I stayed up late last night to watch as much as I could of the second game between the world’s top Go player Ke Jie and the whiz-bang AI bot Alpha Go. Once again the game was streamed on the American Go Association YouTube channel, with expert commentary.
It was a tremendously complex game that found Ke Jie creating a ko battle in an interconnected fight involving precarious groups on the bottom side of the board. Large areas of the board were in flux and Ke Jie’s stones were under great pressure, as he tried to keep the game as complicated as possible. In the end the ko battle was settled in Ke Jie’s favour, but he lost a huge group in the bottom right and there would be no opportunity to make up that loss. Ke Jie finally resigned.
One of the interesting things the commentators discussed before the game is that the current version of Alpha Go uses 10X LESS computing power than the version which played Lee Sedol last year. Alpha Go has the ability to learn and this version was trained by playing millions of games against itself. While it may be possible for the bot to lose an occasional game, the current version has yet to lose and as now played numerous top pros. Fascinating.