Induction Labs Photon-1: Desktop Simulation, Checkers, and Billiard Physics After One Pretraining
Induction Labs released the Imagination Models architecture and the Photon-1 model, which trains on raw video without action labels. Photon-1 can simulate desktop operations, play checkers, and model billiard physics after a single pretraining phase.
Most agents that learn from video are forced to know which action led to each frame. Induction Labs claims this requirement is a bottleneck. Last week, the company introduced Imagination Models—a foundation model architecture that pre-trains on raw video with no action labels at all. The test system was Photon-1, a sparse mixture of experts (MoE) with 106 billion parameters and 5 billion active parameters (106B-A5B). After a single pre-training cycle, the model demonstrates the ability to simulate computer desktop operation, play checkers, and model billiard physics.
Source: MarkTechPost —
original