LeCun's JEPA Roadmap Has a Regulatory Gap

Yann LeCun's war with the mainstream AI establishment is well documented. What gets less attention is where he wants to fight the next battle—and what it would mean for the regulators who aren't watching.

In a detailed interview with Welch Labs, LeCun made his case against vision language action models, the architecture underpinning Physical Intelligence's PI07 robot and most of the industry's current showcase hardware. His critique has two edges. First, behavioral cloning—training robots by having humans demonstrate tasks—is fundamentally unscalable because you cannot collect demonstrations for every variation of every situation a robot will encounter. Second, and more philosophically, VLA models have no explicit world model. They cannot predict the consequences of their actions before taking them. They just act, then observe. As LeCun put it in the Welch Labs interview: "I do not understand how you can even think of building an agentic system without the agentic system having the ability of predicting the consequences of its actions."

His alternative, JEPA—Joint Embedding Predictive Architecture—learns by predicting representations of future states rather than imitating human behavior. The LeWorldModel implementation, demonstrated on a tabletop manipulation task, shows a system that can plan explicitly: simulate candidate action sequences in learned embedding space, score them against a goal state, and select the best one. It is, as LeCun says, "classical optimal control"—something engineers have done for decades, now applied to systems whose physics can't be written down.

The planning horizon is currently modest. Demonstrated implementations can reliably look about five steps ahead, extendable to fifteen with hierarchical world models. That's a long way from cleaning a kitchen. But JEPA's architectural logic has already produced competitive results at the vision encoder and vision-language model layers—V-JEPA 2, trained on one million hours of video without any language supervision, matches or exceeds language-supervised encoders on video understanding benchmarks, according to the paper posted to arXiv as 2506.09985 (readers should note that arXiv IDs are easy to transpose; verify against the Meta research page before citing). VL-JEPA, Meta's full vision-language variant, reportedly outperformed 7-billion parameter generative models on the GQA compositional reasoning benchmark while using just 1.6 billion parameters—though the Welch Labs presentation does not name which 7B models served as the baseline, which matters considerably for evaluating that claim.

These are architecture results. They tell us something about what the approach can do. They do not tell us what happens when LeCun's roadmap reaches the physical world.

The part where it becomes my problem

Here is where LeCun stepped squarely onto regulatory terrain without apparently noticing, or caring. Asked what JEPA-based systems would tackle first, he offered a list: jet engines, chemical plants, power plants. And then: "a patient with a disease like say diabetes—what course of treatment should you follow to control the blood sugar of the patient."

That sentence is a regulatory event.

A system that recommends or controls a course of treatment for a diabetic patient is a medical device under FDA jurisdiction. The FDA's Software as a Medical Device framework, developed in coordination with the International Medical Device Regulators Forum, requires demonstrable evidence of safety and effectiveness. For AI/ML-based devices, the FDA's action plan released in 2021 and updated guidance since has focused heavily on algorithm change protocols—essentially, regulators want to know how a deployed model will change over time and who is accountable when it does.

The behavioral cloning that LeCun criticizes as unscalable is, from the FDA's perspective, one of the few things that creates an auditable training data trail. If a VLA-style medical system fails, investigators can examine the demonstration data, identify what the model was and wasn't shown, and construct a theory of failure. A phenomenological world model trained from raw patient data—predicting future physiological states from historical observations, then planning treatment sequences—has no such trail. The model learned dynamics from data. Which data? What was the distribution? What edge cases weren't represented? These are not hypothetical regulatory questions. They are the exact questions the FDA asks before approving algorithmic decision support tools.

The EU AI Act, which entered force in August 2024, would classify AI systems making treatment recommendations as high-risk under Annex III. High-risk systems face mandatory conformity assessments, transparency obligations, and human oversight requirements. Critically, they must be trained on data that is "relevant, sufficiently representative, free of errors and complete." A world model that learns phenomenological dynamics from hospital records is going to have a difficult conversation with that provision—patient data is rarely any of those things.

VLA's own regulatory vacancy

It would be a mistake to read this as JEPA-specific vulnerability. Physical Intelligence's PI07 can already fold laundry and handle kitchen tasks in what appear to be real residential environments. The choreography question aside, the actual regulatory status of these deployments is not obvious. OSHA's machine safety standards, developed for predictable industrial equipment, were not written with end-to-end learned systems in mind. When a VLA robot fails—drops something, misidentifies an object, applies unexpected force—who bears liability? The operator? The deployer? Physical Intelligence? The question of whether behavioral cloning data carries any provenance or consent requirements has not been resolved; in jurisdictions with broad data protection law, the humans who provided demonstrations arguably have interests that haven't been addressed.

LeCun's critique of VLA generalization is valid as far as it goes. VLA models do generalize—Google's RT2, in a widely cited 2023 demonstration described in secondary coverage as moving a soda can toward an image of a celebrity, showed that language priors learned during pretraining can transfer to novel physical tasks—but the details of that specific demonstration vary across retellings and should be verified against Brohan et al. (2023) before treating it as established fact. What's not in dispute is that VLA generalization degrades as tasks diverge from training distribution, and that nobody has established what "safe enough" generalization looks like for regulated deployment contexts.

The audit problem

LeCun's world model approach has a different version of the same issue. The planning loop is explicit and inspectable in a way VLA is not—you can, in principle, examine what action sequences were considered and why the optimizer preferred one over another. That's genuinely useful for accountability. But the world model itself, the learned representation of how the environment behaves, is still an opaque neural network. When the LeWorldModel demonstration shows the predictor "going off the rails" on longer planning horizons, that instability isn't just a performance limitation. In a regulated context, it's a failure mode that needs a containment strategy.

Hierarchical JEPA—LeCun's proposed solution to the long-horizon problem, where high-level models generate sub-goals for lower-level planners—introduces its own audit complexity. The interface between hierarchy levels is, explicitly, "some embedding space," not semantic or language-based. LeCun said as much in the Welch Labs interview: "It doesn't have to be semantic or certainly not language." That's architecturally elegant. It is also, from a regulator's perspective, a black box inside a black box. The EU AI Act's transparency requirements for high-risk systems require that outputs be interpretable to the humans overseeing them. Embedding space sub-goals are not that.

LeCun's historical analogy—JEPA today looks limited the way early deep learning looked limited before it transformed the field—is not implausible. But early deep learning didn't propose to manage diabetic patients in its first two years of industrial deployment. The research arc that LeCun describes, from push-T tabletop manipulation to jet engine control to clinical treatment planning, compresses decades of regulatory learning into a timeline that assumes the frameworks will either be built or bypassed as needed.

They won't be built in time. That's not pessimism; it's the observable pace of AI-specific rulemaking. The EU AI Act took four years to pass after the Commission's first draft. FDA guidance on AI/ML software has been in iterative development since 2019 with no finalized rule yet in place. The gap between what these systems will be capable of and what regulators are prepared to evaluate is widening, regardless of whether the winning architecture turns out to be JEPA, VLA, or something neither camp has published yet.

LeCun frames the choice between JEPA and VLA as a question of which approach will produce reliable, safe, agentic systems. Regulators will eventually ask the same question—but they'll need an answer they can put in a docket, not an embedding vector.

---

Samira Barnes covers technology policy and regulation for Buzzrag.