Do generative video models learn physical principles from watching videos?