A World Model in Machine Learning is a system that builds an internal representation of an environment and predicts how it changes over time in response to actions. It's an AI that simulates world dynamics, enabling "think before acting" through mental simulations. These models power robots, autonomous driving, and interactive video generation.

Early ideas date to the 1990s with German researchers using Recurrent Neural Networks. In 2022, Yann Le Cun revived the concept, proposing the Joint Embedding Predictive Architecture (JEPA), which predicts abstract concepts rather than next tokens. In 2024, Google introduced its World Model Genie, adopted by California's autonomous taxi Waymo for self-driving simulation.

By 2026, the state-of-the-art includes World Labs' Marble, Alibaba's Happy Oyster, and Nvidia's Cosmos (released June 2026). These models create interactive 3D environments from text prompts. Architectures include JEPA, Recurrent Stochastic State Models (RSSM) like Dreamer, Tree-Search Models like Mu Zero, and Generative Foundation Simulators.

World Models represent a shift from pattern matching to predictive understanding, with significant commercial and research momentum.