Yann LeCun’s current bet is that systems trained only to predict text tokens will not be enough for reliable planning in the physical world. After leaving Meta at the end of 2025, he became executive chairman and co-founder of Advanced Machine Intelligence Labs, or AMI Labs. The company is pursuing world models that learn abstract representations from sensor data and predict the consequences of actions.

That is a serious research program, not proof that large language models are obsolete. Public results such as V-JEPA 2 show progress on video understanding, prediction, and bounded robot planning. They do not yet establish general physical intelligence, broad commercial reliability, or a safe autonomous agent.

As of September 13, 2026, NYU’s May 2026 announcement identifies LeCun as an NYU professor, says he led Meta’s AI research organization, and says he became executive chairman of AMI in January 2026. That public record corrects the original article’s future-tense departure framing.

The short answer: a complementary architecture, not a settled replacement

Language models learn rich patterns from text and code. They can summarize, generate, call tools, and support planning through language. Their next-token objective does not directly require a grounded model of how physical states change after an action. This gap motivates LeCun’s world-model work.

AMI Labs’ official research statement says the company wants systems that understand the real world, maintain persistent memory, reason and plan, and remain controllable. It argues that real-world sensor data contain unpredictable details and that useful models should predict in an abstract representation space rather than reconstruct every pixel.

This is the company’s technical thesis. “Understand,” “reason,” and “safe” are goals, not independently demonstrated properties. The advantage of the statement is that it points to a testable mechanism: learn representations that preserve predictable structure, then use action-conditioned models to forecast consequences and plan.

The framing should not become a false choice between language models and world models. A physical agent may use a language model for instructions and communication, a perception model for observations, a world model for state transitions, a planner for actions, and explicit software for constraints. The important question is which component improves measured performance and control.

What V-JEPA 2 actually demonstrates

Meta released V-JEPA 2 in June 2025 while LeCun was still its chief AI scientist. The paper on arXiv describes self-supervised pretraining on video and images, followed by action-conditioned training with robot trajectories. The authors report results on motion understanding, action anticipation, video question answering, and zero-shot picking and placing with robot arms in two laboratories.

The Meta research announcement describes V-JEPA 2 as a world model and calls its results state of the art on several benchmarks. That is a vendor-affiliated interpretation from the organization that developed the system. The arXiv paper supplies methods, metrics, and authorship, but at the cited version it is a preprint rather than evidence of universal performance.

The bounded result is still meaningful. A model learned useful visual structure from largely unlabeled video and supported planning for specific manipulation tasks without task-specific data from the deployment laboratories. It narrows the claim that representation-space prediction can support action.

It does not show:

  • robust operation across arbitrary robots and environments;
  • long-horizon planning with many interacting goals;
  • reliable causal understanding;
  • safe behavior around people under distribution shift;
  • lower lifecycle cost than alternative architectures;
  • general intelligence.

Those limits do not invalidate the research. They define the next experiments.

AMI Labs turns the thesis into a company

AMI Labs says it will work on industrial control, automation, wearables, robotics, healthcare, and other applications where reliability matters. Its official site names a distributed team and says it plans open publications and open-source work. These are intentions that can be checked over time.

In March 2026, TechCrunch reported that AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation. The report identified Alexandre LeBrun as chief executive and LeCun as chairman, and quoted LeBrun saying commercial applications could take years. Those financing figures are named reporting, not evidence that the research will succeed.

The size of the seed round gives AMI time and compute. It also raises the proof threshold. A research-first company must avoid translating scientific milestones into premature product claims. Investors and partners should expect a staged record:

  1. reproducible representation-learning results;
  2. action prediction that transfers across environments;
  3. planning gains against strong baselines;
  4. reliability under shift, noise, and adversarial conditions;
  5. bounded deployments with explicit safety constraints;
  6. economic value in a real workflow.

Each stage should specify datasets, hardware, baselines, failure cases, and independent reproduction. A high valuation does not move a system forward on that ladder.

Why the Meta departure matters

LeCun spent more than a decade leading AI research at Meta. Leaving a large platform for a research startup changes the constraints around the work. AMI can organize around one technical thesis and recruit directly for it. Meta offered far greater infrastructure, product data, and distribution but had to balance many commercial priorities.

Public reporting confirms the transition; it does not reveal private disagreements, family considerations, relocation decisions, or interpersonal motives. Those details appeared in the earlier version without adequate sourcing and are not needed to explain the strategic move.

The cleaner inference is organizational. A long-horizon architecture can be difficult to prioritize inside a company whose near-term products depend heavily on existing generative models. A dedicated lab can pursue it with fewer portfolio conflicts, while taking on financing, hiring, and commercialization risk.

AMI’s stated commitment to open publications and code deserves scrutiny rather than applause in advance. Observers can track what is released, under which license, with which training-data disclosures, and whether important evaluation tools are reproducible. “Open” has several meanings, and a paper without usable artifacts is different from a system another lab can inspect and run.

How to compare world-model systems

A useful evaluation should not ask whether a model looks intelligent in a video. It should measure whether the model predicts and controls states that matter.

For physical tasks, report:

  • success rate across pre-registered tasks and unseen environments;
  • calibration of predicted outcomes;
  • performance under occlusion, noise, changed objects, and delayed feedback;
  • planning horizon and recovery after a failed action;
  • interventions or safety violations per operating hour;
  • amount and provenance of training and adaptation data;
  • compute, latency, and energy per successful task;
  • comparison with language-model, model-free, and classical-control baselines.

The evaluation must separate perception from planning. If a system fails, researchers should know whether it misread the scene, predicted the wrong transition, chose a poor plan, or could not execute the action. A single end-to-end success number hides the research bottleneck.

Safety also needs an enforcement layer outside learned predictions. A world model can be wrong. Robots and industrial systems require physical limits, interlocks, monitored safe states, human authority, and tested shutdown paths. Learned controllability is not a replacement for engineering controls.

The forecast to track

LeCun’s core forecast is not that language models have no value. It is that reaching robust machine intelligence will require systems that learn persistent world representations, plan through them, and act under constraints. AMI Labs now gives that forecast an institutional test.

Over the next several years, the most informative evidence will be whether AMI can publish results that independent teams reproduce, outperform strong baselines on consequential real-world tasks, and retain performance outside curated laboratories. Commercial partnerships matter only when they produce measurable outcomes and disclose the model’s role.

Progress could also validate a hybrid rather than a winner-take-all conclusion. Language models may remain the interface and knowledge layer while world models supply grounding for physical prediction. Evidence should decide the architecture.

Remaining unknowns

AMI Labs has not publicly demonstrated general physical intelligence or a commercial system with a long production record. Its valuation and funding do not answer technical questions. Public sources do not disclose LeCun’s personal ownership, compensation, private motives for leaving Meta, or family arrangements, and this article does not speculate about them.

V-JEPA 2 results come from a Meta-affiliated research team and bounded evaluations. Independent replication and broader deployment evidence remain important. The term “world model” itself covers different methods, so comparisons need precise objectives and benchmarks.

The evidence supports a restrained conclusion. LeCun helped establish modern deep learning, led Meta AI research, left Meta, and now chairs a well-financed company pursuing representation-space world models. The scientific idea has promising experimental support. Its superiority, generality, safety, and business value remain open questions.

Source and correction note

This revision removes unsupported family and relocation details, anonymous former-employee claims, private motives, and future-tense language about an event that has already occurred. It distinguishes AMI and Meta statements from research evidence and independent reporting, and it narrows V-JEPA 2 conclusions to the tasks actually reported. Status was checked on September 13, 2026.