4. Somebody somewhere will still find a legitimate use case to code in assembly and blow everyone's mind.

5. AI governance is likely needed.

What makes you think you need people for that other than the ones in category 1 and 2?

Building an AI is just like building a website. It can mostly be done by an AI as well.

I think that assumes AGI which I don't believe is possible, or anytime soon.

You don't need AGI for that, just compute. As soon as you formulate the requirements to the model (objective function) and the dataset, methods can be iterated and found automatically, in principle.

If it can program a desktop application it can program and evolve tensorflow/pytorch scripts.

This assumes people will still be willing to make their code public so that AI can train on it.

With the amount of people using GitHub and Copilot (including me) even with all the controversy surrounding it, I'm sure the assumption will hold true. Most people simply don't care enough about their code to make it stop being trained on a neural net, and if they get benefits from it in the form of stuff like ChatGPT or Copilot, then they're even more so not gonna care.