If you think it'll slow down, you can do some of the same stuff you're doing now for lower pay while supervising an AI, maybe?
Current models are already very very capable. If it becomes cheap and very fast, i think it is game over.
[1] https://www.slatestarcodexabridged.com/Meditations-On-Moloch
I talked about this at some length here, including a diagnosis of the structural issue we're facing as well as a path forward: https://news.ycombinator.com/item?id=49461333