Noam Brown — Research scientist at Meta AI (FAIR) who co-created the first superhuman poker AIs (Libratus and Pluribus) and Cicero, an AI that negotiates with humans in natural language to play the board game Diplomacy at a human level.
The gist
Noam Brown walks Lex Fridman through his career building game-playing AI, starting with No Limit Texas Hold'em. He explains Nash equilibrium, counterfactual regret minimization, and why real-time search dramatically boosts performance. He recounts the 2017 Libratus competition where his bot beat four top pros out of $2M over 120,000 hands, and how Pluribus extended this to six-player poker for a fraction of the cost. The conversation then turns to Diplomacy, a seven-player negotiation game that demands cooperation and natural language, requiring human data rather than pure self-play. Brown closes on trust, lying, cheat detection, data efficiency, and what it might take to reach AGI.
Big reveals
- Libratus beat four top heads-up pros over 120,000 hands, winning close to $2 million.
- The bot's surprising last-minute 'overbets' (betting many times the pot) flummoxed pros and have since become standard high-level poker strategy.
- Without test-time search, the strongest Go bot drops from ~5200 ELO to ~3000 ELO; no raw neural net alone is superhuman at Go.
- Brown admits the Libratus competition was so stressful he worked on it nearly nonstop for a year, and only gave the team ~50/50 odds going in.
- Pluribus's final training run cost under $150 on AWS versus an estimated ~$100,000 for Libratus, driven by algorithmic gains (depth-limited search).
- Libratus and Pluribus used no neural networks at all, which surprised many in the field.
- Cicero was deliberately built to minimize lying, because lying made the bot perform worse once other players stopped trusting it.
- A self-play bot trained without human data got destroyed by humans at Diplomacy even in the no-language version, because it couldn't model human conventions.
Things worth remembering
- Heads-up No Limit Texas Hold'em has about 10^161 decision points, more than the number of atoms in the universe squared.
- In any finite two-player zero-sum game there is an optimal strategy that guarantees you won't lose in expectation, no matter what your opponent does.
- The team handed the human pros the bot's hole cards for every hand each night, golden information, and still won.
- Diplomacy was reportedly a favorite game of JFK and Henry Kissinger and was created in the 1950s, set just before World War I.
- In the game's ideal form nobody wins, because if someone nears victory the others should unite to stop them, mirroring the futility of war.
- Cicero's training data came from webDiplomacy.net: ~50,000 games and over 10 million natural-language messages, now being made available to researchers.
- Cicero placed second out of ~19 players who played five or more games across 40 online Diplomacy games.
- Brown argues Diplomacy is a bigger step toward real-world AI than StarCraft or Dota because it runs on open-ended natural language.
- Humans show an 'anti-AI bias': told a bot was present, they spent whole games hunting it and ganging up to eliminate it.
- Data inefficiency is a core gap: a Go AI needs millions of games to learn what a human grandmaster grasps in thousands.
Recommended in this episode
Books, products and media the guest or host genuinely endorsed here — with the buy link.
Affiliate link — we may earn a commission at no extra cost to you.
Skyrim
Bethesda Game Studios (inferred)
“the Creator what I think is the greatest game of all time which is Skyrim and the NPCs there the AI that governs that whole game is very interesting” — Lex Fridman 00:10:53Find it on Amazon