30 Weird Chess Algorithms: Elo World

Tom7 is brilliant for anyone that hasn't seen him before. His Reverse Emulation video ( link ) and Weird Chess Algorithms ( ▶ this video ) are a must watch. His dead pan delivery coupled with the ridiculous amount of effort he puts into things that don't _really_ matter is honestly inspiring.

This reminded me of Tom7's video where he made a bunch of ridiculous engines and pitted them against each other (and against "diluted" versions of Stockfish): ▶ this video link

Anything that catches your eye is good but my favorites are: 30 Weird Chess Algorithms: ELO World ▶ this video Reverse Emulating the NES to give it SUPER POWERS! link

Mine are "Harder Drives"[1], "ELO World"[2], and "Printable x86"[3] [1] link [2] ▶ this video [3] link

It's unknown what role if any BPE plays. I was surprised to discover that the final probability of a move is equal to the probability of each token from the root prompt, i.e. even though "1.Nf3 e5" is encoded as ['1', '.', 'N', 'f', '3', ' e', '5'] the probability of e5 seems unaffected by the fact that Nf3 is 3 tokens as opposed to one. You're right that coming up with a token mapping could help things. It's a bit tricky to do that right now. Your options for fitting a custom vocab seems to be "use sentencepiece to fit a vocab, then modify the gpt-2 codebase to use the sentencepiece library for decoding". I am honestly not sure if the output of sentencepiece is compatible with traditional encoders. What I mean is, it doesn't seem to generate an encoder.json + vocab.bpe file. It seemed to be some other kind of format. So I'm not sure if the tooling that has evolved around OpenAI's encoder format would be applicable there. I really don't know, though. According to this slatestarcodex comment, someone got superior results on solely algebraic notation (which looks like g1f3 instead of Nf3): link Another extension that might help is to periodically inject the full FEN board state. This was the format we were going to try next, which injects the full FEN after every move: link I'm so happy to get to work with GPT-2 1.5B. It's been a lot of fun to train. By the way, if you like this kind of thing, you'll love Elo World. ▶ this video

Only somewhat related, but if you found these ideas amusing, you'd probably enjoy watching tom7's video on unusual chess engine strategies: ▶ this video

Tom7 made a video also: ▶ this video If you haven’t seen them his other videos are well worth watching as well.

Suckerpinch (Tom) recently had a video[1] about predicting the position of the oponents pieces from knowning just white. I don't think he's aware of this. Great video anyway. He trained an NN to predict the pieces and then used Stockfish on the resulting position. [1] ▶ this video

In fact, you don't even need an LLM tournament when you can have tom7's Elo World tournament: ▶ this video

> I figured it's how they were able to weaken a chess engine back in the day; can't adjust the overall strength, so add random blunders. In tom7’s Elo World, he does this (“dilutes” strong Chess AIs with a certain percentage of random moves) to smooth the gradient since otherwise it would be impossible to evaluate his terrible chess bots against something like Stockfish since they’d just lose every time. ▶ this video

"I was initially hoping to get some insight on e.g. which pairs of squares had the fewest moves for a given piece etc." This may not be quite what you were asking for, but it's close, and has the advantage that I can link it right now. Tom7's Elo World chess video has where pieces start and end up, and their survival rates, as a chart: ▶ this video

Tom's really mastered his skill of casual misleading, he adds these "wait, what?" moments to every video. A more notorious example will be his explanation of how optical reflection works, check out his "30 Weird Chess Algorithms" ( ▶ this video ) around 10:20. Edit: spelling.