Reinforcement learning is the branch of AI that learns by doing: try something, get a signal, adjust, try again. It built AlphaGo, cracked six-player poker, and turns out to be running quietly inside your own brain every time a dopamine neuron fires. We combed through our full library of episode summaries to find the conversations that actually explain how RL works and why it matters, not just the ones that namedrop it in passing.
This list runs from the researchers who built the landmark systems (AlphaGo, AlphaZero, Libratus, Cicero) to the theorists trying to define intelligence in a single equation, to the neuroscientists who found the same algorithm already running in your midbrain. Pick based on what you're chasing: game-playing breakthroughs, robotics, the neuroscience angle, or the philosophical deep end.
David Silver: AlphaGo, AlphaZero, and Deep Reinforcement Learning | Lex Fridman Podcast #86
If you only listen to one episode on this list, make it this one. David Silver led the DeepMind teams behind AlphaGo, AlphaZero, and MuZero, and he walks through exactly how self-play reinforcement learning beat Lee Sedol, then discarded human data entirely and got better. The standout moment is 'move 37,' the play that broke centuries of Go convention and proved a machine could produce something a professional called creative. He also admits AlphaGo had persistent knowledge 'holes' that Lee Sedol exploited in the one game he won. Essential listening for anyone who wants the RL origin story from the person who built it.
Read the full episode notesSergey Levine: Robotics and Machine Learning | Lex Fridman Podcast #108
Berkeley's Sergey Levine makes the case that robotics, not chess or Go, is the real proving ground for intelligence, because it forces an agent to handle an open, unpredictable world rather than a clean board with fixed rules. He digs into offline and off-policy RL, the practical problem of a robot breaking every dish while learning to wash them by trial and error, and why any hand-built simulator becomes a bottleneck the moment it stops improving with data. A grounded, technically honest counterpoint to the game-playing headlines.
Read the full episode notesNoam Brown: AI vs Humans in Poker and Games of Strategic Negotiation | Lex Fridman Podcast #344
Noam Brown co-built Libratus and Pluribus, the bots that beat professional poker players, and then Cicero, which negotiates in natural language to play Diplomacy at a human level. He explains why Libratus and Pluribus used zero neural networks, how Pluribus's final training run cost under $150 versus roughly $100,000 for its predecessor, and why a self-play bot with no human data got crushed at Diplomacy because it couldn't model human conventions. The section on why Cicero was built to avoid lying, because dishonesty made it perform worse once players stopped trusting it, is one of the sharpest reveals in the whole list.
Read the full episode notesLeslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
MIT's Leslie Kaelbling gives the clearest explanation on this list of what reinforcement learning actually is underneath the hype: Markov decision processes, partially observable environments, and 'belief space,' where a robot plans not over the world directly but over its own uncertainty about the world. Her airport-navigation example, planning a trip without knowing your gate or who's ahead of you in line, is the best plain-language analogy for planning under uncertainty you'll find in this collection. Good for listeners who want the theory taught by someone who helped build the foundations of the journal that publishes it.
Read the full episode notesHow Dopamine & Serotonin Shape Decisions, Motivation & Learning | Dr. Read Montague
This one flips the usual angle: instead of an AI researcher explaining RL, a neuroscientist explains how the exact temporal-difference algorithm behind AlphaGo Zero is already wired into brains from honeybees to humans. Read Montague argues dopamine isn't a pleasure chemical, it's a learning signal encoding the gap between successive expectations, and that hunger or stress can flip it from rewarding you to punishing you. If you want to understand why RL works as a model of intelligence rather than just a training trick, this is the episode.
Read the full episode notesMatt Botvinick: Neuroscience, Psychology, and AI at DeepMind | Lex Fridman Podcast #106
DeepMind's Matt Botvinick found that meta-learning, learning how to learn, emerges spontaneously inside recurrent neural networks trained with plain reinforcement learning, no special architecture required. He connects that finding directly back to the prefrontal cortex, and covers a paper proposing the brain codes dopamine reward predictions as full probability distributions rather than single numbers. The conversation closes on a genuinely interesting question: what would it take to build AI that's not just capable, but warm.
Read the full episode notesMichael Littman: Reinforcement Learning and the Future of AI | Lex Fridman Podcast #144
Brown professor Michael Littman has been in reinforcement learning since the 1980s, and he traces the field's real lineage: how he tried to reinvent RL from scratch before a colleague handed him Rich Sutton's 1984 TD paper, and how the term 'self-play' comes from his own 1996 PhD dissertation. He's also refreshingly unimpressed by AI doom scenarios, arguing we'll learn to control these systems as we build them. A funnier, more historically grounded entry point than the more technical episodes on this list.
