Superhuman AI for Multiplayer Poker (www.science.org)

Researchers have announced a superhuman AI for multiplayer poker — a notable advance because it tackles a class of games that are fundamentally harder than the two-player zero-sum settings where prior AIs (e.g., in Dota 2 and StarCraft 2) initially beat top humans. Those earlier systems could be exploited over time because the training methods used were not guaranteed to converge to a Nash equilibrium; in true two-player zero-sum games a Nash strategy has an unbeatability property that prevents long-term exploitation. By contrast, multiplayer poker is an imperfect-information, general-sum game where a single-agent Nash guarantee does not exist, so creating robust, widely successful agents requires different solution concepts and training regimes. The work is significant because it shows practical techniques for producing resilient multi-agent policies in non-zero-sum, imperfect-information environments — likely using population-based self-play, approximate-equilibrium search and robustness-focused opponent modeling rather than relying on two-player equilibrium solvers alone. Technically, that means handling richer strategic diversity, coalition dynamics and exploitability trade-offs across many distinct opponents. Beyond poker, these methods advance multi-agent learning theory and have implications for economics, security, and any domain where agents must act under uncertainty against many adaptive adversaries.