LeCun and Balestriero say JEPAs prove the 'Platonic' AI hypothesis. What the numbers show

Randall Balestriero and Yann LeCun have posted a preprint, Proving and Improving the Platonic Representation Hypothesis with JEPAs, to arXiv. It went up on October 29. In it, they claim a formal proof that a family of self-supervised models converges on a shared picture of the world, whatever data type it learns from. They also report that a small text model built on this idea lines up with image models better than language models hundreds of times its size.

Why it matters: if the result holds, it points to a training approach that learns from examples more efficiently than next-token prediction, the objective behind today's large language models. It also hands JEPA, the architecture LeCun has championed, a theoretical argument alongside its benchmark results. The paper is a preprint, and its arXiv page lists no journal or conference acceptance.

The idea in plain terms

The Platonic Representation Hypothesis (PRH) is the conjecture that AI models trained on different modalities, such as text and images, converge on similar internal representations of the world.

According to the abstract, the authors prove a "Platonic Representation Theorem." It says that Joint-Embedding Predictive Architectures (JEPAs) trained on any number of modalities learn the same representation up to a rotation, if their training pairs are "good views" of one scene. A good view means two observations of the same thing. For text, that could be two different captions of the same photo.

The authors then argue that next-token prediction is a JEPA trained on "bad views," meaning a prefix and its continuation from one document. In their account, that is why language models align with vision mainly by getting bigger.

The headline numbers, checked against the paper

All of these numbers come from the authors. None have been independently replicated.

Caveats worth knowing

• "42x" is about examples, not cost. The 42x figure counts training samples on one corpus. Mindverse notes that it doesn't measure compute, energy or performance on downstream tasks. The paper's own text says the gap in tokens is about 14x.
• Alignment is a specific metric. "Beats 72B models" refers to how closely text embeddings match image embeddings, not to writing quality. A text JEPA does not generate fluent text the way an LLM does. Mindverse also points to an ICML 2026 poster warning that some representation-similarity metrics can be biased by model size.
• The theorem is idealised. The full text says the proof assumes a finite world, population optima and an identity predictor. One uniqueness step for the LeJEPA training objective is taken as an assumption rather than proven.
• Paired data still appears. The captioning rotation is fitted on about 30k Flickr image-caption pairs, according to the paper. No network is trained on pairs, though.
• The vision side is weaker. The authors say the image JEPA they trained on video is much weaker than DINOv2, and that scaling good image views is left to future work.

AI Weekly's write-up reports the same core figures. It also flags that the scaling law is said to predict alignment on unseen runs within seed noise, which AI Weekly does not quantify.

Who should care

Researchers working on multimodal models and self-supervised learning should read this paper. The claim that the type of training pair matters more than raw scale can be tested, and others will try. For everyone else, it's an interesting argument from one of the field's best-known LLM sceptics. But it is one preprint with benchmark-specific results, not evidence that next-token LLMs are obsolete.

What to watch: independent reproductions, released code (the arXiv page doesn't link any yet), and whether the result survives peer review at a major venue.