Matt Botvinick: Director of neuroscience research at DeepMind and a cross-disciplinary scientist working across cognitive psychology, computational neuroscience, and AI. Trained as a physician and psychologist before becoming a leader in connecting deep learning with brain function.
What this episode covers
Matt Botvinick argues that psychology, cognitive science, and neuroscience are really one science: understanding how the brain produces behavior. He explains the prefrontal cortex's role in flexible, goal-directed behavior versus habit, and walks through his research showing how meta-learning ('learning to learn') emerges spontaneously in recurrent neural networks trained with reinforcement learning. He connects AI advances back to the brain, including a paper proposing dopamine encodes reward predictions as full distributions rather than single numbers. The conversation turns to human-AI interaction, AI safety, and the deeper question of building AI systems that are not just capable but genuinely warm.
Also referenced (named, not recommended)
Deep Undercover
Jack Barsky
“former sleeper agent for the KGB in the 80s and author of deep undercover which is a memoir that paints yet another interesting perspective on the Cold War era”— Lex Fridman
Parallel Distributed Processing (PDP books)
Rumelhart, McClelland & the PDP Research Group
“are you familiar with the the PDP books from from the 80s”— Matt Botvinick
2001: A Space Odyssey
Stanley Kubrick
“if you look at like 2001 Space Odyssey when that well the monkeys touch the yeah like that's the magical moment”— Lex Fridman
WALL-E
Pixar
“I often think of the movie wall-e yeah that's like dystopian in a totally different ways like the machines are doing everything for us”— Matt Botvinick
Big reveals from this episode
- Botvinick says he has become reluctant to distinguish psychology from neuroscience, viewing them as one science aimed at explaining behavior.
- He reveals he went to medical school and also earned a graduate degree in art history before science, once torn between becoming a surgeon and a psychiatrist.
- His medical-school advisor, later revealed to be a famous psychoanalyst, handed him the still-shrink-wrapped PDP books that pulled him into deep learning.
- His 2018 paper reframes the prefrontal cortex as a meta-reinforcement-learning system where a second learning algorithm emerges from the first.
- He describes the 'almost magical' discovery that meta-learning happens automatically in any RNN with memory trained by RL across related tasks.
- The dopamine paper proposes the brain uses distributional coding, with dopamine representing reward prediction errors as distributions, not single numbers.
- Botvinick frames the ultimate Turing test as building an AI that is genuinely warm and that humans could love, not just capable.
Worth remembering
- Neuroscience can now observe brain activity at the single-unit and even dendritic level, but a huge gap remains between that and high-level function.
- He soothed his own doubts about psychology's 'metaphors' by noting Mendelian genetics productively preceded the discovery of DNA's physical mechanism.
- An early connectionist debate over how the brain forms the past tense ('quasi-regular' rules with messy exceptions) helped launch deep learning in cognitive science.
- World War I brain-damage cases first pushed scientists to map functions of the prefrontal cortex.
- Russian neuropsychologist Luria identified that frontal cortex damage impairs the flexibility to override habits.
- Even primary visual cortex, thought to only detect edges, carries information about behavior and expected reward when you have enough data.
- Jurgen Schmidhuber cheekily proposed 'meta meta meta meta learning' as the path to true intelligence.
- Susan Fiske's research shows humans judge others along two dimensions: capability and warmth, reserving highest esteem for those high in both.
- Botvinick cites WALL-E as a dystopia where machines doing everything for us is exactly what we don't want.