A decade or so ago, I told my poker class that “one can build a bot to crush a chess game, but it will never happen in poker.”Boy was I wrong. Carnegie Mellon Research Lab developed a bot named Pluribus and in July it dominated a six-max no-limit hold’em game where the opponents were all well-known poker pros.
Pluribus won at a 50-big-blind-per-hour rate, which translated to $1K-plus an hour. After reading up on this, I have a few takeaways from my research. I also have a few ideas on how we can learn from Pluribus and perhaps make a few adjustments to our game.
- Pluribus dominated pros and that can’t be overlooked.
- Pluribus isn’t programmed to play certain hands in certain positions. It simply observes a million hands and learns what works. It doesn’t differentiate by hand strength the way we play. It looks at the profitability of each situation and acts accordingly.
- This bot uses fewer computer resources than other bots, which have been designed and programed to beat humans in other games. It only needs about $150 of hardware and drives and cloud computing resources.
- The bot’s success and innovation is a boon to AI and will be applied to many other situation-analysis programs.
- Pluribus raises frequently, donk bets, bluffs and isn’t afraid to bet big.
What can we learn from Pluribus?
- I’m constantly nagging my students to bet bigger, stronger and to not get too hung up on when opponents fold. Often, students ask if they should’ve extracted more value. Hell no. Value comes from putting pressure on opponents and when we lean on them enough, they will respond with calls, which gives you their stack. All poker books stress aggression as part of a winning strategy. Pluribus has learned this technique from observation. Nobody programed this into the bot. In my view, 98 percent of poker players don’t bluff enough. Bet-sizing is an important concept, but isn’t nearly as important as the secret Pluribus has learned. Fold equity is created. Pluribus is happy to scoop up small pots with monster bets and is rewarded with his bet-attack-bet approach.
- Pluribus understands fold equity is pure profit. It will lose some big pots and win some big pots but the uncalled big river bets are where it makes big money. In one hand, Pluribus check-raised all-in and the human poker pro called because he was sure he caught the machine bluffing. Top pair was no good. Action begets action. I’m certain the pro wouldn’t have called that large of a raise from his fellow humans. He was sure Pluribus was bluffing.
- Lastly, when Pluribus semibluffs, it bets big. Who among us will call off our $10K stack with top pair? Aggression pays. — Mark Brement