Simulating everything, sort of: The promise and limits of world models