“Video models like V03 can learn physical concepts such as gravity, buoyancy, lighting, reflection, and refraction emergently by predicting subsequent pixels at a large scale.”