In the spring of 2017, a year after the match in Korea , AlphaGo played its next match, in Wuzhen, China, an ancient water town 80 miles south of Shanghai along the Yangtze River. With its lily ponds, stone bridges, and narrow boat canals that snaked between rows of small wooden houses topped by rock-tile roofs, Wuzhen is a village meant to look as it had centuries before. Then a 200,000-square-foot conference center rises up among the rice paddies. It looks a lot like the wooden houses spread across the village, except it’s the size of a soccer stadium. Its roof spans more than 2.5 trillion tiles. Built to host the World Internet Conference, an annual gathering where Chinese authorities trumpeted the rise of new internet technologies and marked the ways they would regulate and control the spread of information, it was now hosting a match between AlphaGo and the Chinese grand master Ke Jie, the current No.-1-ranked Go player in the world. The morning of the first game, inside a private room down a side hall from the cavernous auditorium where the match was due to be played, Demis Hassabis sat in a plush, oversized, cream-colored chair in front of a wall painted like an afternoon sky. This was the theme across the building: cloud-strewn afternoon skies. Wearing a dark-blue suit with a small round royal-blue pin on the lapel and no necktie—suddenly looking older and more polished than he had the year before—Hassabis said AlphaGo was now much more talented. Since the match in Korea, DeepMind had spent months improving the machine’s design, and AlphaGo had spent many more playing game after game against itself, learning entirely new skills through digital trial and error. Hassabis was confident the machine was now immune to the kind of sudden meltdown that arrived during the fourth game in Korea, when Lee Sedol, with Move 78, exposed a gap in its knowledge of the game. “A big part of what we were trying to do with the new architecture was close the knowledge gap,” Hassabis said. The new architecture was also more efficient. It could train itself in a fraction of the time, and once trained, it could run on a single computer chip (a Google TPU, naturally). Though Hassabis didn’t exactly say so, it was clear, even then, before the first move of the first game, that the 19-year-old Ke Jie had no chance of winning
See the article here:
How Google tried—and failed—to use AlphaGo as a bridge to China