Training JEPA architectures faces the 'collapse' problem: if you simply minimize prediction error in representation space, the system can cheat by predicting constant representations, making the prediction trivial and useless.