The most serious challenge to the large language model's monopoly on AI research is not a better language model. It is an architecture that refuses to generate. It has no chat interface, it cannot write you a paragraph, and parts of it have been downloadable from Hugging Face for over a year. The architecture is JEPA — Joint Embedding Predictive Architecture — LeCun laid it out in a 2022 position paper, and the idea underneath it is older. In 2026 it stopped being purely a research program: one descendant of it is in production on a live dashcam network, and another got a billion-dollar company. Neither can hold a conversation, and that is the point.

This post is the technical tour. I will cover the objective, the model family with their actual numbers, the current research frontier, who is running these systems in production, and what you can download and fine-tune today. Where a claim has a primary source, I link it. Where something is demonstrated and where something is only promised, I keep the two separate, because the gap between them is where most of the sloppy writing about this line lives.

The objective: predict the representation, not the thing

A JEPA has three parts: a context encoder that maps what you can see into an embedding, a target encoder that maps what you want to predict into an embedding, and a predictor that learns to map the first embedding onto the second. That is the whole idea. The model predicts a representation of the target signal from a representation of the context signal, and everything happens in latent space. There is no decoder that turns embeddings back into pixels or tokens, and no reconstruction loss.

LeCun laid this out in his 2022 position paper, A Path Towards Autonomous Machine Intelligence, where the architecture appears as the perception core of a larger proposal. The same paper frames JEPA in energy-based-model terms: a JEPA is an EBM whose energy is the prediction error in representation space. Given a context, the model's energy is low for target embeddings the predictor can reach and high for everything else. Planning, in that framing, is searching for actions that lead to low-energy states. I will come back to that, because it is exactly how the robot version works.

The argument against the alternative — training by reconstruction — is about what prediction pressure does to the model. If the loss forces a system to reconstruct every pixel of a video frame, most of the model's capacity goes toward detail that no amount of context can predict: the exact texture of a leaf, the content of a neighbor's window, the noise floor of the camera. The paper's argument is that this is precisely the wrong budget allocation. A system that predicts in representation space can discard the unpredictable detail and spend its capacity on the structure that survives — objects, motion, cause and effect. (Related work shows JEPAs preferentially capture slowly varying features of the input, which is the same statement from the empirical side.)

Predicting in embedding space has a degenerate solution, though: the encoders can collapse. If every input maps to the same constant vector, prediction loss goes to zero and the model learns nothing. Preventing collapse is the actual engineering problem in this family, and the three known cures define three branches of the literature:

• Asymmetry. The target encoder is updated as an exponential moving average of the context encoder and carries a stop-gradient, so the target cannot race to meet the prediction. I-JEPA and V-JEPA 2 work this way.
• Regularization. VICReg-style variance and covariance penalties keep the embedding distribution from shrinking to a point. Several JEPA variants use this.
• Contrastive methods, which avoid collapse by pushing apart representations of different inputs, and which the JEPA line borrows from and relates back to (Mo et al., 2024).

The newest entry removes the heuristics entirely, and I will get to it below, because it is one of the more interesting papers in the stack.

The family, with receipts

JEPA is not one model. It is a line, and the line has a paper for every step.

I-JEPA (Assran et al., CVPR 2023) is the image version: predict representations of masked target blocks from a single context block, with ViT backbones, no hand-crafted augmentations. It was the first working demonstration that the objective scales.

V-JEPA (Bardes et al., ICLR 2025) extends masked feature prediction to video — predict representations of masked spatiotemporal regions from visible context.

MC-JEPA (Bardes, Ponce, LeCun, 2023) learns content features and optical flow in one shared encoder, which matters because motion is the signal a world model needs.

V-JEPA 2 (Assran et al., 2025) is the release that turned the line into something you could build on. Per Meta's announcement, it is a 1.2-billion-parameter ViT-g trained self-supervised on more than 1 million hours of internet video plus 1 million images, at 256×256 resolution. On the frozen encoder, a lightweight attentive read-out gets 77.3 top-1 accuracy on Something-Something v2 and a then-state-of-the-art 39.7 recall-at-5 on Epic-Kitchens-100 action anticipation. Aligned with a language model at the 8-billion-parameter scale, it reports 84.0 on PerceptionTest and 76.9 on TempCompass.

Meta also shipped three physical-reasoning benchmarks with it — IntPhys 2 (successor to the original), MVPBench, and CausalVQA — and the honest summary is that they are embarrassing for everyone. Humans score 85 to 95 percent across the three. Current video models sit at or near chance on IntPhys 2, the intuitive-physics benchmark. The models can tell you what happened in a video; they largely cannot tell you what would have happened if something had gone differently. That gap is the research front, stated as a measurement.

V-JEPA 2-AC: planning without tokens

The second training stage is where it gets interesting for anyone outside computer vision. Post-training the predictor on fewer than 62 hours of robot video from the DROID dataset — unlabeled, no rewards, no task tags — produces V-JEPA 2-AC, an action-conditioned world model. The frozen encoder never changes. Only the predictor learns, and what it learns is how embeddings evolve when the robot acts.

