Libratus defeated poker pros and heads-up specialists Jason Les, Dong Kim, Jimmy Chou and Daniel McAulay for 14.7bb/100 over 120,000 hands, shocking the poker world. The next version recently battled a team of Chinese pro’s who attempted to use AI researchers in addition to poker knowledge to defeat it.
Previous versions of the AI had massive leaks which allowed poker pro’s to exploit it. The Libratus AI not only fixed the major flaws, it adjusted to the human’s attempts to exploit it. However, the players were not allowed a HUD for the match, which is a massive disadvantage for HU specialists who are used to using specific frequencies in order to exploit their opponents. This is a massive advantage for Libratus, which as a supercomputer has the functions of a HUD built in.
The AI employed overbetting strategies, a disgusting 14% lead range, and a 12% river check raise frequency that is difficult for human’s to do in a balanced strategy. This is combined with multiple bet sizes that would be far too difficult for a human to be able to sustain in a balanced way without turning their ranges face up.
Libratus used counterfactual regret minimization to randomize decisions, reinforcement learning to develop strategies using trial and error, and a system to run through the daily matches and identify its own leaks.
And that is only the beginning. The creators of Libratus were not content to simply beat some of the best HU specialists in the game today. They worked on their AI and developed the next version, Lengpudashi, or “cold poker master”. They beat another team of poker pros heads up to cement their win streak.
What does this mean for the future of online poker?
Doug Polk battled the earlier version of Libratus, called Claudico. Footage of the gameplay is available for free on youtube.
We think that he is one of the most suited people to answer the question of what AI means for online poker, which he did in an AMA on reddit two months ago. When asked how many years poker has before AI will overtake humans and make the game unplayable, he had this to say.
“All heads up formats are for sure going to be solved, or close to solved, at some point. The game trees are way smaller compared to the multihanded no limit.
I talked with some of the guys at CMU, and they were hesitant to even take on 3 handed games. The problem in those games, is that you can play a GTO strategy, and lose. One of the two remaining players can operate as the “banker”, and play in a way that funnels money from themselves and another player to the third player. The combination of this, and the multiplication effect on the game tree, make it not that appealing for people not playing for money to get into botting. So I doubt we will ever see these games turn into this (also I think mtts will always be soft because of the nature fo(sic) the game type).”
However, the problem remains that an AI does not have to be able to beat players who are colluding together to funnel money from the AI to a teammate. It only needs to be able to defeat enough of the field of poker players without requiring the massive resources of a supercomputer in order to make online poker infested and unplayable. This could mean even being able to create a slightly above break-even bot that can play tirelessly over many players, earning small sums which add up over the longterm.
Currently, these bots require huge amounts of resources to run. But Noam Brown, a co-developer of Lengpudashi had the following chilling words to say of the technology required to run such an AI: “It’s surprisingly affordable. Within 5 years, this could be running on smartphones.”
HU poker does not have long to go. Hyper Sit and Go’s could be next, due to the short stacked nature of play that means pre-flop played is already solved for many situations in order to play GTO, and 6-max could fall soon after.
For online players? Get the bankroll and skill needed to start playing 5/10 live. In a decade, who knows if online poker will be available?