The argument
Generative world models such as Genie spend most of their compute drawing the world: every frame, every texture, every leaf. LeCun’s argument is that an agent does not need any of that to make a decision. It needs to know what changes when it acts. A model that predicts the meaning of the next state, rather than its appearance, can ignore everything that is unpredictable and irrelevant, and learn faster from less data.
That is the Joint Embedding Predictive Architecture. Two encoders turn the visible part of the input and the hidden or future part into abstract representations; a predictor learns to map one to the other. The training signal lives in representation space, so nothing ever has to be rendered.
What Meta shipped: I-JEPA to V-JEPA 2
Meta’s FAIR lab built the idea out in steps: I-JEPA for images in 2023, V-JEPA for video in early 2024, and V-JEPA 2 in June 2025, released openly. V-JEPA 2 was pretrained without labels on more than a million hours of internet video. An action-conditioned variant, V-JEPA 2-AC, was then trained on less than 62 hours of robot-arm video from the open DROID dataset, and Meta showed it planning reach, grasp and place moves zero-shot in labs it had never seen.
The 62-hour number is the interesting one. The expensive input in robotics is robot data, and V-JEPA 2 used very little of it, leaning on cheap video for everything else. That is the economic argument from the world models explainer, made concrete.
AMI Labs
LeCun left Meta at the end of 2025 and founded AMI Labs (Advanced Machine Intelligence) in Paris in December. In March 2026 it raised a $1.03 billion seed round at a $3.5 billion valuation, co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, with Nvidia, Samsung, Toyota Ventures, Temasek and Bpifrance among the backers. Alexandre LeBrun is CEO, and the company has offices in New York, Montreal and Singapore.
AMI says it is building systems that understand the world, keep persistent memory, reason and plan, and are controllable and safe, with world models that predict in representation space so agents can plan sequences of actions within guardrails. It has also said not to expect a sellable product for years. That candour is useful: this is a research lab with a very large balance sheet, not a vendor.
How it compares
- Against Genie and Runway: they produce worlds you can see and steer; JEPA produces predictions an agent can plan with. Different outputs, different buyers.
- Against Nvidia Cosmos: Cosmos generates synthetic training data and is open today. JEPA models aim to reduce how much training data a robot needs in the first place.
- Against language models: LeCun’s broader claim is that text alone cannot teach physical common sense. The jury is out, but V-JEPA 2’s robot results are the strongest public evidence for his side.
What to do with it
If you run a robotics team, V-JEPA 2 is open and worth benchmarking as a planning backbone against your current policy, particularly if your bottleneck is collecting robot demonstrations. If you are evaluating AMI Labs as a future supplier, there is nothing to buy yet; watch for published transfer results on real hardware rather than demos.
Related
- World models: the explainer and the full list of world models
- Google Genie vs other world models: the generative approach JEPA argues against
- Nvidia Cosmos: open world foundation models for robot training data
- Embodied AI: where planning models fit in a robot’s stack
Frequently asked questions
What is JEPA?
JEPA stands for Joint Embedding Predictive Architecture, an approach proposed by Yann LeCun. Instead of reconstructing missing pixels or generating the next frame, the model encodes the input into an abstract representation and learns to predict the representation of the hidden or future part. Unpredictable detail such as the exact pattern of leaves never has to be modelled, which makes training more efficient and the learned representation more useful for reasoning and planning.
What is V-JEPA 2?
V-JEPA 2 is Meta’s open video world model, released in June 2025. It was pretrained without labels on more than a million hours of internet video, then an action-conditioned version, V-JEPA 2-AC, was trained on less than 62 hours of robot-arm video from the DROID dataset. Meta showed it planning reach, grasp and place actions zero-shot in environments it had not seen during training.
What is AMI Labs?
AMI Labs (Advanced Machine Intelligence) is the company Yann LeCun founded after leaving Meta, where he was chief AI scientist. It was founded in December 2025, is headquartered in Paris with offices in New York, Montreal and Singapore, and is run by CEO Alexandre LeBrun. It raised a $1.03 billion seed round at a $3.5 billion valuation in March 2026.
Who invested in AMI Labs?
The March 2026 seed round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, with Nvidia, Samsung, Toyota Ventures, Temasek, Bpifrance and others participating.
How is JEPA different from Genie or Sora?
Genie, Sora and Cosmos are generative: they produce pixels you can watch. JEPA models never produce pixels. They are built for an agent to predict the consequence of an action in an abstract space and choose a plan. If you want something to look at, JEPA is the wrong tool; if you want a robot to decide what to do next cheaply, it is a serious contender.
Sources: Meta AI (V-JEPA 2, June 2025) and the V-JEPA 2 paper (arXiv 2506.09985), Silicon Republic and TechCrunch (AMI Labs seed, March 2026), AMI Labs. Figures are as reported by the companies.