Dwarkesh Podcast• Feb 25, 2026Interview

Adam Marblestone – AI is missing something fundamental about the brain

From Dwarkesh Podcast

Adam Marblestone

Executive Summary

This episode provides a deep exploration of the fundamental differences between biological intelligence, specifically the human brain, and current artificial intelligence models like LLMs.
The speaker, Adam, posits that the brain's remarkable efficiency and generalization capabilities stem from principles that AI has yet to fully replicate, such as complex, evolutionarily-derived loss functions and a capacity for 'omnidirectional inference.' The discussion centers on a theoretical framework by Steve Beren, which divides the brain into an innate 'steering subsystem' that provides rewards and drives, and a general-purpose 'learning subsystem' (the cortex).
This framework is used to explain how the brain solves complex learning problems with a surprisingly small amount of genetic information and could offer a roadmap for developing more capable and aligned AI systems.

12 quotes

Concerns Raised

Current LLMs use overly simplistic reinforcement learning techniques, lacking concepts like value functions that are fundamental to biological learning.
The AI field may be neglecting the importance of complex, specific loss functions, which could be a key to the brain's sample efficiency.
It may be possible to create highly capable AI with a minimal set of drives (e.g., curiosity) that lacks human-like social instincts, posing an alignment risk.
The current AI paradigm, while successful, is architecturally very different from the brain, suggesting a potential performance plateau or a missing fundamental component.

Opportunities Identified

Incorporating principles of 'omnidirectional inference' could lead to AI models with superior generalization capabilities.
Studying the brain's 'steering subsystem' could provide a blueprint for building robustly aligned AI with complex, beneficial reward functions.
Multi-agent, co-evolutionary training may be more compute-efficient for developing intelligent agents than training a single monolithic model.
Neuroscience research, particularly large-scale connectomics, could provide crucial architectural and algorithmic constraints to guide the development of next-generation AI.

Key Themes

Brain vs. LLMs: The Architectural Divide

The conversation contrasts the brain's architecture with that of Large Language Models. Key differences highlighted include the brain's use of complex, evolved loss functions, its ability for omnidirectional inference, and its division into innate 'steering' and flexible 'learning' subsystems, which contrasts with the more uniform, unidirectional, and simply-rewarded nature of current LLMs.

This theme is critical for AI researchers looking beyond current scaling paradigms for the next breakthrough in model capability and efficiency.

The Steering Subsystem: Evolution's Reward Function

Drawing heavily on the theories of Steve Beren, the episode explores the concept of a 'steering subsystem'—innate, subcortical brain regions that provide the complex rewards, drives, and heuristics for learning. This system is presented as evolution's compact solution for guiding a general-purpose learner (the cortex) toward survival-relevant goals without needing to pre-specify every behavior.

Understanding this concept is vital for the field of AI alignment, as it provides a potential framework for how to instill complex, robust values in artificial agents.

Omnidirectional Inference as General Intelligence

The speaker proposes that the cortex functions as an 'omnidirectional inference engine,' capable of predicting any subset of variables from any other subset. This is contrasted with LLMs, which are fundamentally unidirectional (predicting the next token), and suggests a more flexible and powerful form of world modeling is at play in the brain.

This suggests a new architectural direction for AI that could overcome current limitations in reasoning, planning, and true understanding.

Amortized vs. Test-Time Compute

The discussion touches on the trade-off between 'amortizing' intelligence into a model's weights during training versus performing explicit computation and sampling at inference time. The brain appears to use a mix of both, while the trend in AI is to distill capabilities that once required inference-time compute (like chain-of-thought) back into the base model.

This highlights a key design choice in AI development that impacts model latency, cost, and capability, with implications for how future systems will balance pre-training with real-time reasoning.

Multi-Agent Scaling

The speaker details a personal experiment suggesting that splitting a fixed training budget among a population of smaller, co-evolving agents can produce a more capable top agent than dedicating the entire budget to a single, larger model. This points to the potential benefits of diversity and competition in training intelligent systems.

This provides a concrete, alternative scaling hypothesis that challenges the prevailing 'bigger is better' approach to model training.

Get started free

Topics

Neuroscience Artificial Intelligence Large Language Models (LLMs)Human Brain Cortex Loss Functions Reinforcement Learning (RL)Evolutionary Psychology AI Alignment Omnidirectional Inference Amortized Inference Steve Beren Jan LeCun Ilya Sutskever Connectomics Multi-Agent Systems AlphaZero

Processed Feb 25, 2026 yt-dlp + mlx-whisper + Gemini