Yann LeCun and Randall Balestriero present theory and experiments on representation alignment when JEPA learns from suitably related views of the same underlying subject. In one comparison, 21.6 million JEPA training samples reached the alignment level achieved by next-token prediction after 900 million samples. The roughly 42-fold difference concerns sample efficiency for that metric, not overall training cost or general reasoning superiority. Synthetic image captions and imperfect fulfillment of theoretical assumptions remain important limitations.