Like almost all machine learning stuff, I expect these things to only be useful in places where it doesn't really matter if the results are correct. When you apply a classifier in real life its critically important to understand its statistical characteristics which is typically done via model characterization, which involves cross validation or boot strapping or whatever. I think the idea that you can just zero-shot or few-shot deploy these things as classifiers and forget about it is incredibly naive unless (as I said earlier) the results just don't really matter.
I've already used LLMs in my work as a data scientist but it requires a ton of work to just make the results tractable (and I have been using GPT4, which behaves pretty well). These smaller language models ain't so regular. Ok, like consider a basic thing you want to do with a classifier: understand its behavior on a held out data set. Since no one knows what is really in the training data (since its so large), its quite hard to understand what the model can generalize about and what it has just accidentally memorized. CF reports that GPT4 doesn't perform nearly as well on even simple programming exercises that are chosen in such a way as to be sure they weren't in the training data.
There is enormous potential for statistical fuck ups here. Prompt engineering, for instance, is an easy place for over-fitting to happen as a prompt is fine tuned on data the prompt engineer has and thus fails to generalize to new data.
I do think there is a lot of value here, but I'm also sure that sloppy use of large language models is going to cause a bunch of trouble in the short to medium term, generate a lot of garbage, pollute a lot of databases, etc, while we figure all this stuff out.
