What Is JEPA? Yann LeCun's Joint-Embedding Predictive Architecture, Explained
— by Vivax
What is JEPA? The Joint-Embedding Predictive Architecture is Yann LeCun's blueprint for AI world models that learn by predicting in latent space instead of…
JEPA — the Joint-Embedding Predictive Architecture — is Yann LeCun's blueprint for AI systems that learn a world model: an internal, predictive understanding of how the world works. Proposed in his June 2022 position paper 'A Path Towards Autonomous Machine Intelligence', JEPA departs from mainstream generative AI in one decisive way. It does not learn by generating data — no pixels, no tokens. Instead, an encoder maps an observation into an abstract representation, and a predictor forecasts the representation of what is missing or what comes next. Because prediction happens in this latent space, the model is free to discard details that are unpredictable or irrelevant — the exact shape of every leaf on a tree, the noise on a camera sensor — and spend its capacity on what actually matters: the state of the world and how it evolves.
Mechanically, a JEPA has three parts: a context encoder that embeds what the model can see, a target encoder that embeds what it must predict (usually a slowly-updated copy of the first, which prevents the network from collapsing to trivial answers), and a predictor that maps from context to target representations. Contrast this with the two dominant generative recipes. Autoregressive language models must commit probability to every next token, and diffusion models must render every pixel — both are forced to model the unpredictable. LeCun's argument is that this is why generative video models need enormous compute to learn what a house cat knows about physics. The efficiency shows up in practice: I-JEPA reached strong ImageNet representations with roughly ten times less pretraining compute than popular pixel-reconstruction methods, training a huge vision transformer in under 72 GPU-hours-scale budgets that generative pretraining could not match.
The architecture stopped being a proposal and became a lineage in three steps. I-JEPA (June 2023, CVPR) applied the idea to still images: mask a region, predict its representation from the surrounding context. V-JEPA (February 2024) moved to video, where masked latent prediction over space and time produced state-of-the-art frozen-evaluation results on motion-understanding benchmarks — with no labels, no text supervision, and no pixel reconstruction. V-JEPA 2 (June 2025) scaled the recipe to 1.2 billion parameters and over a million hours of video, then added a second, action-conditioned training stage on just 62 hours of robot data. The result is the family's headline: deployed zero-shot on robot arms it never saw during training, V-JEPA 2 plans pick-and-place of novel objects at 65–80% success by imagining outcomes in latent space. The lineage has also escaped Meta's labs — medical variants like Neuro-JEPA apply the same latent-predictive objective to multimodal brain MRI.
JEPA is best understood as one of two competing routes to the same destination: the world model. A world model — a term David Ha and Jürgen Schmidhuber revived in their 2018 paper of that name — is an internal simulator: a system that represents the current state of an environment and predicts how it will change over time, including in response to actions. The generative route, taken by NVIDIA's Cosmos world foundation models and DeepMind's Genie line, builds simulators that render the future frame by frame — visually rich, directly inspectable, and expensive. The JEPA route predicts the future's representation instead of its appearance — cheaper, harder to inspect, and by design uninterested in irrelevant detail. Both routes are converging on the same test: can the model support understanding, prediction and planning in the physical world?
The honest part of the story is what JEPA cannot do yet. On Meta's own physical-reasoning benchmarks — IntPhys 2, MVPBench and CausalVQA — humans score 85–95% while today's best video models, V-JEPA 2 included, remain far behind, sometimes near chance. A JEPA has no language interface of its own; it must be aligned with a language model to answer questions. Planning horizons are still short, hierarchical JEPA — the stacked, multi-timescale version LeCun sketched in 2022 — remains unbuilt, and evaluating a model that never generates anything is genuinely harder: you cannot just look at its samples. These are open problems, not footnotes, and they are where the research frontier sits in 2026.
Vivax builds clinical world models in exactly this lineage, and the reasons are practical, not tribal. Medicine is a domain where the JEPA trade fits: most pixels in a clinical scene are irrelevant to the state that matters, labeled data is scarce while passive data is abundant, and the loop a clinician runs — recognize the patient's state, forecast its trajectory, plan the intervention — is the understand-predict-plan loop a world model implements. V-JEPA 2's 62-hour result is the existence proof we care about most: a world model pre-trained on abundant passive data needs remarkably little interaction data to become actionable. That is the economics of medical AI in one sentence — and it is why the question 'what is JEPA?' is, for us, the question of how clinical AI gets built next.