Yann LeCun Leaves Meta to Build a World Model Startup
Meta’s chief AI scientist is leaving the company after 12 years to launch a startup focused on world models. His bet challenges the idea that language models alone can give machines broad understanding.
According to CNBC, Yann LeCun, Meta’s chief AI scientist and one of deep learning’s most influential figures, is leaving the company after 12 years to launch a startup focused on so-called world models. The move brings a fundamental scientific disagreement into the business world: whether large language models are enough to build truly intelligent systems.
CNBC reports that LeCun has confirmed his departure to pursue this line of research. The French-American scientist, who won the 2018 Turing Award alongside Geoffrey Hinton and Yoshua Bengio, has spent years arguing that language models are useful but insufficient as the sole foundation for intelligence capable of understanding and acting in the world.
What world models are
A world model seeks to learn how reality works: what objects exist, how they relate to one another, what happens when something changes and what the consequences of an action might be. It does not simply complete the most likely sentence based on vast amounts of text, which is the central task of a large language model, or LLM.
The idea is easier to understand with a simple example. A system that has learned a world model should be able to anticipate that a cup will fall if it is pushed off a table, that a hidden object still exists even when it is not visible in an image, or that opening a door requires a specific physical sequence. To do this, it needs to handle perception, memory, spatial relationships and prediction.
These capabilities are particularly relevant to robots, autonomous vehicles and assistants that operate beyond a screen. A chatbot can describe how to prepare a room for a task; a machine acting in that room must also interpret the space, anticipate mistakes and adapt its actions.
LeCun’s objection to the LLM path
LLMs have shown that scaling data, computing power and parameter counts produces notable improvements in conversation, programming, translation and information synthesis. Meta has been one of the central players in that race with its Llama family.
But LeCun argues that predicting the next word is not the same as understanding the environment. His criticism does not deny the usefulness of language models; it questions whether simply adding more scale and more data to the same approach can, on its own, solve capabilities such as causal reasoning, long-term planning or learning from physical experience.
The difference matters. LLMs learn primarily from human representations of the world—text, code and labeled or described images. A world model aims to learn regularities directly from video, sensors and interaction. Rather than memorizing that objects generally obey gravity, it should capture patterns that allow it to anticipate new situations.
This approach is not starting from scratch, either. Research in reinforcement learning, computer vision and robotics has spent decades trying to get machines to predict future states and make decisions. What is changing now is the scale of the business opportunity: recent advances in generative models have drawn capital, talent and infrastructure toward attempts to combine perception, memory and planning in more general systems.
A significant departure for Meta and the industry
LeCun was not just another executive at Meta. In addition to guiding the scientific direction of its AI research, he represented a vision different from the one that dominates much of the commercial conversation: the idea that the next frontier will not necessarily be a larger chatbot, but systems capable of building internal representations of their environment.
His new startup is therefore launching with a clearly defined technical thesis. It will have to show that world models can be trained efficiently, generalize beyond the data they have seen and deliver measurable advantages over current models. It will also have to solve a practical problem: learning from the real world is usually more expensive and less abundant than collecting text from the internet.
For Meta, the departure does not change the fact that it continues to compete with language models and multimodal systems. For the market, it opens another avenue for AI investment—one less focused on answering questions and more on building machines that observe, anticipate and act. The decisive test will be whether that scientific ambition translates into products that perform tasks better than they can today, when they still require human supervision.
