According to Gizmodo, AI legend Yann LeCun is leaving Meta after 11 years to create his own independent company focused on developing AI systems that understand the physical world. The 2019 Turing Award winner, who joined Facebook back in 2013 when the company’s AI ambitions were still emerging, reportedly clashed internally as Meta shifted focus toward generative AI and high-profile hires like ChatGPT co-creator Shengjia Zhao. LeCun’s new venture aims to build systems with persistent memory, reasoning capabilities, and complex planning abilities—technology he believes will have “far-ranging applications” beyond Meta’s current interests. Meta plans to partner with LeCun’s startup and will have access to its innovations, though specific deal terms remain unclear. This departure comes after LeCun spent more than a year publicly criticizing LLM research as a dead end for achieving advanced AI capabilities.
Leaving the LLM Party Early
Here’s the thing about LeCun: he’s never been shy about his skepticism toward the current AI gold rush. While everyone else was losing their minds over ChatGPT in 2022, he was telling the Wall Street Journal that AI existential risk talk was “complete B.S.” And last month, he made it clear he had almost no involvement with Meta’s Llama models. He was working in the Fundamental AI Research department, trying to go “beyond LLMs” while the generative AI crowd did their thing in a separate department. Basically, he saw the writing on the wall—Meta was all-in on the chatbot revolution, and he wanted no part of it.
World Models vs Word Models
So what exactly is LeCun pursuing instead? He’s fascinated by “world models”—AI systems that learn from sensory inputs like vision rather than just text. Think about it: current AI can write a decent poem but can’t understand that if you push a glass off a table, it will fall and break. LeCun thinks today’s AI systems are dumber than cats, and he’s probably right. You can already see his approach in projects like V-JEPA-2, which trains on videos to model cause and effect in the physical world. Unlike OpenAI’s Sora, which just replicates video, LeCun’s approach tries to understand why things happen.
The Meta Partnership Puzzle
Now, the partnership angle is interesting. Meta says they’ll work with LeCun’s new company, which sounds a lot like the Microsoft-OpenAI or Google-Anthropic arrangements we’ve seen. Google owns 14% of Anthropic, Microsoft has that special relationship with OpenAI—it’s becoming the standard playbook. But what does Meta get from this? Access to innovations, sure, but also a way to keep tabs on research they might want to acquire later. It’s smart hedging: let the visionary founder chase his moonshot while maintaining a connection if it actually works.
What This Means for AI Development
LeCun’s departure signals something important about where AI might be heading next. The LLM revolution has been incredible, but we’re hitting limitations. These systems don’t really understand the world—they’re just really good pattern matchers. LeCun wants to build AI that can reason and plan, which could have massive implications for robotics, autonomous systems, and basically any technology that interacts with the physical world. He even prefers the term Advanced Machine Intelligence (AMI) over AGI—and given that “ami” means friend in French, it’s a much less threatening branding choice. Will his approach actually work? That’s the billion-dollar question. But after winning the Turing Award in 2019 and spending decades in AI research, he’s earned the right to try something different.
