the "reasoning" models are already optimization, not a breakthrough.
They are not reasoning in any real sense, they are writing pages and pages of text before giving you the answer. This is not super-unlike the "ever bigger training data" method, just applied to output instead of input.