it's not really a matter of what the LLM is capable of generating, it is whether I can understand what it is doing. That's why I feel better limiting it's output to just one line at a time. I want to know what line of code should come next before I let the LLM do it for me. Once I know what I want to write, I hit "tab" and see if the LLM did it right, or did it better, but I can only know if I thought about it myself first.
I have heard the comparison to compilers before, but I don't really agree. Compilers aren't just for translating from one language (e.g. natural language) to another. Compilers are about letting humans express ideas using a programming language. It is a way to shape how programmers express ideas logically. It is a code generator for a calculus. So the **un-**natural language of a compiler is a feature of what compilers are used for, not a bug, or a limitation, or a trade-off between a "steep learning curve" and generating efficient code.
The idea of a programming language is to come up with grammars and semantics that express algorithms. That necessarily requires making the expression of some bad ideas impossible. This isn't a barrier to productivity, it is a way to force you to rethink your idea in a way that makes more sense.
LLMs don't do this, they just generate code to satisfy your demands, even if your ideas are horrible.
