You hand a slightly physical task to an AI. It describes the steps with total confidence, then gets it wrong the moment the situation shifts. The problem is not its vocabulary. It has no idea what happens when an object drops, slides or blocks a doorway. World models go straight at that hole.
A world model is an AI system that builds an internal representation of its environment in order to simulate what will happen in it. Where a language model predicts the next word, a world model predicts the next state of the world: a trajectory, an impact, the result of a gesture. Here is how these models are trained, who is building them, where they will show up, and why they worry researchers more than chatbots do.
A world model is not an LLM with eyes
Large language models like ChatGPT, Claude and Gemini swallowed enormous amounts of text to produce the most plausible next words. That is statistical prediction applied to language. The illusion of understanding comes from having read enough descriptions of the world to give back a coherent picture of it.
Emmanuelle Guyot, Head of Data Science at the Hellowork group, draws the line in the right place in her interview with Blog du Modérateur: an LLM stops at description. It knows that "the object falls" follows "I let go of the object". It does not know how fast, or whether it will bounce.
A world model learns behaviour instead. Its raw material is not text but images, video and sensor streams. Its goal is not to talk about the world but to extract rules from it: gravity, cause and effect, the way objects and situations evolve over time.
Do not confuse a world model with a video generator. A model like Sora produces plausible images of a world. A world model has to answer a different question: if I act here, what happens next? The second ability does not follow from the first.
890 million euros for a bet with no product
In March 2026, Yann LeCun's startup AMI Labs raised close to 890 million euros to build this kind of model. Before that round, the company was valued at 3 billion euros. Yann LeCun argues something simple: LLMs manufacture credible text, they do not understand the physical world.
To put the number in perspective: 890 million raised against a pre-money valuation of 3 billion means roughly 30% of the company's value issued in one go. For a business that sells no product yet, the order of magnitude says investors are betting on a change of architecture, not on revenue.
The idea itself is old. Jürgen Schmidhuber has worked on agents equipped with an internal representation of their environment since the 1990s. In 2018, with David Ha, he published the founding paper "World Models", where an agent learns to drive inside a simulation it generates itself.
How you train a machine to understand the world
Yann LeCun takes his cue from human development. A baby does not learn physics from a textbook. It watches, grabs, piles up sensory experience and works out the rules. The training principle borrows that logic.
In practice, you show the model masses of images, videos and situations, then ask it to anticipate what comes next. With no manual labelling, Emmanuelle Guyot points out. By getting it wrong over and over, the model learns the consequences of actions.
The architecture described by Ha and Schmidhuber holds three parts, and it is still used as a reference:
- a perception module, which compresses raw images into a compact representation;
- a predictive memory, which learns to move from one state to the next;
- a controller, which picks the action based on those predictions.
The cost of that learning is the first obstacle. Emmanuelle Guyot stresses it: these models need far more images and video than an LLM needs text. And you have to supply the right situations, the ones that reveal what physically happens.
Three schools, three definitions of the same phrase
Nobody agrees on the right way to build a world model, or even on what it should be able to do. Three researchers stand for three different answers.
| Researcher | Their definition of a world model | The analogy |
|---|---|---|
| Yann LeCun (AMI Labs) | Grasp the logic of the world, not its appearance. His JEPA architecture trains the model to predict the meaning of a situation in an abstract space, so it can reason before acting. | The mental map of a chess player |
| Fei-Fei Li (World Labs) | Perceive the world as a three-dimensional space you can move through, where objects are seen from several angles and their spatial relations are understood. | A faithful reconstruction of the room |
| Demis Hassabis (Google DeepMind) | A training ground. The AI practises in a virtual environment, piles up mistakes and corrects them with no real-world consequences. | A pilot's flight simulator |
Yann LeCun's approach has a technical name you meet everywhere in the literature: JEPA, for Joint-Embedding Predictive Architecture. Meta published the image version, I-JEPA, then video variants. The bet is to ignore useless detail such as shadows or textures, and predict only the structure of a scene.
You will not get a chat window for world models
LLMs answer universal needs: writing, summarising, talking. Anyone opens ChatGPT and knows what to do with it. World models exist to anticipate and to act, two much narrower uses.
Emmanuelle Guyot therefore does not expect a consumer release. In her view, these models will run under the hood, hidden inside tools, objects and services. A bit like the AI that existed before ChatGPT, which we used without ever seeing it.
