Hot Best Seller

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

Availability: Ready to download

When it was first published in 1972, Hubert Dreyfus's manifesto on the inherent inability of disembodied machines to mimic higher mental functions caused an uproar in the artificial intelligence community. The world has changed since then. Today it is clear that "good old-fashioned AI," based on the idea of using symbolic representations to produce general intelligence, is When it was first published in 1972, Hubert Dreyfus's manifesto on the inherent inability of disembodied machines to mimic higher mental functions caused an uproar in the artificial intelligence community. The world has changed since then. Today it is clear that "good old-fashioned AI," based on the idea of using symbolic representations to produce general intelligence, is in decline (although several believers still pursue its pot of gold), and the focus of the AI community has shifted to more complex models of the mind. It has also become more common for AI researchers to seek out and study philosophy. For this edition of his now classic book, Dreyfus has added a lengthy new introduction outlining these changes and assessing the paradigms of connectionism and neural networks that have transformed the field. At a time when researchers were proposing grand plans for general problem solvers and automatic translation machines, Dreyfus predicted that they would fail because their conception of mental functioning was naive, and he suggested that they would do well to acquaint themselves with modern philosophical approaches to human being. "What Computers Still Can't Do" was widely attacked but quietly studied. Dreyfus's arguments are still provocative and focus our attention once again on what it is that makes human beings unique.


Compare

When it was first published in 1972, Hubert Dreyfus's manifesto on the inherent inability of disembodied machines to mimic higher mental functions caused an uproar in the artificial intelligence community. The world has changed since then. Today it is clear that "good old-fashioned AI," based on the idea of using symbolic representations to produce general intelligence, is When it was first published in 1972, Hubert Dreyfus's manifesto on the inherent inability of disembodied machines to mimic higher mental functions caused an uproar in the artificial intelligence community. The world has changed since then. Today it is clear that "good old-fashioned AI," based on the idea of using symbolic representations to produce general intelligence, is in decline (although several believers still pursue its pot of gold), and the focus of the AI community has shifted to more complex models of the mind. It has also become more common for AI researchers to seek out and study philosophy. For this edition of his now classic book, Dreyfus has added a lengthy new introduction outlining these changes and assessing the paradigms of connectionism and neural networks that have transformed the field. At a time when researchers were proposing grand plans for general problem solvers and automatic translation machines, Dreyfus predicted that they would fail because their conception of mental functioning was naive, and he suggested that they would do well to acquaint themselves with modern philosophical approaches to human being. "What Computers Still Can't Do" was widely attacked but quietly studied. Dreyfus's arguments are still provocative and focus our attention once again on what it is that makes human beings unique.

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

  1. 5 out of 5

    Manny

    Using philosophical arguments from Merleau-Ponty and Heidegger, Dreyfus convincingly demonstrates that there are things people can do, sometimes even without great effort, but which computers are simply incapable of ever being able to achieve. He ends with a list of 20 such items. Thirty-odd years after initial publication, computers still can't do 18 of them - it turns out that Dreyfus wasn't quite right about Grandmaster-level chess and large-vocabulary continuous speech recognition. Maybe the Using philosophical arguments from Merleau-Ponty and Heidegger, Dreyfus convincingly demonstrates that there are things people can do, sometimes even without great effort, but which computers are simply incapable of ever being able to achieve. He ends with a list of 20 such items. Thirty-odd years after initial publication, computers still can't do 18 of them - it turns out that Dreyfus wasn't quite right about Grandmaster-level chess and large-vocabulary continuous speech recognition. Maybe there was a bug in Merleau-Ponty's conceptual analysis. Oh well... if one out of two ain't bad, surely eighteen out of twenty is pretty darn good?

