Four of the best poker players in the world challenged a computer for 20 days.

And for 20 days, they lost.

Libratus, a poker-playing artificial intelligence developed at Carnegie Mellon University, crushed its human opponents over the course of the 120,000-hand Brains vs. Artificial Intelligence tournament this month at Rivers Casino on Pittsburgh's North Shore.

Collectively, the humans never led. Libratus held a 1.56 million chip lead on the humans entering that last day of play Monday. The humans had fewer than 5,000 hands to try to claw their way back.

The humans sunk deeper in their final day, losing more than 200,000 chips and finishing the tournament about 1.77 million behind Libratus. Dong Kim, one of the pros, was the closest to Libratus all tournament, finishing less than 86,000 behind the computer and winning the largest share of the $200,000 pot split between the four humans.

"Its strategy just seemed to improve every day. It seems like you're playing against a tougher opponent every day," said Jason Les, one of the professional poker players challenging Libratus. "You end up getting in this feeling that everything you do is losing. And it's demoralizing."

Heads-Up No-Limit Texas Hold'em poker, the game of choice for the tournament, has been seen as a last bastion of resistance to computer gaming dominance. Computers have beaten the best humans at checkers, chess, "Jeopardy," the game Go and other versions of poker. The humans beat Claudico, an earlier version of Libratus, in the inaugural Brains vs. Artificial Intelligence tournament in 2015.

But CMU professor Tuomas Sandholm and his graduate student, Noam Brown, came back with a more sophisticated and powerful pokerbot this time.

"I thought we had an edge over the humans, but I didn't think it would be this large," Brown said Monday as Libratus finished mopping up against the pros.

Even with fewer than 5,000 hands to play, Brown wasn't divulging any of Libratus' secrets. Those would all be made public when he and Sandholm publish their research. Brown posted on his website Monday a paper on how Libratus made decisions on the turn and the river cards (fourth and fifth cards out of five) that he will present next week at an artificial intelligence conference.

Brown said he won't unleash Libratus on unsuspecting online poker players — it is powered by a $9.65 million supercomputer, so it's a little expensive to run — but will continue to study the data generated during the tournament in an effort to perfect the ability of AI to make decisions with imperfect information.

Poker presents a unique challenge for AI because opposing players keep their cards hidden, forcing the computer to make decisions without knowing all the information. Heads-Up No-Limit Texas Hold'em adds the uncertainty of betting and bluffing to the mix.

When the tournament started, Andrew Moore, dean of CMU's School of Computer Science, said an artificial intelligence proficient with imperfect information scenarios can be applied in business, health, the military and other situations. It eventually could power a cellphone app designed to buy a car for you that knows you're willing to pay $5,000 but hides that from the sales person as it negotiates.

Professional poker player Daniel McAuley said Libratus was impressive. He wasn't about to make excuses but said fatigue was a factor. The pros sometimes played against Libratus for 12 hours a day, often ending after 10 p.m. Then they would grab a quick dinner before meeting to study that day's hands and plan strategy for the next day. After a couple of hours of sleep, they were back in front of their computers.

McAuley rarely left the hotel or casino. He went out for dinner once — last week, to Emporio: A Meatball Joint on Penn Avenue in Downtown, because it had a high Yelp rating. He saw the Duquesne Incline on his short Uber rides to and from the casino and hoped he could check it out before he leaves the city.

Libratus, on the other hand, didn't need to eat or sleep and uses much of the supercomputer powering it to study and strategize while it plays. It appeared to adjust to the humans' strategy as a whole and to individual styles of play. If a player started calling every bet, Libratus stopped bluffing, McAuley said.

When the humans thought they found a weakness, Libratus changed course. McAuley thought they underestimated Libratus at the beginning.

"If we knew it was this good, we wouldn't have made it as obvious as to what we were doing," McAuley said. "I would love a rematch."

---

Copyright ©2025— Trib Total Media, LLC (TribLIVE.com)

---