Just a few years ago, Yann LeCun seemed to be exactly the man in the right place. LeCun is known for pioneering convolutional neural networks (CNN), the powerhouse of modern computer vision, and has spent more than a decade as the intellectual lead for AI research at Meta Platforms. A brilliant researcher at a company with nearly limitless resources on the hottest topic of the century.
What more could you want? Well, if you ask LeCun…you’ll find a better path.
While Meta, OpenAI, and just about everyone else in the AI scene is prioritizing LLM-based AI, LeCun believes that these engines, while linguistically fluent, remain fundamentally “stupid.” So he left, along with many of Meta’s research employees, and founded a startup to create a different type of AI based on a “world model.” Now, his startup called Advanced Machine Intelligence (AMI) has just raised $1.03 billion.
What does Lucan say?
For three years now, the world has been obsessed with large-scale language models (LLMs). We gave them books, questions, and an incredible amount of data. They are already producing some amazing results. However, LeCun believes that LLM is ultimately a statistical illusion. Like librarians who have never actually been out, they don’t understand what they’re actually saying. I know all the words about “gravity” and can talk about it, but I’ve never dropped an object.
Mr. LeCun is not one to keep his mouth shut. He publicly characterized the industry’s obsession with predicting the next token as “LLM-soaked.” He has repeatedly stated that current AI approaches based on predicting the next word or pixel do not produce intelligent agents with a wide range of capabilities.
Instead, researchers want to build a model of the world. If LLM is a “word model,” then the world model LeCun wants to build is a system that can reason, plan, and interact with the three-dimensional reality of the physical world. LLM always uses words as an intermediary. they are seems to be It may seem like you are planning and calculating, but in reality you are not.
The core of their strategy is an architecture called JEPA (Joint Embedding Predictive Architecture). Standard AI tries to predict every pixel in a video, which is a computationally “noisy” task and fails when it encounters something unpredictable like moving grass or moving clouds. JEPA, on the other hand, operates on something called a “latent space,” which is an abstract representation of complex data. JEPA ignores the minutiae and instead focuses on high-level transitions and relationships.
It all sounds very complicated. But think of it this way. When you drive a car, you see a lot of trees, but you don’t have to worry about the shape of the leaves. All you need to do is follow the road and driving rules and be aware of what’s happening around you. This ability to reason, plan, and understand cause and effect is what LeCun calls “bodily intelligence.”
different vision
LeCun worked at FAIR (Facebook AI Research) for 10 years. He succeeded under the philosophy of open science and academic freedom. He championed the public availability of Meta’s Llama AI model, with the goal of democratizing access to powerful artificial intelligence tools.
However, things started to change in the last few years, and Meta focused on competing directly with the big AI companies.
The breaking point arrived in 2025. Meta has reorganized its AI efforts under Scale AI founder Alexandr Wang, 29. LeCun seems to think of Wang as a scaling maximalist, someone who believes that if you throw enough GPUs and data into a model, it will eventually wake up and get smarter. Mr. LeCun disagreed. He argued that although LLM could be extended to consume the sun’s energy, it would still be difficult to truly understand why glass shatters when it hits the floor.
Friction peaked in November 2025. LeCun resigned following allegations that Meta’s generation AI team “fabricated” the Llama 4 benchmark to curry favor with its leaders. It’s not entirely clear whether he took a bunch of FAIR researchers with him or if Meta fired them, but eventually nearly all of the senior executives in Meta’s research department followed him and joined Advanced Machine Intelligence.
Celebrities are listed as investors
It was clear from the beginning that AMI was an elite organization for AI talent. But will investors believe LeCun’s contrarian approach?
As it turns out, they are.
AMI Labs’ investor list is like the “Who’s Who” of the modern economy. The $1.03 billion investment round, co-led by Bezos Expeditions, Greycroft and Cathay Innovation, includes Nvidia, Samsung, Toyota and more. This is not some fake venture capital money that someone has instigated. These are some of the biggest and most serious companies in the world, and they invested $1 billion in a startup that was four months old and didn’t have a single product.
These are also the companies that cannot afford to deal with hallucinations, which are increasingly looking like an unavoidable problem for LLMs. Companies like Toyota and Samsung need AI that can handle factory floors and operating rooms. In such a high-stakes environment, the illusion of an LLM means an unacceptable and catastrophic failure.
Simply put, the tech giants are currently fighting over who can make the best personal assistant. AMI Labs focuses on the things that really make our world run: factories, hospitals, transportation.
AMI is also a challenge to Silicon Valley and China’s AI monopoly. The startup is headquartered in Paris, and LeCun positions AMI as a European champion. ”
Still, not everyone thinks he will succeed. It’s not just Meta, most of the AI giants are betting big on LLM, and many researchers believe that scaling will lead to smarter AI.
The stakes are no more. If LeCun is right, he has cracked the “common sense” that AI has always eluded. We will see robots that can navigate our homes, AI doctors that we can actually trust, and a new era of industrial automation. Ironically, LeCun says one of the first applications could be in Meta’s Ray-Ban smart glasses.
