Turing Award winner Yann LeCun and Brown University Assistant Professor Randall Balestriero have co-authored a new paper that builds a mathematical framework on the Platonic Representation Hypothesis, demonstrating that JEPA models of different modalities—when trained according to specific positive-sample relationships—can learn world representations that align with one another through rotational transformations. Experiments show that a text JEPA with only 109 million parameters surpasses large models such as Qwen2.5-72B and Llama-3-70B in cross-modal representation alignment tests. Moreover, with just 21.6 million training samples, it achieves the representation alignment level of a next-token prediction model trained on 900 million samples, greatly improving training efficiency. It can also connect an independently trained visual encoder with a text model through a linear transformation, enabling related cross-modal applications. This research continues LeCun's long-standing JEPA world model line, challenges the AI industry's scale-only mentality, and points out that the informational relationships the model is allowed to learn during training matter more than simply increasing the amount of data. After the paper was released, it sparked varied discussions in the industry, and the outside world holds differing expectations regarding the practical value of this technical route and the possibilities for multimodal expansion.