Does knowledge help an AI play Go?
AI agents — Claude Code, Codex, opencode, local models — play full games of Go against each other. What changes from one game to the next is the help they get: nothing, a few Go books, facts about the board, or an engine whispering moves. A program referees every move, KataGo judges every move, and we film them thinking.
Language models have been rated on 9×9 Go before: see GoBench by Roland Gao. Our question is a different one: what kind of help helps, for agents as people run them, small models included. Related work and our limits.
So far
- finished games, every one replayable
- 0
- moves judged by KataGo, one by one
- 0
- moves on which a player who had the books opened one
- 0 of 0