alt Hacker NewsLLM's went from generating incoherent garbage, to weird text that seems coherent but isn't (gpt-2), to what we have no with agents passing some of the hardest benchmarks, necessitating benchmark development just to be able to keep up with the rate of progress. They did this in about 5 years.
The sheer rate of progress is what is meaningful. If LLM's agents were frozen now without any further improvement, your position would be must more defensible. But that's not what's happening, we are getting breakthroughs almost on a weekly basis.
It feels to me like the advancements are not accelerating, though, like we're nearer the top of the S-curve than enthusiasts think we are. A lot of the recent advances seem as much a harness improvement as an actual model improvement. We've made exceptional progress in the last couple years, but spent an exceptional amount of money to do it. A moon shot, basically.
Honestly, my gut feeling (worth every bit of what you are paying for it...) is that until we get actual intelligence instead of a good facsimile, we're not looking at an existential problem or a UBI-level problem, even, just a big change in tooling.