If you've ever left a poker table penniless, you definitely don't want to go up against Libratus.

Built by a computer science professor and a graduate student, the artificial intelligence system is handily beating pro poker players in a Texas hold'em tournament in Pittsburgh. Two weeks into the 20-day heads up (or one-on-one), no-limit tournament, Libratus is up by more than a million dollars on its human counterparts.

The A.I. system was designed by Tuomas Sandholm, a professor at Carnegie Mellon, and his student, Noam Brown. It's playing thousands of games per day--and winning most of them.

A.I. systems have already wiped the floor with humans at a number of games. Last year, a system from Google's DeepMind defeated world Go champion Lee Sedol in a five game series. IBM's Watson beat some of Jeopardy!'s most successful contestants. And computers have been thrashing humans at chess, checkers, and backgammon for years.

All these competitions are what Sandholm refers to as complete information games. "You know exactly what the state of the world is when you make your move," Sandholm says. You know what the board looks like and your opponent's score.

But in heads up hold'em, which pits two players against each other, the opposing player's cards are an unknown. The fact that the A.I. can overcome that obstacle and work around the information it doesn't have is why this represents such a breakthrough--beating the best of the best requires levels of reasoning and gamesmanship that computers haven't before achieved.

"Heads up, no-limit Texas hold'em is the benchmark that the A.I. community has converged on," Sandholm says. Last year, it came close: A different A.I. system created by Sandholm beat some skilled players, but faltered when it played the top professionals.

Libratus, on the other hand, is showing it can beat anyone. As of January 26, it was up on its opponents by a combined $1,194,402.

Sandholm and Brown started building the system from the ground up in February 2016. The pair used algorithms that quickly compute strategies given the cards dealt, and others that recognize and act upon mistakes made by the opponent. The A.I. can learn, improving its strategy as the competition goes on.

Libratus decides when to bluff and when not to, and must randomize those actions effectively enough so as to not create a pattern the opponent can detect. The system uses game theory to decide on the best moves to make given the unknown information.

"A.I. gets really interesting in those cases where you have a number of unknowns that's equal to or more than your number of knowns," says Abdul Razack, head of platforms at IT firm Infosys. "I haven't before seen a system that's better than a human at handling the unknown."

Being able to make successful moves with limited information could have implications in areas like weather prediction and financial research. Razack believes similar systems could have helped warn of the 2008 financial crisis before it happened.

"In 2008, the unknowns were this big black box, and people pushed more and more things into it until it exploded," he says. "With technology like this, you can reduce the risk of the unknowns--the unknowns become part of the equation."

While that system might come someday, Sandholm doesn't have that application in mind for Libratus. One area where he sees broad use is in business transactions--for example, telling a company or a person which deals to pursue or proposals to accept. When applied to medicine and biology, the system could help guide treatment plans against diseases like cancer: Apply a certain treatment, take measurements, feed the data to the system, and let it decide what comes next. "It becomes a game against the disease," Sandholm says.

Similarly, the system could be used to fight against fraud or cyberattacks--essentially anything that involves going up against an opponent with unpredictable tactics and unknown resources.

Sandholm admits that as smart as Libratus is, no computer would likely ever be able to nail all of the possible scenarios in poker--there are 10 to the 160th power possible situations. Even a small margin of error would likely have big consequences if this kind of A.I. was deployed to deal with complicated financial transactions or medical treatments.

Still, once Libratus cashes in its chips, it could open up a new world of artificial intelligence.

"These algorithms are not for solving poker," Sandholm says. "They're for solving a broad class of situations in which you have incomplete information."