What we actually built
The technology called artificial intelligence is not intelligent, and the confusion starts there. What has been proven repeatedly is something narrower and more interesting: take a skill we assumed needed human cognition, define a small enough subset of it precisely enough, and it turns out you can do that subset with mathematics.
Deep Blue at chess, Watson at Jeopardy, AlphaGo at Go. Each one took a human skill, bounded it hard, and solved the bounded version. None of them is adaptable. Watson could read unstructured language and it still could not be asked to do anything except play Jeopardy.
Large language models widened the bound enormously. They did not remove it.
The test: can it write his email?
Start with the part of a job nobody enjoys. Given a real incoming email and a one-line instruction, a model returns a competent draft immediately.
It is also obviously not you. Continue a thread you have been writing for a year and the voice shifts, and everyone on it can tell. So the draft needs editing, and now you are editing instead of writing, which is not obviously a saving.
The fix is not a better prompt. He indexed every email he had written, extracted the pairs where a message came in and he replied, embedded them, and at generation time retrieved the five most similar past exchanges and asked for the reply in the voice of those. The output moved from generic to close.
Read that again as a description of work. Scraping an archive, structuring pairs, choosing an embedding, storing vectors, implementing retrieval, assembling the prompt. The model wrote the sentences. Everything that made the sentences worth having was built.
Where it breaks in real applications
Demos work. Applications built on the same thing break quickly enough that it is sometimes faster not to use them at all, or to drop the wrapper and talk to the model directly.
That is not a bug waiting to be patched. It follows from what the thing is: it does not know what it does not know, and it has no way to tell a correct answer from a plausible one. Every production use is therefore a system for constraining it, checking it, and handling the times it is wrong.
So what actually happens to the job
The part of engineering that was typing is worth less. The part that was deciding what to build, bounding the problem, judging whether the output is right, and handling failure is worth more, because there is now more output to judge and more ways to fail.
A confident claim that all junior developers are finished is a claim about typing. The work was never mainly typing.