I’m analyzing a bunch of data this week. Usually that would mean “writing a lot of SQL,” and whilst there is a lot of SQL, I’m barely writing any of it, because this or that LLM is doing it for me. In the past, doing this sort of thing included two important steps: think of the question you want to answer, then translate it into the right SQL syntax. These days, the first step is the same, but the second one absolutely isn’t.

The good side of that, of course, is that you can spend more time on the first part, which was always the most important part. Maybe there’s a bad side? It’s interesting to move from a strict syntax that describes something logical to a floopy syntax where almost anything goes. I think there’s something lost in not “thinking in logic” but I’m not sure if it’s enough that the speed doesn’t make up for it ten times over.

There are moments where I caught myself assuming something was correct based on a cursory glance, only to realize my request had been misunderstood in a way that was important to the analysis. I would say that I’m one of the more “check everything that an LLM gives you very closely” kind of people, and yet here I was. I suspect that the models will eventually get good enough that I’ll just stop thinking about it.