Yann LeCun离开Meta创立AI初创公司，专注领域揭秘

Turing Award得主Yann LeCun将离开Meta，并创办专注于W领域的AI初创公司，LeCun计划创立一家新的创业公司，专注于人工智能领域的研究和发展，他将利用自己的专业知识和经验，推动人工智能技术的创新和应用，这一消息引起了业界的广泛关注，期待LeCun的新公司能够带来更多的惊喜和突破，摘要字数控制在100-200字左右。

Yann LeCun, the Turing Award-winning "godfather of AI," will depart Meta by the end of the year to start a new venture focused on advanced machine intelligence (AMI), the researcher announced on LinkedIn on Wednesday.

His startup aims to build AI systems capable of understanding the physical world, possessing long-term memory, reasoning, and planning complex sequences of actions. LeCun also said Meta will remain a partner in his new endeavor. Meta confirmed the departure earlier in the day.

"We thank Yann LeCun for his 12 years of tremendous contributions to AI at Meta and look forward to continuing to benefit from his cutting-edge research through our partnership," a Meta spokesperson said.

LeCun's exit has already drawn attention from major tech players in Silicon Valley. A prominent venture capitalist said that while several industry giants had been closely monitoring LeCun's startup plans even before the official announcement, the departure may also trigger a wave of entrepreneurial activity among AI researchers previously under his leadership at FAIR.

The separation reflects longstanding strategic differences. As early as 2023, LeCun publicly expressed concerns over Meta's AI roadmap. At an MIT talk, he described large language models (LLMs) as "text artisans in the dark" that generate coherent text but lack physical-world understanding.

He advocated instead for AI "world models" that, like infants, learn through observation. LeCun's proposed Joint Embedding Predictive Architecture (JEPA) trains AI to predict abstracted future frames by masking key video regions—a method divergent from Meta's focus on LLMs.

LeCun had also raised repeated questions about Meta's approach to AI. In 2023, at the NeurIPS conference, he warned that LLMs could trap AI in local optima. In 2024, he criticized Meta's abandonment of open-source initiatives, emphasizing that "as platforms become a central part of communication infrastructure, we need a shared foundation."

The immediate catalyst appears to have been Meta's $14.3 billion acquisition of Scale AI this summer and the creation of a "superintelligence" lab (MSL). Meta recruited Scale AI CEO Alexandr Wang, 28, as chief AI officer, and Nat Friedman, former GitHub CEO, to lead product teams. This restructuring weakened FAIR's internal influence.

During Meta's mass layoffs in October, which affected roughly 1,000 employees, FAIR staff—including reinforcement learning expert Dong Tian—were among those let go, sparking turbulence in the AI community. The changes also reassigned LeCun's reporting line to the new AI leadership, limiting his control over key projects.

LeCun, 65, was born in Paris and earned an electrical engineering degree from ESIEE in 1983 and a computer science PhD from Paris VI (now Sorbonne University) in 1987. He began his career at Bell Labs, leading the development of convolutional neural networks (CNNs), which were implemented in the first automated check-reading systems in the U.S. He joined NYU in 2003, founding its Data Science Center and serving as computer science department chair.

In 2013, Mark Zuckerberg personally recruited LeCun to create FAIR, where he established an "open research" model and helped make the PyTorch framework a global standard in deep learning. LeCun became the first French recipient of the Turing Award in 2018, alongside Geoffrey Hinton and Yoshua Bengio.

LeCun's move comes amid a broader surge of interest in "world model" AI. Fei-Fei Li, a National Academy of Engineering member, recently launched World Labs, which has raised $230 million. LeCun's startup is building on his team's prior work with the PEVA world model, which can coherently predict 10 seconds of future video from a 16-second first-person clip, including object trajectories and lighting changes. The research, described as "milestone-level progress" in world models, has been accepted at NeurIPS 2025.

Sources indicate that LeCun has begun fundraising discussions with top-tier global investors, with preliminary valuations reaching $1 billion. Some core FAIR members are expected to join him. Analysts suggest the new company may embrace an open-source ecosystem, directly challenging Meta's current closed-source strategy.

For Meta, LeCun's departure may mark a definitive end to the FAIR era and resolve years of internal friction over AI strategy. His official exit could also reverberate across the broader AI industry, introducing new uncertainties and competitive dynamics.