In January 2017, an AI called Libratus, built by Tuomas Sandholm and his PhD student Noam Brown at Carnegie Mellon University, defeated four of the world’s best heads-up no-limit Texas Hold’em players in a 20-day competition called “Brains vs. Artificial Intelligence: Upping the Ante” at Rivers Casino in Pittsburgh. Over 120,000 hands ending January 30, Libratus finished ahead of the pros (Dong Kim, Jimmy Chou, Daniel McAulay, and Jason Les) by a collective 1,766,250 dollars in chips, a margin CMU described as statistically significant rather than luck.
What made this significant was the kind of game. Chess, Go, and checkers are games of perfect information: every player can see the entire board. Poker is a game of imperfect information, where players hide cards and bluff, so an AI must reason about what it cannot observe and avoid being exploited. The full results were published in Science under the title “Superhuman AI for heads-up no-limit poker: Libratus beats top professionals” (Brown and Sandholm, Vol. 359, No. 6374, pp. 418-424). The paper reports Libratus beat the humans by 147 milli-big-blinds per hand with 99.98 percent statistical significance.
Libratus combined a precomputed blueprint strategy, real-time refinement of its play during a game, and a self-improvement step that patched weaknesses opponents had probed. As Sandholm put it, “The best AI’s ability to do strategic reasoning with imperfect information has now surpassed that of the best humans.” The result opened the door to applying these techniques beyond poker to negotiation, security, and other strategic settings.