What Is AMI Labs? Inside Yann LeCun's New AI Startup
Yann LeCun left Meta to start AMI Labs in Paris. Here's what the company does, who funds it, and why it bets against today's chatbots.
What is AMI Labs?
AMI Labs (Advanced Machine Intelligence) is a Paris-based AI research company founded by Yann LeCun after he left Meta at the end of last year. Instead of building bigger chatbots, the company is pursuing what LeCun calls “world models,” AI systems designed to understand physical reality rather than just predict the next word in a sentence. It raised a large funding round in March, backed by investors including Jeff Bezos’s investment firm, Nvidia, Samsung, and Toyota.
TL;DR
- AMI Labs was founded by Yann LeCun, one of the original “godfathers of AI,” after he departed Meta where he spent over a decade as chief AI scientist.
- The company raised $1.03 billion in its first funding round in March, one of the largest seed rounds ever seen in Europe.
- Backers include Jeff Bezos’s investment firm, Nvidia, Samsung, and Toyota, a mix of personal capital, chipmakers, and hardware manufacturers rather than typical Silicon Valley venture funds.
- LeCun’s core argument is that large language models hit a ceiling because text alone doesn’t contain enough information to teach a system how the physical world actually works.
- AMI Labs is built around an alternative architecture called JEPA (joint embedding predictive architecture), which predicts abstract representations of what happens next instead of predicting raw pixels or words.
- LeCun frames the move as a bet against the rest of the industry, which he describes as “LLM-pilled” and all racing to do the same thing with the same engineers.
- A recent paper from LeCun and a former Meta colleague offers early mathematical and experimental support for the world-model approach, though it’s a research result, not a shipped product.
Who founded AMI Labs and why does it matter?
Yann LeCun isn’t a random entrepreneur chasing the AI boom. He helped invent the convolutional neural network at Bell Labs in the late 1980s, a technique that became foundational to image recognition and was used to read a large share of handwritten checks in the United States by the late 1990s. That same lineage of work now sits underneath face recognition on phones, medical imaging tools, and the camera systems in self-driving cars.
In 2018, LeCun won the Turing Award, often called computer science’s equivalent of the Nobel Prize, alongside Geoffrey Hinton and Yoshua Bengio. The three are commonly referred to as the godfathers of AI. LeCun went on to found Meta’s AI research lab in 2013 and spent more than ten years as the company’s chief AI scientist while also holding a professorship at New York University.
Leaving that position to start a new company from scratch, in Paris rather than Silicon Valley, is a notable signal. It suggests LeCun believes the disagreement he’s had with the rest of the industry is serious enough to walk away from one of the best-resourced AI labs in the world.
Why did LeCun leave Meta to start AMI Labs?
LeCun has said publicly that the AI industry has become what he calls “LLM-pilled.” At Davos, he described the situation bluntly: nearly every major lab in Silicon Valley is working on the same approach, scaling large language models, and poaching the same engineers from each other. He argued that this creates a trap. No lab can afford to try something different, because if they fall behind on the dominant approach even briefly, they risk losing ground to competitors doing the same thing. LeCun said Meta itself became more committed to this LLM-centric direction after a recent internal reshuffling, which he called a reasonable strategic choice for the company but not the direction he wanted to pursue personally.
That disagreement is the stated reason for founding AMI Labs: a company structured around a different technical bet than the one nearly every other major AI lab is making.
What does AMI Labs actually believe about AI?
The company’s thesis rests on a simple observation LeCun has repeated in talks and interviews. Large language models are typically trained on something like 20 trillion tokens of text, which works out to roughly 10^14 bytes of information. That’s every piece of publicly available text on the internet, more than any human could read in several hundred thousand years.
But LeCun points out that a four-year-old child takes in a comparable amount of information, around 10^14 bytes, just through their eyes in the first four years of life. The child has never read a word, yet understands intuitive physics: that objects fall, that a ball rolling out of sight still exists, that pushing a glass off a table breaks it. A language model trained on the entire internet doesn’t reliably have that kind of grounded understanding.
