“Joint Embedding Predictive Architectures (JEPAs) learn by predicting an abstract representation of an input rather than reconstructing all of its pixels.”