A paper by Randall Balestriero and Yann LeCun shows that, under specified assumptions, JEPAs trained separately on modalities with positive pairs linking the same world states learn representations differing only by rotation. In experiments, pairing different descriptions of a shared scene achieved similar alignment with about 42 times fewer samples than next-token prediction. A 109-million-parameter model also scored above tested language models up to 72 billion parameters on alignment. The public repository provides code and reproduction instructions. The theorem depends on conditions concerning observations and optima, and the benchmarks do not demonstrate superior general language ability or convergence of all models.