“Video world models, pre-trained on billions of internet videos, can act as counterfactual simulation engines for robotics by predicting future pixel states based on input actions.”