Report
PhySim tests AI agents’ predictions about virtual-world physics
PhySim says agents must experiment using sensor readings, without knowing the world’s structure or what the controls do.
TLDR
Built during a Prime Intellect residency, PhySim lets agents run experiments through an opaque interface and scores how well they predict virtual-world dynamics. PhySim reports that GPT-6 Astra earned the highest reward (0.796) among nine predictors in a preliminary evaluation of one world on October 6–7, 2026. Its Blobkit package uses evolutionary search to generate new worlds.
Combined views
51
1 Source, first seen ago
