As Friday night became Saturday morning, Dong Kim sounded defeated.
Kim is a high-stakes poker player who specializes in no-limit Texas Hold ‘Em. The 28-year-old Korean-American typically matches wits with other top players on high-stakes internet sites or at the big Las Vegas casinos. But this month, he’s in Pittsburgh, playing poker against an artificially intelligent machine designed by two computer scientists at Carnegie Mellon. No computer has ever beaten the top players at no-limit Texas Hold ‘Em, a particularly complex game of cards that serves as the main event at the World Series of Poker. Nearly two years ago, Kim was among the players who defeated an earlier incarnation of the AI at the same casino. But this time is different. Late Friday night, just ten days into this twenty-day contest, Kim told me that he and his fellow humans have no real chance of winning.
“I didn’t realize how good it was until today. I felt like I was playing against someone who was cheating, like it could see my cards,” he said after returning to his hotel room to prep for the next day. “I’m not accusing it of cheating. It was just that good.”
The machine is called Libratus—a Latin word meaning balanced—and Kim says the name is an apt description of the machine’s play. “It does a little bit of everything,” he says. It doesn’t always play the same type of hand in the same way. It may bluff with a bad hand or not. It may bet high with a good hand—or not. That means Kim has trouble finding holes in its game. And if he does find a hole, it disappears the next day.
Jason Les and Daniel McAulay, two of the other top poker players challenging the machine, describe its play in much the same way. After the match’s tenth day, all three players said they could potentially pull out a draw but not a win. “It’s pretty clear at this point that an outright human win is off the table,” Les said. “We’re too deep in the hole.” Since then, they’ve fallen deeper into their hole. By Monday night, though the humans won the day’s play, the machine’s lead over its nearest competitor stood at $701,242.
Which means AI is approaching another milestone moment. Though artificially intelligent machines have already topped the best humans at checkers, chess, Jeopardy!, and even Go, no-limit Texas Hold ‘Em is a very different prospect. That’s because it’s an “imperfect information” game. Since some cards are hidden, players can only see part of what’s happening in the game at any give time. To win, they need intuition—an ability to guess what other players are up to. This is particularly true with no-limit Texas Hold ‘Em, where complex betting strategies play out across dozens of hands.
But this match also highlights the role that humans play in the rise of artificial intelligence. Because the machine’s play changes so distinctly from day to day, filling any holes in its game, its human opponents are sure that those Carnegie Mellon researchers are altering its behavior as the match goes on. Tuomas Sandholm, the Carnegie Mellon professor who oversees Libratus, declines to say whether or not these tweaks are happening. Either way, he and his partner, Carnegie Mellon doctoral student Noam Brown, are active participants in this drama. Chances are, they’re altering the machine from day-to-day. And if they’re not, they’re at least being coy in an effort to keep Dong Kim and the other human players guessing—another way of altering the course of the match.
If that seems unfair, well, it’s just how AI works. Humans are always changing it, as they push towards ever greater possibilities, and in many cases, they work right alongside it, because that’s often the best way of realizing those possibilities.
As Kim points out, Sandholm really wants to win. “He is a very accomplished person,” Kim says. “I don’t think he takes losing very easily.” It’s a quality that defines many top researchers in the world of AI, an area where game playing is so often a springboard to something more.
Yes, the modern AI movement—which has spread so quickly across the giants of the internet, including Google, Facebook, Microsoft, and Amazon—is characterized by widespread collaboration. Because so many of the top researchers are academics—or they come from academia—they’re intent on sharing their research in a way that’s altering the corporate culture at many of these companies. But these same researchers are also intent on beating their colleagues to the next breakthrough. Indeed, just before Sandholm and Brown unleashed Libratus in Pittsburgh, rival researchers at the University of Alberta published a paper describing a system that had already beaten many human poker players (though these players weren’t quite at the level of Dong Kim).
Because this very academic breed of competitive sharing is so rapidly mixing with so many corporate dollars, the race toward AI is playing out with unusual speed. Google dollars—and other Google leverage—helped produce a machine that cracked the ancient game of Go ten years ahead of schedule. And the same sort of rapid-fire progress is overtaking the broader tech marketplace.
How AI Works
But the poker match in Pittsburgh also shows that the distinction between AI and human is a blurry one. Humans and AI compete, but they also collaborate—often the best way to reach the best result. Humans are the ones building and constantly rebuilding artificial intelligence systems (at least for now), and many times, they do so in ways that seem to bely conventional notions of artificial intelligence. Inside companies like Google, for instance, linguists are hand-labeling vast amounts of data to help train neural networks to understand natural language.
These days, when AI is put to use, it also operates alongside humans. This is how Facebook identifies hate speech, lewd material, and fake news on its network: AI works to identify this content, but human curators ultimately decide it should stay or go. At Google, researchers are developing AI that can identify disease and illness in medical scans, but this technology won’t operate entirely on its own. It will serve as an extra tool for the world’s doctors.
In Pittsburgh, Dong Kim is tired and frustrated and feeling defeated. It doesn’t quite seem fair that Sandholm won’t reveal how Libratus works, whether he’s changing the machine from day to day or not. But we see this all the time: computers and humans collaborating to build tomorrow’s unbeatable systems. It’s how the game is played.