A text model with just 109 million parameters scored 0.184 on a cross-modal representation alignment test, beating the 72-billion-parameter Qwen2.5 (0.154) and the 70-billion-parameter Llama-3 (0.151), a parameter gap of roughly 660x. The research by Yann LeCun and Randall Balestriero, an assistant professor of computer science at Brown University, was released as a paper on October 29 titled "Proving and Improving the Platonic Representation Hypothesis with JEPAs." The paper addresses what determines whether AI observing the world through images and through text can form similar internal understandings, concluding that what a model is taught to learn during training matters more than simply making it bigger. Starting from the Platonic Representation Hypothesis (PRH) proposed at ICML 2024 by MIT and other institutions, which rested only on empirical curves and intuitive arguments without mathematical proof or an actionable training recipe, LeCun and Balestriero treat different modalities as observations of the same underlying world state. The paper distinguishes "Good Views"—different valid observations of the same scene—from "Bad Views," such as a prefix and its continuation within the same passage or fragments obtained by masking the same input. Within this framework, the team derives a mathematical result: under idealized conditions—a finite world, and each observation sufficient to determine the corresponding world state—if JEPAs of different modalities are all trained with Good