I think the disillusion here is when you find out that these things are mostly smoke and mirrors that tend to generate very suboptimal solutions. For example I am forever having to shoot down ChatGPT generated code before it hits production which is naively and lazily generated and the engineer is not aware of this at all. Yes it can write SQL but does it understand the index strategy on the destination table or the consequences of running thousands of those queries a minute in a production database? No. Does it have a deep understanding of context rather than just syntax (ask a French language translator what they think of it). No. And with trip planning and ideas, which is a speciality of mine, it is always better to find someone local and ask them about it because a huge portion of the information out there isn't something a model can infer about because it isn't written down anywhere. Case in point, I'm off to the Azores islands shortly - it can't generate anything useful for me. I had someone suggest that already. On top of that it is compelled to over explain sometimes to the point of incorrectness, which it likes to do on travel advice. This becomes dangerous. I found someone local to talk to instead.

So personally I think the LLM and question based uses are limited and quite ridiculous from an objectivity perspective. They barely scratch the surface of the problem and the hit rate is less than I'd get doing the legwork myself, which I do not find useful. And the information is hard to verify and judge if it's safe without deeper contextual knowledge. This is a complete mire.

The other use cases, like classification and search are where I see it winning but those aren't sexy and don't play on our human emotions as well enough to part with cash. They will be left after the hype of course.

Really I think people are buying into a romantic fantasy at the moment.