“AI systems can be used for self-improvement with human oversight and engineering safeguards without allowing for fully autonomous self-improvement.”