Over the last few days, an unknown Go player named “Master” has won 60 of 61 online matches against some of the best players in the world. Google has now fessed-up, admitting that “Master” is actually the AlphaGo AI, and that it has been secretly playing humans in order to test an improved version.
Google’s AlphaGo, developed by British AI lab DeepMind, made history last year when it became the first AI to beat a professional Go player, namely Lee Sedol. We haven’t heard very much about AlphaGo since the end of the five-game tournament, but it now appears that this computer program has been very busy lately.
Since 29th December, Go players using the online Tygem platform have been getting annihilated by an unknown player named “Master” in accelerated, time-controlled games. And we’re not talking chumps; players defeated included Korea’s Park Jung-hwan (ranked no. 3), and Japan’s Iyama Yuta (ranked no. 5), just to name a few. By 4th January, Master had tallied 60 wins, including a tie and no losses. Naturally, speculation began to emerge that Master was probably some sort of AI.
Turns out this speculation was true. Earlier today, DeepMind founder Demis Hassabis admitted via a tweet that Master is a new version of AlphaGo:
— Demis Hassabis (@demishassabis) January 4, 2017
It’s not immediately clear why DeepMind felt it necessary to disguise AlphaGo, but it was probably done (1) to prevent the human players from playing any differently than they normally do, (2) to avoid being recognised in the event that the new version was a disaster, and (3) to add some master moves to AlphaGo’s database of millions of moves (the more AlphaGo plays, the better it gets).
DeepMind hasn’t disclosed the nature of AlphaGo’s upgrades, but as Hassabis’ tweet states, we can expect to see the improved version in action later this year when its set against human opponents in regular, full length matches. But if these latest test results are of any indication, this won’t be much of a tournament. [New Scientist, Quartz]