Poker AI
Poker AI represents one of the most significant achievements in artificial intelligence: mastering games of imperfect information. Unlike chess, where both players see the full board, Texas Hold'em poker involves hidden cards, bluffing, deception, and reading opponents — making it a fundamentally different and arguably harder challenge for AI.
Noam Brown & the Breakthrough Systems
The defining breakthroughs came from researcher Noam Brown, working with Tuomas Sandholm at Carnegie Mellon University. Brown developed the core algorithms behind two systems that changed the field. Libratus (2017) defeated four top professional poker players in a 20-day, 120,000-hand heads-up no-limit Texas Hold'em competition. The system earned the Marvin Minsky Medal for Outstanding Achievements in AI — one of the field's highest honors. Pluribus (2019) went further, beating elite professionals in six-player games — the standard format of most real poker. The margin of victory was decisive: over the equivalent of millions of dollars in expected value. Pluribus appeared on the cover of Science magazine and was runner-up for Science's Breakthrough of the Year in 2019. These weren't narrow wins; the AI dominated professionals who had made their livings playing the game.
Technical Approach: Imperfect Information Game-Solving
The technical approach differs fundamentally from AlphaZero's approach to chess. Poker AI relies on counterfactual regret minimization (CFR) — an algorithm that iteratively approximates Nash equilibrium strategies by simulating every possible decision point and minimizing the regret of not having chosen differently. Because the game tree for no-limit poker is effectively infinite, the AI uses abstraction techniques to group similar situations and search algorithms to refine its strategy in real-time during play. Brown's key insight was combining offline game-solving with real-time search — allowing the AI to compute finer-grained strategies during actual play rather than relying solely on pre-computed solutions.
From Poker to Diplomacy: Multi-Agent Reasoning
Brown extended his imperfect-information game-solving research to an even more complex domain. At Meta's FAIR lab, he led the development of CICERO — the first AI to achieve human-level performance in the board game Diplomacy, which requires not just strategic reasoning but natural language negotiation with multiple players. CICERO combined a language model for communication with strategic planning algorithms descended from the poker work, demonstrating that the principles of imperfect-information reasoning could scale to domains requiring both strategic thinking and natural language interaction. This connects directly to the broader challenge of AI in Diplomacy and multi-agent negotiation systems.
Implications Beyond Games
Many real-world problems involve imperfect information, bluffing, and strategic deception: negotiation, cybersecurity, military strategy, financial trading, diplomatic relations. Poker AI demonstrated that machines can reason effectively in environments of uncertainty and hidden information — a capability that directly informs autonomous negotiation systems and any application where agents must operate with incomplete knowledge while anticipating adversarial behavior. Brown himself previously worked at the Federal Reserve Board on algorithmic trading research, and the connection between game-theoretic AI and financial markets is direct: the same mathematics that finds optimal poker strategies can model strategic interactions between trading algorithms.
Brown, now a Research Scientist at OpenAI, is applying the principles from poker and Diplomacy to broader AI challenges — including multi-step reasoning, self-play, and multi-agent AI systems. Named to MIT Technology Review's 35 Innovators Under 35, his trajectory illustrates how game-solving research feeds directly into building more capable AI agents that can reason, plan, and negotiate in complex real-world environments.