Look at VLM mechanistic interpretability papers vs just pca on JEPA trained weights.

JEPA gives you interpretability for free.

I have not personally inspected them and my view is maybe a more exaggerated/dramatic claim of those working in the JEPA sphere