A simulated world is also a world, and I can easily imagine if it would have been trained on real world data than it would have learner some of the real world's

A simulated world is also a world, and I can easily imagine if it would have been trained on real world data than it would have learner some of the real world's physics the same way, as the big video gen models already showing some of that. But all these models still seem very sample inefficient, they need lot of data to learn some basic rules of the world(s), and even then they are far from and human-like model that's includes math and logic for model the world more accurately...