Only one LLM is good at chess

https://dynomight.net/chess/

Very interesting post outlining the results of letting LLMs play chess. All models were “terrible” except for turbo-instruct which was excellent.

Was it tuned for chess? Were games a big part of its training set?

Edit: seems like1 post-training is dulling chess ability, but it can be recovered with fine-tuning?

  1. https://x.com/willdepue/status/1857308465624146311?s=46 ↩