Platonic Representation Hypothesis is claimed to be a theorem

A post on X claims that JEPAs trained separately on any modality provably learn the same representation up to one rotation. It also claims a 109M JEPA-Text model aligns with vision models better than 72B LLMs. In a separate quote post on X, a commenter suggests the result explains why NTP aligns by treating it as multi-view learning with “bad views.”