Home Business Facebook won’t release the code for its AI poker bot that can beat human pros because of ‘the potential impact on the poker community’

Facebook won’t release the code for its AI poker bot that can beat human pros because of ‘the potential impact on the poker community’

by admin2 admin2
9 views
Facebook won’t release the code for its AI poker bot that can beat human pros because of ‘the potential impact on the poker community’

Poker will never be the same again.

On Thursday, Facebook and Carnegie Mellon University announced that a joint team of researchers have managed to build artificial intelligence-powered software capable of beating some of the world’s best poker professionals in games of six-player no-limit Texas Hold’em poker.

Poker’s complexity, its multiple participants, and the limited information available to players have long made it a major milestone that AI researchers have been working towards. It’s a fiendishly difficult problem to solve, and one that has other real-world applications, from self-driving cars to negotiations.

But now, like Go and Chess, previous goals for AI development, researchers have managed to develop bots that are capable of “superhuman” performance at the game — and like the games before it, the achievement will likely radically transform the game itself.

In an interview with Business Insider, Facebook AI research Noam Brown said that the company isn’t releasing the code of the bot, named Pluribus, publicly because of concerns about the potential impact on the poker community, but discussed how it might ultimately impact the game in the years to come.

“It’s going to change the way that professional poker is played,” he said. “I think we’ll see some of the approaches” it used subsequently employed in the real world.

One example, given in the researchers’ paper published in the journal Science: The “donk betting” technique, traditionally derided by poker players, was regularly incorporated by Pluribus in its winning strategies.

“Pluribus disagrees with the folk wisdom that ‘donk betting’ (starting a round by betting when one ended the previous betting round with a call) is a mistake,” the researchers wrote. “Pluribus does this far more often than professional humans do.”

It also worked to balance its bluffs, making it harder to predict when it was bluffing or not, and varied bet sizes wildly in a way that risk-averse humans are less likely to — also making it harder to read.

An example of where conventional wisdom has been proven right: The idea that “limping” is bad. “Limping (calling the ‘big blind’ rather than folding or raising) is suboptimal for any player except the “small blind player who already has half the big blind in the pot by the rules, and thus has to invest only half as much as other players to call.”

Pluribus played extensively against a suite of human pros, over the space of 12 days and 10,000 hands of poker. But this is still a tiny number of games relative to the total number being played all around the world every day. As the technology becomes more widely available, it will offer fascinating new insights into unconventional and successful strategies for human-versus-human play.

It may also drastically upend online poker games, eroding trust in the format as players grow wary of playing against unseen opponents lest they’re playing against superhuman AI.

The impact of superhuman-level software has had a massive impact on other gaming fields. In chess, it helped birth a new generation of prodigies like Magnus Carlsen, who grew up playing AI opponents and incorporate computer games as an essential part of their training routines. Go grandmasters have already begun drawing lessons from Google’s formidable AlphaGo software.

The effects on poker will undoubtedly be similar — even if Facebook’s decision to keep the code private holds back the tide for a little while yet.

“We’ve chosen not to release the code in part because of the potential impact on the poker community, and how this might impact” the game, Brown said.

Running Pluribus was remarkably cheap — costing only around $150-worth of cloud computing resources to train the model — and the researcher said you could likely run similar software on an iPhone with only mild dips in performance.


Got a tip? Contact this reporter via encrypted messaging app Signal at +1 (650) 636-6268 using a non-work phone, email at rprice@businessinsider.com, Telegram or WeChat at robaeprice, or Twitter DM at @robaeprice. (PR pitches by email only, please.) You can also contact Business Insider securely via SecureDrop.


Read more:

Read More

You may also like

Leave a Comment