Facebook has developed an AI that can beat humans at poker, but it doesn't want to reveal its code

Facebook has developed an AI that can beat humans at poker, but it doesn't want to reveal its code

Poker will never be the same again. On Thursday, Facebook and Carnegie Mellon University announced that a joint team of researchers has succeeded in developing an artificial intelligence-based software capable of beating some of the best no-limit and six-player Texas Hold’em poker professionals. The complexity of poker, the multiplicity of its participants and the little information available to players have long made it an important milestone for AI researchers. This is a fiendishly difficult problem to solve, and one that has other issues in the real world, concerning for example autonomous cars or negotiation.

But now, like the game of go or chess, the old ideals of the development of artificial intelligence, researchers have managed to develop robots capable of a “superhuman” performance – and as with previous games, success will probably transform the game itself in a radical way. In an interview with Business Insider US, Noam Brown, an artificial intelligence researcher at Facebook, said the company would not release the bot’s code, named Pluribus, because it was worried about its potential impact on the poker community. He added that Facebook was considering its impact on gaming in the coming years.

The Pluribus AI drastically varies the size of its bets, unlike humans

“It’s going to change the way professional poker is played,” he says. “I think we’ll see some of the approaches” used by AI next in the real world. An example, given in the researchers’ article published in the journal Science: the “donk betting” technique, traditionally mocked by poker players, was regularly used by Pluribus in its successful strategies. “Pluribus diverges from the popular belief that the ‘dong bet’ (starting a round by betting when one has completed the previous bidding round with a call) is a mistake,” the researchers wrote. “Pluribus does this much more often than professional humans.”

He also made sure to balance his bluffs, which made it more difficult to predict when he was bluffing or not. It drastically varied the size of bets while humans, often risk averse, are less likely to use this method – which also makes it harder to decipher.

Also read — Facebook officially unveils the ‘Libra’, its new cryptocurrency that will allow the whole world to pay with its smartphone

Conversely, here is an example of a time when popular belief turned out to be right: “limping” is a bad technique. “Limping” (paying the ‘big blind’ rather than lying down or raising) is a suboptimal technique for any player except the “small blind player who already has half of the big blinds in the pot, and therefore has to invest only half of the sums that other players have to pay to follow”.

This breakthrough could upend online poker

Pluribus has played a lot against human professionals, in the space of 12 days and 10,000 poker hands. But this is still a small number of games compared to the total number of games played every day around the world. As technology becomes more accessible, it will offer fascinating new insights into the unconventional and successful strategies of human-to-human play.

It could also upend online poker games, shaking confidence in the format, as players are wary of playing against invisible opponents for fear of playing against a superhuman AI. Superhuman-level software has had a massive impact on other games. In chess, he helped the birth of a new generation of prodigies like Magnus Carlsen, who grew up playing against AI. It could also lead to making video games an essential part of training programs.

The great masters of the game of Go have already begun to learn lessons from Google’s great AlphaGo software. The effects on poker will no doubt be similar – even if Facebook’s decision to keep the code private will hold up the wave for a while yet. “We chose not to release the code in part because of the potential impact on the poker community, and how it could impact the game,” said Noam Brown.

Using Pluribus was remarkably inexpensive – about $150 in cloud computing resources to train the model. The researcher added that you could probably use similar software on an iPhone with only a slight drop in performance.

This article, written by Rob Price, was first published on Prime.

Receive our latest news

Every day, the main Business Insider news

Go to our cases Get a free quote