  2. 4 out of 5

    Dan Raghinaru

    “Being is essentially different from a being, from beings” - stated Heidegger. In other words, what gives beings their being is not itself a being. Translated into AI language, this means that what makes some things programmable is not itself programmable and moreover it cannot be specified, described, or even named. When the digital computer was invented, an entire generation of programmers from MIT and other leading institutions optimistically postulated that General AI was within reach and rus “Being is essentially different from a being, from beings” - stated Heidegger. In other words, what gives beings their being is not itself a being. Translated into AI language, this means that what makes some things programmable is not itself programmable and moreover it cannot be specified, described, or even named. When the digital computer was invented, an entire generation of programmers from MIT and other leading institutions optimistically postulated that General AI was within reach and rushed to accomplish it. Unfortunately, they naively followed the standard metaphysics and its excessively rational definitions of what human and intelligence mean - metaphysics developed and perpetuated by most of the philosophers over the last two thousands of years. This first AI attempt failed; however some byproducts were created. The second AI wave also failed; while we are now in the middle of the third wave. Dreyfus's philosophical “critique of artificial reason” was perfectly on the mark, devastating, and fulfilled almost completely in several years. According to Dreyfus, none of the four assumptions employed by the AI workers (biological, psychological, epistemological, and ontological) were justified and compared the AI workers with the old alchemists. Instead of paying any attention to him, the AI workers – most of them his peers at MIT – insulted and completely avoided him on the campus. According to Dreyfus, the digital computer logic and its working requires these particular four assumptions – since others cannot be implemented on it. Consequently, the AI workers adopted the four assumptions as self-evident and built their careers and dreams on them. It seemed obvious to them that a philosopher cannot understand their work and cannot criticize them; consequently, they refused to listen to him and optimistically persevered in their work despite increasing difficulties and failures. Dreyfus was exactly in the right place to point out their basic metaphysical assumptions; that is, he was a continental philosophy professor at MIT and he was just invited into the most advanced AI program at that time - RAND. Pointing out metaphysical assumptions - deeply rooted into our cultural, scientific, and technological worldviews for hundreds of years - to programmers is inevitably going to lead to huge and insanely funny miss-communications as presented in this book and in its reception by the AI community. I never laugh so much while reading a book as technical and philosophical as this one. But here is a sober reflection from this book: while it is not possible to create a digital program or machine to match humans, this constant and ubiquitous programming and digital worldview may reduce humans to match the existing digital programs and machines. In Heidegger's terms, this is called “enframing”: as everything else that is, man will be eventually turned into a resource by technology and for the further use of technology. It seems to me that most of the present AI work dropped the original strong pretense to build a General AI and they are instead focusing on limited, but highly practical and successful use of neural networks trained on big data. However, there are some old-fashioned AI workers out there that still predict that “the singularity is near”; that is in a never-changing 20 years horizon (i.e. not short enough to compromise themselves with a failed prediction, but long enough to sustain enthusiasm and to build a career for themselves). The AI field changed a lot since the first wave - called “good old-fashioned AI” - that Dreyfus criticized in this book. Since some assumptions and approaches were dropped in the AI field for good, the corresponding critiques in this book no longer apply. However, I believe that the main arguments presented in this book still stand against General AI understood as a “rational conscience”/“singularity” and prove its impossibility. Interestingly enough, Nature published an article against General AI a few months ago; the author used some of Dreyfus's old arguments to prove its impossibility (https://www.nature.com/articles/s4159...). This is just the old 1972 book with a “new”/1979 introduction.

  3. 4 out of 5

    Andrew

    If one earns one's bread in the world of Internet People too long, one will encounter a large number of people who seem inherently suspicious of the concept of humanity and go into long diatribes disparaging the weakness of the human mind without technological augmentation. Turns out that not only are they assholes who ruin your lunch break, they are also on very epistemologically shaky ground. Dreyfus' argument, along with John Searle's critique, are both devastating attacks on the concept of ar If one earns one's bread in the world of Internet People too long, one will encounter a large number of people who seem inherently suspicious of the concept of humanity and go into long diatribes disparaging the weakness of the human mind without technological augmentation. Turns out that not only are they assholes who ruin your lunch break, they are also on very epistemologically shaky ground. Dreyfus' argument, along with John Searle's critique, are both devastating attacks on the concept of artificial intelligence. Granted, Dreyfus has gotten some egg on his face as some of the things he considered impossible back in the '70s have since been proven attainable, but the majority of his argument remains sound: that a computer, while it can be trained to learn tasks heuristically, cannot conceive of meaning (among other failings) and is therefore not an intelligence. It also casts serious doubt over the entire program of cognitive science as it is now practiced. Look out kids, a lot of those Ted talks aren't as accurate as they seem.

