Adam [00:14]: "I think evolution may have built a lot of complexity into the loss functions... Actually many different loss functions for different areas, turned on at different stages of development."

2. The Structure of the Cortex and Omnidirectional Inference

• [02:30] Discussion of the cortex as a general-purpose prediction engine, vs. the narrow forward-prediction of LLMs.

• Adam proposes that the cortex can flexibly predict any subset of variables given any other, unlike LLMs, which always predict the next token.

  • Quote – Adam [02:37]: "What if the cortex is just natively made so that any area of cortex can predict any pattern in any subset of its inputs given any other missing subset?"

• Marblestone relates this idea to concepts like probabilistic energy-based models (Yann LeCun) and omnidirectional inference.

3. The Role of Evolution and the Steering Subsystem

• [06:25] Focus on how evolution instills innate “reward functions” and drives.
• Explains Steve Byrnes’ theory that the brain wires complex, learned representations (like recognizing “Yann LeCun is upset with me”) to innate biological reactions like shame, via dedicated “steering subsystems” (e.g., amygdala, hypothalamus).
  • Quote – Adam [07:30]: "Evolution has to encode this desire to not piss off really important people in the tribe... without knowing in advance all the things that the learning subsystem... is going to learn."
• [10:56] Explains how the cortex generalizes: high-level features (like “spider” in language) can trigger innate reactions due to predictive learning, not direct supervision.
• The steering subsystem contains “hardwired” reward heuristics, while cortical systems learn predictors for those innate metrics, producing broad, flexible generalization.

4. Architectural Differences: AI vs. Brains

• [15:05] Dwarkesh asks if omnidirectional inference could be trained into current AI with new masking/objective strategies.
  • Adam: Possibly, but the brain’s inference flexibility may be more fundamental. Foundation models (multimodal) are a partial, but limited, analogy.
• [17:26] Discusses embedding/representation bottlenecks in current AI, and how inductive biases or architectural tweaks (e.g., through evolution) may enable better multimodal transfer and abstraction in brains.

5. Evolution as Meta-Learning, Compression, and Sample-Efficiency

• [28:08] Analogy between evolution and pretraining: evolution seems to code for very little (narrow genome) yet achieves massive generalization.
  • Key is that reward/loss functions can be compact and program generalizable learning without specifying every future scenario.
• [29:15] Quote – Adam: "The reward function in Python, the reward function is literally a line. And so you just have a thousand lines like this, and that doesn't take up that much space."
• The diversity and number of specialized cell types are higher in “steering” (reward) regions than in the cortex, suggesting more “bespoke” innate wiring for species-specific drives.

6. RL, Reward Mechanisms, and Model-Based vs. Model-Free Learning

• [43:02] Dwarkesh queries distinctions between model-based/model-free RL in the brain.
• Adam notes that certain brain areas (basal ganglia) may implement simple, “model-free” RL, while the cortex builds model-based predictive world models and can simulate “value functions”—but with enormous complexity.
  • Quote – Adam [43:02]: "There are probably parts of our brain that are just doing very simple RL algorithms... But then the cortex world model can contain a model of when you do and don't get rewards."
• Interesting speculation about hierarchical RL: evolution = model-free RL, basal ganglia = model-free, cortex/culture = model-based.

7. Biological Constraints and Hardware/Software Tradeoffs

• [51:32] Explores whether brains would work better or worse if “uploaded” onto silicon, or vice versa.
• Brains have strengths (energy efficiency, memory-compute colocation, stochasticity), but can't be copied/read-written arbitrarily.
  • Quote – Adam [51:39]: "It cannot be copied… But otherwise maybe it has a lot of advantages. It also tells you that you want to somehow do the code design of the algorithm…"
• Some aspects (e.g. neurons as natural samplers) might be ideal for energy-based/probabilistic computation.

8. Limits of Interpretability in Brains and Neural Networks

• [64:33] Question: Would a perfect connectome “solve” intelligence, or just produce another uninterpretable model?
• Adam argues for focusing on higher-level architecture, loss functions, and learning rules as key to understanding, rather than hoping for full interpretability at the neuron-by-neuron circuit level.
  • Quote – Adam [65:05]: “What I think we should do is we should describe the brain more in that language of things like architectures, learning rules, initializations...”

9. Neuroscience Moonshots: Mapping the Connectome

• [71:32] Practical discussion of brain mapping efforts (connectomics).
• Technological costs are dropping, but still require billions for human connectome, millions for mouse. Scaling, molecular annotation, and activity mapping are key technical frontiers.
• Funding sources could be philanthropy, government, or AI-sector investment if a direct link to safer or more efficient future AI is demonstrated.

