AI Policy Network's @peterwildeford warns why the tail risks of AI alignment matter even when models are aligned most of the time: "The situation is not so completely bleak and hopeless. It does seem that AIs are quite aligned for the most part. It is nonetheless these edge cases that I am worried about because it does seem like when things do go off the rails, they can go really off the rails." "These AIs are already getting out of containers, not just because humans left the containers…