I think you are reading too far into this. The title of the technical paper is “ Video generation models as world simulators”.

This is “just” a transformer that takes in a sequence of noisy image (video frame) tokens + prompt, and produces a sequence of less noisy video tokens. Repeat until noise gone.

The point they’re making, which is totally valid, is that in order for such a model to produce videos with realistic physics, the underlying model is forced to learn a model of physics (a “world simulation”).