Found first: a primary source the press has not covered yet.
Randall Balestriero and Yann LeCun have converted the Platonic Representation Hypothesis from open conjecture into a formal theorem, proving that Joint-Embedding Predictive Architectures (JEPAs) trained independently on any number of modalities converge to the same representation, up to a rotation, when trained with "good views." The paper, submitted October 29, 2026, is available on arxiv. It also identifies why next-token prediction requires massive scale to achieve cross-modal alignment and points to view design as the decisive variable.
What the source says
The authors define "good views" as two observations of the same scene, and "bad views" as parts of a single document. Their Platonic Representation Theorem proves that JEPAs converge to identical representations across modalities under good views. Next-token prediction, they show, is a JEPA operating on bad views, which accounts for the alignment gap between language models and vision systems. A 109M-parameter text JEPA outperforms language models up to 72B parameters on cross-modal alignment tasks; on zero-shot ImageNet-100, the text JEPA achieves 64% top-5 accuracy against at most 56% for language models up to 8B parameters. Validation spans datasets from 145,000 captions to 940 million web images, and good views require 42 times fewer training samples than next-token prediction to reach equivalent alignment.
Why it matters
The Platonic Representation Hypothesis previously rested on empirical observation: neural networks trained on different modalities appeared to converge toward a shared representation of reality. This paper replaces that observation with a proof and names the condition that must hold. For practitioners, the implication is that positive-view design is the lever for cross-modal alignment, not data volume. The authors also derive a scaling law that predicts alignment on unseen runs within seed noise, giving the theorem quantitative predictive power.