This market resolves to YES if I lose a game of chess against an LLM by December 31, 2027 (inclusive).
Only games played with standard chess rules count (not 960, crazyhouse, etc.). If the LLM makes an illegal move, it loses automatically. The LLM must not have any outside assistance. Only games where I am seriously trying to win count, though it would count if I lose due to playing too fast, not paying attention, etc. I won't bet in this market.
I'm a National Master (~2000 FIDE). So far no LLM has come close to beating me. I generally try playing any models that seem to be a big step up in capabilities.
Update 2025-10-03 (PST) (AI summary of creator comment): - The LLM may be finetuned.
No outside information/tools that would be disallowed in tournament chess (e.g., engines, opening databases).
No scratch pad/notes for intermediate calculations during the game.
Other tools will be judged case-by-case using a human-equivalence heuristic (allowed only if the human-equivalent is allowed in tournament play).
Update 2025-11-18 (PST) (AI summary of creator comment): Regarding what counts as "seriously trying to win":
If the creator loses while playing casually because they were confident they could beat the LLM that way, this would count and resolve YES
If the creator is experimenting or playing in a style they believe is suboptimal enough to meaningfully decrease their odds of winning (e.g., testing sharp theoretical lines against a strong opening player), this would not count because they wouldn't do that if seriously trying to win
1,000People are also trading
@traders I just played two games against Gemini 3. It managed to get a draw against me!
Low sample size, but by my estimation its opening knowledge is dramatically higher than any previous LLMs I've played. It seems to be well over 2000 FIDE at openings. It's still very weak at tactics and board vision - it hallucinated in its chain of thought many times, but didn't make any illegal moves. At one point it incorrectly called me out for making an illegal move. Tactically it might still be below 1000.
The game it drew against me, it got a completely winning position out of the opening, which I managed to save by getting a perpetual check.
If I had been actually calculating instead of just playing by feel, I don't think it would've had a chance to draw. But playing by feel has been enough to easily crush all LLMs until now.
@DanielJohnston Hey, would you resolve this Yes if an LLM beats you while playing casually? Did you just have your guard down because previous LLMs sucked so much?
@CraigDemel Good questions!
I wrote: "Only games where I am seriously trying to win count, though it would count if I lose due to playing too fast, not paying attention, etc." So if I lost playing casually because I was very confident I could beat it like that (like I was for these two games), I think I'd have to resolve "Yes." But if I'm just experimenting/playing in a style I believe is suboptimal enough to meaningfully decrease my odds of winning (for example, playing a sharp theoretical line against Gemini 3 now that I know its opening strength), then that wouldn't count because I wouldn't do that if were "seriously trying."
I definitely had my guard down! At the least, my definition of "seriously trying" against Gemini 3 has gone up dramatically.
What qualifies as a LLM for the purpose of this question? e.g. can the LLM be fine-tuned or have access to tools?
Also, does the game end upon the LLM making an illegal move?
@fluttershy The LLM can be finetuned. It can't access any outside information that would be disallowed during a tournament game, such as engines, databases, etc. Otherwise, it would only be fair for me to get to use Stockfish too! And that would defeat the whole point lol. It can't use a scratch pad because humans aren't allowed to write notes to themselves during games. I might have to make a judgment call on other tools as they come up, but the heuristic I'll try to use is if the human-equivalent thing is allowed or not.
Yes, if the LLM makes an illegal move that will be counted as a forfeit.