World Models And The Convergence Claim

There are two ways to build a system that predicts.

The first is to predict the next piece of the data. Given these tokens, what is the next token. Given these pixels, what are the next pixels. The training signal is the data itself, which is why it scales so well: every document on the internet is a labelled example, for free.

The second is to predict in representation space. Encode what you have seen, encode what comes next, and train the encoder so that the first predicts the second — not the raw bytes of it, but its representation. This is what a Joint Embedding Predictive Architecture does, and the distinction is not cosmetic.

The argument for the second approach is about what the first is forced to spend capacity on. A system predicting raw output has to model everything in the signal, including everything irrelevant. Predict the next frame of a video and you must predict the exact texture of the carpet, the precise noise in the sensor, the leaf that happens to move. None of that is the structure of the world. It is the surface the world happens to present. A model that predicts representations can discard it, because the representation is learned alongside the prediction and nothing forces it to retain the carpet.

Whether that argument is correct is an open research question and this article does not settle it. What it does do is take one recent result in that programme and separate what was proved from what has been reported — because the gap between them is instructive, and the same gap will appear in the next result, and the one after that.

The theorem

The Platonic Representation Hypothesis is the idea that models trained on different modalities — images, text, audio — converge toward the same internal representation, because there is one world underneath and each modality is a view onto it. It was a hypothesis with suggestive evidence behind it.

A recent paper proves a version of it. Here is the version it proves.

Assume there is one finite world. Assume each modality is a rendering of that world. Assume every modality's encoder is trained with the same positive-pair distribution. Assume the minimizers of that objective are unique up to rotation. Then, at the population optimum, every modality's encoder equals one common world embedding, differing only by a single orthogonal map.

Read that list of assumptions again, because it is the result. This is a conditional statement about an idealised setting, and it is a real contribution — the earlier work was an empirical observation in search of a reason, and now there is a reason. But the headline version, that AI models have been shown to converge on reality, is not what was demonstrated. A theorem with five antecedents is a statement about what follows if those antecedents hold.

The part the coverage dropped

The paper extends the argument to language models and explicitly marks that extension as heuristic.

Next-token prediction can be written in the same form — the positive pair being a prefix and its continuation — but under that framing the solution is identified only up to an invertible linear map, not up to a rotation. The alignment argument depends on rotations. It does not transfer cleanly. The authors say so, in the paper, and describe the reading as their own.

So the claim that this proves large language models converge on a shared model of reality asserts precisely the thing the authors declined to assert. That is not a subtle misreading. It is the load-bearing caveat, removed.

This is worth noticing as a pattern rather than as a complaint about one paper. The caveat that gets dropped is almost never a minor one, because minor caveats are not interesting enough to drop. What gets dropped is whatever stands between the result and the headline.

The finding that is actually useful

Underneath the theorem is an empirical result that will change what some teams do on Monday, and it is much less glamorous.

Positive pairs — the two things you train the encoder to agree about — can be constructed two ways. You can take one observation and corrupt it twice: two crops of an image, two maskings of a sentence. The paper calls these bad views. Or you can take two genuinely different observations of the same underlying thing: two separate captions describing one scene, two frames from one video clip. These are good views.

Good views win, and the paper explains why rather than merely reporting it. Any capacity an encoder spends representing what is specific to one observation — this crop's framing, this sensor's noise — is capacity that no rotation can align with another modality. Two corruptions of one observation share that observation's idiosyncrasies, so the encoder is rewarded for keeping them. Two independent observations of one scene share only the scene, so the idiosyncrasies get dropped.

The reported efficiency gain is large: on one benchmark, a good-view model trained on 21.6M samples matched next-token prediction trained on 900M, which is where the forty-two-times figure comes from. Worth holding that number loosely — it is one metric on one benchmark, and in the same paper the same model loses to an existing contrastive baseline on a retrieval measure. But the direction is well-argued and the mechanism is explained.

The practical translation, for anyone training a representation of anything: two real observations of the same thing beat two distortions of one observation. If you are working with physical sensing, that means two camera angles or two sensor types capturing one moment, rather than augmentations of a single capture. It costs nothing to prefer, and now there is a reason to.

What it excludes

The predictor in the analysis is the identity. The theory covers encoders that map observations into a space where matched pairs land near each other.

It does not cover learned, action-conditioned predictors — a model that takes a state and a proposed action and predicts the resulting state. The paper says so directly.

That exclusion matters more than anything else here, because an action-conditioned predictor is exactly what a world model for a robot or an agent is. The entire point of such a model is to answer "if I do this, what happens." A convergence result about perceptual encoders says nothing about whether that prediction will be any good, or whether two systems learning it will arrive anywhere near each other.

So the result sits inside the world-model research programme, and it is evidence that the representation-space approach has structure worth having. It is not evidence that world models work, and anybody citing it as support for an agentic or robotic architecture is citing a paper that explicitly placed their case outside its scope.

How to read the next one

The specific result will be superseded. The reading method will not. Four questions, in order:

What are the antecedents? Every theorem has them. Find the assumption list before you find the conclusion, because the conclusion is only as strong as the weakest item on it. "Unique up to rotation" and "population optimum" are doing enormous work in the result above.

Which claim is proved and which is argued? Good papers mark the difference and the marking is usually one sentence deep in a section nobody quotes. Here it was the word heuristic attached to the language-model extension.

What is explicitly excluded? Limitations sections are the most honest part of a paper and the least read part. The exclusion of action-conditioned predictors is in plain language and almost none of the coverage mentions it.

What is the empirical claim's scope? One benchmark is one benchmark. A number that holds on a single metric, against a single baseline, with a single encoder, is a finding and not a law — especially when the same table shows a loss elsewhere.

None of this requires reading the mathematics. It requires reading the conditions, which are in prose, and taking seriously that the authors wrote them down for a reason. A result reported without its conditions has not been simplified. It has been changed into a different claim, and usually a false one.

The discipline here is the same one that applies to a model's output, a vendor's benchmark or your own system's metrics: a number without the conditions that produced it is an assertion. The conditions are the result.