“Video world models are pre-trained video foundation models that can predict future frames in pixels given an action as input, effectively acting as counterfactual simulation engines.”