Researchers at Induction Labs have developed an imagination model architecture that can learn from raw video without needing action labels. Their test system, Photon-1, is a 106 billion parameter model that has demonstrated the ability to simulate desktop environments, play checkers, and model billiard physics. These skills were achieved from a single pretraining run, suggesting that the model's architecture is capable of extracting relevant information from video data. This breakthrough has implications for the development of more efficient and effective AI models.