Exactly my (and my coworkers) experience. AI generally amplifies the skillset, both in the good and the bad.

One fantastic usecase for me just recently was writing up a concept for an authentication daemon. With codex this is like a conversation where I pick from the suggestions, cross reference them with normal web-search and decide on a final draft which I then discuss with colleagues.

This "conversational" planning with integrated web-search (aka plan mode) is insanely useful. Also reviewing already written code with AI is purely beneficial in my opinion.

In my opinion the main caveat of AI is, you eventually have to be smarter than then tool. So for example if Codex suggests I should use tech-stack X then I must research and fully understand why this is actually good and still have to compare to other solutions. I think this is where the problem lies, some people skip this step which leads to so so many problems, and that's fatal. You MUST be smarter than the AI after your conversation and fully understand and be able to critique what it said.

This is my experience. I'll use LLMs as a sounding board for architectural decisions and to bring discussion points up to the team, and we talk through assumptions and pros and cons. And then once we have the architecture in place, LLMs are pretty good at implementation.

I agree with this assessment but even among us seniors with accumulated knowledge it has the dangerous potential of getting out from under your feet and produce large amounts of code that you don't have full comprehension of.

I can generally make it produce excellent well-tested code. Far better than I could do in the same time on my own. But it's a challenge to keep on top of knowledge about everything it made.