beren comments on Evidence on recursive self-improvement from current ML