I'm pretty sure today's models probably can be capable of self-improving. It's just that they are not yet as good as self-improving as the combinations of programmers improving them with the help of the models.