The Collapse of the Pixel Cage: Yann LeCun Opens the Real Door to World Models with 15M Parameters and Two "Locks"

*Full English translation of a Chinese forum post from zhichai.net. Note: the post references events and a paper dated 2025–2026; facts below are as stated by the original author.*

Sora's Sunset: When the Whole Industry Celebrated Pixels, One Person Walked Away

When news broke that OpenAI would discontinue Sora, the author recalls the 2024 launch that made the entire AI community celebrate: "the world model has arrived!" Millions of parameters and massive video data predicted every pixel with cinematic fidelity.

Yet Yann LeCun, Turing Award winner, had warned since 2022 that "modeling the world for action by generating pixels is as wasteful and doomed to failure as the largely-abandoned idea of 'analysis by synthesis'" (his post drew 530K views on X). Meanwhile the arms race continued: Google DeepMind's DreamerV4 learned Minecraft through pixel reconstruction, and "bigger is better" became the default religion.

In November 2025, LeCun left Meta to co-found AMI Labs, raising a $1.03B seed round at a $3.5B valuation (investors reportedly including NVIDIA and Bezos), betting everything on true world models. Then, in March 2026, the quiet release of the LeWorldModel paper changed the conversation.

LeCun's 62-Page Prophecy: Why "Seeing Every Pixel" Blinds You to the World

LeCun's 2022 position paper *A Path Towards Autonomous Machine Intelligence* argued that AI should predict in abstract representation space, not pixel space. His analogy: simulating a soccer ball's flight requires mass, velocity, and gravity—not modeling the texture of every hexagonal panel.

Generative models do the opposite: they are forced to reproduce every pixel, including noise irrelevant to physics—like a student memorizing every word on an exam instead of the core formulas.

> Representation collapse is the most dangerous trap of the JEPA family. The model finds a cheat: map all inputs to the same point, so "predicting the future" is always trivially correct. Like a student who always answers "I don't know" and never learns anything.

The Minimalist Philosophy of JEPA: Keep the Skeleton, Not the Pixels

LeWM uses a Joint-Embedding Predictive Architecture: it predicts future states in abstract embedding space, keeping "where objects are and how they move" while discarding textures, lighting, and colors.

The previous best JEPA variant, PLDM, used 7 loss terms and 6 hyperparameters to prevent representation collapse—like locking one door with seven locks that jam each other. LeWM uses one lock: SIGReg.

SIGReg's mathematical foundation is the nearly century-old Cramér-Wold theorem: to verify whether a high-dimensional distribution is close to a standard normal, you don't inspect the full space—project onto 1024 random 1-D directions and apply the Epps-Pulley normality test to each. Like a health checkup via blood pressure and heart rate rather than opening up the body.

Only two forces remain:

• Prediction loss (contractive force, letting embeddings converge for prediction)
• SIGReg (expansive force, keeping embeddings diverse)

Result: training variance dropped from PLDM's ±5.0 to ±2.83; hyperparameter search complexity from O(n⁶) to O(log n).

The Data: 96% Success, 48x Faster Planning

| Dimension | LeWM (2 losses) | PLDM (7 losses) | DINO-WM (frozen, 124M-image pretraining) | |---|---|---|---| | Push-T success rate | 96% | 78% | 74% | | Training variance | ±2.83 | ±5.0 | — | | Hyperparameters | 1 | 6 | — | | Search complexity | O(log n) | O(n⁶) | — | | Planning speed | baseline | — | 48x slower |

From scratch, with only a 5M-parameter encoder, LeWM beat DINO-WM pretrained on 124M images. Why does deleting anti-collapse tricks make things *more* stable? PLDM's 7 losses are like seven people pulling one cart with seven ropes in contradictory directions. LeWM has two clear forces—contraction and expansion—like natural selection and mutation in evolution.

Emergence Nobody Asked For: Temporal Latent Path Straightening

The paper's most striking finding: during training, latent trajectories became increasingly "straight"—cosine similarity between consecutive latent velocity vectors rose from near 0 to about 0.6. No loss term asked for this; SIGReg constrains per-timestep distribution shape only.

This is temporal latent path straightening—pure emergence. PLDM *specifically designed* a time-smoothing loss (L_time-sim) plus an extra hyperparameter to encourage this, yet LeWM's trajectories are straighter.

In 2019, neuroscientist Hénaff et al. found human brains also represent complex temporal dynamics as nearly straight latent trajectories (the "temporal straightening hypothesis"), an evolutionary byproduct, not a design. A 15M-parameter model trained on one GPU for a few hours converged on the same representational principle as a brain shaped by hundreds of millions of years of evolution—suggesting straightening may be a universal inductive bias of efficient temporal prediction.

Testing AI Like an Infant: Violation-of-Expectation Experiments

Developmental psychology's classic violation-of-expectation (VoE) paradigm: 4-month-old infants stare longer at physically impossible scenes (a ball passing through a wall), indicating physical intuition.

The authors tested LeWM the same way, measuring prediction-error spikes ("surprise") across three trajectory types: normal, physically impossible (object teleports), and visually changed (object changes color):

• Teleportation: LeWM is significantly surprised (p < 0.01, in all three environments)
• Color change: almost no reaction

LeWM learned that "objects don't vanish" matters but "objects change color" doesn't—exactly LeCun's claim: understanding is knowing which details to ignore.

The reverse experiment is even sharper: forcing pixel reconstruction onto LeWM exploded training variance from ±2.83 to ±7.54. Seeing more clearly meant understanding more shallowly.

What Is He Really Against? The Debate Isn't Over, but Minimalism Has Proven Itself

LeWorldModel is restrained—no direct attacks on Sora or Dreamer—but its existence is a statement, co-authored by the man who said "wasteful" for four years.

Reality check: LeWM is validated only on simple continuous-control tasks (Push-T, Reacher, Two-Room). In complex 3D environments (OGBench-Cube), DINO-WM remains stronger. And JEPA has a structural weakness: without reconstruction, you cannot visualize what the model "imagines." If it develops physical hallucinations, you can't see them—generative models at least render their imagination for inspection.

But LeWM proves the minimalist route is viable—and possibly better: two losses beat seven. Emergent straightening beats designed straightening. An AI that doesn't look at pixels understands physics better than one forced to.

The Future of Physical Intuition

An easily missed appendix detail: SIGReg's internal parameters (number of projection directions M, quadrature nodes) have "almost no effect" on final performance. Its success rests not on tuning but on the correctness of a theorem proven nearly a century ago.

The entire industry is trying to replicate God's retina. He only wants to borrow a bit of physical intuition.

Minimalism, often, is the deepest complexity.

References

1. Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, Randall Balestriero. *LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels*. arXiv:2603.19312v2, 2026. 2. Yann LeCun. *A Path Towards Autonomous Machine Intelligence*. 2022 Position Paper. 3. Hénaff et al. *Perceptual Straightening of Natural Videos*. Journal of Neuroscience, 2019. 4. Spelke & Kinzler. *Core Knowledge*. Developmental Science, 2007. 5. Cramér, H. & Wold, H. (historical statistical foundations) and modern JEPA-family extensions—based on LeWorldModel's engineering implementation of the Cramér-Wold theorem.