Facebook's artificial intelligence team has created what it describes as a "superhuman" poker champion, a robot with the ability to beat the world's leading human professionals.
Facebook is announcing the bot AI, called Pluribus, as a breakthrough: the first capable of defeating six players, in a game that involves "hidden" information, the cards that have not yet been revealed.
The team's research also makes humble reading for any poker player proud of their ability to detect a "tell".
"We think of bluffing as a very human trait," said Noam Brown, the principal investigator of Facebook's artificial intelligence team, speaking with BBC News.
"But what we see is that deception is actually a mathematical behavior. When the bot 'bluffs', do not see it as deceptive or dishonest, it's the way to earn more money. "
Mr. Brown said that neither he nor Facebook had plans to use AI in real poker games. In fact, the firm has said that it is not publicly disclosing much of the code for fear of having a negative impact on the poker community. I would give a spokesperson, I would provide examples of techniques to other researchers working in AI. A research paper has been published in the journal Science.
Beyond poker, Brown would not be asked what practical use could Facebook have in mind for technology.
"My research is focused on advancing the fundamentals," he said.
"Everything from cybersecurity to detecting fraud, and navigating traffic with cars that drive by themselves, this research could be fundamental."
Facebook created a model in which the AI would play hands against itself "trillions" of times, once programmed with the rules of the game, a technique called "learning by reinforcement". In other words, the practice makes (almost) perfect.
One notable achievement, Brown said, was the relatively low amount of computing power needed to train and execute Pluribus.
The AI only needed $ 150 in cloud computing resources to function. Similar efforts, from the artificial intelligence research store of Google Deepmind, have been based on supercomputers that consist of more than 5,000 specialized processors, at a reported cost of millions of dollars. The reduction in computational power required for artificial intelligence experiments is considered a key obstacle to the development of technology, since the computing power currently required exceeds the speed at which processors are increasingly efficient.
Mr. Brown said that only 20 hours of learning were needed to program artificial intelligence to the capacity of a global poker professional. On average, the bot earned $ 1,000 per hour when playing against five humans. As part of the announcement of the new technology on Facebook, Facebook cited several human poker champions who had been invited to play against the AI.
"Pluribus is a very difficult opponent to play," said Chris Ferguson, a champion of the World Series of Poker.
"It's very difficult to hold it in any kind of hand."
The work will likely make the team popular with AI researchers who have been trying to create an AI that plays poker for more than a decade, but Brown said it was unclear how such advances in the business of poker could be seen. the casinos.
"I was in Las Vegas last week," he joked. "But they did not know about these results, I'll have to go back."
Follow Dave Lee on Twitter @DaveLeeBBC
Do you have more information about this or any other technological history? You can communicate with Dave directly and securely through the Signal encrypted messaging application at: +1 (628) 400-7370