At the risk of misrepresenting the book, since I don't have it in front of me, here's what bothered me most: stating early that AI is basically an effort to approximate an optimal Bayesian agent, then much later showing that a Bayesian approach permits AI to scope-creep any human request into a mandate to run amok and convert the visible universe into computronium. That doesn't demonstrate that I should be scared of AI running amok. It demonstrates that the first assumption -- we should Bayes all the things! -- is a bad one.
If that's all I was supposed to learn from all the running-amok examples, who's the warning aimed at? AFAICT the leading academic and industry research in AI/ML isn't pursuing the open-ended-Bayesian approach in the first place, largely isn't pursuing "strong" AI at all. Non-experts are, for other reasons, also in no danger of accidentally making AI that takes over the world.
The book comparison was probably a cheap shot (on the other side of it, Einstein didn't need popular interest/approval for his ideas to matter; I think that is a positive).
I think as much as anything the comparison is worthless because we can look backwards at Einstein.
It is part of the foundation of quantum mechanics.
Superintelligence: Paths, Dangers, Strategies is in the range of 100,000 words (348 pages * roughly 300 words per page).
I'm not familiar with it, but looking around it isn't even clear if it even lays out any sort of concrete theory.