Without shadow IT.
This is LeCun’s explanation for why AI can pass a bar exam or solve calculus problems but still can’t power a reliable household robot or a true level-five self-driving car. Language, in his framing, is a special kind of data where the words themselves are the substrate of reasoning, which is why models excel at math and code. The physical world, by contrast, is continuous, messy, and much harder to compress into text.
How does AMI Labs’ technical approach differ from LLMs?
AMI Labs’ research centers on an architecture called JEPA, short for joint embedding predictive architecture. The idea addresses a specific failure mode LeCun has described: if you train a system to predict a video frame by frame, pixel by pixel, it has to guess details that are fundamentally unpredictable, like the exact appearance of a stranger who might walk into a room. Forcing a model to predict the unpredictable wastes capacity and muddles its learning.
JEPA instead encodes what it observes into a compressed, abstract representation before making predictions in that representation space, rather than in raw pixel or word space. The system learns to discard details that don’t matter and focus on the parts of a scene that actually carry predictive information, similar to how a sports commentator describes “the striker running toward goal” rather than cataloguing every blade of grass on the field.
One known risk with this design is “collapse,” where the model cheats by producing the same bland summary for every input, since that makes its predictions technically consistent but useless. Training techniques referenced in LeCun’s recent research, including methods called SigReg and VICReg, are used specifically to prevent that shortcut and force the model to learn something meaningful.
Is there evidence this approach actually works?
AMI Labs’ research direction got a notable data point in a paper titled “Proving and Improving the Platonic Representation Hypothesis with JEPA,” published in late October and co-authored by LeCun and Randall Balestriero, a Brown University professor who previously worked with LeCun at Meta. The paper builds on the “platonic representation hypothesis,” an idea introduced by MIT researchers in 2024 observing that AI models trained on very different types of data (images versus text) tend to organize information internally in increasingly similar ways as they scale, as if converging on the same underlying map of reality.
LeCun and Balestriero provided a mathematical account of why this happens with JEPA-style models and under what conditions, centered on the concept of “good views,” where a model learns by comparing two different observations of the same underlying reality (a photo of an event and a text description of it, for instance). The paper also argues that next-word prediction, the mechanism underlying today’s large language models, is a comparatively “bad view” of reality, because consecutive words in a sentence often carry limited information about the actual state of the world. That may help explain why language models align only weakly with how image-based models represent the world, even at enormous scale.
Beyond the theoretical argument, reporting on AMI Labs’ research has highlighted experiments in which a world-model system with a far smaller parameter count outperformed language models many times its size on specific evaluation tasks, consistent with LeCun’s broader claim that understanding can substitute for raw memorized scale.
Frequently Asked Questions
What does AMI Labs stand for?
- 1Secure
- 2Regulation-compliant
- 3Auditable
- 4Scalable
AMI stands for Advanced Machine Intelligence. It’s the name Yann LeCun gave to the company he founded in Paris after leaving Meta.
How much funding has AMI Labs raised?
AMI Labs raised $1.03 billion in its first funding round in March, with backers including Jeff Bezos’s investment firm, Nvidia, Samsung, and Toyota. It’s described as one of the largest first funding rounds Europe has ever seen.
What is a world model in AI?
A world model is an AI system designed to build an internal understanding of how the physical world behaves, so it can predict what happens next in a scene or situation, rather than relying primarily on patterns learned from text.
Why does Yann LeCun think large language models are limited?
LeCun argues that text alone doesn’t contain enough grounded information about physical reality. He compares the roughly 10^14 bytes of text used to train large models to the comparable amount of visual information a four-year-old child processes, arguing the child develops real-world understanding that text-only models lack.
What is JEPA and how is it different from a standard language model?
JEPA (joint embedding predictive architecture) predicts outcomes in an abstract, compressed representation space rather than predicting raw pixels or exact words. This lets it ignore details that are impossible to predict and focus on the parts of a situation that actually matter for understanding what happens next.
