Matt Botvinick: Neuroscience, Psychology, and AI at DeepMind | Lex Fridman Podcast #106

In this episode, Matt Botvinick discusses the fascinating intersection of neuroscience, psychology, and artificial intelligence. He begins by addressing a fundamental question: how much do we actually understand about the brain? Despite centuries of research, Botvinick emphasizes that our knowledge remains surprisingly limited, particularly regarding consciousness and subjective experience. The brain contains roughly 86 billion neurons with trillions of connections, and we are still in the early stages of mapping these networks and understanding how they give rise to thought and behavior.

The conversation then shifts to psychology and the paradoxes that emerge when studying the human brain. Botvinick explores how cognition cannot be understood as purely internal to the brain but must be viewed as a function of the dynamic interaction between the brain and its environment. This perspective challenges traditional views that treat the brain as a self-contained computing device.

A significant portion of the discussion focuses on the prefrontal cortex, which Botvinick describes as crucial for flexible and goal-directed behavior. This brain region appears to serve as an integrative hub that allows organisms to override immediate impulses and plan for future contingencies. The episode then explores how information is processed in the brain and introduces the concept of meta-reinforcement learning, which has profound implications for understanding both biological and artificial intelligence. Meta-reinforcement learning refers to the ability of systems to learn how to learn, adapting their learning strategies across different tasks and contexts.

Botvinick provides valuable insights into dopamine, clarifying that it functions more as a learning signal than a simple reward signal. This understanding helps explain how the brain drives behavioral adaptation and maintains motivation toward goals. The discussion then broadens to examine how insights from neuroscience can inform artificial intelligence research and vice versa. Rather than viewing neuroscience and AI as separate fields, Botvinick argues they should be deeply integrated, with each informing the other's understanding of intelligence.

The episode addresses the human side of AI, exploring what it means to create artificial systems that can meaningfully interact with humans. This leads to a thought-provoking final question about whether we can create an AI that humans can genuinely love. Throughout the conversation, Botvinick demonstrates how computational approaches can illuminate fundamental principles of brain function while remaining grounded in empirical neuroscience. His work at DeepMind exemplifies how bridging these disciplines can generate novel insights into intelligence itself, whether biological or artificial.