  4. 5 out of 5

    Bookworm

    What Computers Still Can’t Do (1992) is an evolution of Hubert Dreyfus’s original work, What Computers Can’t Do (1972). Today, the ideas coming out of GOFAI research (Good Old Fashion Artificial Intelligence), which is based on the notion of using symbolic representations to replicate intelligence, are being replaced by more complex models of the brain/mind. In the revised edition, Dreyfus has added an introduction presenting an overview of the developments that have occurred in the field of Art What Computers Still Can’t Do (1992) is an evolution of Hubert Dreyfus’s original work, What Computers Can’t Do (1972). Today, the ideas coming out of GOFAI research (Good Old Fashion Artificial Intelligence), which is based on the notion of using symbolic representations to replicate intelligence, are being replaced by more complex models of the brain/mind. In the revised edition, Dreyfus has added an introduction presenting an overview of the developments that have occurred in the field of Artificial Intelligence (AI) since the publication of his original manifesto. Dreyfus also assesses how the perspectives of neural networks and connectionism have transformed the field. That said, What Computers Still Can’t Do presents a similar philosophical analysis to the one contained in the original work, which triggered an avalanche of outrage in the AI community upon its release in the seventies. Dreyfus's philosophical inquiry tracks the history of developments in AI and is concerned with exposing the incorrect assumptions (psychological, epistemological, and ontological) made by researchers working in the field. Throwing light on the kind of discursive moves used to understand the human brain/mind as reducible to the processes of a digital computer, Dreyfus concentrates on the two subfields of Artificial Intelligence: Cognitive Simulation and Artificial Intelligence. These two fields, he argues, have led to the examination of two distinct but interrelated questions: (1) Does a human being in “processing information” actually follow normal rules like a digital computer? (2) Can human behavior, no matter how generated, be described in a formalism which can be manipulated by a digital machine? Dreyfus notes that given the difficulties AI experienced during Phase I and Phase II of its development (e.g. failure of GPS), cognitive simulation nevertheless assumed that the information processes of a computer revealed “the hidden information processes” of a human being. He also asks a pertinent question still relevant today: Why do those working on artificial intelligence assume that there must be a digital way of performing human tasks? He writes: Those who think that a formalization of intelligent behavior must be possible, seem to base their arguments on the ontological assumption that the world can be analyzed into independent logical elements and an epistemological assumption that our understanding of the world can then be reconstructed by combining these elements according to heuristic rules. Dreyfus puts forward four models of human information processing to highlight the differences between human information processing and that of a computer: fringe consciousness, ambiguity tolerance, essential/inessential discrimination, and perspicuous grouping. In the context of GOFAI research, for example, each of these kinds of human processing includes a symbolic analogue that cannot be mapped directly onto the essential intelligence of human beings who are able to demonstrate the four models mentioned above. The symbolic analogues relevant to digital computers are as follows: heuristically guided search, context-free precision, trial and error search, and character lists. Dreyfus claims that descriptive or phenomenological evidence, when considered separately from traditional philosophical prejudices, suggests that non-programmable human capacities are involved in all forms of intelligent behavior. The key point Dreyfus makes in this text is that the complex information processing humans are capable of doing, and human cognitive processes more generally, cannot be reduced to the systematic rule-driven workings of a digital computer. In this capacity, Dreyfus argues, AI is limited by its assumption that the world can be explained in terms of elementary atomistic concepts, a view that dates back to the Greeks (Plato). That said, today AI has evolved by leaps and bounds. Developing computer systems that are increasingly “intelligent” (e.g., Angelina), the field of AI has become one of the most significant components in technological research and development. Although Dreyfus was incorrect about the implementation of GPS, his arguments are still relevant for contemporary analyses of the tendency toward the mechanization of the human brain/mind in all things neuro. What Computers Can’t Do (1972) and What Computers Still Can’t Do (1992) are necessary reads for those who are quick to jump on the “intelligence explosion” bandwagon (e.g., Ray Kurzweill) that predicts society is headed for a technological singularity (or the singularity), a hypothetical moment in history (2045) when AI has evolved to surpass human intelligence.

