Ha ha. I suppose an inverted j curve technology would offer short term productivity but yield long term slow-downs. I can see aspects of that - perhaps using agents people write code that quickly adds the next feature but overall is not maintainable. Or they quickly write a project plan but in the long term it doesn’t result in good payoffs. I’ve already observed both of these.
I think the hardest part to figure out will be delineating the illusion of productivity from actual productivity.
In the other hand - we do have empirical research (from the economist in this article) even from 2023 showing LLMs ability to offer real productivity in certain tasks : https://www.nber.org/papers/w31161. And the models have become much more better since then.
So it’s probably a question of if we can use them for what they are good for without being lulled by the sirens song of fake productivity.