>But then, looking forward years in time, people start leaning more heavily on that and stop posting to Stack Overflow and the well of information for AI to train on starts to dry up

for coding tasks I'd imagine it could be trained on the actual source code of the libraries or languages and determine proper answers for most questions. AI companies have seen success using "synthetic" data, but who knows how much it can scale and improve