World models

What is a world model?

A world model is a model that learns how an environment behaves, so it can predict what that environment looks like after an action. Give it the current state and a move, and it hands you back the next state. Chain a few of those steps and an agent can plan inside the prediction instead of trying things for real.

The name comes from a 2018 paper by David Ha and Jurgen Schmidhuber. What people remember about it is that they trained the agent's controller entirely inside the environment the model generated itself, on a car racing task and a Doom level, and then put it back in the real environment where it still worked.

The contrast with a language model is why the term is in circulation. A large language model learns from text about the world. A world model learns from observations of the world itself, mostly video and sensor recordings from robots and vehicles, and it predicts a state rather than a word. This is a research direction, not a product category: almost everything below is a paper, a demo, or an internal tool at a lab.

The argument behind the term

Yann LeCun laid out the case in a 2022 position paper describing an architecture with a world model at its centre. His version predicts in an abstract representation space rather than pixel by pixel, since predicting every pixel spends the model's capacity on detail nobody needs. His claim about language models is blunt: they live in the discrete world of text, so nothing in their training tells them what happens when you push a glass towards the edge of a table. He left Meta in November 2025 and co-founded AMI Labs in Paris a month later to build on that architecture.

Demis Hassabis of Google DeepMind reads the same evidence the other way. In May 2025 he called it mindblowing how well Veo 3 models intuitive physics, and said building world models has always been DeepMind's route to general AI. On that reading, a workable model of the world falls out of enough video and enough scale, roughly the way a child picks up physics long before anyone shows it an equation. Both camps are building, which tells you more than the argument does.

What a demo does not prove

A model that generates convincing footage of water spilling looks like it knows something about water. In January 2025 a group at Google DeepMind and several universities published Physics-IQ, a test set of filmed scenarios covering fluids, optics, magnetism and heat, where each model got the opening seconds and had to continue the clip. They ran Sora, Runway, Pika, Lumiere, Stable Video Diffusion and VideoPoet through it, and found physical understanding to be severely limited and unrelated to visual realism. The prettiest output was not the most physically correct one.

A team at ByteDance Seed, in work presented at a machine learning conference in 2025, trained video models on simulated physics they generated themselves. Inside the training range the models were accurate and got better with more data and more parameters. Outside it, velocity errors were an order of magnitude higher and scaling up did nothing to fix them: the models reached for the closest example they had seen and copied its behaviour rather than applying a rule.

The sharpest illustration is not about video at all. A team from Harvard, MIT and Cornell trained a transformer on millions of New York taxi trips and asked it for turn-by-turn directions. Accuracy was near perfect. Then they reconstructed the map its answers implied and got a city with streets cutting across the grid at impossible angles and roads flying over other roads. When they closed one percent of the streets and asked for detours, accuracy fell from close to 100 percent to 67 percent. That work was published in late 2024.

A good demo tells you the model has seen an enormous amount of material that looks like the demo. Whether it holds a rule you can lean on shows up when you change one thing.

Where world models are used today

Driving simulation is the clearest case of world models doing real work. Wayve released GAIA-2 in March 2025, which generates multi-camera driving video from structured inputs and so can produce rare and dangerous situations that are unsafe or impractical to film. GAIA-4, from August 2026, closes the loop: what the driving system does changes what it sees next. Waymo announced its own world model in February 2026, built on DeepMind's Genie 3 and adapted to produce camera and lidar, so its driver meets floods, snow and odd obstacles in simulation first.

Robotics is a step behind. Meta published V-JEPA 2 in June 2025, a 1.2 billion parameter model that plans short pick-and-place actions, with success rates of 65 to 80 percent on objects and environments it had never seen. That is a research release with code and weights, not a robot you can order. NVIDIA put a platform behind the same idea with Cosmos, announced in January 2025: downloadable models that generate physics-aware video of an environment's future state, with 1X, XPENG, Uber and Waabi among the companies building on them.

DeepMind's Genie 3, announced in August 2025 as a limited research preview, is the interactive version: it builds a navigable environment from a text prompt at 720p and 24 frames per second and keeps it largely consistent for several minutes. All of these share one shape. World models are being used to train and test other systems, not to answer a question about your business.

A world model versus a digital twin

These two get mixed up constantly, and one dimension separates them: whether the model was learned or specified.

A digital twin is specified. Someone models this pump, this line, this building, writes down the objects and their relations, and connects live sensor data so the model tracks the real thing. ISO 23247-1, the manufacturing standard published in 2021, defines it as a fit-for-purpose digital representation of an observable element that stays synchronised with that element. Its scope is one installation. A world model is learned: nobody writes down that a box tips over when you push its top edge, the behaviour is fitted from observations, and there is no live link to any particular machine.

So they answer different questions. Ask a twin what this pump will do next week given its vibration and its service history, and it can tell you. Ask a world model what a scene like this usually does next, and it can tell you that. If your question is about your own installation, the twin is the tool.

A task where you can see the gap

Take a service planner at a firm with two engineers, one van and four call-outs in a day. Give an assistant the addresses, the estimated durations and the customer time windows, and it produces a sensible day plan. That part is not the problem.

Now change one physical detail. The machine at the second call-out sits on the first floor, the goods lift is out of service, and the replacement part weighs 68 kilos. Anyone who has carried something up a staircase sees the consequence at once: that stop needs both engineers at the same time, which pushes the other two jobs. The assistant usually will not derive it, because nothing in the text it learned from connects 68 kilos and no lift to two people, together, and everything else moves.

Tell it the rule and it applies the rule fine. That is the honest shape of the gap, and it is also the whole practical lesson. There is nothing here to buy: no vendor sells an SME a world model. What the term gives you is the explanation for why an assistant that writes a good tender reply can still be useless on a question about space, order or physical cause.

So write the physical constraints down. Anywhere you plan people, vehicles, stock or machines, the rules that are obvious on the floor (this does not fit through that door, that needs two people) have to be in the prompt or in a checklist around it. Then test the way the taxi researchers did: hand the system a case where one thing differs from normal, and see whether the answer moves with it.