Planning is then embarrassingly simple, and I mean that as a compliment. You specify the goal as an image. Encode the current state and the goal state. Sample candidate action sequences, use the predictor to imagine where each one lands in embedding space, and keep the candidates whose predicted embeddings land closest to the goal embedding. Execute one step, re-observe, re-plan. There is no language model in the loop, no token decoding, and no gradient descent at planning time — it is gradient-free sampling in latent space, run as model-predictive control.

Deployed zero-shot on Franka arms in two labs the model had never seen — no data collected from those robots, no task-specific training — it completed pick-and-place of novel objects at 65 to 80 percent success (Meta's blog; the paper's per-object breakdown averages 80 percent for cups and 65 percent for boxes across the two labs). A world model trained on internet video, adapted with 62 hours of robot footage, planning in embedding space. Nothing here required a prompt.

The frontier moved after LeCun left

LeCun left Meta at the end of December 2025. The line kept moving, in several directions at once, and the newest papers are the ones I would watch:

• VL-JEPA (Delong Chen et al., Meta FAIR, with LeCun as co-author) — vision-language JEPA. Given a video and a query, it predicts the embedding of the answer instead of generating tokens autoregressively, decoding to text only when text is required. It matches or beats generative VLMs on classification, retrieval, and video QA with half the trainable parameters, its selective decoding cuts the number of decoding operations by 2.85×, and its non-autoregressive design makes real-time streaming video-language tasks practical — the generative stack has structural trouble with that latency profile.
• LLM-JEPA (Hai Huang, Yann LeCun, Randall Balestriero) applies the objective to language itself: keep the next-token loss for generative power, add a JEPA loss in embedding space. Across finetuning on NL-RX, GSM8K, Spider, and RottenTomatoes, it beats standard finetuning on models from the Llama 3, OpenELM, Gemma 2, and OLMo families, and it is robust to overfitting in a way plain finetuning is not. Code is public at rbalestr-lab/llm-jepa. This one deserves attention from anyone training or finetuning language models, because it is the first sign the objective is not vision-specific.
• LeJEPA (Balestriero and LeCun, "Provable and Scalable Self-Supervised Learning Without the Heuristics") is the theory paper. It argues the optimal embedding distribution is an isotropic Gaussian and derives SIGReg — Sketched Isotropic Gaussian Regularization — a single objective that provably prevents collapse with no stop-gradient, no EMA teacher, and no contrastive pairs. Implementation is a few dozen lines, code is public at galilai-group/lejepa, and it converges from ResNets to ViTs past a billion parameters. If the heuristics were the reason JEPA-style training felt fragile, this paper removes the excuse.
• H-JEPA — hierarchical JEPA, layered predictors over multiple timescales — remains at the proposal stage from the 2022 paper. Nobody has shipped a credible public implementation at scale. It is on the map, not in the repos.
• Apple's research group published an analysis of why self-distillation networks like JEPAs avoid learning noisy features — the implicit bias of deep linear self-distillation nets (paper, research summary). Independent theoretical support from outside the Meta orbit is worth more than another benchmark win inside it.

And Meta itself continued the line post-LeCun: V-JEPA 2.1 adds a dense predictive loss and deep self-supervision, producing dense features for depth, tracking, and manipulation — including a 20-point improvement in real-robot grasping success over V-JEPA 2-AC, 7.71 mAP on Ego4D short-term object-interaction anticipation, and 40.8 recall-at-5 on EPIC-KITCHENS. The institutional bet did not leave with its chief advocate.

In production: Nexar's BADAS

The most concrete evidence that this line is more than a research program is BADAS 2.0 (April 2026), Nexar's collision-anticipation system. It fine-tunes a V-JEPA 2 ViT-L backbone (300M parameters) end-to-end on 178,500 labeled dashcam videos — about 2 million training clips — mined from Nexar's network, which captures over 100 million miles of driving footage a month. The data pipeline is its own lesson: BADAS 1.0 ran as an active-mining oracle over millions of unlabeled drives to surface rare, safety-critical scenarios for human annotation, which is how the corpus grew 5× without exhaustive search.

Results, from the paper: Kaggle-competition mean average precision went from 0.925 to 0.940, early-warning recall — the fraction of collisions detected before they happen — is 91.3 percent for the full model, and distillation produces edge variants (Flash, 86M parameters, 4.8 ms; Flash-Lite, 22M, 2.8 ms) that run in real time on device. The paper's SSL ablation is the part I would show any skeptic of pretraining economics: a randomly initialized ViT-S scores 0.693 AP on the long-tail benchmark; the identical architecture with V-JEPA-style self-supervised pretraining on 2.25 million unlabeled driving videos scores 0.974 before any task training. The pretraining is the product.

Robotics labs have reproduced the planning loop on small hardware — hobby-scale SO-101 arms running cross-entropy-method sampling against a frozen V-JEPA 2 feature space — which is the replication signal you want to see before taking a paper's system claims seriously.

