Brown, Noam
▶ Watch on YouTubeCore ideas
Brown grounds the conversation in game theory fundamentals: any finite two-player zero-sum game (like heads-up poker or chess) has a Nash equilibrium strategy that guarantees a player will not lose in expectation regardless of the opponent's play, and his early poker bots (Libratus, and its successor Pluribus for six-player play) were built to approximate this equilibrium through counterfactual regret minimization, a form of self-play in which the system repeatedly plays against copies of itself and updates its strategy based on how much it "regrets" not having taken alternative actions in past situations. He explains why poker's hidden information makes it, in his view, a harder and more interesting game than chess or Go: a player must reason not just about which action to take but about the probability of taking it, since a strategy that becomes too predictable (e.g., always bluffing or never bluffing) can be exploited, whereas the value of a chess opening does not depend on how often you play it.
He describes the 2017 Brains vs. AI Libratus match against four elite heads-up players (a five-year project he calls one of his life's most stressful and rewarding experiences) as the moment that settled a long-running debate in the poker world between "Game Theory Optimal" play and exploitative, read-the-opponent play: Libratus, making no attempt to model or exploit its specific opponents, won roughly $1.8 million in chip equity over 120,000 hands. He recounts the players pooling their expertise and even being given full hand-history logs of the bot's cards each night, initially clawing back some ground before ultimately losing badly — and describes how the bot's tendency toward large, unusual bet sizes ("overbets"), originally added almost as an afterthought, proved so effective at forcing humans into difficult decisions that overbetting has since become a standard tactic in professional play. A key technical throughline across his poker work is the outsized importance of real-time search (planning several moves ahead at decision time rather than relying purely on a pre-computed strategy): he notes that even AlphaZero's Elo rating collapses from about 5200 to roughly 3000 if its Monte Carlo tree search is removed at play time, and that his own "depth-limited search" innovation — searching only a few moves ahead rather than to the end of the game — cut Pluribus's training cost from roughly $100,000 (Libratus) to under $150, using no neural networks at all, a result he attributes to algorithmic rather than hardware improvement.
The conversation's second half covers Cicero, Brown's system for the seven-player alliance-and-betrayal board game Diplomacy (created in the 1950s, reportedly favored by JFK and Henry Kissinger, and designed so that the "ideal" outcome is no single winner because other players cooperate to check anyone approaching victory). Brown explains why Diplomacy is a fundamentally different and, in his view, more difficult AI challenge than prior adversarial game breakthroughs (chess, Go, poker, StarCraft, Dota): it requires open-ended natural-language negotiation and genuine cooperation with other agents, so self-play from scratch fails outright — a bot trained purely against copies of itself develops an internally consistent but human-incomprehensible strategy and gets "destroyed" when placed among real humans, analogous to a self-driving car trained only against other self-driving cars that all happen to drive on the wrong side of the road. Cicero's solution combines a language model (pretrained on general internet text, then fine-tuned on roughly 50,000 human Diplomacy games with over 10 million messages from webdiplomacy.net) with a separate strategic-reasoning module that computes "intents" (planned actions for itself and desired actions for other players) via self-play and search regularized to stay close to an "anchor policy" modeling actual human play; the language model then generates messages corresponding to those intents, filtered to avoid nonsensical or actively deceptive statements, since the team found that lying measurably hurt the bot's long-run score by destroying other players' trust in it. Cicero placed second out of 82 human participants across 40 games (ranked among the 19 who played five or more games), excelling both at building durable alliances and at tactical calculation, and the team observed that a purely self-play-trained bot, even without any language component, still needed to model human irrationality explicitly — for example, expecting that human players might vindictively "throw" a game out of anger at being betrayed or exploited, a dynamic that has no equivalent in purely adversarial games and can only be learned from human data.
Brown discusses several broader implications: that making game AIs deliberately human-like (rather than superhumanly alien in style, as with modern chess engines) opens promising directions — such as training bots that play in the style of, or at a calibrated skill level relative to, a specific human, useful both for training partners and for opponent preparation — while also creating a harder cheat-detection problem, since cheat-detection today relies partly on identifying moves that look "too computer-like." On ethics, he flags the discomfort of deliberately building a system capable of deception (even though Cicero was tuned to minimize outright lying), an observed anti-AI bias in experimental settings where human players united to identify and eliminate a known bot regardless of its behavior, and open questions about how such techniques might eventually generalize to real-world negotiation, diplomacy, or human-AI trust more broadly — a connection he considers plausible in principle but currently limited by the same data and reward-specification challenges that constrain applying reinforcement learning to real-world, physically embodied domains. He closes with career advice emphasizing a strong foundation in math, statistics, and computer science combined with a willingness to pursue an atypical path (his own being rooted in game theory rather than a standard ML curriculum), and reflects, prompted by a question about "solving" how to live life optimally, that the harder problem for both AI systems and humans is not optimizing toward a reward function but correctly specifying one that avoids unintended, harmful shortcuts.
Excerpts
On why poker's hidden information makes probability itself part of the strategy, unlike chess:
“The value of an action depends on the probability that you're going to play it... if you're the kind of person that never bluffs and suddenly you bluff, there's a really good chance that bluff is going to work... you contrast that with a game like chess — it doesn't matter if you're opening with the Queen's Gambit ten percent of the time or a hundred percent of the time, the expected value is the same.”
On why a bot trained purely through self-play, with no human data, fails catastrophically against real people in Diplomacy:
“It's kind of like if you develop a self-driving car and it's trained completely from scratch with other self-driving cars — it might learn to drive on the left side of the road, that's a totally reasonable thing to do if you're with these other self-driving cars... but if you put it in an American city, it's going to crash.”
On why Cicero was built to avoid deception despite Diplomacy's reputation as a game of lies:
“We actually found that lying would make the bot perform worse in the long run — it would end up with a lower score, because once the bot lies, people would never trust it again, and trust is a huge aspect of the game of diplomacy... if you talk to top diplomacy players, what they'll tell you is that diplomacy is a game about trust, and being able to build trust in an environment that encourages people to not trust anyone.”