“Techniques like sigreg or vicreg are necessary to prevent modal collapse when training JEPA-style models where the loss function minimizes the distance between predicted and actual latent states.”