Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12
Lex Fridman (host), Tuomas Sandholm (guest)
In this episode of Lex Fridman Podcast, featuring Lex Fridman and Tuomas Sandholm, Tuomas Sandholm: Poker and Game Theory | Lex Fridman Podcast #12 explores aI Mastering Poker Reveals Future of Game Theory and Strategy Lex Fridman interviews Tuomas Sandholm about Libratus, the first AI to decisively beat top professionals in heads-up no-limit Texas Hold'em, and what that implies for AI, game theory, and real-world strategy. Sandholm explains how imperfect-information games differ fundamentally from perfect-information games like chess and Go, and why abstractions, equilibrium computation, and belief modeling are central. They discuss extensions to learning-based methods, exploitation and collusion in multi-player settings, and moving beyond benchmark games into business, military, and market design applications. Sandholm also reflects on mechanism design, real-world impact (kidney exchanges, auctions), and why he is optimistic about AI’s societal benefits despite theoretical concerns about misalignment.
AI Mastering Poker Reveals Future of Game Theory and Strategy
Lex Fridman interviews Tuomas Sandholm about Libratus, the first AI to decisively beat top professionals in heads-up no-limit Texas Hold'em, and what that implies for AI, game theory, and real-world strategy. Sandholm explains how imperfect-information games differ fundamentally from perfect-information games like chess and Go, and why abstractions, equilibrium computation, and belief modeling are central. They discuss extensions to learning-based methods, exploitation and collusion in multi-player settings, and moving beyond benchmark games into business, military, and market design applications. Sandholm also reflects on mechanism design, real-world impact (kidney exchanges, auctions), and why he is optimistic about AI’s societal benefits despite theoretical concerns about misalignment.
Key Takeaways
Imperfect-information games require fundamentally different AI techniques than perfect-information games.
In poker, players must reason not only over hidden cards but also over belief distributions about what others know and think, so you cannot just learn a simple state-value function as in Go; strategy and beliefs are co-determined via equilibrium concepts.
Get the full analysis with uListen AI
Libratus’s success came from deep game-theoretic reasoning, not opponent modeling or deep learning.
The system computed near–Nash-equilibrium strategies over abstracted versions of the massive game tree, avoided reliance on past human data, and proved that purely equilibrium-based play can reach superhuman performance in high-stakes poker.
Get the full analysis with uListen AI
Abstraction design is critical and non-trivial in large imperfect-information games.
Both information abstraction (grouping hands by current and future potential) and action abstraction (choosing representative bet sizes) must be carefully constructed; finer abstractions are not always better and can paradoxically yield worse real-game play.
Get the full analysis with uListen AI
Exploiting opponents is powerful but dangerous against strong adversaries.
A game-theoretic strategy is unexploitable but may leave money on the table; hybrid approaches start from equilibrium play and selectively adjust in parts of the game tree where data reveal consistent opponent weaknesses, while staying close enough to avoid counter-exploitation.
Get the full analysis with uListen AI
Moving beyond two-player zero-sum dramatically increases conceptual and computational complexity.
With more players or non-zero-sum payoffs, multiple equilibria, coordination issues, and collusion (e. ...
Get the full analysis with uListen AI
Computational game theory is poised to impact real-world strategic domains.
Sandholm’s startups aim to bring these tools to business strategy, markets, security, and military planning, analogous to how machine learning moved from lab benchmarks to pervasive industrial use; he expects much of future strategy work to be algorithmically supported.
Get the full analysis with uListen AI
Mechanism design and automated mechanism design are powerful but bounded by impossibility results.
While some classes of objectives cannot be satisfied by any mechanism, automated design can still find ‘islands of possibility’ in specific instances—e. ...
Get the full analysis with uListen AI
Notable Quotes
“A game-theoretic strategy is unbeatable, but it doesn’t maximally beat the other opponents.”
— Tuomas Sandholm
“Until you’ve seen [an AI] over-perform a human, it’s hard to believe that it could.”
— Tuomas Sandholm
“We don’t need any data. It’s all about rationality.”
