Noam Brown

What he works on

Noam Brown is a research scientist at OpenAI. His homepage lists his research interests as reasoning, reinforcement learning, self-play and multi-agent AI.^[1] He is best known for a series of systems that reached superhuman strength in games with hidden information and more than two players — two-player poker, six-player poker, and the board game Diplomacy — and, more recently, for work on models that reason before they answer.^[1][4]

Education and earlier jobs

Brown earned a PhD in computer science at Carnegie Mellon University.^[1] He has also worked in the International Financial Markets section of the Federal Reserve Board, researching algorithmic trading, and later at Meta's FAIR lab.^[1]

That mix is unusual for an AI researcher: a central bank's markets desk, an industrial research lab and a university game-theory group all appear on the same path. The thread running through them is a interest in decisions made under uncertainty and against other actors, which is what trading and multiplayer games have in common.

Poker: Libratus and Pluribus

Libratus, which Brown built with his CMU advisor, defeated top human poker professionals and received the Marvin Minsky Medal for Outstanding Achievements in AI.^[1]

The better-known multiplayer system is Pluribus, co-developed with Tuomas Sandholm at a time when Brown was a Carnegie Mellon PhD student and a research scientist at Facebook AI.^[2][3] Pluribus plays six-player no-limit Texas hold'em.^[2][3] Adding players makes poker much harder than the heads-up game: each additional opponent brings more hidden cards and more possible responses, so a program cannot simply work out one optimal line. Pluribus defeated professionals Darren Elias and Chris Ferguson, with each pro playing 5,000 hands against five copies of the AI.^[2][3] In a separate test, 13 professional players who had each won more than $1 million playing poker played five at a time, for 10,000 hands in total.^[2] Pluribus was on the cover of Science and was runner-up for the journal's 2019 Breakthrough of the Year.^[1]

CICERO and Diplomacy

Meta describes CICERO as the first AI to achieve human-level performance in Diplomacy, and says that in 40 games against human players it achieved more than double the average score of the human players.^[4] Diplomacy is a game of negotiation as much as of movement: players must persuade, promise and sometimes deceive each other in ordinary language before orders are resolved. Meta credits Brown, as a researcher at Meta AI, with explaining how CICERO brings together strategic reasoning and natural language in that game.^[4]

Reasoning models

Brown's homepage describes him as a foundational contributor to reasoning models, naming OpenAI o1 (also called Strawberry) as an example.^[1] A reasoning model is one trained to work through intermediate steps, spending more computation before it commits to an answer, rather than producing the first plausible response. Self-play — systems improving by playing against copies of themselves — is one of the techniques his homepage lists among his interests.^[1]

Why it matters

Games have been a useful testing ground for AI because they give a clean score and an unforgiving opponent. Poker and Diplomacy are a step past chess and Go: the AI cannot see everything, its opponents adapt to it, and in Diplomacy it has to use language to build coalitions. Systems like Libratus, Pluribus and CICERO show that such problems can be attacked with search, game-theoretic reasoning and learning rather than hand-written rules. Brown's later work on reasoning models carries the same theme — decision making under uncertainty, with computation spent where it helps — into general-purpose models rather than games.^[1]