I am not so sure that we disagree actually.

Jobs that LLMs manage to replace - within the engineering ecosystem at least, which is a HUGE caveat I need to make there (as it differentiates between e.g. that vs. graphic design artists) - are jobs that did not need to exist in the first place. Although I am not certain that I could name any particular role there - maybe Project Manager could be the closest, but even there they become super useful for large & complicated setups yet don't many companies seem to use them regardless of the need?

If you can remove an entire position and replace their tasks with a glorified autocorrect, then I would definitely wonder how important those tasks were to begin with!?

And if they are important, then I would not rely upon an LLM to perform them correctly.

Though my worry is less than LLMs will perform so well as to replace humans, and more that the humans will be fired and replaced with LLMs anyway, which will then do such a fantastically shitty job of replacing the role formerly performed by a human being that the load upon all the humans will be increased. i.e. not computers aiding humans to accomplish more effectively, but humans propping up computers that are constantly misbehaving simply for the reason that it makes the C-suites "feel" more in charge to have such surveillance systems working for them.

Which is a simple extension of the current model where, when misapplied, takes as much if not more time to fuss over Jira tickets than to actually accomplish any real work. Do you Stall the ticket or close it and create a follow-up, thereby breaking the cohesion of the comments, despite how Jira is nearly impossible to find things in once a Sprint is over anyway - no matter what decision is made for you by the ones in charge, somehow it always manages to be your fault for not doing it right. Except now substitute the person doing that with an LLM, with even less capacity (as you said, zero actually) to understand why it is even doing what ends up having been done.