The AI Poker Champions

There are all kinds of games that computers are better at than humans these days. Chess. Jeopardy.

Go. An artificial intelligence, or AI, is a computer system designed to solve the problems you’d normally expect to need a human brain to solve. And the latest advance? AIs are getting real good at beating humans at poker. I’m Stefan Chin. And you might recognize me from my riveting performance on the SciShow Quiz Show.

And today I’m here to tell you about how artificial intelligence is taking over the world. In January, an AI called Libratus beat 4 expert human players after playing about 120,000 games of poker. And in a paper published yesterday in the journal Science, a separate research group announced that their AI, called DeepStack, beat 10 out of 11 expert human players after playing about 45,000 games. Both AIs played a version of poker known as Texas Hold ‘em, where each player gets two facedown cards that only they are allowed to look at. There are also five face-up cards that everyone can see, and three rounds of betting.

Most of the games that AIs have conquered so far, like the strategy game Go, are what are known as perfect-information games, meaning that all the players have the same information about the game. For example: in chess or Go, both players can see all the pieces on the board, so they’re making decisions based on the same information. But Texas Hold ‘em is an imperfect-information game.

Since players can’t see each other’s face-down cards, they don’t all have the same information. That makes things much more complicated, because you have to make guesses about the other player’s hand. Like, say your opponent raises the bet.

Is that because they actually have great cards? Or are they bluffing because they think you’re bluffing? And do they think you’re bluffing because in the last round of bets you thought they were bluffing? Those kinds of brain-bendy questions come up in imperfect-information games all the time, and these two new AIs each used different techniques to figure out the most likely answers. They both only played against one opponent per game, which helped because the more players there are, the more ways the game can play out.

But they also both played the no-limit version of Texas Hold ‘em, meaning that the players could bet however much they wanted. And that made things harder, because when you can bet whatever you want, the results of each round affect the way you bet in the later rounds, so the game has way more possible outcomes. Specifically, there were 10^160 possible outcomes for each game. That’s a 1 with 160 zeroes after it. It’s a number so big, that there’s no way even the most powerful computer could actually consider all of those possibilities. For the Libratus AI that won against 4 people in January, researchers first had it play literally trillions of games against itself.

They programmed it to learn from those games so it could work out the best strategies in different situations, based on how the rest of the game would play out. Then, they unleashed Libratus on the four human players in a massive tournament that lasted 20 days. At first, the human players found some weaknesses in the AI’s gameplay, and for the first six days or so, they weren’t losing too badly. But the researchers also designed the AI to learn from its games against the human players. So every night, it would refine its strategies before the next day’s games.

And around the seventh day, the AI started beating the humans by a wider and wider margin. By the end of the tournament, it had won more than $1.2 million. On the other hand, the researchers behind the DeepStack AI, designed it to use neural networks. Neural networks involve layers of processors working together to solve a problem, with each layer using the results of the other layers in its calculations. It’s a strategy that’s modeled after the way brains work, and it’s being used in some of the most advanced AIs in the world.

Like Libratus, DeepStack trained itself by studying random games — although it only looked at about 11 million of them. But Deepstack wasn’t designed to consider how a move would affect the whole rest of the game before deciding on a strategy. Instead, it looked at how different decisions would affect only the next few moves, then used what it had learned about the game to calculate whether those next moves brought it closer to winning. So DeepStack tries to forecast how the next part of the game might go, without trying to predict the whole thing.

And when the researchers had DeepStack play against 11 expert human Texas Hold ‘em players, it outperformed 10 of them across thousands of games. So even though Libratus and Deepstack were designed very differently, both AIs mastered a complicated, imperfect-information game. And now there’s one more thing that computers are better at than humans. But this is a big step toward some broader advancements, too.

There are lots of real-world situations where you have to make decisions even though you’re missing some information, just like in Texas Hold ‘em. And the success of these two AIs means we’re on our way to creating systems that can analyze those kinds of problems better than a human can. For things like deciding on the best treatment for a disease, that’s an awesome plus. And these AIs could also be useful for things like stock trading and diplomacy. One thing’s for sure, though: the future is gonna have some amazing AI poker players.

And I for one welcome our new robot overlords. Thank you for watching this episode of SciShow News, and thanks especially to all of our patrons on Patreon who make this show possible. If you want to help us keep making episodes like this, just go to­. And don’t forget to go to and subscribe!