possible bug in the Route 5 bot?

https://twixt-commentator.duckdns.org/game/blank/105307

That's a very good point, especially about the opening and swap decision. The current model was trained on the full-size board, and the smaller boards are adapted from it rather than independently trained. Local tactical patterns seem to transfer reasonably well, but I agree that opening values and swap decisions may not scale in the same way because the link geometry itself does not scale with the board.
I'm actually planning to run a small 16x16-specific training experiment, so this gives me a very useful test case.
Peyrol's five-minute move time is probably a separate issue, though. Master taking a few seconds is much closer to what I normally expect, so I'd also like to find out which backend his browser is using and whether it is falling back to CPU.

My PC is old and tiny. I don't even have a slot for a GPU. I use Firefox under Linux Mint. If you email me what you want, I can send you files and console output.

Thanks — that explains the very long thinking time quite well. On an older PC without a GPU, the AI may simply be running entirely on the CPU, so several minutes per move is plausible.

That may be separate from the playing-strength issue, though. Alan is seeing fairly weak play on the small board even on an iPhone where the AI runs quickly, so I want to investigate the small-board adaptation independently.

I may take you up on your offer and ask for a small amount of browser console output and system information. That would help me confirm which backend Route 5 is actually using on your machine.

I sent you an email.