Metamagical Themas Review

Metamagical Themas: Questing for the Essence of Mind and Pattern, Hofstadter. Non-fiction.


Hofstadter, of Gödel, Escher, Bach fame, here collects his Scientific American columns, along with other writings. He explores the nature of intelligence, the meaning of “I” (the pronoun), what it means to recognize “i” (the letter), and generally the nature of cognition. He is (or was, as of 1985, when this was published) an unrepentant believer in strong AI, and is sometimes harshly critical of the direction of AI research. I share many of his objections to the AI establishment that gave us "Expert Systems" (really "novice systems"), and am happy to have such a cogent voice articulating what is right and wrong with how we think about thinking.
A recurring theme in the work is the madness represented by the nuclear arms race, and I can’t help but wonder whether he continues to see nuclear conflict as the most pressing danger that we’re ignoring, or if (as I’d like to imagine) he’s more troubled by the attack on our civil rights that began in earnest just over five years ago.
I first started this book when it was new, and it took me a good few months to get through it this time (though it was not my bus reading), but I’m glad to have taken the time.
A digression on "novice systems": I attended a talk given by Hubert Dreyfus sometime probably in 1985 or 1986. He was billed as an AI critic, but he primarily criticized the same dead-end (if your goal is to create a thinking machine) research avenues that Hofstadter criticizes in MT, implying they were the best that AI researchers had to offer. Dreyfus didn’t address in his talk the approach advocated by Hofstadter (I will summarize this as saying that you don’t necessarily need to model every synapse, but you do need to allow/force the cognitive functions to emerge from lower-level, largely deterministic (Hofstadter says non-deterministic, but I don’t know whether we disagree or are talking about different things or different aspects of the same thing), interactions that you do model—Dreyfus seems to believe that you would also need your machine to have a body and culture like ours, but I don’t buy that), but I was interested to find myself agreeing so frequently with someone whose fundamental premise was diametrically opposed to my own, at least nominally. One point Dreyfus made that I found particularly compelling was that expert systems (really just rule-based decision engines, however elaborate) have more in common with the way novices perform tasks than the way experts perform them. Experts don’t consult a set of rules to determine what to do next: they have completely incorporated any rules that they started with, and have expanded them with the experience they gained as they were becoming experts. Even when an expert gives you a rule for why she did something, it is almost always derived a posteriori, rather than truly representing the decision factor. Hence "novice systems".