Hacker News Reader

Summarizing here for my dear friends, the guy on the other end managed to fine tune a model gpt-3.5-fine-tune against stockfish vs stockfish games to perform at 1200 elo level against stockfish.

I have been proved wrong I should have quit while I was ahead. /s

View on HN

I think your summary is not really correct, maybe I'm missing something. As far as I can read, the approximate takeaways are these:

* LLM chess play is hyper sensitive to the harness being used (see: the section on regurgitation), and only mildly sensitive to fine tuning

* The best ELOs the author was observing were from 1500 to 1750 or so (circa 2024/25). Not grandmaster, but no longer incompetent monkey either.

Reply on news.ycombinator.com