LeCun lays out numbers supporting the AMI Labs bet — but the measurements are the researchers' own

LeCun lays out numbers supporting the AMI Labs bet — but the measurements are the researchers' own

A new paper by Yann LeCun and Randall Balestriero claims that a text model of just 109 million parameters produces representations that align better with image embeddings than far larger LLMs. But the result rests on the authors' own measurement method, and no independent group has repeated the comparison — while one of the authors has just founded a company built on exactly this architecture.

What's new

On October 29, the paper "Proving and Improving the Platonic Representation Hypothesis with JEPAs" was posted to arXiv as 2610.39302. It is written by Randall Balestriero, an assistant professor at Brown, and Yann LeCun, and it makes a claim concrete enough to be checked: a text model with 109 million parameters, trained as a joint-embedding predictive architecture (JEPA), produces representations that align better with the image embeddings of DINOv2 and CLIP than any language model the authors tested — including Qwen2.5-72B, a model with 72 billion parameters.

The number carrying the claim: measured against Meta's DINOv2 ViT-B, the small model scored 0.184 on the authors' alignment metric, versus 0.154 for Qwen2.5-72B. All figures in this story come from the paper as reported by Startup Fortune (32aa9055); this coverage is based on secondary reporting, not on the paper itself.

Why it matters — and for whom

The timing is more than coincidentally interesting. LeCun left Meta after twelve years as chief AI scientist to found AMI Labs, a Paris-based company built around world models and the V-JEPA architecture rather than large language models, according to Reuters and other outlets that covered the departure, as reproduced by Startup Fortune. The company is reported to have raised over one billion dollars before any product exists, with Alexandre LeBrun as CEO and LeCun as executive chairman.

A new academic paper arguing that JEPA architectures converge toward shared representations across modalities — and that this happens more data-efficiently than next-token prediction — thus constitutes direct technical support for a heavily funded industrial bet by the same author. That makes the result more significant, and it also makes it necessary to read it with extra restraint. What is pointed out here is only the coincidence of interests, not any wrongdoing by any of the parties.

The argument: good and bad views

The paper grounds its claim in what is called the platonic representation hypothesis — the idea that good models, regardless of modality, converge toward similar representations of the same underlying reality. Balestriero and LeCun argue, as relayed through Startup Fortune, that convergence has very little to do with scale and almost everything to do with what the model is shown during training.

The key is the distinction between "good" and "bad" views. Next-token prediction, the mechanism behind all common large language models, does not qualify: predicting the next word from the words before is, according to the paper, a JEPA trained on bad views — two slices of the same document, not two observations of the same reality.

From this distinction the authors derive two things. First, a theoretical conclusion: training on good views, according to the paper, provably converges toward shared representations across modalities. Second, an empirical figure: good views are claimed to require 42 times fewer training samples than next-token prediction to reach equivalent cross-modal alignment. Both are presented here as the authors' claims — not as established facts.

The caveats are substantial

It is worth saying plainly, as the reporting itself does: one paper is not a verdict. Three factors limit what can be concluded.

First, the empirical basis rests on the authors' own choice of alignment metric and their own benchmark setup. A different metric or a different lineup of models could give a different picture.

Second, there is no independent replication yet. No group outside the authors has repeated the comparison between the 109M model and the 72B models.

Third, the paper's details, including the proof sketch for good views and the reported figures, are known here only through secondary reporting from Startup Fortune. This is not an argument against the result, but it sets the limit on what is known at this point.

This also means it would be wrong to read the paper as a verdict on scale. The difference between 0.184 and 0.154 on one author-defined metric, without replication, is a hypothesis worth testing — not proof that 72 billion parameters are superfluous.

What to watch

Two things will determine whether the result holds. The first is independent replication: whether other labs repeat the alignment comparison with their own measurement methods and their own model pairs, and find the same gap between small JEPA models and large LLMs. The second is whether the claim of 42-times data efficiency survives outside the authors' setup — if it does, that would be the most actionable part of the paper, since it would then specify how expensive it actually is to train models that learn shared representations.

Until then, the honest summary is this: two prominent researchers, one of whom now leads a company built on exactly this architecture, have presented a theoretical argument and an unproven figure, both pointing in the same direction. The question for the coming months is whether anyone else can independently find their way to the same result.

Sources: Startup Fortune (32aa9055).