AMI Labs

AMI Labs is a company in Paris that builds world models. A world model is a network trained to predict how a scene will change, so that a machine can try actions in imagination before taking them. Yann LeCun, a Turing Award winner and until 2025 Meta's chief AI scientist, announced in November 2025 that he was leaving Meta to start it. Alexandre LeBrun, who had run the health-care AI company Nabla, is its chief executive officer.^[a]

On March 10, 2026, the company announced a seed round of $1.03 billion.^[3] In July its chief executive described it as pre-product.^[4]

LeCun had spent 12 years at Meta, five as the founding director of its Fundamental AI Research laboratory (FAIR) and seven as chief AI scientist. In a post of November 2025 he said that he would leave at the end of the year to start a company that would continue the AMI research program, and that Meta would be a partner in it. He has said that large language models are not a path to human-level intelligence. He described the aim as machines that understand the physical world and can plan complex sequences of actions.^[1]

A TechCrunch report on who was behind the company, published on January 23, 2026, as its website newly launched, named LeBrun as chief executive officer and LeCun as executive chairman. LeCun keeps his professorship at New York University, where he teaches one class a year. LeBrun had founded Wit.ai, which Facebook acquired, and then worked under LeCun at FAIR before co-founding Nabla. Nabla's board agreed to his move to chief AI scientist and chairman there in exchange for privileged access to AMI's world models.^[2]

Laurent Solly, a former Meta vice president for Europe, is chief operating officer. The March announcement also named Saining Xie as chief science officer, Pascale Fung as chief research and innovation officer and Michael Rabbat as vice president of world models.^[5]

The seed round, co-led by five venture firms, came to $1.03 billion, or about €890 million, and LeCun called it probably the largest seed round for a European company.^[3][6] At the start there were about 12 employees, in Paris, New York, Montreal and Singapore. Meta is not among the investors.^[7]

The website adds persistent memory, reasoning, and controllability and safety to the aims, and names the first uses as "industrial process control, automation, wearable devices, robotics, healthcare, and beyond". No products are listed. The company says it will publish its research and release some of its work as open source.^[8] The site's one stated belief is that "real intelligence does not start in language. It starts in the world."^[8]

LeBrun told TechCrunch that "world models" would be "the next buzzword", and that within six months every company would be calling itself one to raise money. He said that AMI starts from fundamental research and that commercial work in fields such as health care could take years.^[5] In July 2026 he said the company had nothing to sell yet and was looking for industrial partners, especially in South Korea, because a world model needs real-world data to be tested against.^[4]

LeCun set out the design in "A Path Towards Autonomous Machine Intelligence", a position paper posted to OpenReview on June 27, 2022.^[9] Summaries of it describe an agent built from six modules.^[10] Perception estimates the current state of the world. The world model predicts how that state will change, and it also fills in what perception cannot see, such as an object hidden behind another. A cost module scores the agent's "discomfort". Built-in drives such as hunger or pain feed it, and so does a critic that is trained. Short-term memory holds recent states and their costs, the actor turns predictions into actions, and a configurator adjusts the other modules for the task at hand. LeCun called the configurator a mystery that remained to be solved.^[10]

The agent has two ways of acting. In the fast mode, perception drives action directly, like a reflex. In the slow mode, borrowed from Daniel Kahneman's account of human thinking, the actor searches over possible sequences of actions, runs each through the world model, and keeps the sequence with the lowest predicted cost.^[10] Searching this way means running the world model many times, so it has to be cheap. LeCun's proposal for learning one was the joint embedding predictive architecture, which predicts the hidden part of an input as a representation and can leave out detail that no model could predict. AMI's site gives the same reason for avoiding generative models of sensor data, which is largely unpredictable.^[8] Meta's V-JEPA 2 of June 2025 carried out the planning step on robot arms in two laboratories, choosing actions by their predicted result against a goal image.^[11]

Two preprints with LeCun as an author deal with collapse, the failure in which a JEPA outputs the same vector for everything. LeJEPA, by Randall Balestriero and LeCun, was posted on November 11, 2025. Its regularizer, SIGReg, pushes the embeddings toward an isotropic Gaussian distribution, which its authors say removes the need for stop-gradients and teacher-student networks. With a frozen ViT-H/14 backbone it reached 79 percent linear-probe accuracy on ImageNet-1k.^[12] LeWorldModel, posted on March 13, 2026, applies the same idea to a world model that learns from raw pixels. Its model is small, about 15 million parameters with two loss terms, and trains in a few hours on one graphics processor. Its authors say it plans up to 48 times faster than world models built on foundation models.^[13]

Molecular Reality Corporation proposes a world model of its own for nanopore current. An essay of June 2026 in the voice of the company's puppet Keltar chooses a JEPA as its predictor, and the company's guide files the plan among proposals: "Both are proposals, not products, and I want that said first."^[14][15]

• Brain lab
• Joint embedding predictive architecture
• World model
• Meta AI