There is zero evidence that the current approach will ever lead to a self-improving model, or that current GPU/TPU infrastructure is even capable of running self-improving models.