“Engineering safeguards that maintain human oversight can enable the benefits of self-improving AI systems while mitigating the risks of uncontrolled self-improvement.”