with data privacy decisions I like to think about how the policy/reasoning would sound in a NYT headline or congress testimony. your stance does not imbue trust for me.
0 employment contracts or user facing ToS I've read make a distinction between sharing schema vs data with unvetted third parties. In my opinion, schemas reveal pretty valuable info about an application (and they're absolutely valuable in aggregate because you can use them to train AI data scientists).
Maybe I sound like a curmudgeon but since I have the option of running the AI locally with a 5 percentage point accuracy loss, I absolutely will. If GPT-4 was 100% that would be different because you could build totally different things, but 83% has most of the same design problems as 78%.
These small SQL LLMs are indeed worthless, since LLM performance on SQL queries is quite limited right now, every % of accuracy matters.