The fields she names fall into two families:
- acting in the physical world: robotics, self-driving vehicles, industrial control;
- immersive environments: video games, education, simulation-based training.
Those fields are moving, mostly in labs and supervised pilots. Cosmos, NVIDIA's platform, serves as a training environment for robotics and autonomous driving. Genie 3, presented by Google DeepMind, generates playable worlds where the physical rules hold. Neither one is a product you install at home in the evening.
Agentic AI is the real tipping point
An AI agent today is an LLM handed a list of steps, which it runs using tools. It works, but its autonomy runs out fast. It has no picture of the environment it acts in, so it cannot tell what its action will break.
Emmanuelle Guyot sees the most direct benefit for digital professionals right there. The instruction changes nature: you give a goal, you describe an environment, and the agent figures it out. It analyses, decides, acts.
The jump is qualitative. An agent that anticipates the effects of its actions can plan, back out, pick the least risky path. That is a different category from today's assistants, whose autonomy rests mostly on how well we word our instructions. If you already build agents, the gap shows up quickly in a tool like n8n: our course on automations and AI agents with n8n covers exactly this mechanics of goals and guardrails. On the model side, agentic ability is progressing among LLM vendors too, as shown by the agentic focus of Claude Sonnet 5.
What gets risky when AI acts without us proofreading
With an LLM, the human stays in the loop by design. The model produces text, we read it, we decide whether to believe it. That control layer costs nothing: it is the product itself.
Emmanuelle Guyot points to its partial disappearance. Since a world model exists to decide and to act, it gains autonomy and responsibility that we will not always be able to check. A mistake made by a robot or a vehicle is not a hallucination you fix by rereading a paragraph.
Misplaced trust is the sneakiest risk here. These systems will be introduced as AI that thinks and understands the world. Emmanuelle Guyot draws the logical conclusion: we will trust them even more than we trust LLMs, at the exact moment we have less grip to control them.
There is a second point, discussed far less. A model that learned not only physics but also human behaviour, psychology and individual beliefs would concentrate knowledge no single person holds. Perfect ground for manipulation and for privacy abuse.
LLMs or world models? The question is badly framed
Emmanuelle Guyot does not expect a replacement. LLMs explain, converse, connect with humans. World models anticipate how a situation will evolve, reason about consequences, help with planning. Two goals that do not overlap.
The direction is combination: one module that understands instructions in plain language, another that simulates and plans. NVIDIA already went that way with Cosmos, which pairs a model simulating the physical world with a vision-language component that reads the instructions.
For Yann LeCun and other researchers, world models are a step towards artificial general intelligence, an AI able to understand and do anything. The road ahead is huge. The longest part, according to Emmanuelle Guyot, is gathering the right learning situations: you would have to list the entire world.
What this changes for you, even working solo
Stop waiting for an interface before you pay attention. There will be no ChatGPT moment for world models. The shift will reach you through tools you already use: a simulator, a game engine, a planning block inside an automation platform. Read the release notes of your own stack rather than the announcements of new chatbots.
Describe your agents' environment, not just their steps. That habit prepares the ground, and it pays off already. Write down what the agent may touch, what is off limits, what happens on failure. The day a planning module lands in your stack, that description work is done.
Put a human checkpoint on every irreversible action. Sending an email to a list, making a payment, deleting a database: none of those three can be replayed. The more autonomy your automations gain, the more that manual approval is worth. It is one config line today and one avoided disaster later.
Frequently asked questions about world models
What is the difference between a world model and an LLM?
An LLM predicts the most likely continuation of a text sequence. It knows descriptions of the world, not the world. A world model learns from images, video and sensor streams, and predicts the next state of a situation: a trajectory, a collision, the effect of an action. The first one talks, the second one anticipates.
Will world models replace ChatGPT?
No, according to Emmanuelle Guyot, Head of Data Science at the Hellowork group: the two approaches chase goals that are too different. Research is heading towards combining them, one module that understands plain language and one that simulates and plans. NVIDIA already follows that route with Cosmos.
Can you try a world model yourself?
Not the way you open a chatbot. The existing systems, such as Cosmos from NVIDIA or Genie 3 from Google DeepMind, target research, robotics and autonomous driving, with gated access. Most real-world deployments are still prototypes or pilots inside controlled environments.
What is the main obstacle to building world models?
Data. You need examples of situations that physically show how the world behaves, in volumes far above the text that language models ingest. Emmanuelle Guyot describes this as the longest part of the job: you would have to list everything you want the model to understand, which means the entire world.