Read the full episode notesJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Juergen Schmidhuber co-created LSTMs and has been chasing recursively self-improving machines since his 1987 diploma thesis. He distinguishes real meta-learning, a system that can inspect and rewrite its own learning algorithm, from today's narrower transfer learning, and explains his 'power-play' framework where a system invents its own problems to solve rather than waiting to be handed one. His claim that curiosity and consciousness are just byproducts of a compressor trying to compress itself is one of the more genuinely strange ideas in this whole list.
Read the full episode notesYann LeCun: Deep Learning, ConvNets, and Self-Supervised Learning | Lex Fridman Podcast #36
Turing Award winner Yann LeCun makes the contrarian case against pure reinforcement learning: today's model-free RL would need millions of driving hours, and kill thousands of simulated pedestrians, to learn what a human picks up in 20 to 30 hours because humans build world models first. He backs it with hard numbers, AlphaStar's StarCraft training equaled roughly 200 years of self-play, and argues self-supervised learning, not RL alone, is the real path to machines that reason. Essential for understanding RL's limits, not just its wins.
Read the full episode notesMarcus Hutter: Universal Artificial Intelligence, AIXI, and AGI | Lex Fridman Podcast #75
Marcus Hutter built AIXI, a mathematically optimal but incomputable RL agent that folds prediction, planning, exploration, and curiosity into a single equation. He explains it through Solomonoff induction and Kolmogorov complexity, and offers one of the sharpest cautionary notes in this list: the elevator-control example, where a naive reward function makes an elevator pick people up but never drop them off, showing how easily RL reward design goes wrong. Dense, but the clearest theoretical treatment here of what RL is actually trying to formalize.
Read the full episode notesDemis Hassabis: DeepMind - AI, Superintelligence & the Future of Humanity | Lex Fridman Podcast #299
DeepMind CEO Demis Hassabis traces the direct line from AlphaGo through AlphaZero and MuZero to AlphaFold, showing how the same self-play reinforcement learning ideas that mastered Go were repurposed to solve a 50-year problem in protein folding used by over 500,000 researchers. He also covers DeepMind's RL-based plasma control for nuclear fusion. Less technical than some entries here, but the best big-picture view of where RL research actually leads once it leaves the game board.
Read the full episode notesIlya Sutskever: OpenAI Meta-Learning and Self-Play | MIT Artificial General Intelligence (AGI)
This MIT lecture has Ilya Sutskever walking through OpenAI's RL research: hindsight experience replay, sim-to-real transfer via domain randomization, and self-play as a way to 'turn compute into data.' The concrete detail that lands hardest is the OpenAI Dota bots going from random play to world-champion level over about five months. Sutskever's framing of backpropagation as 'the miraculous fact on which the rest of AI stands' sets up the deep learning side of RL well for listeners newer to the field.
Read the full episode notesDeep Reinforcement Learning (John Schulman, OpenAI)
If you want the actual mechanics rather than the philosophy, this is the one. OpenAI's John Schulman, creator of TRPO, gives a straight technical lecture on policy gradients versus Q-learning, why RL is uniquely fragile to oversized update steps (a broken policy then collects its own bad training data), and when RL is honestly overkill compared to simpler optimization methods. A good pick for listeners who want the textbook chapter, delivered by the person who wrote part of the textbook.
Read the full episode notesDeep Learning State of the Art (2020)
A broad survey lecture covering the year's biggest RL milestones in one sitting: OpenAI Five's 45,000 years of simulated Dota self-play, AlphaStar reaching StarCraft Grandmaster under human-like constraints, and Pluribus's poker win described by pros as nearly unreadable. Not as deep on any single topic as the dedicated researcher interviews above, but useful as a fast-paced recap that ties RL's game-playing wins to the same-era progress in transformers and self-driving.
Read the full episode notesMIT AGI: Cognitive Architecture (Nate Derbinsky)
This one approaches RL from an unusual angle: cognitive architecture. Nate Derbinsky covers Soar, a system built for extreme efficiency (a 50-millisecond cycle ceiling even with billions of rules loaded), and tells the story of deliberately adding forgetting to it. In a reinforcement-learning dice game, the forgetting mechanism cut memory needs by more than half while playing just as well as a system that forgot nothing, because fewer than 1% of its value estimates were ever even updated. A good closer for anyone curious how RL fits into the bigger AGI research picture beyond deep learning.
Read the full episode notesThat's our list, but it barely scratches our library. Browse the full collection of episode summaries on Episode Notes to find more conversations on AI, neuroscience, and the algorithms quietly running your own brain.