Writing World models
Language was the easy part
Yann LeCun's new company has raised $1.03bn to build world models. Fei-Fei Li and Google DeepMind are working on them too. The next AI has to understand a kitchen as well as a sentence.

Language models learned about the world by reading about it. It was a brilliant shortcut and it took them remarkably far. It also left them strangely helpless in front of a glass of water. A model that can explain fluid dynamics cannot reliably predict where the water goes when the glass tips, because it has never watched a glass tip. A world model is meant to close that gap: an AI that learns how the physical world behaves, so it can predict what an action will do before it is taken.
Yann LeCun has made this argument for years, often to audiences that wanted to hear about chatbots. In November 2025 he said he would leave Meta, where he had led AI research for more than a decade, to found Advanced Machine Intelligence Labs in Paris. In March 2026 AMI raised $1.03bn at a pre-money valuation of $3.5bn, the largest seed round ever raised by a European company. The company says it wants AI that learns from reality, not just from language.
Three ways to build a world
He is not alone. Fei-Fei Li, who built ImageNet and so helped start the deep-learning boom, emerged with World Labs in September 2024. In November 2025 it launched Marble, its first commercial product, which turns text, photos or video into persistent 3D worlds that can be edited and exported. In February 2026 it raised another $1bn. Google DeepMind took another route with Genie 3, shown in August 2025: describe a place in a sentence and it generates an interactive world you can move through in real time, at 24 frames per second, consistent for a few minutes. Since January 2026 Google AI Ultra subscribers in the US can try it in a prototype called Project Genie, 60 seconds at a time.
The approaches differ. LeCun wants models that learn how the world behaves and use that to plan. His architecture, JEPA, makes its predictions in an abstract representation of what it senses rather than token by token. World Labs builds explicit 3D spaces. DeepMind generates the world frame by frame. All three share one conviction: an AI that is going to act in the physical world needs an inner model of it, the way a child learns that a ball rolls and an egg breaks long before it learns either word.
Why it matters outside the lab
World models sound like research. They will reach companies faster than that. AMI names industry, robotics and healthcare as its first fields. A machine that can predict the result of an action before taking it can be trained in simulation, makes fewer mistakes and can work next to people. Most robots today repeat movements someone programmed. World models are what will let them handle the glass that is not quite where it was yesterday.
They also need data the internet does not have: how objects behave in real hands, with forces, slips and breakages. That data has to be recorded first.
Where we work
OptAI®'s lab works on both ends of this: on world models and world action models, which turn a prediction into a movement, and on the data they learn from. It is early, and much of it will fail. The direction is clear all the same. Language was the easy part, because we had already written it all down. The world has to be shown to them.
Sources
- TechCrunch: Yann LeCun's AMI Labs raises $1.03 billion to build world models, 9 March 2026
- Futurum: Yann LeCun's AMI raises $1bn seed round
- heise: World model instead of LLM, Yann LeCun's startup receives 890 million euros
- Maginative: World Labs emerges from stealth with $230 million, 13 September 2024
- SiliconANGLE: World Labs launches Marble, 12 November 2025
- World Labs: Marble, our frontier multimodal world model, 12 November 2025
- World Labs: Funding, 18 February 2026
- Google DeepMind: Genie 3, a new frontier for world models, 5 August 2025
- Google: Project Genie, experimenting with infinite, interactive worlds, 29 January 2026
More writing
LabThe smallest model that does the jobNVIDIA's researchers argue most work inside AI agents is better done by small models. For Swiss companies that is good news on cost, speed and privacy.Author: Rea Stark
ZürichZürich has the labs. Now it needs the usersGoogle, OpenAI, Anthropic, Meta and Microsoft all do frontier AI research in Zürich. Only a third of Swiss SMEs build it into their work.Author: Rea Stark
IntegrationThe plumbing is the productRivals adopted Anthropic's MCP. OpenAI and 19 investors put over $4bn into a firm that places engineers with customers. AI's value is in the connections.Author: Rea Stark