A turn-based strategy game, an AI that learned it by playing against itself, and the workflow that trains it with as little human input as possible.
Runs in your browser. Made for a laptop or desktop screen.
Two sides on a hex map, three robot types, income from the cities and factories you can connect to your base, and supply lines that decide which units can still fight back. Small enough to learn in a few minutes, deep enough that the AI keeps finding new lines.
A convolutional policy network of about 1.1 million parameters, trained with PPO by playing against itself on a GPU copy of the rules, over more than 170 training rounds since May 2026. The version on this site runs entirely in your browser; nothing is sent to a server.
Most of the engineering, analysis and bookkeeping is done by LLM coding sessions working under a written set of rules. I set the direction, approve every training round and own the game design. Language models also play the game as scouts, looking for what the AI does not know.
An MCP server that turns the game into text, what the scouts found that no test did, and what they are not good for.
The contract every session works under, how a training round is judged, and three times the instruments were wrong.
Keeping a TypeScript game and its GPU training copy identical, down to the rounding, and the browser AI as a third copy.
The AI beats every scripted opponent I have written, in every game. It is not unbeatable. It expands along one familiar route, and on randomly generated maps a copy of it nudged to commit to a different expansion beats it about four games in five. A human who learns its habits can beat it too. Training continues, and so does this log.
The source code is private; these pages describe it.