Skip to content
Lex Fridman PodcastLex Fridman Podcast

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

Matt Botvinick is the Director of Neuroscience Research at DeepMind. He is a brilliant cross-disciplinary mind navigating effortlessly between cognitive psychology, computational neuroscience, and artificial intelligence. Support this podcast by supporting our sponsors: - The Jordan Harbinger Show: https://www.jordanharbinger.com/lex - Magic Spoon: https://magicspoon.com/lex and use code LEX at checkout EPISODE LINKS: Matt's papers: https://scholar.google.com/citations?user=eM916YMAAAAJ PODCAST INFO: Podcast website: https://lexfridman.com/podcast Apple Podcasts: https://apple.co/2lwqZIr Spotify: https://spoti.fi/2nEwCF8 RSS: https://lexfridman.com/feed/podcast/ Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4 Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41 OUTLINE: 0:00 - Introduction 3:29 - How much of the brain do we understand? 14:26 - Psychology 22:53 - The paradox of the human brain 32:23 - Cognition is a function of the environment 39:34 - Prefrontal cortex 53:27 - Information processing in the brain 1:00:11 - Meta-reinforcement learning 1:15:18 - Dopamine 1:19:01 - Neuroscience and AI research 1:23:37 - Human side of AI 1:39:56 - Dopamine and reinforcement learning 1:53:07 - Can we create an AI that a human can love? CONNECT: - Subscribe to this YouTube channel - Twitter: https://twitter.com/lexfridman - LinkedIn: https://www.linkedin.com/in/lexfridman - Facebook: https://www.facebook.com/LexFridmanPage - Instagram: https://www.instagram.com/lexfridman - Medium: https://medium.com/@lexfridman - Support on Patreon: https://www.patreon.com/lexfridman

Lex FridmanhostMatt Botvinickguest
Jul 2, 20202h 0mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

DeepMind’s Matt Botvinick on brains, behavior, learning, and AI’s future

  1. Matt Botvinick discusses how cognitive psychology, neuroscience, and AI are converging into one science of mind and behavior, with the brain viewed as a computational system mapping perception to adaptive action.
  2. He explains why high‑level psychological constructs (like attention and memory) and low‑level neural mechanisms (spikes, synapses, dopamine) must ultimately be linked, using deep learning both as a scientific model of brain function and as an engineering tool.
  3. Key research topics include prefrontal cortex as a meta‑reinforcement learning system, dopamine as a distributional value signal, and how agents can learn to learn across families of tasks.
  4. He closes by arguing that building AI that genuinely benefits humans requires serious work on human–AI interaction, social choice, values, and even warmth and “lovability,” not just raw capability or safety constraints.

IDEAS WORTH REMEMBERING

5 ideas

Treat psychology and neuroscience as one integrated enterprise focused on behavior.

Botvinick argues that the point of neuroscience is to explain what the brain is *for*—producing adaptive behavior from perception—so high‑level constructs like attention and memory must eventually be grounded in physical neural mechanisms, not kept in separate silos.

Use deep learning as both a brain model and an AI engine.

Artificial neural networks, particularly recurrent and deep architectures trained with reinforcement learning, often reproduce neural response patterns seen in biology, suggesting they’re reasonable approximations of how brains compute while also being powerful engineering tools.

Prefrontal cortex may implement meta‑learning through recurrent dynamics.

By training recurrent networks across families of related tasks, a slow reinforcement‑learning process over synaptic weights can create fast, emergent ‘learning to learn’ in activity patterns—an effect Botvinick proposes as an analogue of how prefrontal cortex supports rapid, flexible learning and cognitive control.

Represent value as a distribution, not just a single expectation.

Distributional reinforcement learning maintains full reward distributions instead of collapsing them into averages, enabling richer internal representations and faster learning; Botvinick’s work suggests dopamine neurons may implement a distributional code for value, with different cells encoding optimistic or pessimistic prediction errors.

Flexibility and task‑general abilities are the central missing pieces in current AI.

Human intelligence is marked by quick adaptation, abstraction, and the ability to switch between many tasks; modern deep RL systems excel in narrow domains but lack this general cognitive flexibility, making it a primary target for future research.

WORDS WORTH SAVING

5 quotes

To me, the point of neuroscience is to study what the brain is for. The brain, as far as I can tell, is for producing behavior.

Matt Botvinick

Remaining forever at the level of description that is natural for psychology, for me personally, would be disappointing. I want to understand how mental activity arises from neural activity.

Matt Botvinick

If you have a system that has memory and it’s trained with reinforcement learning across related tasks, this kind of meta‑learning just happens. You can’t stop it.

Matt Botvinick

We shouldn’t only be asking what can go wrong. We also have to bring into focus the question of what it would look like for things to go right.

Matt Botvinick

It’s not out of the question that we could build systems that end up learning what it is to interact with humans in a way that’s gratifying to humans. Honestly, if that’s not where we’re headed, I want out.

Matt Botvinick

Unity of psychology, cognitive science, and neuroscience as one science of behaviorPrefrontal cortex, cognitive control, and meta‑reinforcement learningDeep learning as a model of brain computation and emergent meta‑learningDopamine, temporal‑difference learning, and distributional value codingModularity versus distributed, graded organization in the brainHuman–AI interaction, values, and AI’s impact on societyFlexibility, abstraction, and multi‑task generalization in future AI systems

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome