Resources
JEPA Hub
JEPA is an approach to training an AI model — similar to the way generative or diffusion models are approaches to training. JEPA isn't a product or even a specific technology implementation. There will be many, many products and companies built that train their models as JEPA models. While we are the first to bring a product built on a JEPA-style model to market commercially, we won't be the last. This hub is one place where anyone can learn about JEPA more broadly and why it represents the best path to what comes after LLMs.
Joint Embedding Predictive Architecture is the research lineage Primate Intelligence builds on. Papers, companies, analysis, and talks on world models that predict in representation space instead of pixel space.
Details on Primate's implementation of a JEPA model can be found on our technology, product and blog pages.
Papers 17
The JEPA family, in order of publication.
The founding position paper. Introduces the Joint Embedding Predictive Architecture (JEPA) as the centerpiece of a broader argument for world-model-based AI over pure autoregressive generation.
The first working JEPA model. Learns image representations by predicting masked-patch embeddings in latent space instead of reconstructing pixels — no hand-crafted augmentations.
Unifies optical flow estimation and content-feature learning in a single shared encoder trained with the JEPA objective.
Extends JEPA to video: learns by predicting masked spatio-temporal regions of a clip in representation space, producing strong off-the-shelf video features with no labels or fine-tuning.
Carries the JEPA masked-latent-prediction principle from images into the audio spectrogram domain.
Applies JEPA-style latent prediction to sensor time-series, compressing high-dimensional signals into low-dimensional semantic embeddings for bandwidth-constrained remote control.
Scales V-JEPA to 1M+ hours of internet video and adds an action-conditioned variant (V-JEPA 2-AC) that plans robot manipulation zero-shot in new environments from a small amount of unlabeled robot footage.
A JEPA architecture adapted specifically for general time-series representation learning, extending the masked-latent-prediction recipe beyond vision and audio.
Integrates V-JEPA representations with multimodal trajectory distillation to address mode collapse in end-to-end autonomous driving policies.
Unifies imitation learning and self-supervised learning by jointly predicting action sequences and latent observation sequences — up to 40% improvement in world-model understanding and 10% higher task success vs. the strongest baseline. Published in IEEE Access.
Asks whether the JEPA recipe that works so well for vision can improve language model training. LLM-JEPA outperforms standard LLM training objectives across Llama3, OpenELM, Gemma2, and Olmo, while being robust to overfitting — direct evidence embedding-space objectives beat input-space reconstruction outside vision too.
Vision-language model that predicts continuous text embeddings instead of autoregressively generating tokens. Beats CLIP, SigLIP2, and Perception Encoder on 16 video classification/retrieval benchmarks with 50% fewer trainable parameters, and matches InstructBLIP/QwenVL on VQA at just 1.6B params — the clearest evidence yet that JEPA-style training generalizes beyond pure vision into multimodal understanding.
Shapes a JEPA world model's representation space so the goal-conditioned value function approximates a distance between state embeddings, significantly improving planning performance over standard JEPA on control tasks. Presented at the World Modeling Workshop 2026 (Mila).
Systematic ablation of architecture, training objective, and planning algorithm for JEPA-based world models (JEPA-WMs) — combines findings into a model beating both DINO-WM and V-JEPA-2-AC baselines on navigation and manipulation. Accepted at TMLR.
Extends masked joint-embedding prediction from image patches to object-centric representations, forcing interaction-dependent prediction instead of shortcut solutions. ~20% absolute gain on counterfactual VQA reasoning and 100x fewer latent features needed for comparable planning performance. Accepted at ICML 2026.
First JEPA that trains stably end-to-end from raw pixels with just two loss terms (down from six), using only ~15M params trainable on a single GPU. Plans up to 48x faster than foundation-model-based world models while staying competitive on 2D/3D control — and its latent space reliably detects physically implausible events.
First formal generalization theory for JEPA world models — frames JEPA pretraining as conditional spectral graph learning, connects pretraining error to downstream planning regret, and proves a finite-sample generalization bound. Theoretical grounding for why latent-space prediction beats input-level prediction.
World Model Landscape: JEPA vs. Other Approaches 3
The definitive long-form pieces mapping where JEPA sits against video diffusion, symbolic, and spatial-3D approaches.
Maps the five competing research traditions racing toward a universal world model — the Cinematographer's (video diffusion), Dreamer's (imagination training), Physicist's (simulation), Roboticist's, and Architect's (JEPA) roads — and argues they're converging around video, except for JEPA, which bets the opposite: that pixel prediction is a dead end and abstraction wins. Names V-JEPA 2's 16-second planning speed vs. Cosmos's 4 minutes as the core empirical stake in what it calls "the LeCun Bet."
A deep-dive on AMI Labs chief scientist Saining Xie's argument that language is 'a poison' for building world models — that LLMs absorb a lossy, communication-optimized compression of reality rather than reality itself, and that a true world model needs learned representations with language as just one output interface, not the foundation.
McCormick's widely-read primer, written with General Intuition's Pim de Witte: why action-conditioned world models compute physical dynamics (crowds, contact, motion) that are computationally intractable for traditional simulators or LLMs, tracing the idea from Schmidhuber's 1990 paper through Ha & Schmidhuber's 2018 'World Models' to today's funding wars between LeCun's AMI Labs, Fei-Fei Li's World Labs, and DeepMind.
Blogs & Thought Leadership 5
Essential reading on JEPA and world models.
Meta's own explainer on why I-JEPA matters: representation learning that mimics how humans build abstract models of the world instead of memorizing pixels.
A clear, well-sourced technical explainer walking through the full JEPA lineage — I-JEPA, MC-JEPA, V-JEPA, V-JEPA 2 — and how each extends the core idea.
Mainstream coverage of the V-JEPA 2 launch and why the industry is watching world models as the next infrastructure layer for physical AI.
Long-form conversation where LeCun makes the case for joint embedding architectures over autoregressive LLMs as the path to systems that actually understand the physical world.
The competing (and complementary) thesis for why the next frontier of AI is spatial/world understanding, not just language — from the founder of World Labs.
Videos & Talks 9
Yann LeCun and others making the case for world models over LLMs, on camera.
The clearest public explainer of why LeCun raised ~$1B for AMI Labs to bet against autoregressive LLMs — walks through what a JEPA world model is, how it differs from next-token prediction, and why he thinks pixel-level generation is a dead end. Start here.
Follow-on to Part 1, going deeper on the JEPA architecture itself — energy-based self-supervised learning, representation-space prediction, and the hierarchical planning story LLMs can't do.
Focused walkthrough of the JEPA architecture as self-supervised world modeling — the core mechanics of predicting in representation space instead of pixel space.
Covers VL-JEPA, Meta's vision-language JEPA variant, and how it diverges from standard LLM architectures under LeCun's guidance.
Traces the JEPA roadmap from self-supervised learning through to LeCun's AGI milestones — a useful timeline view of how the architecture evolved.
Kilcher's paper-explainer breakdown of the original V-JEPA paper — unsupervised representation learning from video via feature prediction alone, no pixel reconstruction.
2.5-hour deep technical conversation. Timestamped sections on the limits of autoregressive LLMs, video prediction, JEPA, JEPA vs. LLMs, and hierarchical planning — the most complete public explainer of LeCun's reasoning on video.
LeCun's clearest post-Meta statement: LLMs are "a dead-end on the path to human-level intelligence" despite being useful products, because they don't build a world model. Covers why he left Meta and what AMI Labs is betting on instead.
LeCun and JP Vert discuss how LLMs translate — or fail to translate — into physical-world understanding, and the infrastructure world models will need.