Kernel is far too big to throw at a single LLM, just like a human. It is.. I think somebody said 40mil lines of code.

So you have to chunk it down to workable pieces, and then if you collect enough data of that process you can look at higher level stats.

Same for every other existing piece of software, btw.