The ChatGPT of 4 years from now will get its answers from the copilot-enabled editors of people who know how to do the thing. If we're lucky it'll also be paying those people royalties, but I'm not holding my breath.

How does it distinguish between people who know how to do the thing and people who use copilot, but are terrible at their jobs?

vscode is owned by microsoft and could send runtime results back.

so is github...

It gives bad advice to users, from which it learns that that advice didn't work, and only the good stuff gets used going forward (may take some time to retrain and incorporate user feedback). I expect a proliferation of advice that works well enough to convince you that your problem is solved, but which might be subtly broken in ways that you have to come back to.

Tracing the fix-it sessions backwards in time to the broke-it sessions (such that the AI can learn from its mistakes) might be tricky unless we somehow manage to connect copiloting sessions through git history. Imagine a big database of changes where the AI is waiting to see whether the code persists, or if somebody comes along and fixes it.

ChatGPT can evolve. If it doesn't know the answer you could be identified as an "expert" in the field and you could be asked for an answer. This is how ChatGPT will help connect humanity with all our experts. Basically we'll all be training the model and getting information from the model. We're going to have to get serious about thinking through oversight so bad actors can't corrupt valid information.

LLMs can navigate sources and generate both documentation and potential answers directly from sources.

Many questions on SO are exactly about things that are not covered by the docs of the tool/language/library in question.

ChatGPT can answer the homework and lazy developer questions that comprise 99% of SO, so maybe SO will just have the hard & interesting questions.

One can hope, anyways.

This is wrong especially for programming where a program can execute a program. GPT-x will learn by itself from programming. No data training will be needed.