Yann LeCun, Meta's longtime Chief AI Scientist and a pioneer of deep learning, is reportedly preparing to leave the company to launch his own AI startup.
The Financial Times reports that LeCun is already in early talks with potential investors. Neither Meta nor LeCun has commented on the news. The Turing Award winner has not given a timeline for his departure, and there are no official details about the focus of his new company. His position at New York University will remain unchanged.
LeCun's departure comes as no surprise
LeCun's move follows months of growing frustration inside Meta. He was reportedly unhappy with new internal publication rules that require more extensive internal review before research can be published - a policy several team members see as a threat to academic freedom.
Meta's recent restructuring also hit the core AI research group FAIR, which faced layoffs and lost influence to the product-focused TBD Lab, led by Alexandr Wang. LeCun was reportedly required to report to Wang.
LeCun has also, without any real need, publicly distanced himself from Meta's latest Llama models, which have struggled to gain traction. Another sign of his dissatisfaction.
Politics likely added more strain. LeCun has been a vocal critic of the current US administration, while Meta CEO Mark Zuckerberg has shifted the company toward Trump-aligned policies, a clear break from Meta's earlier stance.
LeCun's next move: Challenging the LLM status quo
LeCun is a longtime critic of the industry's push toward large language models, arguing that the focus on LLMs is one-sided and overhyped. While some companies see large transformer models as the path to human-level intelligence, LeCun has repeatedly questioned whether this architecture is really suited to building true intelligence.
With his own startup, LeCun appears ready to pursue approaches outside the current LLM mainstream - especially those he's developed with his Joint Embedding Predictive Architecture (JEPA). Instead of generating text or images, JEPA aims to build systems that learn by observing the physical world and develop abstract models, giving them the ability to reason and plan.
