Meta’s Yann LeCun to Launch Physical A.I. Startup After Declaring LLMs a ‘Dead End’

<img decoding="async" class="lazyload size-full-width wp-image-1599246" src="data:image/gif;base64,R0lGODlhAQABAAAAACH5BAEKAAEALAAAAAABAAEAAAICTAEAOw==" data-src="https://observer.com/wp-content/uploads/sites/2/2025/11/GettyImages-2197497048.jpg?quality=80&w=970" alt="Man in beige jacket and dark framed glasses" width="970" height="647" data-caption='Yann LeCun’s exit marks another major shake-up at Meta, as the company doubles down on advanced A.I. systems and superintelligence research. <span class=”lazyload media-credit”>Stephane de Sakutin/AFP via Getty Images</span>’>

Yann LeCun, Meta’s chief A.I. scientist, is reportedly leaving the tech giant to launch his own A.I. startup, according to the Financial Times, which cited sources familiar with the matter. His departure comes as Meta increasingly reorients its efforts toward developing advanced forms of A.I. A hiring spree and internal restructuring have prioritized Meta’s research on superintelligence, placing LeCun’s long-term work at Meta’s Fundamental AI Research (FAIR) lab on the back burner.

Meta and LeCun did not respond to requests for comment from Observer.

The French-American computer scientist, who moved to the U.S. in the 1980s to work at AT&T Bell Laboratories, is known for his pioneering research in machine learning. LeCun joined Meta in 2013, where he helped launch the company’s FAIR team and became its chief A.I. scientist in 2018. The same year, he won the Turing Award for his contribution to the breakthroughs in neural networks.

LeCun previously reported to Meta’s chief product officer Chris Cox, but following the company’s recent restructuring, he now reports to Alexandr Wang, head of Meta’s superintelligence division. Wang was recruited earlier this year after Meta invested more than $14 billion in his data labelling startup, Scale AI.

Meta’s aggressive talent push coincided with a company-wide restructuring that is consolidating much of its A.I. research under TBD labs, a group focused on achieving A.I. that surpasses human capabilities. Three other divisions within Meta’s A.I. organization are dedicated to products, infrastructure and FAIR.

FAIR, which focuses on long-term, exploratory research, has contributed to early versions of Meta’s Llama model. LeCun’s departure marks another setback for the group, which in April lost its leader, Joelle Pineau. Pineau is now head of research at Canadian A.I. startup Cohere.

World models pick up steam

LeCun’s next venture will center on “world models,” according to the Financial Times. He is reportedly in early funding discussions for a startup focused on training A.I. systems to understand the physical world rather than merely generating language.

His move into world models isn’t surprising given his skepticism about the sustainability of large language models (LLMs). LeCun has previously called LLMs a “dead end” to reaching human-like A.I., instead advocating for systems that can perceive their environments and grasp physical concepts such as gravity.

He isn’t alone in this pursuit. Stanford’s Fei-Fei Li has raised about $230 million for her startup World Labs, which similarly seeks to give A.I. “spatial intelligence.” Google DeepMind has explored world models through its Genie releases, and Nvidia is pushing into physical A.I. with products like its Cosmos world models.

LeCun believes world models, not LLMs, are the key to developing A.I. that can reason, plan complex actions, and make predictions. “We’re never going to get to human-level A.I. by just training on text,” said the researcher during a Harvard talk in September. “Despite what you might hear from some of the more optimistic-sounding CEOs of various A.I. companies in Silicon Valley, it’s just not going to happen.”