Maybe the statistical approximate just coincidentally lined up with the actual solution?

Yes, right, and this can happen. I didn't say they are a bad approximation. LLMs may be the most advanced and sophisticated statistical models ever created (if there are other examples of statistical models that are more sophisticated, I would love to learn about them). But given what an LLM actually objectively is, a statistical model of the next token to follow from a sample of language, what other explanation could there be?

We need to keep in mind what the tools that we are working with actually are.

As a code generator, they can produce great results, especially simple stuff like generating a script or some function implementation. Once you get to software engineering tasks like designing system architecture and designing maintainable code, it starts to fall apart really fast. You end up doing all of the work for it in natural language and just using the LLM for a usecase that it is actually great for: translation, from detailed spec to code.

This is almost the perfect task for them. I think if them as pattern matches. They have patterns for libraries in their training, and you gave it a technical spec, and it pattern matched it across to a library.

On top of that, you can verify it and reuse it. But regenerating it every time wouldn't be a good use both for the cost, and the risk of subtle issues that don't get noticed. Same argument as for any library.

Probably because I've been doing this so long, I often find it easier and more precise to describe things in code or pseudo code than common English, which often aides my use of LLMs.

On another note, I'm curious what you're making. My "when I get time" project is to use an old STM32 drone flight controller to do some basic robotics, which will only be possible because I think the LLM will pattern match me out of trouble getting an embedded C program compiling after 20 years out of that game.