A New Overlord: Google's AI DeepMind Masters Chess and Dominates the Game in Just Four Hours

Thousands of Indian players play chess at a university ground Amit Dave/Reuters

Google's AlphaZero program, after teaching itself chess in a matter of hours, beat the world's greatest computerized chess player. Playing 100 matches against the former world champion chess program, Stockfish 8, Google's program came out on top.

AlphaZero developed its skills completely from scratch: as researchers write in a a pre-printed, unreviewed paper, the program was given no knowledge of chess except for the basic rules. As Wired notes, humans have long been given a run for their money by computer players at several different games. AlphaZero was taught to play three games: chess, Shogi (a Japanese game similar to chess) and Go.

According to Wired it took AlphaZero just four hours to become a chess champion, two hours for shogi, and eight hours to defeat the world's greatest Go-playing computer program.

It was big news last May when that Go-playing program, which AlphaZero has beaten, took down the world's greatest human player of the game.

"From a scientific point of view, it's the latest in a series of dazzling results that DeepMind has produced," University of Oxford computer scientist Michael Wooldridge told the BBC.

Men in Hungary play chess at a bath house. Bernadett Szabo / Reuters

This is far from the first time in recent memory that artificial intelligence has beaten a game champion. Earlier this year, Google's AlphaGo program bested the world's greatest human player at the strategy game Go.

There is a storied history of artificial intelligence programs taking on the greatest human game players. In the 1990s, Russian chess master Garry Kasparov was defeated by IBM's supercomputer, Deep Blue.

According to Wired, DeepMind is looking at the computer game Starcraft as the next frontier.

Though it's certainly exciting that a computer is able to attain these feats, it's not exactly a reason to be praising our new "overlords" just yet.

As The Guardian notes, playing these games is a single, "discrete" task. And according to Wired, while AlphaZero is a step towards developing AIs that are "less specialized," and able to generalize rather than solve just one particular problem, there's still a ways to go before these programs can do that on a human level.

Wired highlights the response of NYU AI and games researcher Julian Togelius, who writes that, while "this is excellent work and I'm very impressed," AI that can actually generalize is a while off.