Video generation is shifting from creating simple, appealing clips to building complex virtual environments that follow real-world physics and allow for interaction. The authors from Carnegie Mellon University and Nanyang Technological University introduce a four-generation framework to categorize the evolution of video-based world models. These generations progress from basic visual realism to sophisticated simulations that can reason and plan, culminating in models that handle both predictable and stochastic events. The first generation focuses on generating short, realistic videos but lacks depth in understanding physical laws and task execution. The second generation adds interactive elements, enabling users to influence the generated content through commands or actions. By the third generation, models can simulate real-time scenarios with complex interactions, producing coherent sequences that adapt to user input. The final generation integrates probabilistic reasoning, allowing these models to simulate rare events and uncertainties, functioning as general-purpose simulators. Key characteristics of these evolving models include real-time responsiveness, multi-scale planning, and intrinsic physical fidelity. The authors emphasize the importance of both physical and mental world models, where the former simulates external dynamics and the latter captures cognitive processes. Such comprehensive models have implications for fields like robotics, autonomous driving, and gaming, as they enable systems to learn and generalize from diverse inputs, ultimately enhancing their ability to interact with complex environments.