Pluribus AI Defeats Professional Poker Players in Multiplayer Format

Artificial intelligence has conquered one of the most complex arenas of human strategy by becoming the first bot to defeat professional players in multiplayer poker. Developed by Noam Brown of Facebook AI Lab and Tuomas Sandholm of Carnegie Mellon University, the AI known as Pluribus mastered no-limit Texas hold 'em in a six-player format, marking a significant leap in machine learning.

Over 20,000 games, Pluribus repeatedly trounced top-tier professionals. The testing was split into two distinct phases: 10,000 games where the bot faced five human opponents, and another 10,000 where a single human played against five copies of the AI. According to data published in the journal Science in 2019, Pluribus won an average of over 30 milli big blinds per game. In monetary terms, the bot averaged winnings of $1,000 per hour, or approximately $5 per hand.

The Complexity of Multiplayer Games

Pluribus is a descendant of Libratus, an earlier AI that beat humans in heads-up, two-player poker. While two-player games can be solved using Nash equilibrium—a mathematical state where no player can improve their outcome by changing strategy—multiplayer games are computationally too complex for such an approach. To overcome this, the developers implemented a method of "cutting the decision tree." Rather than calculating every possible outcome to the end of the game, Pluribus looks only a few moves ahead, allowing it to manage the vast number of variables inherent in a six-player table.

A New Strategic Approach

The bot's dominance resulted from a distinct strategic style that differed from human experts. Analysis shows that Pluribus avoided "limping"—simply calling the minimum bet to see a flop—and utilized "donk betting" more frequently than professional humans. This clinical approach left human players feeling overwhelmed. Poker pro Jason Les described the experience as feeling "very hopeless," noting that it didn't feel like there was anything a human could do to win. Conversely, pro Jimmy Chou noted that playing the bot provided new insights to incorporate into his own game.

Real-World Implications

This achievement is significant because poker is a game of "imperfect information," where players must account for hidden cards and the possibility of deception. Unlike "perfect information" games like chess, where every piece is visible, poker mimics real-world decision-making. The developers intend to apply these methodologies to high-stakes fields including cybersecurity, financial market analysis, autonomous vehicles, and fraud detection.

What's Next

While Pluribus has set a new benchmark for AI in strategic gaming, the focus now shifts to how these deception-handling algorithms can be scaled. The efficiency of the project was notable; training the bot took only eight days and cost between $144 and $150 in cloud computing resources, suggesting that high-level strategic AI does not necessarily require prohibitive infrastructure.