Sure, it would be nice if we had an LLM that could help us with the more unusual questions. But for that we would either need much better training data, full of all those unknown answers (which does not exist), or a totally different system that can actually think for itself instead of just mimicking existing human output.
LLM's are excellent at stuff that has been done before and they are also fantastic for transposing code from one language to the other.
Looking up code in old projects takes time and brain RAM. Spelling out the steps in pseudo code and then prompt guiding an LLM through figuring out the required code often works great for smaller tasks.
I've also learned many new tricks by just asking it why it made certain choices.