I’d like to draw some parallels between AI math and AI chess: 1. There are certain moves that chess engines make that always leave you questioning “why’s it doing that” that just no human would’ve thought of. I’m shocked this has the potential to exist in math (but not really). There are also algorithms to identify computer-like behavior in chess players, somewhat similar to AI plagiarism detectors in essays 2. Ppl who play chess at the highest level can trick smaller engines by playing anti-computer-like chess and win at a higher rate than if they were to use their regular playing style. There are some fundamental differences here but this really makes me think of prompt injecting an LLM or even just hallucinations in general 3. We’ve indexed a lot of chess games and trained engines on them but it really wasn’t enough. Ultimately it was making the engine to play against itself that led to it becoming very good at the game… in the same way that feeding LLMs the entire internet isn’t enough and setting up really good RL environments is what’s really advancing the frontier There’s probably a lot more that I’m missing here