The money: AMI Labs

LeCun announced his departure from Meta on November 11, 2025, worked through the end of December, and founded AMI Labs. In March 2026, Reuters reported the company raised $1.03 billion at roughly a $3.5 billion valuation. Its CEO is Alexandre LeBrun. As of late October 2026 the lab is strictly pre-product — no models, no API, no benchmark results — which I state plainly so nobody reads the funding as a product announcement.

The thesis is the one from the 2022 paper, sharpened: LLMs have no persistent memory, no world model, and no planning capability (LeCun cites PlanBench-style evidence), because language is a narrow and lossy channel into intelligence. The bet is that world models learned in representation space are the road to systems that reason and plan, and that the road has years, not months, of research left.

Counterarguments deserve their paragraph, and here they are honestly. Peter Norvig has argued that language models already are world models — implicit ones, learned through language — and that the field needs tweaks, not a restart. Scaling maximalists hold that continued scaling of autoregressive models subsumes the problem; every capability LeCun attributes to world models, they say, shows up in sufficiently large sequence predictors with the right scaffolding. And the VLA camp — vision-language-action systems like Physical Intelligence's π0 — bypasses explicit world models entirely, training policies directly on robot trajectories, and currently owns most of the commercial robot-foundation-model momentum. There is also a competing landscape to keep straight: generative world models (DeepMind's Genie 3, World Labs' Marble — World Labs was acquired by AMD for roughly $8.2 billion in stock in September 2026, which tells you what strategists think this category is worth), NVIDIA Cosmos 3 (open-weight, Mixture-of-Transformers), and model-based RL in the Dreamer line (Danijar Hafner's research program). JEPA-style predictive-representation models are one of three architectural families in the world-model race, not the only one.

What you can run today

Everything below is downloadable. No waitlist, no API key, no product meeting.

The V-JEPA 2 checkpoints are on Hugging Face: facebook/vjepa2-vitl-fpc64-256 (the ViT-L that BADAS fine-tuned), facebook/vjepa2-vith-fpc64-256, facebook/vjepa2-vitg-fpc64-256, and the 384-resolution vitg-fpc64-384. The transformers integration loads them through the standard interfaces:

from transformers import AutoVideoProcessor, AutoModel

model_id = "facebook/vjepa2-vitl-fpc64-256"
processor = AutoVideoProcessor.from_pretrained(model_id)
model = AutoModel.from_pretrained(model_id)

# videos: list of frame tensors, e.g. 64 frames @ 256x256
inputs = processor(videos=videos, return_tensors="pt")
outputs = model(**inputs)
features = outputs.last_hidden_state  # embed everything you need from here

What the compute budget looks like, honestly:

• Feature extraction plus a probe (linear or attentive read-out on frozen features) runs on a single GPU. This is the recommended entry point and how most of the paper's understanding results were produced.
• Backbone fine-tuning is a multi-GPU job — BADAS 2.0 trained with gradient checkpointing on 8 GPUs — but it is reachable by a well-funded lab or a serious team, not just Meta.
• Pretraining from scratch on internet-scale video is a frontier-lab job. Do not plan on it; plan on the frozen features, which the evidence says carry most of the value.

For the planning loop, the shape is a dozen lines of pseudocode against the frozen encoder:

goal = encoder(goal_image)
state = encoder(current_frame)
for step in range(horizon):
    candidates = sample_actions(state)          # CEM / random shooting
    predicted = [predictor(state, a) for a in candidates]
    scores = [distance(p, goal) for p in predicted]
    execute(best_action(candidates, scores))
    state = encoder(observe())

For a curated entry point, AbdelStark/awesome-jepa indexes the papers, checkpoints, and code across the whole family, and the EB-JEPA library packages image, video, and action-conditioned JEPA experiments into a single-GPU tutorial.

Why track this if you are not building robots

Three reasons, in increasing order of speculativeness.

First, the production evidence is real. A JEPA-derived system is fine-tuned and distilled onto edge hardware, making sub-five-millisecond collision predictions on a commercial dashcam network today. That is not a roadmap slide; it is a bill of materials.

Second, the economics are different in kind. LLM inference is metered in tokens generated per request. A predictive world model running against a video stream is metered in time — it encodes continuously whether or not anyone asks it a question. If this family grows into production systems the way LLMs did, it will not inherit the request/response pricing model, and cost management for it has barely been invented. Anyone whose job involves forecasting AI capacity should want a vocabulary for that before it arrives.

Third, and this is the actual lesson of the last three years: the space is bigger than the chat box, and it changes fast. A billion dollars of new capital, a production system on real roads, and a planning result on real robot arms all came out of a line most practitioners still file under "LeCun's crank opinion." None of it showed up through the channels we watch — no model release, no benchmark tweet, no product launch. It showed up in arXiv feeds and a dashcam vendor's engineering blog.

I am not claiming JEPA replaces anything. V-JEPA 2 cannot write a sentence, the physical-reasoning benchmarks still humiliate every model taking them, and AMI Labs has shipped nothing yet. What I am claiming is narrower: the generative paradigm has a serious, funded, downloadable alternative with production traction in at least one domain, and pretending the alternative does not exist is not a strategy. It is a blind spot with a paper trail.

Track it. The checkpoints are public, and the next Jevons moment does not schedule itself.