AMI Labs

AMI Labs, short for Advanced Machine Intelligence, is an AI research company developing world models: systems that learn abstract representations of sensor data and predict how situations evolve under actions. Its research program aims to support persistent memory, reasoning and planning, with controllability and safety as design goals. In March 2026, AMI announced $1.03 billion in financing to support long-term research, global hiring and development.[3][4]

AMI's approach learns representations of continuous, high-dimensional sensor data while ignoring unpredictable details, then makes predictions in that representation space. Action-conditioned world models incorporate proposed actions into those predictions, allowing an agent to anticipate consequences and plan sequences toward a task, subject to safety guardrails. These are intended properties of the research program, rather than demonstrated guarantees of reliability or safety.[3][4]

AMI names industrial process control, automation, wearable devices, robotics and healthcare as intended application areas. It also proposes working with industry partners, product developers and academic researchers through open publications and open source. Its stated premise is distinctive: “real intelligence does not start in language. It starts in the world.”[3]

The research direction predates the financing announcement. HV Capital describes AMI's technology as based on Yann LeCun's Joint Embedding Predictive Architecture, or JEPA, which he had been developing since 2022. That earlier work supplies technical context for AMI; the company's documented healthcare partnership followed in December 2025.[5][6]

On December 18, 2025, Nabla announced an exclusive strategic partnership with AMI, describing it as a new research company founded by Yann LeCun. The agreement gave Nabla first access to AMI's emerging world-model technologies. Nabla positioned the partnership as a path toward FDA-certifiable agentic healthcare systems—an ambition, not an announcement of regulatory certification or a completed deployment.[6]

Nabla also announced a planned, board-supported transition for co-founder and CEO Alex LeBrun to become AMI's CEO while taking the roles of chief AI scientist and chairman at Nabla. Nabla co-founder and COO Delphine Groll would continue leading daily operations during the search for a permanent CEO. By the March 2026 financing announcement, LeBrun was identified as AMI's CEO.[6][4]

On March 10, 2026, co-lead investor Cathay Innovation announced that AMI had raised $1.03 billion, approximately €890 million, at a $3.50 billion pre-money valuation, approximately €3 billion. HV Capital identifies the financing as a seed round. The named co-leads were Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Jeff Bezos Expeditions; the announced use of funds was long-term research, global hiring and development of reliable intelligent systems.[4][5]

The March announcement named Yann LeCun as chair, Alexandre LeBrun as CEO, Saining Xie as chief science officer, Pascale Fung as chief research and innovation officer, Michael Rabbat as vice president of world models and Laurent Solly as COO. AMI operated in Paris, New York, Montreal and Singapore, establishing a footprint across three continents from the outset.[4]

The LpWorldModel paper, submitted on August 24, 2026, provides a concrete example of AMI's collaborative research. Its authors are Yilun Kuang, Yash Dagade, Quentin Le Lidec, Lucas Maes, Randall Balestriero and Yann LeCun. Kuang, Le Lidec and LeCun list NYU and AMI affiliations; Maes lists Mila and AMI; Balestriero lists Brown University and AMI; Dagade lists Duke University. The work also acknowledges support through AMI Labs and NYU computing resources.[7]

LpWorldModel (LpWM) asks whether sparse representations make action-conditioned dynamics easier to predict. It uses a JEPA encoder to map observations into non-negative sparse codes, with many coordinates exactly zero. Rectified Distribution Matching Regularization matches those features to a rectified generalized Gaussian distribution, inducing sparsity and preventing the representation from collapsing. A predictor estimates the next code from the current representation and actions; planning searches for an action sequence whose predicted final representation approaches the encoded goal.[7]

The authors report that, on PushT—a task involving pushing a T-shaped block to a target pose—sparse LpWM outperforms dense LeWorldModel by “up to 57% in planning success at intermediate predictor capacities.” The benefit depends on predictor capacity: the highest-capacity comparisons performed similarly, while the simpler Wall navigation environment left little room to distinguish sparse and dense representations. The predictor-complexity sweeps used one training seed and three planning-evaluation seeds.[7]

The paper also explores what sparse codes represent. In a navigation environment divided into force-field zones, the active coordinates captured discrete dynamics regimes while their magnitudes captured finer state within each regime. In robot-arm cube manipulation, however, sparsity alone largely tracked motion rather than contact. Adding temporal Jaccard regularization, which discourages rapid changes in active coordinates between frames, made the support track slower physical changes such as object interaction. The result qualifies the interpretability claim: sparsity alone does not ensure semantically meaningful dynamics.[7]