What Computers Still Can't Do: A Critique of Artificial Reason

I would love to see a working "HAL", and still hope we progress in the development in AI. However, the discovery of a General Problem Solver has always been elusive, and never matched the corresponding development for the raw physical abilities our computers now have. Why is that? And why is this still true? Mr. Dreyfuss attempts to answer this with a refreshing accountability of the scientific method. He compares the historical development of AI to the theoretical expectations, and apparently not without resistance. He manages to challenge the "salesmanship" of following trends, and thus avoiding assumed results. The book provides many examples of its logic, and carefully draws the conclusions. When I first read it I was tremendously impressed with its (if I may) "insights". I feel it is a must read; not just for those in the AI world, but for those interested in the scientific method as well.

Mr. Shallit, in a previous review, sounds exactly like a throwback to the previous era when glorified computer technicians started to take themselves seriously and bestowed upon themselves the bogus title of "computer scientists". When Dreyfus critique of strong AI came out they were enraged that someone was trying to punch the air (and as it turned out it was full of hot air) out of their over inflated clown balloon. After decades of sheer nonsense, and of confusing metaphor with the real thing (their preposterous proposition of the "complexity critical mass" for intelligence is akin to thinking that a sufficiently complex computer simulation of an atomic explosion will actually explode: I guess we should all be really worried...) strong AI proponents just gave up. Now they are much quiet and many have given Dreyfus some concessions after being proven completely wrong (old habits die slowly). No serious GCT (Glorified Computer Technician) of today would profess faith in strong AI.

Dreyfus' book is about the history of failure of Artificial Intelligence researchers such as Marvin Minsky to embody intelligence at the human level. It is easy to read, but is rather exasperating. It is like riding in a truck driven by Minsky, and other AI researchers, where they are trying to make a long trek across a country without roads. They keep getting caught in swamps, blowing tires, and hitting trees, all the time shouting "We're almost there!" Meanwhile Dreyfus is a dog in the bed of the truck continually barking at dangers, and the folly of the drivers. Amidst Dreyfus' continuous cacophony of sarcastic cynicism there are some important points on what assumptions are doomed to failure, which he made quite clear by tedious repetition.Basically there are two types of mistakes made by Minsky and many others:1. believing they were getting close to understanding human thought,2. repeatedly announcing same to the world.The philosophy of Dreyfus in the first 300 pages is largely concerned with fallacious assumptions made by AI researchers. Finally in the last 50 pages (350 page book) he settles down and gives us some interesting concepts that should be understood if we are to seek AI at the human level. He develops the concept of "nonformal behavior" - which we humans usually learn by generalizing examples and following intuition without use of formal rules. Examples: chess at the gestalt master level, and disambiguation of broken sentences. Dreyfus acknowledges the possible importance of neural network architectures, but dismisses them as outside the scope of his critique. He touches on the poor idea of AI trying to program a full functioning adult, and further carries out a critique of machine learning ("reinforcement learning").The most important point he makes is that of nonformal behavior -- the non-logical almost Zen-like process that humans must go through. The irony is that we have to struggle with our nonformal thinking to do simple formal tasks such as long division; whereas the computer must struggle with its built-in hard logic to attempt nonformal tasks such as pattern recognition.The book is for the most part quite dated, but nevertheless, it is very worthwhile reading for anyone in a serious pursuit of machine intelligence. My criticism of his style is just that. I have only a minor criticism of the intelligent content and his restrictions in scope.

Mr. Shallit compares the critique of cognitive science by Professor Dreyfus to 'creation science'. He remarks that Dreyfus is not a computer scientist. This is true. But many 'cognitive scientists' aren't either, cognitive science being an interdisciplinary pursuit engaging philosophers, psychologists, linguists, neuroscientists, anthropologists and sociologists. It is unfortunate that Dreyfus allowed himself to polemicize by using the word 'alchemy' to characterize his opponents, but he has, by far, been the victim of unargued diatribes against his work. The fact is: most of the salient issues in cognitive science are logical and conceptual, NOT technological. Here, Dreyfus broke new ground (although I would have preferred his treatment to have been more Wittgensteinian than Heideggerian). Phil Agre's brilliant book on computation and human experience (Agre IS a computer scientist) shows that SOME AI-workers have found aspects of Dreyfus's work very telling. But, of course, the issues are, again, not empirical but logical in this field. See, for example, Graham Button et al., "Computers, Minds and Conduct" (Polity/Blackwell, 1995) which picks up where Dreyfus left off. Shallit remarks that Dreyfus has been 'refuted': where? by whom?The fact is that cognitivism is hotly contested by serious thinkers in many disciplines, but Shallit's name-calling (and the comparison of cognitivism's serious critics to creation scientists) smacks of an abdication from serious engagement and argument.Dreyfus's revised edition is a fine piece of work, worthy of serious intellectual discussion and confrontation. His many aarguments against Fodor, Chomsky, Simon and others have great merit. It is unfortunate that some folks simply close their eyes and argue from authority. But appeals to (even 'scientific') authority wear thin when left to stand alone!!

This is an absolute classic that everyone interested in or working in AI should have read. It is one of the very view books on a computer related subject that is over 25 years old and still useful today. That alone might tell you something. I find it interesting that many AI-workers seem to be actually afraid of this book. They should not. It may give the reader a far better sense on limits, use and future of AI work. I would also recommend this book to people outside the AI world and who are interested in what role the digital computer may play in our lives. But the book is not about bits, so if you don't like technical mumbo-jumbo, this is still a book for you. The book is very well written. Some readers may find it a difficult book, as it also contains some philosophical issues. But some readers may find themselves in a bookstore asking for the work of Wittgenstein or Heidegger and actually understanding what they read (and like what they read) after having read this book. I have only one complaint. The introduction to the 1992 MIT Press edition is in fact an afterword. It assumes that you already are familiar with the history of the subject. So, if you read this book, you should start with the Introduction to the 1979 edition instead and keep the Introduction to the MIT Press edition definitely for last