— Tuomas Sandholm
“In many situations in AI, you really have to build the big systems and evaluate them at scale before you know what works and doesn’t.”
— Tuomas Sandholm
“I think the big breakthrough is to show that most of business strategy or military planning will actually be done strategically using computational game theory.”
— Tuomas Sandholm
Questions Answered in This Episode
How might techniques from Libratus scale to complex, multi-player, non-zero-sum environments like real-world markets or geopolitical conflicts?
Lex Fridman interviews Tuomas Sandholm about Libratus, the first AI to decisively beat top professionals in heads-up no-limit Texas Hold'em, and what that implies for AI, game theory, and real-world strategy. ...
Get the full analysis with uListen AI
What are the most pressing open problems in making game-theoretic strategies interpretable enough for high-stakes domains like military or regulatory decision-making?
Get the full analysis with uListen AI
Where is the boundary between safe, rational equilibrium play and aggressive opponent exploitation in settings where adversaries are adaptive and potentially deceptive?
Get the full analysis with uListen AI
How could automated mechanism design realistically influence large, politically constrained systems such as elections, healthcare markets, or climate policy instruments?
Get the full analysis with uListen AI
Given Sandholm’s optimism about AI safety, what specific scenarios—if any—does he see where value misalignment or objective misspecification could cause serious real-world harm?
Get the full analysis with uListen AI
Transcript Preview
The following is a conversation with Tuomas Sandholm. He's a professor at CMU and co-creator of Libratus, which is the first AI system to beat top human players in the game of Heads-up No-Limit Texas Hold'Em. He has published over 450 papers on game theory and machine learning, including a best paper in 2017 at NIPS, now renamed to NeurIPS, which is where I caught up with him for this conversation. His research and companies have had wide-reaching impact in the real world, especially because he and his group not only propose new ideas, but also build systems to prove that these ideas work in the real world. This conversation is part of the MIT course on Artificial General Intelligence and the Artificial Intelligence podcast. If you enjoy it, subscribe on YouTube, iTunes, or simply connect with me on Twitter @lexfrid. And now here's my conversation with Tuomas Sandholm. Can you describe, at the high level, the game of poker, Texas Hold'Em, Heads-up Texas Hold'Em, for people who might not be familiar, uh, this card game?
Yeah, happy to. So Heads-up No-Limit Texas Hold'Em has really emerged in the AI community as a main benchmark for testing these application-independent algorithms for imperfect information game solving. And this is a game, uh, that's actually played by humans. You don't see it that much on TV or casinos, because, uh, well, for various reasons, but, uh, uh, you do see it in some expert-level casinos, and you see it in the best poker movies of all time. It's actually an event in the World Series of Poker. But mostly it's played online and typically for pretty, uh, big sums of money. And this is a game that usually only experts play. So if you re- uh, go to your home game on a Friday night, it probably is not gonna be Heads-up No-Limit Texas Hold'Em. It might be, uh, No-Limit Texas Hold'Em in some cases, but typically for a, a big group when it's not as competitive. While heads-up means it's two players, so it's really like me against you. Am I better or are you better? Much like chess or, or, or Go in that sense, but an imperfect-information game, which makes it much harder, of course. I have to deal with issues of, uh, you knowing things that I don't know, and I know things that you don't know, instead of pieces being nicely laid on the board for both of us to see.
So in Texas Hold'Em, uh, there's, uh, two cards that you only see? The-
Yes.
... they belong to you?
Yeah.
And then there is they gradually lay out some cards that add up overall to five cards that everybody can see.
Yeah.
So the imperfect nature of the information is the two cards that you're holding in your hand.
Up front, yeah. So as you said, you know, you first get two cards in private each, and then you, uh, there's a betting round. Then you get three cards in public on the table, then there's a betting round. Then you get the fourth card in public on the table, there's a betting round.
Install uListen to search the full transcript and get AI-powered insights
Get Full TranscriptGet more from every podcast
AI summaries, searchable transcripts, and fact-checking. Free forever.
Add to Chrome