This is really great! I’m excited to see this sort of progress being made, and I’m really glad the KataGo exploit is on your radar.
I’m curious if you might improve interpretability of a similar network by providing some “off-board” space for it to use, so that it doesn’t have to use e.g. the top row of the board for information on passing at the expense of squishing the board state down.
Also, to what extent do you think you can scale this sort of approach to a much larger network?
This is really great! I’m excited to see this sort of progress being made, and I’m really glad the KataGo exploit is on your radar.
I’m curious if you might improve interpretability of a similar network by providing some “off-board” space for it to use, so that it doesn’t have to use e.g. the top row of the board for information on passing at the expense of squishing the board state down.
Also, to what extent do you think you can scale this sort of approach to a much larger network?