AMI Labs and JEPA: LeCun's Bet on World Models Over LLMs
On December 19, 2025, Yann LeCun confirmed what the industry had been speculating for months: he launched a new company dedicated to world model development. The startup, identified as AMI Labs, reportedly seeks a valuation above $5 billion. LeCun will not serve as CEO. What he will bring is the technical thesis he has defended publicly since 2022: LLMs cannot reason, and the future of artificial intelligence runs through architectures that learn models of the physical world.
"LLMs produce statistically plausible text. They do not understand the world. They cannot plan. They cannot learn the way animals learn — by observing the physical world. A machine that only predicts the next token will never reach human-level intelligence." — Yann LeCun, A Path Towards Autonomous Machine Intelligence (2022)
The Three Structural Limitations of LLMs
LeCun's position is not new. Since his 2022 paper he has argued that LLMs present three limitations that no scaling can resolve. These are not data limitations, parameter limitations, or compute limitations — they are paradigm limitations.
No world model
An LLM predicts the next token in a sequence. It maintains no internal state of the world. It does not "know" that objects fall, doors open, or fire burns — it has merely seen those phrases co-occur many times.
No continuous planning
LLMs operate in discrete token spaces. Physical and decision environments are continuous: (s, a) ∈ S × A. Chain of thought simulates planning but cannot navigate continuous state spaces or optimize trajectories with dynamic constraints.
Correlation ≠ causality
Token prediction pretraining maximizes statistical coherence, not causal understanding. A model can correctly answer "if I drop the ball it falls" without having learned any causal mechanism — only text co-occurrence across trillions of documents.
JEPA → W-JEPA
Joint Embedding Predictive Architecture does not generate in observation space. It learns abstract representations and predicts future states in that space — enabling multi-step planning, causal reasoning, and physical understanding without producing pixels or tokens.
JEPA Architecture: The Mathematics of the Paradigm
The central intuition of JEPA is to minimize distances in representation space, not to reconstruct signals. The encoder learns to compress x into s_x — discarding irrelevant perceptual details. The predictor learns to predict s_y from s_x. The key: y is never reconstructed in pixel or token space.
Inference: LLM vs. W-JEPA in Planning
The operational difference is critical. An LLM generates tokens one by one: each token conditions the next but cannot "go back" to modify a prior decision. W-JEPA operates in a continuous state space where the predictor can optimize a complete trajectory before executing the first action — the equivalent of mentally planning the route before starting to walk. V-JEPA 2 already demonstrated emergent physical reasoning (pendulum prediction, falls, collisions) without explicit physics supervision — purely as a consequence of predicting in latent space.
The JEPA Family: From Image to Physical World
JEPA vs LLMs vs Diffusion Models: Comparative Analysis
| Dimension | JEPA / W-JEPA | LLM (GPT / Claude) | Diffusion (Sora / DALL·E) |
|---|---|---|---|
| Training objective | Predict representations (embedding space) | Predict next token (max log-likelihood) | Reconstruct full signal (denoising) |
| Operation space | Continuous (latent) | Discrete (vocabulary) | Continuous (pixels) |
| Causal reasoning | Structural — in the architecture | Emergent — fragile, spurious causality | Absent |
| Multi-step planning | Native in latent space (trajectories) | Chain of thought — no continuous state | N/A |
| Sample efficiency | High — 2-6× vs. diffusion (V-JEPA 2) | Low — scale (trillions of tokens) compensates | Medium |
| Integrated physics | Core W-JEPA objective | No — linguistic correlation of physics | Partial (appearance, not mechanics) |
| Natural applications | Robotics, simulation, autonomous control | NLP, code, symbolic reasoning | Content generation |
AMI Labs: Spinout Structure and Context
Not the CEO
Classic scientific spinout pattern: the technical founder leads the vision, an external operator manages the company. LeCun provides the thesis; another executive turns it into business.
Meta FAIR
I-JEPA and V-JEPA were developed at Meta FAIR. W-JEPA applications (robotics, hardware-in-the-loop) do not fit Meta's advertising business model — hence the spinout.
$5B+
A round of this size implies tier-1 fund backing. The bet: that W-JEPA can be trained with efficiency comparable to current LLMs in domains where LLMs have structural limits.
19 dic. 2025
Described by press as "the worst-kept secret in tech." Industry speculation anticipated the announcement by months — confirming that capital interest preceded the public confirmation.
W-JEPA Production Stack: Reference Architecture
Timeline: From JEPA to AMI Labs
Key Takeaways
- JEPA does not generate: it predicts distances in abstract representation space. Eliminates reconstruction of irrelevant perceptual details and achieves 2-6× greater sample efficiency than equivalent diffusion models (V-JEPA 2).
- LLMs' three structural limitations are paradigm limitations, not scale limitations: no world model, no continuous-space planning, statistical correlation instead of causality. No parameter or token scaling resolves them.
- V-JEPA 2 demonstrated emergent physical reasoning (pendulums, falls, collisions) without explicit physics supervision — validating that predicting in latent space is sufficient to learn mechanics.
- AMI Labs seeks $5B+ with LeCun as chief scientist (not CEO). The Meta FAIR spinout is natural: W-JEPA for robotics and hardware-in-the-loop is incompatible with Meta's advertising business model.
- For production decision systems, W-JEPA enables what LLMs cannot: multi-step latent trajectories, consequence simulation, and counterfactual reasoning in continuous state spaces.