A world model is an AI system that learns the underlying dynamics of an environment — how objects, space, and physics behave over time — so it can simulate, predict, and interact with that world rather than just generate text or images. World models are widely positioned as the path beyond text-only LLMs.
How it works
Where an LLM predicts the next token of text, a world model predicts the next state of an environment — the next frame, the next physical configuration — conditioned on actions. This lets it serve as an interactive simulator for training robots, agents, and self-driving systems. Leading approaches include Google DeepMind’s Genie 3 (real-time interactive world model, Aug 2025), NVIDIA’s Cosmos platform (CES 2025), OpenAI’s Sora 2 (Sept 2025), and Yann LeCun’s JEPA architecture, which he left Meta in late 2025 to pursue full-time.
Why it matters
World models are the clearest “next paradigm” bet in lab language — the foundation for embodied AI and physical robotics, and a candidate route to more general intelligence than scaling LLMs alone. See After the Agent for how this fits the 2026 roadmap.
Related terms: Embodied AI · Foundation Model · Frontier Model · All glossary entries