Pluribus: six-player poker
On 11 July 2019 Noam Brown and Tuomas Sandholm (Carnegie Mellon and Facebook AI) described in Science Pluribus, a program that beat professionals at six-player no-limit Texas hold'em: over 10,000 hands against five professionals at once it won an average of 48 thousandths of a big blind per hand.
Why it matters
Earlier game wins, at checkers, chess, Go and heads-up poker, were in two-player zero-sum games, where a Nash equilibrium guarantees at least a draw. With several players there is no such guarantee, and Pluribus showed strong play is possible without it, and on ordinary hardware.
The first series had 13 professionals, five each day, over 12 days, with a $50,000 prize split by result; the standard error on 48 is 25. In the second, Chris Ferguson and Darren Elias each played 5,000 hands against five copies of the program: together Pluribus won 32 thousandths per hand (p = 0.014), but against Ferguson alone 25 at p = 0.107, which is not significant. The blueprint strategy took 8 days on a 64-core server, 12,400 core hours and under 512 GB of memory, about $144 at the cloud prices of the time; in play the program ran on two Intel processors. The record does not claim when the hands were played: none of the three sources read gives dates of play.