World Model

A world model is a system's internal picture of how its environment works — what objects persist, how they move, what follows from an action. Given that picture, a system can predict consequences before acting, and can practice inside its own simulation rather than only in the real world.

The idea matters most where acting is expensive or dangerous. A robot that can predict the result of a grasp can rehearse thousands of attempts internally; one that cannot must try them physically. The same principle underlies video generation, where producing a coherent clip requires some representation of objects persisting and physics roughly holding across frames.

Whether today's language models contain world models is genuinely contested. They demonstrably encode structured relationships that look like understanding of how things work, and they also fail in ways suggesting the representation is thinner than it appears. Treat confident claims in either direction with suspicion.

This is the concept underneath most of the current push into robotics and physical AI, and it is the reason video generation and embodied AI are converging — both need a system that knows what happens next. It is also the crux of a serious research disagreement about whether scaling language models produces genuine understanding or a very good approximation of it.

Mostly a lens for reading research and vendor claims rather than a knob you turn. When a robotics or video vendor claims a world model, the question worth asking is what it predicts and how that prediction was validated — the phrase is doing heavy marketing duty at the moment.