Oh, I have nothing against the fact you want to share feedback or test some new approaches.
I just wanted to see what I do from a different perspective (sorry of reality check for me).
Personally, I build (try to build) tools that are aiming practical use and value creation. Maybe it’s my modular thinking or some of the experience, but I try to get tools with “connectors” on edges and an ability to hook them together or replace one another within the system. With this approach, for me it is easier to handle logic in code/told workflow and use AI for text related tasks.
Seeing the limits of the AI (transformer models in this case), I cannot afford running them for tasks that are either complex (as in multi-step) or precise (like in calculations) in the app workflows because the error rate is over the one I can accept.
So for me, AI stays perfect tool to deal with semantics, and code is to deal with the logic. So far, proves to be a viable approach for me.
And sure, I constantly run “tests” - sort of stress use of AI for beyond the edge functionality just to see where the new limits are and where it breaks…