I think what you've identified is the "certain kinds of learning" I mention in the post.

If I want to learn about the structure of ELF binaries, I don't figure it out from first principles. I can't: the decisions that went into its design where highly specific to its history and the sensibilities of its creators, and in a modern context a lot of those decisions are near enough random. These are the sorts of things you should use specs, docs, Google, and maybe an LLM for.

What I'm talking about is, I guess, 'constructive' learning. I might not be able to build my own ELF parser through guesswork, but I probably can write a binary format of my own and a simple OS kernel to read it, given enough time. It probably won't look much like ELF, but the process of discovery that'll happen will teach me about why ELF came to be and what problems is solving, and also about the wider design space that sits around it.

My suggestion is that you should not use an LLM's help for the latter.

I like to think of it this way: if you bolted me to a chair in a room with nothing but a laptop, a text editor, and a compiler, what might I be able to create if you left me for 10 years? What couldn't I create because I wouldn't have sufficient implementation-specific information?