Simulating everything, sort of: The promise and limits of world models
Summary
The article surveys world models—AI systems that aim to simulate or interact with the physical world—through interviews with MIT's Vincent Sitzmann, Runway's Anastasis Germanidis, and World Labs' Ben Mildenhall. It explains definitions, current approaches (video-based diffusion versus latent 3D representations), and practical applications in robotics, 3D asset generation, and scientific simulation, while highlighting data challenges, interface uncertainties, and the bets and funding shaping the field.