In this case the machine, named Libratus, is using artificial intelligence (AI) technology developed at nearby Carnegie Mellon University, a hot bed of AI and robotics research.
The tournament kicked off Wednesday with odds makers favoring the human players 4 or 5 to 1 over Libratus. “But we ended up ahead,” Tuomas Sandholm, CMU professor of computer science told Fortune over the phone, sounding delighted. And that was before Thursday afternoon’s action, at which time Libratus was extending its lead. There’s a live feed if you need play by play.
But still, Sandholm knows the war has just begun. The “Brains Vs. Artificial Intelligence: Upping the Ante” contest calls for 120,000 hands of Heads-Up No-Limit Texas Hold ’em poker to be played over 20 days. Believe it or not that is possible, Sandholm said, with play starting at 11 a.m. and ending at 7 p.m. each day.
And, if the human players—Jason Les, Dong Kim, Daniel McAulay, and Jimmy Chou—prevail, they will share $200,000 in prize money.
In a similar tournament of 80,000 hands of poker played last year, the machine, an earlier iteration of the AI lost, but just barely, to the humans. Team CMU is hoping for a different outcome this time.
This marathon contest is part of the university brain trust’s attempt to show that AI, which has already prevailed against humans in tic-tac-toe, checkers, chess, and Go—can do the same in this particular game.
Here’s why this is a tougher challenge: In poker a single player gets an incomplete view of the overall game. For one thing, players have no idea who holds the outstanding cards, and for another, opponents purposely obfuscate their strategy. In short, they bluff.
“Poker requires intentional deception and withholding of information, and that makes it very hard for computers,” Andrew Moore, dean of CMU’s School of Computer Science, told Fortune during an interview last week.
Sandholm agreed. “What’s so hard is there is imperfect information unlike in chess, checkers or tic-tac-toe,” he noted over the phone on Thursday. “You don’t really know the state of the game and you don’t really know what’s happened in the past since the other players are shielding it from you. You don’t know their cards.”
AI seems to have figured out other, less difficult versions of poker—AI technology from the University of Alberta, Edmonton, Canada, beat humans at Head’s Up Limited Hold ‘Em, for example. But the current task is harder.
As technical journal IEEE Spectrum put it:
Running through all those scenarios requires heavy-duty number crunching. The CMU team used “tens of millions of hours of CPU time” at a large Pittsburgh supercomputer to create the models used in this work, Moore said. A CPU or central processing unit is a common measure of computer power.
But in the end, this is about much more than winning poker. It’s about the use of computation and game theory, as espoused by mathematician John Nash (the subject of the 2002 multi-Oscar-winning film A Beautiful Mind) in many applications. Game theory is the practice of applying mathematical models to the interactions among two or more players or participants. It is applicable to what we literally know as games (poker and chess, for example) but also to political, financial, and social interactions.
For more on AI, watch:
Anything involving negotiations could use this sort of AI model to finagle the best deal, Moore said.
“The reason this poker tournament is important and not just scientifically interesting is that it will show we can program computers to be deceptive,” he said. That means a computer, perhaps even your smart phone and the virtual assistant running on it, can act as your agent in transactions.
Get Data Sheet, Fortune’s technology newsletter.
Say you need a ride home. Your bot will know what you’re willing to pay but won’t necessarily disclose that to Uber (uber) or Lyft. Instead it will negotiate with them to get the best price by not tipping your hand. Now imagine applying that AI to multi-million-dollar trade deals or other financial transactions stock transactions or real estate deals.
“That’s the sort of computing you’d have to do. And thanks to John Nash, there will be a mathematical way to do it.” Moore added.
The possible problems that could arise from willfully deceptive AI models, are fodder for another story but that’s for another day.
Editor’s Note: (Jan. 12, 2017 9:20 p.m.) This story was updated to correct Tuomas Sandholm’s quote. He said the team ended up ahead, not that it won, after the first day.