World model
Also known as: AI world model, world modeling
A world model doesn't just produce outputs: it maintains an internal representation of how things work and can simulate what would happen next given a set of actions. This is closer to how humans think about the physical and social world. For an AI agent, a world model enables planning by predicting consequences of different actions before committing to one.
In 2025, world models became a major research focus. Google DeepMind, Meta, Nvidia, Tencent, and others all released world models, primarily trained on video data. Sora (OpenAI's video generation model) is sometimes described as implicitly learning a world model because it generates physically plausible scenes. Game-focused world models like those built for open-world simulations can generate interactive 3D environments on the fly.
For builders, world models are currently most relevant in robotics, gaming, and simulation contexts. But the underlying idea: that models which understand causality and dynamics are more capable reasoners than models that only predict the next token, is influencing how researchers think about the next generation of general-purpose AI architectures.