Meta's former AI chief raised $1 billion to prove chatbots wrong

Yann LeCun has spent years arguing that Silicon Valley’s favorite AI products are, intellectually speaking, really just very impressive party tricks. Now, the former Meta $META AI chief has raised $1.03 billion for a startup called AMI to prove that the road to something like real machine intelligence doesn’t run through ever-larger chatbots (with ever-bigger spending), but through “world models” that can understand how reality works.

$1 billion is a giant sum for a company founded four months ago, with no product and no interest in pretending one is around the corner. But while much of the AI industry has spent the past few years treating LLMs as the obvious route to bigger and better “intelligence,” LeCun has been one of the field’s loudest dissenters. 

In an interview with WIRED, the Turing Award winner called the idea that LLMs will simply scale their way to human-level intelligence “complete nonsense.” AMI, short for Advanced Machine Intelligence, is his chance to stop being the guy yelling from the wings of the stage and become the guy building his own spotlight.

LeCun left Meta in late 2025 after spending 12 years building Meta’s AI research operation; he remains one of the most prominent voices inside Big Tech arguing that the industry is getting a little drunk on autocomplete. 

Essentially, the former AI chief seems to think that Meta had to chase the LLM race, and doing so pulled the company away from the work he thinks actually matters. He told WIRED that Meta’s push to “catch up with the industry on LLMs” was “not my interest,” and said he eventually went to CEO Mark Zuckerberg and told him he could do this “faster, cheaper, and better outside of Meta.” Meta’s Superintelligence Labs is now led by former Scale AI CEO Alexandr Wang.

Meta is largely a company obsessed with consumer products and AI assistants, but LeCun had a different target in mind: systems that can reason about the physical world and be sold into enterprise settings where fluent nonsense isn’t an acceptable error mode. So this latest venture is LeCun trying to prove, with over a billion dollars of other people’s money, that the industry may be marching down the wrong road.

Chatbots made the market. LeCun is betting that machines that can model reality will matter more. Whether that turns out to be profound or just extremely expensive remains the fun part.

What are “world models”?

AMI wants to build “world models” — AI systems can understand the world, reason through it, and stay controllable and safe while doing it, not just predict the next word in a sentence. LeCun’s framework for that is JEPA, or Joint Embedding Predictive Architecture, which aims to learn abstract structure. That means a bigger emphasis on reasoning, planning, spatial understanding, and learning from video and real-world data, which LeCun has argued is a more plausible route to human-level intelligence than today’s large LLM obsession.

AMI is reportedly initially targeting organizations that manage complex systems with consumer applications (such as home robots) further down the road. And heck, maybe even a partnership with Meta; LeCun has left open the idea of partnering with his former employer later on to commercialize the tech, potentially even in products such as Meta’s Ray-Ban smart glasses.

“World models” are having a real funding moment. Last month, Fei-Fei Li’s World Labs also raised $1 billion to work on “spatial intelligence,” which is adjacent terrain: AI that understands and generates 3D environments for robotics, AR/VR, and other real-world applications. Nvidia $NVDA has similarly described “world models” as systems for understanding real-world dynamics and generating training environments for robots and autonomous vehicles.

Silicon Valley’s newest expensive hobby may be teaching machines that the world exists outside of a prompt.

AMI’s goal is a lofty one, and LeCun seems perfectly aware of it. So does CEO Alexandre LeBrun, who told TechCrunch that this is “not your typical applied AI startup” that ships in a quarter and starts bragging about revenue by the end of the year. “‘World models’ will be the next buzzword,” he predicted, adding that “in six months, every company will call itself a world model to raise funding.”

He may be right. But AMI has at least one thing most buzzwords don’t: a clear reason for existing. LeCun thinks today’s dominant AI products are useful but fundamentally limited, especially in settings where sounding plausible isn’t good enough. That helps explain why the first commercial lanes being discussed are manufacturing, robotics, biomedical work, and healthcare, high-stakes environments where hallucinations can get dangerous, expensive, or both. Nabla, LeBrun’s healthcare startup, is the first disclosed partner, which gives the AI venture a concrete healthcare-adjacent lane from day one. 

AMI’s investor roster has plenty of big names. Nvidia is in. So are Samsung, Toyota $TM Ventures, Cathay Innovation, Temasek, and Bezos Expeditions. As are Mark Cuban, Marcel Dassault, Eric Schmidt, Xavier Niel, and Tim and Rosemary Berners-Lee. That’s a pretty clear clue about how AMI is being read by the market: as a possible infrastructure layer for physical-world AI. The company is based in Paris, with offices planned in New York, Montreal, and Singapore, and its early bench pulls from Meta and DeepMind.

There is, of course, an obvious chance that this becomes the most elegantly argued expensive science project in Europe. LeCun is asking investors to back a long-horizon technical thesis in a market that rewards speed, spectacle, and software that can dazzle on command. But that’s also what makes AMI interesting; it’s one of the clearest signs yet that some of AI’s biggest names and deepest pockets aren’t content to keep feeding the same machine and calling the result something like destiny.

Silicon Valley has spent the AI boom treating chatbot fluency like proof of intelligence. LeCun just raised a billion dollars to argue that the harder, more valuable trick is understanding the world in the first place.