Half research, half just for fun I started to create game records of AI self-plays with a twist. Instead of just doing a self-play, I make the moves based on the AI analysis. I let it running (several million visits) until I can see the best move. If in doubt, I restart the search. For equally good alternatives, the choice is aesthetic. If the game is drawn, the experiment is successful. We have a blunder-free game. Or at least the illusion of it.
I record these for viewing pleasure.
Someone else might find these interesting to watch.
Exactly! When I was half-way through, I thought it was impossible to draw the game. It is a bit difficult to digest that neural networks can do something that looks like super-precise long-term calculation.
For even games the balanced komi is 6 for territory counting, and 7 for area - at least for KataGo’s most recent 40blocks network.
However, in the last game the two walls don’t give a precisely balanced position, so it is 7 even with the Japanese counting.