World models: AI that learns how the world works

Ask a language model what happens to a glass standing at the edge of a table when someone bumps into it, and it will answer correctly: the glass falls and probably breaks. But it knows this because it has read thousands of sentences describing it, not because it pictures the table, the glass and gravity. For writing an email, that is not a problem. For steering a robot, driving a vehicle or simulating a factory, it becomes one.

That is what world models are about, and they have become one of the most active research topics in artificial intelligence over the past two years.

An old idea making a strong comeback

A world model is a system that learns how an environment changes over time. You show it a situation and an action, and it predicts what happens next. The idea is not new: we all do this constantly without thinking. Before crossing the road, we picture the path of the oncoming car. Before putting down a cup, we know it will stay on the table and not on the sloping edge of a book.

In AI, the term was popularised in 2018 by research in which an agent learned to play inside a "dreamed" version of a video game, simulated by its own model. Since then, several researchers, including Yann LeCun, have argued that language models alone will not lead to an intelligence able to act in the real world, and that we need systems that first learn by watching, the way a child observes before speaking.

What changed recently

The topic left the lab in 2025. Several major players presented models trained on huge amounts of video to learn how the world behaves. Some generate interactive environments in real time, where you can move around and objects react consistently. Others are designed for robotics and self-driving: they produce realistic training situations or help a robot plan its movements before carrying them out.

There are two broad families. The first predicts the next images directly, which makes for spectacular demos but is expensive to compute. The second predicts in a more abstract space, ignoring irrelevant detail such as the exact shape of a tree's leaves, to focus on what matters: where objects are heading, what is stable, what is about to fall.

What it is actually used for

• Robotics: a robot that can simulate the result of a movement before making it learns faster and breaks fewer things.
• Vehicles and drones: thousands of rare situations, like a snowy road or a pedestrian stepping out, can be generated without having to film them.
• Industry: a model that has learned the dynamics of a production line can anticipate drift or test a setting before applying it.
• Simulation and training: environments that respond believably to the user's actions, for practising situations that are hard to reproduce.

How it connects to our work

This approach ties in with work we already do. In predictive maintenance, we use models that learn a machine's dynamics to spot the moment it starts to drift. That is a kind of world model, at the scale of a single piece of equipment. The digital twin points the same way: a virtual copy is only truly valuable if it behaves like the original.

For serious games and virtual reality, world models promise scenarios that are less scripted, where the environment reacts plausibly to what the learner does instead of following a branching script written in advance. We are not there yet for production projects, but it is a direction we are following closely.

What to keep in mind

World models are still a young technology. Public demos hold together for a few minutes before consistency breaks down, the models are very expensive to train, and the question of reliability, which is critical as soon as robots or vehicles are involved, is far from settled. As always, a simulation is only worth something if you can measure how closely it matches reality.

In short

A world model learns how the world changes in order to predict what happens next and plan an action. After language models, it is one of the most serious paths towards giving AI an understanding of the physical world. The first applications are in robotics, self-driving and industrial simulation, and they point to more realistic digital twins and immersive experiences in the years ahead.