AI won't take your job (yet)

There's not a week when someone claims "AI will take programmers jobs". I dare to disagree, on a few levels.

There's a space for AI software builders - if you need an internal tool, something to help you with work, to write some algorithmic pieces of code, and to improve speed and performance overall. It does do that on single files or small pieces of code, but not for entire projects.
It's not even close when it comes to "AI software engineering agents". I'll show you why.

I have a Ruby on Rails project from 2017 that I want to restart. There are a few steps to do, as I would like to migrate it to the newest ruby version, the newest rails and update dependencies.
So I tried Claude Code, OpenAI Codex, Gemini (app), ChatGPT with 4.5 and Claude Sonnet 4.5 and Sonnet 4.7

All of them failed miserably.
The first issue? The newest Rails release is 8.0.2 - all models are choosing up to 7.1.2.
Same with Ruby.
Same with puma.
It also got stuck on trying to make changes in one file over and over again. (circleci configuration)
I'm not even talking about changes in the configuration, adjusting code to breaking changes, fixing tailwind versions, etc.
This is something that a mid-level engineer with a few years of experience could fix in a few days.
And this is something that AI is not capable of doing right now.
So until it's getting up-to-date with information about libraries, knowledge about legacy systems, it's fine for vibe-coding, some greenfield hobby projects, and building PoCs - which is not bad, if you need to validate your idea quickly, but it's absolutely not ready for production yet.

At the same time, I see a trend not to hire junior developers because "we can do it with mid/senior engineers and AI". But what will happen when those seniors are gone? You know we need to help junior developers grow to get a new pool of seniors, right?

I won't say that it never will happen that AI will get to the point where it's senior-level engineer as I've seen tremendous advancements in the last three years, but from my experience every new model is trying to outsmart all of the previous ones, and the best one for 99% of my cases is still Sonnet 3.5 from Anthropic.
If I want to solve simple issue I want simple solution, not overengineered solution that's way too complex to solve the same thing I would do with good old stackoverflow 10 years ago.