Dispatch News of 29 October 2026 4 min
The Platonic Representation Theorem, explained three ways
A new proof says text and image models can learn one shared map of the world. The training pairs decide it.
In one sentence How you pair training data, not model size alone, decides whether two models see one world.
Imagine you and your cousin stand in the same garden. You draw a map of it. Your cousin writes a story about it. Nobody peeks at the other's work.
Later you lay the map next to the story. Every tree in the story has a spot on the map. You only turn the map a little to make them match.
Two scientists, Randall Balestriero and Yann LeCun, proved something like this about computer programs. Their paper came out on October 29, 2026.
One program learns from pictures and one learns from words. If each learns from two looks at the same thing, they end up with the same map inside. One turn makes the maps line up.
The trick is the pairs you give them. Two captions of one photo make a good pair. Two halves of one sentence make a weak pair. The weak pair teaches more about the words than about the garden.
In the tests a small program with good pairs matched pictures better than giant chat programs.
Try this at home. Ask someone to draw your room from the door while you draw it from the window. Lay the drawings side by side. Turn one until the bed and the desk line up. That turn is the rotation the scientists talk about.
Randall Balestriero and Yann LeCun posted the paper on arXiv on October 29, 2026. Both list AMI Labs. Balestriero also lists Brown University and LeCun lists New York University.
The Platonic Representation Hypothesis said models of different modalities converge as they scale. It had no proof and no recipe. This paper turns it into a theorem for JEPAs, the core of T-05.
The mechanism is short. A JEPA with whitened second-order anti-collapse learns the spectral embedding of its positive pairs. Suppose two modalities render one world and their positive pairs link the same world states. Then the two training problems are one problem in different coordinates. Their solutions match up to one rotation.
The view decides the outcome:
- Good views are two observations of one scene, such as two captions of one image.
- Bad views are corruptions of one observation. Spans of one caption keep its word choice.
- Bad views make the encoder spend directions on that nuisance. No rotation can align those with another modality.
The paper reads next-token prediction as a JEPA with bad views. Its positives are a prefix and its continuation in one document. That is why language models align with vision only through scale. This bears on T-18.
The numbers, all on alignment with frozen image encoders:
- A good-view JEPA with 21.6M samples matches next-token prediction with 900M samples. That is the 42x figure.
- A 109M-parameter text JEPA scores .184. Qwen2.5-72B scores .154.
- One fitted rotation turns separate encoders into a captioner with CIDEr .838 on COCO.
Where it breaks matters most for this atlas. The analysis uses an identity predictor. World-model JEPAs predict through a learned predictor, often conditioned on an action. The authors say their analysis does not cover that case. See T-08.
Two more caveats come from the authors. The theorems hold over an idealized finite world. At scale, two of three texts per image come from a LLaVA captioner built on CLIP vision. Without it, good views reach .060 on WIT against .107 with it.
Alignment is a geometry score. It is not task accuracy or planning quality. The paper links public code and a live demo.
What to measure next:
- Good views on action-conditioned predictors.
- Whether better alignment carries into better planning.
Balestriero and LeCun proved when text and image models learn the same picture of the world. The paper appeared on October 29, 2026. Their tests suggest well-paired data can beat sheer model size.
The property is alignment. Do a text model and an image model see the world the same way? The usual answer was to scale. Here a 109M-parameter text model beat every language model tested, up to 72B parameters.
For an enterprise the lesson is data design. A work order, its photos and a field note describe one event. Those are good views in the paper's sense.
The risks are plain:
- This is theory with lab tests. It is not a product.
- The proof does not cover world models that predict the result of an action.
- The large tests used machine-written captions. Part of the gain may come from that captioner.
The decision today is to learn, not to buy. Ask your data team whether records of the same event stay linked.
A question for the board: are we paying for model size when better-paired data could do the same job?