According to Sifted, Meta’s outgoing chief AI scientist Yann LeCun announced at the AI-Pulse conference in Paris on Thursday that he is building his new AI startup from Europe. The Turing Award winner, who will leave Meta at the end of this year, stated the company will focus on developing “world models,” a non-generative AI architecture that understands the physical world, unlike current text-based LLMs. He called Silicon Valley “completely hypnotised by generative models” and said scaling current tech to achieve general intelligence is “bullshit.” The startup, dubbed AMI (advanced machine intelligence), will be a global entity with bases worldwide, including Paris, and will partner with Meta, though operate independently as its application scope is beyond Meta’s interests.
LeCun’s big bet
Here’s the thing: LeCun is making a massive, public wager against the entire direction of the current AI industry. While everyone from OpenAI to Google is pouring billions into making better chatbots and image generators, he’s saying they’re all missing the point. He’s basically arguing that generating plausible text is a parlor trick compared to building a system that understands how the world actually works. And you know what? He has a point. Our best LLMs can write a sonnet about a robot, but they can’t tell that robot how to walk across a room without bumping into a chair. That’s a pretty fundamental gap.
The European gambit
His choice to base this out of Europe, particularly Paris, is fascinating. It’s not just a homecoming; it’s a strategic statement. He’s explicitly saying the talent and environment needed for this moonshot aren’t in the “hypnotised” Valley. He’s betting he can attract top researchers who are tired of the generative AI gold rush and want to work on what he sees as the real problem. But let’s be skeptical for a second. Silicon Valley isn’t just hypnotised; it’s funded. Deeply. Completely. Building a capital-intensive, long-term research startup outside that ecosystem is a huge risk. Can he secure the kind of sustained, patient funding needed to compete with the near-infinite budgets of Big Tech? That’s the billion-dollar question.
World models and hard reality
So what’s a “world model”? Think of it as an AI with a persistent, internal simulation of physics and cause-and-effect. Instead of predicting the next word, it predicts what might happen in a room if it pushes a cup off a table. This is the kind of foundational intelligence that could revolutionize robotics, autonomous systems, and complex simulation. The potential is staggering. But the history of AI is littered with brilliant researchers who thought they had the next architectural paradigm that would unlock “true” intelligence. Remember symbolic AI? Remember expert systems? LeCun’s own approach, which he calls a “joint embedding predictive architecture,” is still largely theoretical. Moving from a compelling idea to a practical, scalable technology is a canyon to cross, not a gap to bridge.
The Meta factor
The Meta partnership is a double-edged sword. On one hand, it provides instant credibility and likely some initial resources and data access. Zuckerberg “really likes” the project, which means doors will open. But on the other hand, being “beyond what Meta was interested in” is corporate-speak for “too speculative, too long-term, and too far from our core ads business.” Meta’s patience with FAIR and fundamental research is legendary, but even they have their limits. This spin-out suggests the project was too blue-sky even for them. Now, as an independent entity, AMI will face the relentless pressure to show progress, attract venture capital, and eventually build a product. That pressure has a way of bending even the most pure research toward short-term gains. Can LeCun resist that gravity? We’re about to find out.
