[jononor]

The AI does not loose its utility once it has replaced SO. Further improvements would need to come from other sources than more "free" text data though. There are several avenues for this: A) better learning efficiency - it is known that models are not samples efficient now B) pay people to create more data at large scale - already happening at the big LLM companies (though maybe less do for code right now) C) learn from the prompting interactions that customers - extremely desirable, as people would effectively pay you to improve your model. Companies are gonna keep that close to their chest D) multi-modal learning and transfer between modalities E) Self-learning, where ML model iteratively improved itself. Analogous to AlphaGo, GANs etc. This is the holy grail, but also possibly the hardest. Code has the benefit of being formal languages with available verifiers, which should be useful here. F) Specific go code: Using functional feedback, not just textual. For example the large amount of unit tests available on GitHub

So I believe there are many avenues for further improvement. I still think it will be very hard though, and until we hit the next breakthrough, will be very resource intensive and possibly quite slow going.