Yann LeCun turned up at RAISE Summit 2026 one day after his birthday, batted away the "Godfather of AI" label ("Godfather in New Jersey means you work for the mafia"), and then spent half an hour explaining why he thinks the entire LLM era is a detour.
Seven months after leaving Meta to co-found AMI Labs, the convolutional neural network pioneer was blunt about the limits of today's chatbots. LLMs can pass the bar exam and write code, he said, but "when it comes to real signals, video sensors from industry or whatever, LLMs are completely useless, essentially." That gap explains why we still have no level-five self-driving cars and no domestic robots. Machines cannot yet do what a 10-year-old can, he argued, "or even what a cat can do."
The problem, in his telling, is baked into how generative models work. Predicting the next word in a sentence is tractable. Predicting the next frame of a video at the pixel level is not, because most of what happens in the physical world is unpredictable. Video generators paper over this by inventing plausible details. "The very idea of a generative model does not work in the real world," LeCun said.
His answer is JEPA, an architecture he has pushed for years, which learns an abstract representation of a video and makes predictions in that space, discarding the unpredictable details. AMI Labs' V-JEPA models can already flag physically impossible events in video, he said. And data is not the constraint: the roughly 10^14 bytes of public text on the internet is about what reaches a child's visual cortex in the first four years of life. "A four-year-old has seen as much training data as the biggest LLMs."
So why leave Meta after 12 years? LeCun said the world-model project had Mark Zuckerberg's backing, but in 2025 Meta's AI organization pivoted to catching up on LLMs, driven by people convinced that scaling them leads to AGI, a belief he publicly rejected. The bigger issue was fit: the applications he sees, from industrial process control to power plants and pharmaceutical manufacturing, are B2B, and Meta is not a B2B company. AMI Labs is headquartered in Paris with offices in New York, Montreal and Singapore, investors split roughly 40% Europe, 33% U.S. and 27% Asia, and zero presence on the U.S. west coast or in China. "That's by design for now," he said.
LeCun saved his sharpest warning for AI sovereignty. With Llama fading and U.S. export controls hitting Anthropic's Mythos and Fable models, he sees "no credible provider of open AI platforms from the West anymore. They're all Chinese," and many companies are wary of depending on models that could be switched off by political decisions. His fix is Project Tapestry, which held its kickoff meeting in Paris two months ago: a distributed training scheme letting countries and institutions train a shared foundation model on their own data and hardware, exchanging only model parameters, never the data itself. The stakes, as he framed them, are cultural. If all information flows through a handful of proprietary assistants from the U.S. or China, "this is the end of local culture."
On Europe's chances, LeCun was upbeat. Silicon Valley, he said, is stuck in "a trench that everybody is digging," unable to deviate from LLMs for fear of falling behind. World models are the opening. "Europe has not lost that race."