mstembera

What I'm really wondering (and communicating poorly) is if the LC0 team has published anything about how transformers apply specifically to chess. [Edit] What puzzles me most is that transformers...

This seems interesting https://github.com/Ergodice/lczero-training/blob/attention-net-body/README.md

I wonder if it would make most sense to write ```C++ const bool betterVotingValue = thread_voting_value(th) > thread_voting_value(bestThread) || (newThreadPV[0] == bestThreadPV[0] && newThreadPV.size() > bestThreadPV.size()); ``` It's simpler and...

Yes my suggestion above does disable thread_voting_value which wasn't my intention so i retract it. However, I now realize that with this patch we are back to allowing truncated pv's...

> according to the logic in the current master, we still allow truncated PVs if a move has an unparalleled best votes and the best thread is truncated, In master...

Regarding the test links you posted I am not against anything that passes elo gaining bounds. This patch has not done that yet it reintroduces truncated pv's.

This isn't a new "viewer experience improvement" patch though because it has truncated pv's. If someone wants the original patch(currently in master) tested for no regression I'm not opposed.

Lol We want to minimize truncated pv's as much as possible w/o losing elo. This patch is a step backwards toward that goal.

As will this patch except it also will in additional cases. Nothing imaginary.

Hmm I'm even less sure. Maybe open a PR for the fix as an RFC and see what the maintainers and others say?