Yann LeCun got a hometown-hero welcome at VivaTech, then spent his fireside chat with Steven Levy dismantling the industry consensus, one heresy at a time.
Start with his diagnosis of Silicon Valley. LeCun has a name for people who believe scaling LLMs leads to superintelligence: "They're called LLM pilled. They took the pill, and now they're hypnotized." Because every lab is poaching the same engineers and chasing the same benchmarks, nobody can afford to deviate, and the result is a monoculture built on what he calls "a bit of a delusion." He's careful to say LLMs are useful, even superhuman at translation and some coding. But "there is simply not a path to human-level intelligence" through them. AMI Labs, he said, is about "leapfrogging" — then paused to note, delighted, that French has no word for leapfrogging "which is crazy because we're frogs."
The Meta backstory got its fullest airing yet. The world-model work, called AMI internally at FAIR and led by Joëlle Pineau until her departure last year, had backing from Mark Zuckerberg and CTO Andrew Bosworth. But in 2025 Zuckerberg created TBD Lab to bridge FAIR and the Llama organization, and FAIR got pulled into its short-term objectives. Add talk of restricting publications, and LeCun was done. He insists there's no bitterness: "Have good relations with Mark and Buzz."
He also told the inside story of Meta's biggest open-source moment. After weekly leadership debates over whether Llama 2 should be open, with LeCun and Bosworth arguing vehemently for release, Zuckerberg finally said: "Yeah, we're going to open source Llama 2. Tell me how." A whole startup industry followed. Now that Meta has clammed up and the open frontier belongs to Chinese labs, Levy noted that people no longer look to Meta as the company carrying the torch. LeCun's reply: "Carrying the PyTorch."
The sharpest material came when the conversation turned to AI safety. Companies like Anthropic, he charged, are "basically lobbying to essentially make open source illegal." His counterargument reached back to the 15th century: the Catholic Church and the Ottoman Empire tried to resist the printing press, which nonetheless triggered mass literacy, the Enlightenment, and the end of feudalism. LLMs, in his view, are the same kind of technology, a means of disseminating knowledge rather than generating it. "If you're going to block this because you think it's dangerous, you are being a medieval obscurantist, and I think that's insane." He sees "a big arrogance and superiority complex in the idea that only a few are capable of controlling AI," and shrugged off doomsday security scenarios: email brought spam and scams, and we fixed those too. He even got in a jab at Dario Amodei's most famous phrase: "We're not going to have a country of geniuses in a data center next year."
On Project Tapestry, the federated-learning scheme he's pushing through the AI Alliance, LeCun added a democratic argument to the cultural one: AI assistants will mediate everyone's information diet, and every model carries implicit bias, like every newspaper. Without a diversity of assistants built on an open foundation model, "culture is in big trouble. Democracy is in big trouble." Asked when Tapestry happens, he pointed to national models already live, from Switzerland's ETH-EPFL effort to the UAE's K2 and India's BharatGen, and gave Levy a three-word answer: "It's happening now."