Induction Labs Photon-1 Simulates Desktops, Plays Checkers, and Models Billiard Physics From One Pretraining Run

A foundation model was trained on raw video of computer screen recordings without any labels identifying which actions produced each frame. The model learns to predict future frames in a compressed representation space, which researchers argue teaches it to understand and complete tasks implicitly. The approach achieved better performance than a baseline model while requiring significantly less training compute and lower inference costs. The model also generalized to domains not seen during pretraining, such as playing checkers and simulating billiard physics, demonstrating that learning from video prediction alone can produce capabilities beyond its original training domain.