World Models must live in Parallel Worlds

World models learn spatio-temporal representations of a world, enabling them to predict future states, and support interaction, navigation, and simulation capabilities. For generative models to become effective agents in the physical world, they must develop and use world models. We posit that world models must be capable of counterfactual simulation - the ability to reason about what if scenarios. By simulating alternative realities, world models will be more capable, safe and creative when faced with novel, out-of-distribution scenarios. Furthermore, they can transcend mere pattern matching to achieve a true causal understanding of the world, a capability central to human intelligence, and a prerequisite for the next generation of AI agents.