So speaking of "mastery":
I wanted to know how to clone a single folder in a Git repository. Having done this before, I knew that there was some incantation I needed to make to the git CLI to do it, but I couldn't remember what it was.
I'm very anti-AI for a number of reasons, but I've been trying to use it here and there to give it the benefit of the doubt and avoid becoming a _complete_ dinosaur. (I was very anti-vim ages ago when I learned emacs; I spent two weeks with vim and never looked back. I apply this philosophy to almost everything as a result.)
I asked Qwen3-235B (reasoning) via Kagi Assistant how I could do this. It gave me a long block of text back that told me to do the thing I didn't want it to do: mkdir a directory, clone into it, move the directory I wanted into the root of the directory, delete everything else.
When I asked it if it was possible to do this without creating the directory, it, incorrectly, told me that it was not. It used RAG-retrieved content in its chain of thought, for what that's worth.
It took me only 30 seconds or so to find the answer I wanted on StackOverflow. It was the second most popular answer in the thread. (git clone --filter=tree: --depth=0, then git sparse-checkout set --no-cone $FOLDER, found here: https://stackoverflow.com/a/52269934)
I nudged the Assistant a smidge more by asking it if there was a subcommand I could use instead. It, then, suggested "sparse-checkout init", which, according to the man page for this subcommand, is deprecated in favor of "set". (I went to the man page to understand what the "cone" method was and stumbled on that tidbit.)
THIS is the thing that disappoints me so much about LLMs being heralded as the next generation of search. Search engines give you many, many sources to guide you to the correct answer if you're willing to do the work. LLM services tell you what the "answer" is, even if it's wrong. You get potential misinformation back while also turning your brain off and learning less; a classic lose-lose.