[precompute]

Okay, first, I haven't read that book. I'm going off the cherrypicked examples in the OP.

If a LLM was indeed used, the output was likely massaged to a degree that wouldn't be immediately obvious.

Now, my 2c: Writing is sometimes atrocious and sometimes authors jam in stupid things to maintain flow. If a person could make the connection between "Silk", "Spider", "weaving" and "grandmother" then another person could as well (even when one is "verifying" and another "proving"). And using those properly and in context and succinctly is far beyond most LLMs, and would require a fair amount of gambling, which would be out of character for someone whose writing prowess has been verified pre-LLMs.

As for what I mean by "directly reference the previous thing": LLMs can jam in well-known (to them) phrases and sentences and structures onto an idea/request. However, they are unable to loop upon the particulars of that idea/request in a coherent manner, which leads to slopification at large output sizes, and shows us the ceiling of the quality of writing a LLM can output.