In Depth
A world model is an AI system's internal simulation of how the world works. It captures the rules, dynamics, and relationships of an environment, allowing the model to predict what will happen next, evaluate hypothetical scenarios, and plan sequences of actions. The concept draws from cognitive science, where humans maintain mental models that enable imagination, planning, and understanding of physical causality.
In AI research, world models take many forms. Video prediction models like Sora implicitly learn physics and spatial relationships. Reinforcement learning agents build world models to plan ahead without needing to take real actions (model-based RL). Yann LeCun has advocated for world models as a path toward more capable and efficient AI, arguing that learning a model of the world is a prerequisite for human-level intelligence.
World models are considered a key frontier in AI research because they could enable more sample-efficient learning, better generalization, and more robust reasoning. An AI with an accurate world model could mentally simulate the consequences of actions, understand cause and effect, and plan complex multi-step strategies. This is distinct from current large language models, which learn statistical patterns in text rather than building causal models of reality.