One trench in Silicon Valley

Late in 2025 LeCun left Meta, where he had led AI research for more than a decade, and helped launch Advanced Machine Intelligence — AMI Labs — with a Paris headquarters and nodes in New York, Montreal, and Singapore. In March 2026 the lab announced a $1.03B round at about a $3.5B pre-money valuation. Investors included Bezos Expeditions, Nvidia, Samsung, Toyota Ventures, and a long list of European and Asian funds.

His public reason for leaving is almost an engineering diagnosis. The valley digs one trench: large language models. Companies poach the same people, buy the same accelerators, and cannot afford to deviate — falling behind the shared roadmap looks more expensive than the risk of a dead end. For a public company that focus can be rational. For a researcher who believes human-level intelligence requires understanding the physical world, it is a cul-de-sac.

Separate rhetoric from the bet. LeCun does not claim ChatGPT, Claude, or Gemini are useless. He claims something sharper: scaling language models alone will not yield AGI and will not replace a household robot. Those are different jobs. The industry spent too long measuring progress with one ruler — parameter count and token count.

How much “world” a child sees versus how much text an LLM swallows

LeCun likes a simple calculation. A typical large language model trains on about 20 trillion tokens. At roughly three bytes per token that is on the order of (10^{14}) bytes — comparable to the visual information that travels the optic nerve in a child’s first four years (about 16,000 waking hours at roughly 2 MB/s).

These are order-of-magnitude numbers, not a lab protocol. The point is the contrast. From that volume of experience a human extracts everyday physics: object permanence, causality, the fragility of glass, the inertia of a ball. From a similar byte budget of text, a model extracts language statistics. Both systems “saw” enormous data. The data and the training goal are not the same.

Hence LeCun’s favorite contrast. A teenager can learn to drive in dozens of hours of practice. Level-5 autonomy still collides with infrastructure, long-tail cases, and huge corpora. A home robot that flexibly cleans a stranger’s kitchen remains a demo, not a product. Language models, meanwhile, already pass bar exams and draft working code. Success lands where reasoning lives in language; lag appears where you need a stable model of the physical world.

For engineers this is a useful filter. If your product is policy search, report drafting, or a coding assistant, LLMs are already the right tool. If the product is a warehouse manipulator, a turbine digital twin, or a household robot, “a bigger chat model” is usually not enough: you need dynamics, planning, and memory of environment state.

What JEPA is without the marketing gloss

A Joint Embedding Predictive Architecture does not have to reconstruct every pixel or every word. Observation (X) (and optionally action (A)) is encoded into a representation. A future or related observation (Y) is encoded too. Prediction happens in that representation space.

Picture a football match. A generative model can waste capacity on stadium seat color and a spectator’s shirt pattern — details that barely help you understand that a striker is closing on goal. JEPA tries to keep what helps the forecast: who is where, who is moving, what is likely next. Unpredictable noise can be dropped from the representation, so the forecast is coarser in detail and sharper in scene meaning.

That is close to how a person catches a ball. The brain does not simulate every blade of grass and every gust of wind. It keeps an approximate trajectory and updates it in flight. LeCun wants machines to learn a similar compression: an internal world model, not endless surface generation.

Historically the idea grew into I-JEPA, V-JEPA, and control-oriented branches. In 2026 the line continued with compact world models such as LeWM (~15M parameters, trainable on one GPU in hours) and hierarchical H-JEPA, where different levels watch different timescales.

From Plato’s hypothesis to a theorem about representations

In 2024 MIT researchers stated the Platonic Representation Hypothesis: as networks get stronger, their internal organizations look more alike — even across different training data. The intuition is elegant, but without a proof it did not say when similarity must appear or how to accelerate it.

Balestriero and LeCun turn the observation into a claim for JEPAs. In an idealized model, separately trained JEPAs on different modalities converge to one embedding up to rotation when their positive pairs relate the same world states and minimizers are unique up to rotation. When pairs disagree, misalignment is bounded by that mismatch.

The practical split is good views versus bad views. A good view is two observations of one scene: a photo of a dog catching a frisbee and a caption of that moment, two camera angles, an image and an action. A bad view is a corruption of one observation: masked text spans, random crops, the continuation of the same document. The first pulls the representation toward the world. The second pulls it toward modality quirks.

Against that backdrop, next-token prediction looks like a JEPA with bad views: the positive pair is a prefix and its continuation inside one document. The paper’s careful but sharp claim follows: language models align with vision mostly through scale because the training objective is poorly aimed at a shared world representation.

The 109M vs 72B experiment: what actually won

The numbers that circulate in videos should sit next to the metric definition.

On corpora of texts tied to web images (up to hundreds of millions of images), a 109M-parameter text JEPA reached about 0.184 on a PRH-style alignment benchmark. Qwen 2.5 at 72B scored about 0.154; Llama 3 at 7B about 0.151; CLIP’s text encoder lower still. The size gap is hundreds of times. On sample efficiency with the same corpus, good views reached comparable alignment with about 21.6M training samples where next-token prediction needed about 900M — roughly 42× more examples.

Another applied gesture from the paper: after separate training, encoders can be linked by one rotation and caption images without heavy joint fine-tuning. On a pair-free benchmark the text JEPA also posted strong zero-shot numbers against language models up to a few billion parameters.

Now the boundaries — more important than the headline.

1. The win is measured on text↔image representation alignment and related metrics, not on MMLU, HumanEval, or product chat quality.
2. Independent labs still need to reproduce the full pipeline. The theorem holds under stated conditions (whitening, spectral gap, pair design).
3. Scaling JEPA to frontier size is an open question. AI history is full of methods that looked brilliant small and lost their edge large.
4. World models in the literature can be brittle to small visual shifts — background color, lighting. For a robot that is a safety issue, not a footnote.

