But what exactly are world models?

World models represent a critical advancement in AI, functioning as internal simulations that allow systems to predict how physical environments change over time based on specific actions. These models diverge into two primary architectures: generative approaches, which produce pixel-based visual outputs for data augmentation or interactive environments, and predictive architectures, which operate in abstract latent spaces to capture fundamental physical laws. NVIDIA’s Cosmos family exemplifies the generative school, leveraging physics-first data to improve autonomous vehicle safety, while Meta’s V-JEPA explores predictive, modality-agnostic representations. Beyond synthetic data generation, these models enable real-time planning and model-based reinforcement learning, allowing robots to practice in simulated environments before physical deployment. By moving beyond simple pattern recognition, world models provide the common sense and spatial reasoning necessary for AI to navigate complex, real-world scenarios safely and competently.