LLMs often can’t beat a decent amateur at chess, and sometimes lose to total beginners. Researchers and hobbyists testing models like ChatGPT have documented them making blatantly illegal moves, moving pieces that aren’t there, or confidently “winning” games they’ve actually lost track of. This is because large language models are built to generate fluent text: they’re essentially guessing what a chess move “sounds like” rather than actually tracking the board, which is why they shine at conversation but stumble over a game that demands precise, step-by-step logic, and are genuinely bad at playing chess.
More: o3 OpenAI wins Kaggle Game Arena AI Exhibition Chess
Here’s the twist: while LLMs flub chess as players, they’re surprisingly capable chess builders. This has fueled the rise of “vibe coding,” where someone describes what they want in plain English and lets the AI handle the implementation, iterating through conversation until a working engine emerges. The result is a strange inversion of an LLM that might blunder away its queen in a casual match, but can nonetheless hand you the blueprint for a program that plays ruthlessly well.
Entire vibe coded engines have now reached a strength to be featured in the world’s premier chess engine competition TCEC. This will provoke a change in the ruleset , followed by a decision of the tournament director on vibe coded engines, just days before the start of the jubilee 30th edition of the competition.