I completely disagree that TAOCP is a "reference work" or that invariably there are "by now better places to start". If you care to learn about any of the topic

I completely disagree that TAOCP is a "reference work" or that invariably there are "by now better places to start". If you care to learn about any of the topics in TAOCP, to the level of detail it contains, it is usually one of the best places to read. Knuth has taken the entire mass of published literature on the topic, passed it through his "interestingness" filter, and boiled it down to usually the best writing on the topic. For example, suppose you want to learn about Binary Decision Diagrams. I have before me TAOCP Volume 4A, which has 57 pages of text, 22 pages of exercises, 58 pages of answers, and a wealth of references. (Section 7.1.4, you can see the draft Fascicle 1B online: http://www.cs.utsa.edu/~wagner/knuth/ http://www.cs.utsa.edu/~wagner/knuth/) There are simply no other books or papers where you can learn so much about BDDs and ZDDs this well or easily. The writing is crisp and clear, and is really intended to teach, not to be a reference to look up. Of course, a lot of people have simply no need to learn about zero-suppressed binary decision diagrams, so they can just simply not read TAOCP. But that's a different claim from saying that it's a reference work or not a good read on its topics.