Life In 19x19

gowan wrote:But a good "teaching" program would have to understand how humans can think about the game to explain mistakes in a useful way. Rather than simply saying move x is better than move y it would have to explain why move x is better, not just say the probability of winning is higher with x than with y.

Yes and no. It's true that a program that understands "this position is thick" is the best thing that's imaginable. But a program that can find points where your move leads to a sharp drop in evaluation is quite helpful even without that possibility. Similarly, if you think a particular line refutes a move, it can show you why it doesn't.

I can also imagine a MCTS program where you say "what about aji here?" and it finds lines where the opponent attacks and shows you refutations. Doing that well would be almost as much a problem of UI as programming. As time goes on, expect more of that sort of thing to be developed, but right now, the race for strength is so pressing that it'll take time for those ideas to be implemented.