What I wonder then is...is it going to be the final word on the evolution of software development? AI has been trained on practices that were developed by people, by professionals and amateurs, by discussions on discussion boards, at tech events, through written books and then disseminated across open source projects. If no one writes code anymore then how to write and build software projects will never evolve beyond the 2022 state of affairs.

That data goes into pretraining the initial model. The agentic post-training relies on the model making stochastic exploration. So in a sense it learns via trial and error. By the accounts I have read, this second stage is the dominant reason we have seen the increase in model capabilities in the past year. It is still not clear how innovative the models can be, but they are not just relying on static data.