Adam Marblestone argues that current AI systems lack the brain’s evolution-shaped, richly structured learning and reward mechanisms, which enable humans to learn efficiently and flexibly. He suggests that understanding and emulating the brain’s architecture—particularly the interplay between general learning (cortex) and innate steering (subcortical structures)—could lead to major advances in AI, and advocates for large-scale neuroscience efforts to uncover these principles.
Adam Marblestone discusses the fundamental differences between artificial intelligence (AI) systems, like large language models (LLMs), and the human brain, focusing on why current AI still falls short of human-level intelligence despite being trained on vast amounts of data. He argues that the key to understanding this gap lies in neuroscience, particularly in how the brain’s architecture, learning algorithms, and especially its complex, evolution-shaped loss and reward functions differ from those used in machine learning. Marblestone suggests that evolution has encoded a rich set of innate cost functions and developmental curricula in the brain, allowing different regions to learn efficiently and adaptively, something current AI systems lack.
A central theme is the distinction between the brain’s “Learning Subsystem” (primarily the cortex) and the “Steering Subsystem” (subcortical structures like the hypothalamus and amygdala). The Learning Subsystem is highly general and capable of omnidirectional inference—predicting any subset of variables from any other subset—while the Steering Subsystem provides innate drives, reflexes, and reward signals. Marblestone, referencing Steve Byrnes’ theories, explains how the cortex learns to predict and generalize the responses of the Steering Subsystem, allowing humans to flexibly connect learned concepts (like social status or danger) to innate emotional and behavioral responses.
Marblestone also explores the limitations of current AI training paradigms, which typically use simple, mathematically convenient loss functions (like next-token prediction or cross-entropy) and lack the rich, multi-stage, and context-dependent reward structures found in biological brains. He speculates that more sophisticated, evolution-inspired cost functions and architectures could dramatically improve AI’s learning efficiency and generalization. He also discusses the concept of amortized inference in neural networks versus the potentially more flexible, sample-based probabilistic inference that might occur in the brain.
The conversation touches on the practical challenges of neuroscience, such as mapping the brain’s connectome and understanding the molecular and cellular diversity of different brain regions. Marblestone advocates for large-scale, technology-driven neuroscience projects to uncover the brain’s wiring and learning rules, arguing that such efforts are essential for both scientific understanding and the future development of AI. He draws parallels to the Human Genome Project, emphasizing the need for technological innovation to make brain mapping scalable and affordable.
Finally, Marblestone discusses the implications of these ideas for AI safety, alignment, and the future of science and mathematics. He highlights the potential of formal methods (like the Lean proof assistant) to automate and verify mathematical reasoning, and speculates about the future of provable, interpretable AI systems. Throughout, he stresses the importance of interdisciplinary research and scalable scientific infrastructure, suggesting that breakthroughs in understanding the brain could unlock new paradigms in AI and transform our approach to intelligence, creativity, and collaboration.