How do you make a chess bot blunder believably? Mixing Maia and Stockfish isn't enough
space64-llc · reddit · 2026-09-29
A developer building chess AI opponents shared an interesting critique from a playtester: the bot moved its queen away from an attacked rook, then put it right back in danger — "blunders need to feel reasonable, with intention behind them."
Their current setup mixes Maia and Stockfish with custom move-selection rules, but often feels like two different players taking turns. The author's analysis: humans don't just search and evaluate — they recognize familiar shapes, focus attention on weak squares, and carry plans across moves. Mistakes come from attention and stubbornness (e.g., continuing a three-move setup after it stops making sense), not random weakening. The open problem: modeling believable blind spots and continuity of thought without the bot becoming predictably flawed or feeling like it's letting you win. They're soliciting similar work in chess or other games.
More from Fun
- Claude autonomously builds a 'p(doom)' Beat Saber clone in one afternoon — burny_tech · 2026-09-29
- Weekend project: an art gallery built with Astra and ChatGPT Sites — blakesamic · 2026-09-29
- ArtButSports pairs 1891 Sargent painting with sports moments, goes viral — chrisalbon · 2026-09-29
- Early 'Opus 5.5' animation test drops: Into the ANTVERSE — ChrisGPT · 2026-09-29
- Driver swaps FSD subscription for open-source sunnypilot on comma four, 8 months in — SuB8u · 2026-09-29
- A $20-plan Sonnet 5.5 built PULSE//BREACH, a Rez-inspired music rail shooter — AIandDesign · 2026-09-29