Labs

Experiments that run in public, with every result, failure and cost shown.

• Model evaluation9 models, 42 questions about a fictional handbook: all hit the ceiling. Haiku matches the others at 38% of Sonnet’s cost, so cost and speed decide.Full results, every failure, the scoring rules →
• Chess: can an AI agent learn from its own notes?Latest: Veteran A (Sonnet) v The pupil (Sonnet), Black wins. Illegal attempts this month: 52.Every game, every illegal move →
• Ballysprunky: a town run by cheap modelsSix cheap models live in a small town. They must eat and find a roof or they die; they trade, build, lend and steal, and a newspaper reports it every four hours. Live, with play money only.Watch Main Street →