David J Wu

If you're on windows, you might be failing to find the relevant DLLs (despite compiling against those libs, windows might not know at runtime where to find the DLLs). Try...

What does "combine training runs" mean? > This makes it unfarmable for elo points unless you're legitimately above its rank. This sounds difficult to achieve. My strong intuition is that...

@poptangtwe This is an interesting rules proposal, and although I'd need to think about it more, I do think it may be a meaningfully distinct and interesting rules variation to...

`writetrainingdata` *was* designed for human sgfs, not AI self-play data. :) But you should think of `writetrainingdata` not as a command that you would ever simply run, but a massive...

If we're still interested in this, it doesn't only have to be me to try all the stuff! Too much of a bottleneck. :) A way that anyone else could...

Possibly relevant in the sense that it affects the model code and training data format - KataGo has added some new targets to help learn Japanese rules and is also...

Thanks, yes, this would probably be a good idea. It would take a bit of work to implement at this point because for distributed training because it would require a...

We didn't ever implement this feature, unfortunately. It might still be a good idea, although one major downside is that it might well double or triple the disk space size...

> I would have expected each of those variants to correspond to a specific network, trained on dedicated training data. And so, I would have expected b18c384nbt-humanv0.bin.gz to be like...

Not a bug. KataGo deliberately tolerates moves that violate the rules if the GTP controller asks it to make the move, because in practice there are many GTP controllers that...