Yes, absolutely. I certainly don't need an LLM to do something like that.

When I ask for the weather, I want to know exactly what the Met Office says the weather is. Not what an LLM guesses the Met Office might have said, with a non-zero chance of introducing a hallucination.

This habit of inserting LLMs into otherwise deterministic tasks is frustrating. Why take something that can be solved accurately and deterministically (like parsing the Met Office's data) and make it probabilistic, error-prone, and unpredictable with an LLM?

LLMs are appropriate for problems we cannot solve deterministically and accurately. They are not appropriate for problems we can already solve deterministically and accurately.

I'm pretty sure the idea is to use an LLM to parse the natural language into a query, not for guessing the weather.

I didn't assume either that the LLM is to guess the weather. I said that using LLM for parsing the Met Office's data is maybe not such a good idea if you can do it deterministically.

The idea was "forward $@ to a LLM to make a structured request", not to parse a structured response.

Fun fact: The fire from a strawman this size could warm a small town for many days.

Impressively snide for such a bad point.

I take the request:

> $ weather in san francisco, today evening?

To be an example of some free-form written request without any special format. Parsing that input seems like a reasonable job for an LLM, right? Otherwise we will have the typical adventure game problem of “use thumb on button” doesn’t work because it expected “finger,” or whatever.

Exactly, this is what I meant. No matter how much or for which reasons one might dislike LLMs, you can't deny that they are the best general NLP tools we have right now.

I’m quite confused as to how you could have possibly been misunderstood, and kinda wonder if it is just some folks who wanted to find an interpretation that makes you wrong.