AlphaGo versus Ke Jie

• 1. Introduction

  ›

  AlphaGo versus Ke Jie unfolded across three games in Wuzhen, China, in May 2017. The event formed part of the Future of Go Summit. It pitted a Google DeepMind computer program against Ke Jie, the world's top ranked Go player. AlphaGo won every game. But the true story sits underneath that scoreline. This same program had already beaten Ke Jie online months earlier. Behind the version of AlphaGo playing in Wuzhen, DeepMind was already holding a stronger program secret from the public. What was DeepMind hiding, and what did it cost Ke Jie to face a machine that had already proven itself unbeatable?

• 2. Ranked Number One

  ›

  Since late 2014, Ke Jie had held the number one spot in Rémi Coulom's worldwide ranking system for Go players. That same ranking repeated across the sport's other major authorities. Ke Jie also sat at number one under the Korea Baduk Association, the Japan Go Association, and the Chinese Weiqi Association.

  Yet before the Future of Go Summit even began, AlphaGo Master had already beaten Ke Jie three games to zero. That defeat came during a run of sixty straight online wins over top professionals at the end of 2016 and the beginning of 2017. That version of AlphaGo was called Master, and its exact rating and hardware would soon come into focus.

• 3. The Machine's Secret

  ›

  Four tensor processing units on a single machine gave AlphaGo Master an Elo rating of 4,858. DeepMind said this version played roughly three stones stronger than the AlphaGo that had beaten Lee Sedol.

  Yet Master was not DeepMind's strongest program. The company already held AlphaGo Zero, a version stronger still. Nature had received DeepMind's paper on AlphaGo Zero on the 7th of April 2017, weeks before Ke Jie sat down to play. DeepMind kept Zero's existence hidden from the public until Nature published the paper in October 2017. None of that history was visible on the board, where three real games were about to begin.

• 4. Three Games, Three Outcomes

  ›

  On the 23rd of May, AlphaGo played white and won the first game by half a point, a margin that took 289 moves to settle.

  Two days later, on the 25th of May, Demis Hassabis sent a tweet about an hour into the second game. It said AlphaGo's own evaluations showed Ke Jie playing perfectly. Ke Jie later lost ground on the lower part of the board. About four hours in, AlphaGo simplified the position, and the outcome became clear. AlphaGo, playing black, won by resignation after move 155.

  On the 27th of May, Ke Jie, playing white, resigned in the third game. At the moment of resignation, AlphaGo still had roughly an hour and a half left on its clock. Ke Jie had only about 32 minutes remaining. Off the board, real money and a decision about AlphaGo's own future were already riding on that outcome.

• 5. The Aftermath

  ›

  Google DeepMind put $1.5 million on the table for the winner of the match. It set aside $300,000 for the losing side simply for taking part. AlphaGo won that prize outright, and afterward the program retired from competition entirely. DeepMind did not stop its research, though. The company simply redirected its work toward other areas.

  The Chinese Weiqi Association subsequently awarded AlphaGo a professional 9-dan title. Whatever that title meant on paper, back home in China, the public did not get to watch the match happen live.

• 6. A Broadcast Ban

  ›

  In China, authorities barred the match from being livestreamed. Coverage still reached audiences there, both online and through national television broadcasts on Zhejiang TV. For most of the country, that single channel was the only way to watch the three games unfold.