It is fairly clear to me that self-learning is already possible. Give an agentic LLM the ability to add things to its own context window, or fine tune itself, and it will.