I'm not sure this is a useful way to approach "magic". I don't think I can build a production compiler or linker. It's fair to say that I don't fully understand them either. Yet, I don't need a "full" understanding to do useful things with them and contribute back upstream.

LLMs are vastly more complicated and unlike compilers we didn't get a long, slow ramp-up in complexity, but it seems possible we'll eventually develop better intuition and rules of thumb to separate appropriate usage from inappropriate.

if only the opposite were true!