  5. 4 out of 5

    Joshua Stein

    The book is a bit dated, and it really shows when Dreyfus talks about the conteporary limitations of computers. He disparages chess playing computers in a time before Deep Blue, and so it is important to keep in mind that there are large portions of the critique that seem to have overstepped the appropriate boundaries, and that some of those criticisms have been scaled back in the wake of contemporary successes in certain forms of artificial intelligence. For those interested in the take that man The book is a bit dated, and it really shows when Dreyfus talks about the conteporary limitations of computers. He disparages chess playing computers in a time before Deep Blue, and so it is important to keep in mind that there are large portions of the critique that seem to have overstepped the appropriate boundaries, and that some of those criticisms have been scaled back in the wake of contemporary successes in certain forms of artificial intelligence. For those interested in the take that many continental philosophers of mind have on the older-school understandings of A.I. and computational approaches to mind in general, Dreyfus is a good introduction. His colleague, John Searle, is arguably more widely read on the subject and more influential in the philosophical literature, but Dreyfus is much more clear in this particular area, expressing a level of comfort with and awareness of his philosophical history, as well as his current context in the continental tradition. The context isn't totally articulated, though, in the sense that there has been some movement in what is meant by "A.I." since Dreyfus originally wrote the book. It used to mean something fairly overtly computational, centered on input and output; it used to mean, almost exclusively, software. IT no longer means that, and part of the reason is an understanding of the commentaries on context-free approaches. This is one such commentary, but Dreyfus often commits himself to the position that computers will, in principle, never be able to do certain things, which turns out to be predicated on an oversimplistic concept of "computer." This can hardly been seen as the fault of Dreyfus, though perhaps it is a failure of imagination. On the other hand, Dreyfus manages to simultaniously provide an interesting view of continental philosophy of mind that is explicitly understood in terms contemporary technology. This makes for a deeply relevant account of an area of thought that is often disparaged as frustratingly limited in terms of its relevance; Dreyfus, and several of the thinkers who started pubishing around the same time or not long after this book, gave continental philosophy of mind a new sense of importance in the philosophical community, and in academia more broadly, and this is a nice way to become introduced to that in the context of some thoughts on technology.

  6. 5 out of 5

    Danirainbow

    I'm probably biased towards Dreyfus' perspective in this book because I've grown fond of him from listening to his recorded Heidegger lectures at UC Berkeley. Despite his harsh and occasionally smug tone in this book, I've always found him a joy to listen to. He's clearly an expert on both AI technology and on continental European thought--not an easy mix to find!) What surprises me about this book is that it isn't more widely read, given that I believe it to be the most destructive critique on I'm probably biased towards Dreyfus' perspective in this book because I've grown fond of him from listening to his recorded Heidegger lectures at UC Berkeley. Despite his harsh and occasionally smug tone in this book, I've always found him a joy to listen to. He's clearly an expert on both AI technology and on continental European thought--not an easy mix to find!) What surprises me about this book is that it isn't more widely read, given that I believe it to be the most destructive critique on the possibility of human-level AI, hands-down. Maybe it's because the book is somewhat dated, but I find Dreyfus' critique far more convincing than Searle's famous Chinese Room, which is taught far more often.

