DeepMind’s AlphaGo is secretly beating human players online

Google DeepMind’s Go-playing AI has done it again. After beating top player Lee Sedol at the ancient Chinese game in 2016, the AlphaGo AI has been secretly taking on more of the world’s best players – and beating them.

For the last few days, an unknown player called “Master” has been thrashing players on an online Go platform called Tygem. Master beat the world number one player Ke Jie twice, and won 50 out of 51 games that it played, drawing the one it didn’t win outright due to an internet connection time out.

There was a lot of speculation about who could be behind the Master account, with many people suspecting that artificial intelligence was making the moves but unsure who could have developed it.

Master revealed

On Thursday, DeepMind founder Demis Hassabis revealed in a tweet that Master was in fact a new version of AlphaGo.

“We’ve been working hard at improving AlphaGo, and over the past few days we’ve played some unofficial online games,” Hassabis said, adding that DeepMind is “excited by the results.”

AlphaGo hit the headlines last year when it beat Sedol, one of the world’s best Go players, 4-1 in a five-game match. Never before had a machine beaten a 9-dan professional, the best Go ranking available.

The AI initially trained its artificial neural networks from a database of 30 million moves played by expert human players. It then honed its skills by playing thousands of games against itself. Based on this experience, AlphaGo was able to evaluate how different moves would impact its chance of winning at each stage of the game in order to decide how to play. The details of DeepMind’s latest tweaks to the system have not yet been revealed.

According to Hassabis’s tweet, we can expect more human-AI Go games this year. “Now that our unofficial testing is complete, we’re looking forward to playing some official, full-length games later this year in collaboration with Go organisations and experts,” he said.