It's totally valid if the models want to pack words tight during their thinking process, as long as the final conclusion (which is the interface to user) is written in HUMAN LANGUAGE, then I don't care whether the model thinks in alien language