We're on the cusp of models that post-train themselves to get better. It's not speculative anymore. The self-improvement loop is actually closing..