KataGo v1.7.0 is released, https://github.com/lightvector/KataGo/r ... tag/v1.7.0 supporting CUDA 11 (although users have had mixed results with CUDA 11 being better or not) and merging a bunch of code that allows running a community-distributed run, in preparation for starting such a run in the near future. In fact, the server code for a distributed run is also just about finished, and seems to be working. So mostly I just need to move around a bunch of data, do some preprocessing to generate a bit of necessary meta stuff we'll need for the run, and then open up the server.
Secondly: recently KataGo has found what seems to be an improvement to MCTS search, maybe adding around 20 to 60 Elo to many different nets and configurations. And loosely sort of contradicting what I myself said earlier in a different thread about the applicability of "blind" heuristic methods relative to neural nets.
(although, careful thought about that issue was part of the design of this method). This is not part of a release yet (it's not in v1.7.0) but is pushed to master and you can try it if you know how to compile KataGo. The method is described here: https://github.com/lightvector/KataGo/b ... correctionAnd lastly: the rest of that doc is a doc that I just pushed to tie up some loose ends by describing most of the remaining notable techniques implemented in KataGo's training that aren't documented elsewhere.
Enjoy!