10. Brain-Data Augmented Training and Behavior Cloning

• [79:12] Adam explains the idea (seen in Gwern’s blog post) of using neural data as auxiliary labels in AI training (beyond standard supervised learning), hoping this would sculpt more brain-like representations and aid generalization.

11. Automated Mathematics, Lean, and RL-Driven Proof Systems

• [83:28] As a board member of Lean, Adam discusses the prospects for AI-augmented, formalized math proof ("RL-verify everything!") to automate much of "cleverness" in math and software verification; full creativity may still require more AGI-like capabilities.
  • Quote – Adam [85:25]: “...RL-verifying the crap out of math proving is basically going to work... We will be able to have things that search for proofs and find them in the same way that we have AlphaGo..."
• Application to math, software, and possibly a “provable world model” paradigm.

12. Representation, Symbolic vs. Subsymbolic, and Binding Problems

• [98:18] Adam: The representational scheme of the world model in the brain is likely “a huge mess”, not symbolic in the traditional sense, but formed by architectures and learning dynamics.
• Variable binding, abstraction, and geometry (as in spatial or face maps) are all discussed.
  • Quote – Adam [99:42]: “My hunch is it’s going to be a huge mess... I don’t expect it to be pretty in there.”

13. The Need for Scale Science & the Gap Map

• [105:03] Adam shares the “gap map”: a meta-survey across scientific fields identifying bottlenecks and infrastructural gaps that could be solved by focused "moonshot" projects.
• Surprising cross-disciplinary needs (e.g., not just neuroscience or genomics, but also math and proof infrastructure) can emerge, often solvable at a relatively modest cost compared to their transformative impact.

Notable Quotes & Moments

• Adam [00:14]: “This might be the quadrillion dollar question… I don’t claim to know the answer.”
• Adam [02:37]: "What if the cortex is just natively made so that any area of cortex can predict any pattern in any subset of its inputs given any other missing subset…"
• Adam [07:30]: “Evolution has to encode this desire to not piss off really important people in the tribe or something like this in a very robust way, without knowing in advance all the things that the learning subsystem… is going to learn.”
• Adam [29:15]: “The reward function in Python... is literally a line. And so you just have a thousand lines like this, and that doesn’t take up that much space.”
• Adam [43:02]: “There are probably parts of our brain that are just doing very simple RL algorithms... But then the cortex world model can contain a model of when you do and don’t get rewards.”
• Adam [65:05]: “We should describe the brain more in that language of things like architectures, learning rules, initializations, rather than trying to find the Golden Gate Bridge circuit and saying exactly how does this neuron actually…”
• Adam [85:25]: "RL-verifying the crap out of math proving is basically going to work... And we will be able to have things that search for proofs and find them in the same way that we have AlphaGo..."
• Adam [99:42]: “My hunch is it’s going to be a huge mess and we should look at the architecture as the loss functions... I don’t expect it to be pretty in there.”

Timeline of Important Sections

| Timestamp | Topic / Notable Moment | |--------------|-----------------------------------------------------------------------------------------------| | 00:00 | The big question: How does the brain outperform AI? | | 02:30 | Human cortex, omnidirectional inference, and comparison to LLMs | | 06:25 | Innate reward functions; Steve Byrnes' "steering subsystem" theory | | 10:56 | Generalization in the cortex: word triggers innate responses | | 15:05 | Could omnidirectional inference be trained into AI? | | 28:08 | Evolution as a meta-learner; why genomes can be so compressed | | 43:02 | Model-free vs. model-based RL in the brain | | 51:32 | Biocomputation vs. AI hardware – strengths and weaknesses | | 64:33 | Interpreting the connectome: Is connectomics enough? | | 71:32 | Timeline and feasibility of mouse and human connectome mapping | | 79:12 | Behavioral cloning and brain-data-augmented training for AI | | 83:28 | Prospects for automated math through Lean and RL-verification | | 98:18 | How does the brain represent its world? Symbolic vs. messy subsymbolic representations | | 105:03 | The Gap Map: Infrastructure moonshots for science |

Final Thoughts

Adam Marblestone and Dwarkesh Patel synthesize neuroscience, evolution, and AI to highlight fundamental, unresolved questions at the heart of brain-inspired intelligence and future AI. The brain’s unique blend of architectures, cost functions, and meta-optimization via evolution distinguishes it from today’s AI, and mapping these phenomena—literally and conceptually—could be key to the next paradigm shift. The conversation underlines the importance of cross-field investment, technological moonshots (like connectomics and formalized math), and humility about the remaining mysteries of mind and intelligence.