Yann LeCun Raises $1 Billion to Build AI That Understands the Physical World

Advanced Machine Intelligence (AMI), a caller Paris-based startup cofounded by Meta’s erstwhile main AI idiosyncratic Yann LeCun, announced Monday it has raised much than $1 cardinal to make AI satellite models.

LeCun argues that astir quality reasoning is grounded successful the carnal world, not language, and that AI satellite models are indispensable to make existent human-level intelligence. “The thought that you’re going to widen the capabilities of LLMs [large connection models] to the constituent that they’re going to person human-level quality is implicit nonsense,” helium said successful an interrogation with WIRED.

The financing, which values the startup astatine $3.5 billion, was co-led by investors specified arsenic Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Other notable backers see Mark Cuban, erstwhile Google CEO Eric Schmidt, and French billionaire and telecommunications enforcement Xavier Niel.

AMI (pronounced similar the French connection for friend) aims to physique “a caller breed of AI systems that recognize the world, person persistent memory, tin crushed and plan, and are controllable and safe,” the institution says successful a property release. The startup says it volition beryllium planetary from time one, with offices successful Paris, Montreal, Singapore, and New York, wherever LeCun volition proceed moving arsenic a New York University prof successful summation to starring the startup. AMI volition beryllium the archetypal commercialized endeavor for LeCun since his departure from Meta successful November 2025.

LeCun’s startup represents a stake against galore of the world’s biggest AI labs similar OpenAI, Anthropic, and adjacent his erstwhile workplace, Meta, which judge that scaling up LLMs volition yet present AI systems with human-level quality oregon adjacent superintelligence. LLMs person powered viral products specified arsenic ChatGPT and Claude Code, but LeCun has been 1 of the AI industry’s astir salient researchers speaking retired astir the limitations of these AI models. LeCun is good known for being outspoken, but arsenic a pioneer of modern AI that won a Turing grant backmost successful 2018, his skepticism carries weight.

LeCun says AMI aims to enactment with companies successful manufacturing, biomedical, robotics, and different industries that person tons of data. For example, helium says AMI could physique a realistic satellite exemplary of an craft motor and enactment with the shaper to assistance them optimize for efficiency, minimize emissions, oregon guarantee reliability.

AMI was cofounded by LeCun and respective leaders helium worked with astatine Meta, including the company’s erstwhile manager of probe science, Michael Rabbat; erstwhile vice president of Europe, Laurent Solly; and erstwhile elder manager of AI research, Pascale Fung. Other cofounders see Alexandre LeBrun, erstwhile CEO of the AI wellness attraction startup Nabla, who volition service arsenic AMI’s CEO, and Saining Xie, a erstwhile Google DeepMind researcher who volition beryllium the startup’s main subject officer.

The Case for World Models

LeCun does not disregard the wide inferior of LLMs. Rather, successful his view, these AI models are simply the tech industry’s latest promising trend, and their occurrence has created a “kind of delusion” among the radical who physique them. “It's existent that [LLMs] are becoming truly bully astatine generating code, and it's existent that they are astir apt going to go adjacent much utile successful a wide country of applications wherever codification procreation tin help,” says LeCun. “That’s a batch of applications, but it’s not going to pb to human-level quality astatine all.”

LeCun has been moving connected satellite models for years wrong of Meta, wherever helium founded the company’s Fundamental AI Research lab, FAIR. But he’s present convinced his probe is champion done extracurricular the societal media giant. He says it’s go wide to him that the strongest applications of satellite models volition beryllium selling them to different enterprises, which doesn’t acceptable neatly into Meta’s halfway user business.

As AI satellite models similar Meta’s Joint-Embedding Predictive Architecture (JEPA) became much sophisticated, “there was a reorientation of Meta’s strategy wherever it had to fundamentally drawback up with the manufacture connected LLMs and benignant of bash the aforesaid happening that different LLM companies are doing, which is not my interest,” says LeCun. “So sometime successful November, I went to spot Mark Zuckerberg and told him. He’s ever been precise supportive of [world exemplary research], but I told him I tin bash this faster, cheaper, and amended extracurricular of Meta. I tin stock the outgo of improvement with different companies … His reply was, OK, we tin enactment together.”