> history has, so far, mostly proven that the end result of better tools is better experts

I’m inclined to agree with the general sentiment. However, it is not a given that LLMs are better tools. You don’t really get better at them in the same sense as before, you just type something and pray. The exact same thing you typed may produce exactly what you wanted, and aberration, or something close with subtle mistakes that you don’t have the expertise to fix. Other tools made you better at the craft in general.

> You don’t really get better at them in the same sense as before, you just type something and pray.

I very much disagree with this. I've spent the last 18 months working with LLMs daily at my work (I'm not exaggerating here) and while the models themselves have certainly gotten better, I have personally learned how to extract better results from the tools and improved my proficiency at applying LLMs as part of my overall toolchain and workflow.

LLMs are very much like many other tools in some ways, that the more you learn how to use them, the better results you will get.

I do agree that their non-deterministic nature makes them less reliable in some contexts as well though, but that's a trade-off you work into your approach, just like other trade-offs.