gensai.ai
In March 2026, Yann LeCun, the French Turing Award laureate and founder of AMI Labs, closed the largest seed round ever raised by a European startup: 1.03 billion dollars, at a 3.5 billion pre-money valuation. Toyota, Nvidia, Samsung and Jeff Bezos are on the cap table.
What justifies such a bet? An intellectual position LeCun has defended for years, now funded at scale: world models — models capable of understanding the physical world, where LLMs, in his view, merely produce plausible text.
Fei-Fei Li, a pioneer of computer vision and founder of World Labs, summed up the ambition this way: "If AI is to be truly useful, it must understand worlds, not just words."
A formulation that says it all — and opens a fundamental question for every organisation investing in AI today.
What exactly is a world model, how does it differ from the LLMs we use daily, and what are the concrete implications for companies structuring their AI architectures?
Do LLMs really understand the world?
An LLM (large language model) is trained on billions of texts. It learns to predict the most plausible continuation of a sequence of words. If the result often seems convincing, it is because human texts abundantly describe our reality: its rules, its physics, its causes and effects.
But that description remains indirect. An LLM knows that a dropped object falls because it has read thousands of sentences saying so. It does not understand gravity. It has a statistical representation of it.
This is precisely the crux of the problem Yann LeCun points to. At the NVIDIA GTC conference he stated unambiguously: "LLMs are too limited. Scaling them will not get us to AGI."
And at the World AI Cannes Festival in February 2026: "LLMs are not appropriate" for building an AI that genuinely understands the real world.
This is not a criticism of the quality of today's LLMs; it is a criticism of their underlying architecture. Producing plausible text and modelling physical reality are two fundamentally different objectives.
What exactly is a world model?
A world model starts from a different ambition: modelling the rules of the world itself. Its causality. The dynamics of objects and situations. The consequences of actions in a physical environment.
Where an LLM processes language, a world model processes reality.
The approach LeCun defends rests on an architecture he developed: JEPA (Joint Embedding Predictive Architecture). The intuition is close to that of a chess player: "We don't ask it to reproduce the world realistically, but to grasp its logic" — to anticipate the consequences of an action before carrying it out.
In practice, the model observes massive volumes of images and video, then learns to predict what comes next, without systematic human annotation. Through repeated exposure, it infers the rules governing situations: that a falling glass breaks, that a ball bounces differently depending on the surface, that a door only opens one way.
Three schools, one shared intuition — how do you build a world model?
There is no consensus yet on the most effective approach. Three major figures embody distinct visions.
Yann LeCun (AMI Labs) defends the abstract mental map: understanding the logic of a situation in order to anticipate and plan, without having to reconstruct its appearance. JEPA, his flagship architecture, trains to predict the meaning of a situation rather than imitate its pixels.
Fei-Fei Li (World Labs) bets on spatial intelligence: AI must perceive the world as a real three-dimensional space in which it can move and observe objects from different angles. Her Marble platform, launched in November 2025, already generates navigable 3D environments from text, images or video. "World models will play a decisive role in solving the data problem for robotics," she writes in her manifesto.
Demis Hassabis (Google DeepMind) conceives the world model as a simulator — a training ground where AI accumulates experience without real-world consequences before acting in the physical world. An approach closer to reinforcement learning techniques than the previous two.
Three schools, one shared intuition: to act effectively in the world, an AI needs an internal model of that world, not just a textual description of it.
What are the concrete applications for businesses?
World models will not replace your generative AI tools tomorrow. Their natural applications are industrial and physical, and that is where they genuinely change the game:
- Robotics: a robot able to anticipate the consequences of its movements in an unstructured environment — warehouse, construction site, operating theatre — without needing to be reprogrammed for every new configuration.
- Autonomous vehicles: understanding the dynamics of an intersection in real time, anticipating a pedestrian's behaviour, adapting the trajectory to unforeseen conditions that the training data did not cover.
- Industrial simulation: testing production line configurations, reducing time-to-production costs, training operators on digital twins that behave in a physically coherent way.
- Immersive training and video games: generating credible training environments for crisis situations, surgical procedures or safety scenarios, with physics that holds up.
World models and agentic AI: why the combination is decisive
AI agents — systems able to chain autonomous actions to complete a complex task — are currently limited by a structural factor: they do not model the consequences of their actions. They act, observe the result, correct. A slow, costly loop, and sometimes a risky one in physical environments.
An agent coupled with a world model can anticipate before acting. It mentally simulates several scenarios, evaluates the likely consequences and chooses the most appropriate action. Decision-making autonomy changes in nature, and in level.
This is precisely what makes the combination strategic for sectors where errors are expensive: industry, healthcare, logistics, security. And it is also what raises legitimate questions about human control of these systems, their auditability and their compliance with the AI Act, whose most demanding provisions take effect in August 2026.
World models and LLMs: competitors or complements?
The question deserves to be put clearly, because it shapes investment choices. In reality, the two approaches do not address the same problems.
An LLM excels at everything language-related: summarisation, content generation, reasoning over textual data, conversational interaction. A world model excels at everything involving action in a physical or simulated environment.
Hybrid architectures — LLM for the interface and language reasoning, world model for environment modelling — appear to be the direction in which the most advanced players are converging. The point is not to choose between the two. It is to understand which one answers which problem, and not to over-invest in a technology whose operational benefits are still 12, 24 or 36 months away depending on the sector.
What businesses should take away today
World models will not replace LLMs in the short term. But they signal something important: AI is gradually moving out of the sphere of language and into that of physical action. And this shift is being funded at levels that suggest the first industrial applications are close.
For sectors deeply rooted in material processes — industry, logistics, healthcare, technical training — this evolution deserves close attention. Not to anticipate immediate deployment, but to calibrate current architectures so they remain compatible with what is coming.