Pluribus (poker bot)

• 1. Introduction

  ›

  Pluribus entered the record books in 2019 as "the first bot to beat humans in a complex multiplayer competition." Created jointly by Facebook's AI Lab and Carnegie Mellon University, it plays no-limit Texas hold'em. Before it arrived, the AI community had named defeating multiple opponents at once as "the widely recognized main remaining milestone" in computer poker. The puzzle had resisted every earlier approach. What combination of ideas finally cracked it? And when professionals sat across from Pluribus at an actual table, what kind of opponent did they encounter?

• 2. The Nash Equilibrium Problem

  ›

  In two-player zero-sum card games like heads-up hold'em, a Nash equilibrium strategy leaves no single opponent able to exploit you. Add more players, and that guarantee dissolves. A theoretically unexploitable approach in a one-on-one match offers no such protection at a table with multiple opponents. Pluribus faced precisely this constraint. Its creators were direct about the solution they found: the method Pluribus uses lacks "strong theoretical guarantees." What it had instead was empirical success. Against professional opponents, it won. Building that empirical edge, without a formal safety net, pointed toward a training approach far cheaper than what comparable milestones in the field had required.

• 3. Eight Days and $144

  ›

  Through eight days of competing against itself in a process called offline self-play, Pluribus built the base strategy it would carry into every live match. At market computing rates, that entire computation cost about $144. The developers noted this was much smaller than what contemporary superhuman game-playing milestones, such as AlphaZero, had required. Once Pluribus moved from training into live play, the learning did not stop. It continued adapting its strategy in real-time during each online session, adding fresh adjustments to the foundation its self-play had already produced. The habits that emerged from that process would prove stranger than anyone might have predicted.

• 4. Donk Bets and No Limping

  ›

  Across all competitions, Pluribus won an average of over 30 milli big blinds per game. It reached that margin with a playing style that departed in measurable ways from professional norms. Pluribus avoids "limping," the practice of calling the big blind rather than raising or folding. Experienced players often limp in certain situations; Pluribus largely abandoned the move. Even more distinctive was its use of "donk betting": ending one round with a call, then opening the next round by betting. Human experts treat this sequence with suspicion. Pluribus employed it more frequently than professionals typically do, and its margins confirmed the approach was working. The players who faced those margins across the table had their own ways of describing what it felt like.

• 5. Voices From the Table

  ›

  Playing five professionals simultaneously, Pluribus won an average of $5 per hand, which translated to roughly $1,000 per hour. Facebook called this a "decisive margin of victory." The professionals who faced it were direct about what made it hard to beat. Jason Les said he felt "very hopeless. You don't feel like there's anything you can do to win." Chris Ferguson said "Pluribus is a very hard opponent to play against. It's really hard to pin him down on any kind of hand." Jimmy Chou saw it differently. "Whenever playing the bot," he said, "I feel like I pick up something new to incorporate into my game." In The Wall Street Journal, science editor Daniela Hernandez called Pluribus "advanced at a key human skill - deception." That characterization, from a game played for real money against professionals, would prove central to the decision its creators made about the code behind it.

• 6. Code Withheld

  ›

  Following the competitions, the developers chose not to release Pluribus' source code. The reason they gave was direct: they feared the system would be used to surreptitiously cheat against human poker players in online matches. A bot that had already beaten professionals, running undetected inside online rooms, posed a specific risk. The developers were not willing to create that risk by distributing what they had built. The concern was precise: not casual cheating, but surreptitious play. A machine passing as a person, at online tables where opponents would have no way to tell the difference.