Go-playing humanity has suffered a decisive defeat: On Tuesday night, Google’s artificial neural network AlphaGo beat Lee Sedol, the best Homo sapiens had to offer, completing a four-to-one series that felt one-sided from the start. Go joins checkers, chess, Scrabble, and Jeopardy on the list of games conquered by artificial intelligence. What will humanity lose at next?
We know the next game won’t look like chess or Go, because there’s simply nothing like them left for an artificial intelligence to tackle. We’ve run out of those sorts of games.
“Go was the highest mountain of deterministic perfect information games,” says artificial intelligence programmer Bruce Wilcox, who’s built several Go-playing algorithms. (He’s since moved on from Go to create more intelligent artificial conversations.) “It involved projecting the future using legal moves, to steer the board toward an evaluation in your favor.”
A deterministic game is, essentially, a game in which a given input always and exactly equals a known output. In chess, to move a rook three spaces you move it exactly three spaces. Nothing funky happens along the way — God doesn’t play dice with a chessboard. Contrast that with drawing a card in, say, Magic: The Gathering. You know you get one card, and you might know the probability of drawing a given card, but you can’t know exactly what you’re going to end up with. Maybe it’s an island, maybe it’s a Lovecraftian tentacle-beast. Moreover, in Magic or, sure, poker, there’s hidden information. Other players grip unknown cards.
There is no bluffing in Go or chess.
Imperfection does not mean it’s impossible for computers, just tougher. The University of Alberta, which has a dedicated Computer Poker Research Group, declared its computer scientists had “solved” limited Texas Hold ‘em in the journal Science in January 2015. It’s an impressive algorithm, relying on a “billion” aritifical hands to make its decisions, but there were some caveats. Namely, poker can’t be solved, unlike tic-tac-toe, but it can probabilistically win over and over again; the Canadians say for every $1,000 big blind, the best a player walks away with is $1 a hand.
Predicting when, exactly, computers will beat poker pros in no-limit, multiple player games is a fool’s errand. Go was considered a decade off, at least. Until, of course, AlphaGo stomped its way into history. As Elon Musk wrote on Twitter, “Many experts in the field thought AI was 10 years away from achieving this.”
That algorithm can also only handle two players. A full table, or even a no limit game, is beyond what computers can currently muster. In a no-limit game, a Carnegie Mellon University program soundly lost to pros in a Vegas exhibition last May. Though one researcher told the L. A. Times the matches weren’t statistically significant — a draw — by the end of the tournament the four poker pros had a cumulative lead of $700,000 in chips. There’s a ways to go before an A.I. can mimic the life of the poker playing mind.
Computers will have to bluff. “Games with hidden information, like poker, still have an interest due to the psychology of misleading your opponent,” Wilcox points out. Welcome to our world.