The section about closed timelike curves is, I think, a good example of why computational complexity has metaphysical consequences but struggles to bridge the two - see what Aaronson writes about the "grandfather paradox".