Just for those people who hang out more here than reddit or L19:
KataGo today has uploaded the final neural nets from its latest run, and which are a big step up from any previous nets due to a learning rate drop. They can be found here: https://github.com/lightvector/KataGo/releases/tag/v1.4.5
See the above link for the release and the new nets and executables and/or see this reddit post for more details.
For anyone interested in seeing some games by this latest neural net, the recent series of games from kata-bot-test in the last day or so are the new 40x256 net duking it out against RoyalZeroSlow while limited to about 5000 visits.
kata-bot has also been updated with it as well, and the recent games between it and JBXKataBot2 from 2020-06-19 to 2020-06-20 are also the new net stomping on its former older version (but still losing a few games!) as JBXKataBot2 had only been using the prior released network. Note that JBXKataBot2 also sometimes interleaved this with games against other bots too, so you might have to skip through the history at points.
Although it could definitely go longer and improve more, the run has reached a reasonable stopping point and so I’m stopping it for now due to the cost of continuing. It’s been a great run, and I hope people continue to enjoy it. Thanks to everyone who gave their enthusiasm and support, and for those who contributed or helped test things out.
And if you’re curious about where KataGo might be headed in the future after this, I did include some thoughts in that reddit post above so feel free to check that out. In particular, - in the future maybe look out for a community-distributed run, and/or message me if you have solid web dev experience that might help make that possible. But other than that, I do plan to continue working on KataGo.
Yay.