Yann LeCun’s New Startup AMI Raises $1 Billion to Build ‘World Model’ AI

Yann LeCun, Meta’s former chief AI scientist and a Turing Award–winning pioneer of modern AI, has launched his first commercial venture since leaving the social media giant and it’s starting with more than $1 billion in funding.

The Paris-based startup, Advanced Machine Intelligence (AMI), wants to build AI systems that understand the physical world, not just language. The company says the new capital values AMI at $3.5 billion and will fund its push to develop sophisticated “world models” AI systems that learn how the world works so they can reason, plan, and act within it.

A bet on world models, not bigger chatbots

LeCun has long argued that human intelligence is grounded in the physical world and that language is only a thin layer on top. In an interview, he dismissed the idea that simply scaling today’s large language models (LLMs) can deliver human-level intelligence.

“The idea that you’re going to extend the capabilities of LLMs to the point that they’re going to have human-level intelligence is complete nonsense,” he said. While he acknowledges that LLMs are becoming powerful tools for generating code and could be useful across a wide range of applications, he draws a hard line on their limits. According to LeCun, success in code generation “is not going to lead to human-level intelligence at all.”

AMI positions itself as a direct challenge to the current industry playbook followed by leading labs such as OpenAI, Anthropic, and Meta itself, all of which lean heavily on scaling LLM architectures. LeCun’s stance carries particular weight because of his central role in shaping modern AI and his reputation for speaking bluntly about its direction.

AMI (pronounced like the French word for “friend”) says it aims to create “a new breed of AI systems that understand the world, have persistent memory, can reason and plan, and are controllable and safe.” The company plans to operate globally from day one, with offices in Paris, Montreal, Singapore, and New York. LeCun will continue his academic role as a professor at New York University alongside leading the startup.

The more than $1 billion financing round was co-led by investors including Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, the investment firm of Jeff Bezos. Other backers include entrepreneur and investor Mark Cuban, former Google CEO Eric Schmidt, and French telecom billionaire Xavier Niel.

AMI is targeting sectors that already generate large volumes of data and where detailed models of the physical world can generate immediate value. LeCun points to manufacturing, biomedical applications, and robotics as early focus areas. One example he gives: building a realistic world model of an aircraft engine and working with the manufacturer to optimize it for efficiency, reduce emissions, or improve reliability.

The founding team brings together several leaders LeCun previously worked with at Meta and beyond:

• Michael Rabbat, former director of research science at Meta
• Laurent Solly, Meta’s former vice president for Europe
• Pascale Fung, former senior director of AI research at Meta
• Alexandre LeBrun, former CEO of AI healthcare startup Nabla, who will serve as AMI’s CEO
• Saining Xie, a former Google DeepMind researcher, who will be the company’s chief science officer

LeCun has been exploring world models for years. At Meta, he founded the Fundamental AI Research (FAIR) lab, where his team advanced approaches including Meta’s Joint-Embedding Predictive Architecture (JEPA). As those models matured, he concluded their most compelling uses would come from selling them to enterprises — a direction he says did not align neatly with Meta’s consumer-focused business.

He describes a strategic shift inside Meta as the company moved to prioritize LLMs and align with what other major AI players were doing. That change pushed his world-model research out of the spotlight. In November 2025, he says, he told Meta CEO Mark Zuckerberg that he believed he could pursue the work “faster, cheaper, and better outside of Meta” by sharing development costs with other companies. According to LeCun, Zuckerberg was supportive of that move.

Meta is not an investor in AMI, but LeCun says the two sides are discussing potential collaboration. One possibility he raises is using AMI’s world models to power assistants in Meta’s smart glasses.

Although AMI is openly positioned as a critique of LLM-centric AI, LeCun is not dismissing their usefulness outright. Instead, he describes them as the latest in a line of promising technologies whose success may have created “a kind of delusion” among their creators about how far they can go toward general intelligence.

AMI plans to build its technology as open source. LeCun argues that AI is too powerful for any single company to control and that no one founder or lab should decide how the technology can be used. The question of who sets those boundaries has sharpened in recent months, including around moves by the US government.

He points to a recent dispute in which the Pentagon moved to blacklist Anthropic after the company attempted to set restrictions on how the US military could use its AI. On this point, LeCun finds himself unexpectedly aligned with the US government, despite his history of criticizing the Trump administration.

“I don’t think any of us, whether it’s me or Dario [Amodei], Sam Altman, or Elon Musk, has any legitimacy to decide for society what is a good or bad use of AI,” he says. Technology, he notes, can be used for “good things or bad things,” and in more authoritarian states, AI could more easily be turned toward harmful uses.

LeCun has wrestled with these questions before. He notes that convolutional nets — neural networks he helped pioneer, inspired by how the human visual system processes images — underpin the face recognition tools used by governments that monitor their own citizens. He says that although he helped create the underlying technology, decisions about its use should be made through democratic processes rather than by individual researchers or executives.

He recalls that some AI leaders tried to prevent the use of AI in autonomous weapons about a decade ago. Since then, he argues, the technology has also been employed to defend liberal democracies in Europe. As one example, he cites Ukraine’s growing use of autonomous drones in its conflict with Russia.

AMI plans to move quickly. LeCun says the startup will release its first AI models in the near term, though he doesn’t expect them to attract widespread public attention initially. Early work will focus on specific partners, including companies such as Toyota and Samsung, to learn how best to apply world-model technology in real-world settings.

Over time, AMI aims to build what LeCun calls a “universal world model”, a general system that can serve as the foundation for broadly useful, generally intelligent AI across industries, regardless of sector. “It’s very ambitious,” he says, acknowledging the scale of the goal.