JEPA (Joint Embedding Predictive Architecture) is one of the most convoluted research topics that is difficult to follow because it involves predicting high-dimensional representations in a learned latent space rather than tokens or noisy pixels.