On September 20, Yann LeCun reiterated his view that autoregressive LLMs alone will not lead to human-level AI. He advocated reasoning in continuous representation spaces rather than token space, and self-supervised learning with Joint Embedding Predictive Architecture (JEPA). Jürgen Schmidhuber replied by challenging JEPA's originality, claiming its family of techniques is actually his team's 1992 Predictability Maximization (PMAX) work. He linked his March 31 post, which describes PMAX as using two networks to learn informative internal representations that are predictable from related inputs.