Induction Labs' Photon-1 learns world simulation without action labels

Induction Labs released "imagination models," a foundation model architecture pretrained on raw video with no action labels attached. Their model, Photon-1, is a sparse 106-billion-parameter mixture-of-experts (MoE) system that can simulate desktop interfaces, play checkers, and model billiard-ball physics from a single pretraining run.

Why it matters: Removing the need for action-labeled data matters because labeled interaction data is the scarce, expensive ingredient in training world models and embodied agents -- most video on the internet has no action labels. If this generalizes, it could unlock a far larger pool of training data for agents that need to reason about cause and effect in digital or physical environments.