Over the course of a 12-day gambling spree, an AI named Pluribus sat at a table with five of the world’s best human poker players for 10,000 hands of Texas Hold ‘Em—and dominated. If each of their virtual chips had been worth $1, the computer would have cleaned the humans out at a rate of about $1,000 an hour.
In the field of AI research, this is a clear breakthrough: Game theorists and computer scientists have been trying to design a machine that could beat humans at poker for decades. Mastering six-player Texas Hold ‘Em shows that the AI can handle complicated situations with hidden information—an arena of deceit once thought to be the exclusive domain of human cunning.
What is less clear is what return the research’s funders—including Facebook and, indirectly, the United States Army—might hope to get on their investments.
The gambling machine was built by Carnegie Mellon computer science professor Tuomas Sandholm and his PhD student Noam Brown, who reported their work in a paper in Science today (June 11). Since November, Brown has worked as an AI researcher at Facebook, which bankrolled the $56,000 in prize money divided among the experiment’s 15 poker players. Meanwhile, Sandholm’s game theory research, applied in this poker bot, was in part funded by a grant from the US Army Research Office (ARO). The National Science Foundation and the Pittsburgh Supercomputing Center also contributed to the research.
What could get such seemingly misaligned funders involved in AI poker research? “The techniques we’re using are very general and they can be applied in a wide variety of settings,” Brown said. “This is really fundamental for getting AI into the real world, because so many real world situations involve hidden information.”
But when asked what kind of real-world applications Facebook might be interested in, Brown demurred. “I think right now that this project is just focused on pure research,” he said. “And this is not an anomaly. Companies have pure research labs that are just focused on developing fundamental research techniques.”
Facebook, however, makes one thing crystal clear in its press release on the poker bot: What the research will not be used for. Pluribus’s “poker-specific code, written as a collaboration between Carnegie Mellon and Facebook for the current study, … will not be applied to defense applications.”
That might seem like an odd thing to have to clarify about a card-playing computer, but Sandholm has launched another company, Strategy Robot, Inc., dedicated to selling military applications of his lab’s research. (Sandholm has another start-up, Optimized Markets, that uses his lab’s research to create better strategies for selling ad campaigns; he said his company does not work with Facebook to sell ads.) In fact, while giving a talk to Microsoft engineers last year about a predecessor to Pluribus—a poker-playing robot named Libratus, which Facebook did not fund—Sandholm’s final slide listed “war gaming,” “force planning & acquisition,” and “strategic, operative, and tactical level planning in adversarial military settings” among potential applications for the AI.
“During the Libratus part of the project, the defense component was front and center,” said Facebook spokesperson Ari Entin, “so we just wanted to let people know where we stood now that Facebook is involved in the project.”
“The incremental code developed for Pluribus, on which Facebook was a collaborator, is poker-specific,” Sandholm said, “so that code cannot be used for any other application except poker.”
But while the code is tailor-made for poker, he said the underlying concepts can be applied to developing a wide range of AI applications, from automating business negotiations to directing cybersecurity efforts.
In fact, his grant from the Army Research Office is for an even more out-there application: using AI to help steer the evolution of biological organisms. In a 2015 paper (pdf), Sandholm proposes “the wild idea of steering evolution/adaptation strategically,” using game theory to find ways to defeat disease and engineer bacteria that can clean up environmental toxins.
Purush Iyer, who leads the ARO’s research on a branch of AI and game theory dubbed “adversarial reasoning,” says the Army has given Sandholm’s Carnegie Mellon lab a grant of more than $100,000 per year for the past three years to fund his lab’s wide-ranging AI work. (WIRED reports that the Army also has a two-year, $10 million contract with Strategy Robot.)
Iyer said the Army is interested in the underlying mathematics and game theory behind Sandholm’s work, as well as its potential medical applications. “In a battlefield there could be soldiers that are getting infected due to whatever the local conditions are, and you would want to be able to fix them pretty quickly so that they can get on their feet very quickly,” he said.
He said this kind of AI research could also be used to model dogfights, or help policy makers navigate precarious geopolitical situations, like Iran’s apparent sabotage of oil tankers in the Strait of Hormuz. “At the end of the day, we have a limited budget, and we want the maximum punishment for the adversary that’s trying to attack us, whether it be on the battlefield or whether it be in the air or whether it be in the cyber domain,” Iyer said.
But Noam Brown isn’t interested in inflicting maximum punishment on any adversary. He just wants to push the field of AI forward. And he wants to stress that his code really couldn’t be used for defense purposes—unless Russia, China, Iran, North Korea and Venezuela were to challenge the Pentagon to a high-stakes game of Texas Hold ‘Em. “At the end of the day, this is a poker bot,” he said. “It plays six-player poker.”