Architecture and positive-pair design can beat crude scale on alignment. That is not proof the industry “went the wrong way” on every other axis.

Hierarchy, robots, and a cat with 250 million neurons

LeCun often compares modern robots to a house cat. A cat has on the order of 250 million neurons and still jumps onto a narrow sill, dodges obstacles, and plans short movement sequences — without trillions of text tokens. The metaphor is deliberately provocative, and the direction is clear: action intelligence grows from a model of the environment, not from an encyclopedia.

LeWM (LeWorldModel) shows the engineering style of that bet: a compact JEPA world model trained end-to-end from raw pixels with two loss terms, planning in latent space faster than heavy foundation world models on several control tasks. H-JEPA adds floors: each level predicts farther ahead in its own space and temporal stride; planning runs top-down, and on Visual AntMaze the authors report success rising from 18% to 73% with less planner compute.

Other groups build neighboring pieces. Google DeepMind develops the Genie family for interactive worlds. Fei-Fei Li’s World Labs works on space modeling from a few photos. Nvidia publishes simulators and world models for robotics; university collaborations show policy transfer from fast soft-body simulators into reality.

The 2026 picture is no longer “LLMs versus everything.” A second race is forming: who can teach a machine to hold world state, predict action consequences, and plan across timescales. Language models do not vanish in that race — they become the interface, a skill library, and an explanation layer. AMI and LeCun still put a world model at the core of autonomy.

What this changes in the compute economy

If a coherent world representation can be learned with far fewer parameters and examples, more than academic rankings shift. Intelligence can live on a phone, glasses, a robot board computer, or a plant controller — without a permanent pipe to a data center full of accelerators.

AMI Labs says early commercial uses are likelier in industry: digital twins of power plants, aircraft engines, chemical facilities. There a world model is an equipment-state simulator and early-warning system, not a home assistant with a vacuum. A mass-market household robot sits farther down the roadmap.

Large LLM builders can simply absorb pieces of the idea: better views, latent-space prediction, hierarchical horizons. Then a “JEPA win” is not a vendor swap but an objective swap inside the same corporations. Product teams should not root for a lab brand; they should know which job they optimize — plausible text or stable environment forecast.

Related reading on this site: agentic engineering, AI hardware, local MoE and memory.

Four gaps to human-level intelligence — and what JEPA closes

LeCun usually lists several large holes on the path to human-level intelligence: understanding the physical world, long-term memory, reasoning, and planning complex action sequences. He often calls today’s attempts to “teach LLMs to reason” via prompt chains and external tools too thin: they imitate a train of thought in text without holding a stable model of the environment.

The JEPA line mainly advances the first item — and only partly. It offers a more efficient way to build world representations and connect modalities. H-JEPA adds a slice of multi-timescale planning. Persistent memory as a subsystem and adult-like general reasoning do not appear from these papers alone.

That is the sober filter for AGI headlines. Even if independent labs fully confirm the alignment numbers tomorrow, we get a strong architectural brick, not a finished general intelligence. Bricks matter: walls do not stand without them. Confusing a brick with a house is clickbait habit, not engineering.

For teams building agents on top of LLMs, the translation is concrete. Tools, RAG, and “thinking” prompts handle knowledge work. The moment an agent must hold the state of a warehouse, shop floor, or apartment and choose actions minutes ahead, you need a world-model loop — even a simplified simulator or explicit state graph — not another text wrapper around the same chat model.

What to do today if you build on LLMs

Do not throw GPT-class models out of production tomorrow. Do stop measuring every advance on the single axis of “more parameters.”

1. Split product jobs. Language, documents, code — LLM strength. Physics, long action plans, robustness to lighting and background — where world models, simulators, and classical control belong.
2. Treat multimodal alignment as its own metric. If you glue text and image, measure representation consistency, not only caption quality by eye. Good-view thinking helps now: train on pairs that are two observations of one scene, not only corruptions of one file.
3. For robots and digital physics, start with a simulator and compact dynamics. Public LeWM / H-JEPA branches and neighboring simulators offer an entry point without a frontier-lab budget.
4. Keep skepticism toward clickbait. “600× smaller and smarter” almost always means “better on a chosen axis.” Read the protocol: metric, corpus, baselines, what was not measured.
5. Watch reproducibility. Until independent repeats and scale-ups land, treat the numbers as a strong directional signal, not a replacement for your roadmap.

A one-week minimum: list three user scenarios for your product and mark each “language,” “world,” or “hybrid.” For hybrids, decide early where you need an environment-state model and where RAG plus an LLM is enough.

FAQ

Further reading

• Balestriero, LeCun — Proving and Improving the Platonic Representation Hypothesis with JEPAs
• Zhang et al. — H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning
• LeWM / hierarchical world models — AMI Labs / Planning-for-WMs repositories and notes
• AMI Labs updates: amilabs.xyz/updates
• On this site: agentic engineering 2026, ML frameworks, computer vision without framework lock-in

Conclusion

Videos titled “AGI found, transformers archived” catch the nerve of the moment and almost always overclaim the conclusion. The sober 2026 formulation is narrower. Language models will remain infrastructure for working with knowledge. Systems that act in the physical world as confidently as they talk about it will likely need a different learning goal: build a shared compressed world representation and predict consequences inside it.

JEPA, the Platonic representation theorem, LeWM, and H-JEPA are not the end of that story. They are evidence and prototypes that the lever “which views of the world we treat as the same” can beat the lever “how enormous the model is.” If that advantage survives scale, the next major AI shift may come not from another giant data center, but from a smarter training objective.