  7. 4 out of 5

    Chris Harris

    This discussion of the state of the field of Artificial Intelligence and "Cognitive Simulation" or CS was written in 1972 and revised in 1992. Dreyfus argues that AI and CS will not succeed by programming computers with sets of formal rules as this is not how human consciousness operates. It's only in the closing chapter of the book that Dreyfus starts talking about alternative ways to achieving AI than programming a fully-developed, adult-level AI from scratch, suggesting that computers might f This discussion of the state of the field of Artificial Intelligence and "Cognitive Simulation" or CS was written in 1972 and revised in 1992. Dreyfus argues that AI and CS will not succeed by programming computers with sets of formal rules as this is not how human consciousness operates. It's only in the closing chapter of the book that Dreyfus starts talking about alternative ways to achieving AI than programming a fully-developed, adult-level AI from scratch, suggesting that computers might fare better if we program them to learn as children do. The subsequent successes of machine learning seem to indicate that this is the path to take and two of the prime examples that Dreyfus gives of areas where computers were failing (playing chess and go) have seen spectacular successes in recent years, with computers now able to beat any human player. And yet I was fascinated by how the book's examination of alternative approaches shies away from the underlying problem of what we actually mean by consciousness; the characteristics and qualities of "the thing applying the rules" are clearly pivotal, but they're backgrounded other than Dreyfus's suggestion that an AI needs to be embodied to make sense of the world. General Problem Solving AI (or "Good Old-Fasioned Artificial Intelligence, GOFAI for short) still seems to be thirty years away, just as it was back in the 1960s. The question of how to build a mind has yet to be answered, and this book shows just how much work remains to be done.

  8. 5 out of 5

    Garth Eaglesfield

    An update of the original incendiary critique of AI, still completely relevant

  9. 4 out of 5

    Carl

    Bought this a while back and keep meaning to get to it, but I a bit ignorant in Cognitive Science and computers, so it's been too intimidating so far. This is primarily a critique of AI research back 30-40 years ago, from what I hear, though it has been updated for this edition (though this edition is old by now as well, considering the speed with which research advances in the sciences compared to philosophy). Bought this a while back and keep meaning to get to it, but I a bit ignorant in Cognitive Science and computers, so it's been too intimidating so far. This is primarily a critique of AI research back 30-40 years ago, from what I hear, though it has been updated for this edition (though this edition is old by now as well, considering the speed with which research advances in the sciences compared to philosophy).

  10. 4 out of 5

    Adriano Gaved

    Nobody should even talk about Artificial Intelligence without having read thi book! Furthermore, I found very strong and original the way he uses both phylosophical arguments and historical facts to make his points across.

  11. 4 out of 5

    Ari

    Clever

  12. 5 out of 5

    Jesse

    Best philosophy of cognitive science book I've read. Dreyfus is harsh, but his words proved prophetic. Best philosophy of cognitive science book I've read. Dreyfus is harsh, but his words proved prophetic.

  13. 5 out of 5

    Seth Graham

    Dreyfus I think is correct, and he is the most endearing person in interviews.

  14. 5 out of 5

    Odie

  15. 4 out of 5

    Bradley Beth

  16. 5 out of 5

    Marcin Milkowski

  17. 4 out of 5

    Jordan Peacock

  18. 5 out of 5

    Fernando Pasquini Santos

  19. 4 out of 5

    etzel

  20. 5 out of 5

    Juanpi

  21. 5 out of 5

    Justin Ohms

  22. 5 out of 5

    Van Hellsing

  23. 5 out of 5

    Phillip

  24. 5 out of 5

    Allen Severino

  25. 5 out of 5

    Etrit Syla

  26. 5 out of 5

    Ramanirudh

  27. 5 out of 5

    Malcolm Shute

  28. 5 out of 5

    Rachael

  29. 5 out of 5

    Patrick Merlevede

  30. 5 out of 5

    I

Add a review

Your email address will not be published. Required fields are marked *

Loading...
We use cookies to give you the best online experience. By using our website you agree to our use of cookies in accordance with our cookie policy.