All my devs who try to use ChatGPT or similar tools end up wasting more time that those who just jump into the code and try to work everything out themselves. How long till this balance changes is not known to me.
If you can phrase all programming in the style of a Stack Overflow or Leetcode problem, AI will solve it easily. But that's what's slower than jumping in. A middle ground is writing tests.
It does really badly on Swift and is benchmarked on Python. So there's still blind spots. It can't do frameworks like Ren'Py either.
This is true in the way that Google Search is able to solve Leetcode problems.
GPT-4 struggles with even simple real world codebases.
It would be great if stackoverflow and/or similar companies do the developer survey in the industry.
I don't think LLM will replace software engineers anytime soon, as a whole, but removing the need for the bottom 10% is not a stretch at all. That has massive implications for salary, imo, except for the top end.