Artificial Intelligence: The Very Idea

Rating:
Summary: Don't judge this book by its cover...
Review: Don't judge this book by its cover-or at least by its title. Haugeland's Artificial Intelligence: The Very Idea does not adequately serve as a general introduction to the conceptual underpinnings and philosophical background of the quest to create an artificial mind. Rather, it focuses on one specific approach to how natural and man-made thought works: "thinking...essentially is rational manipulation of mental symbols." (p. 4) Haugeland plows forward with this as his core assumption, barely noting that some AI researchers see thought from a very different perspective (for example, the connectionists) and others find the whole enterprise fraught with theoretical difficulty (such as Dreyfus).

So Haugeland's story is that of a particular theory of mind that held predominance for several decades (what the author himself dubs "good, old-fashioned artificial intelligence" or "GOFAI", p. 112) but is now gradually being superceded. His introduction to this story concludes with a description of the Turing test and a justification for its use, and a brief statement of the efficacy of describing a system in different-even contradictory-ways through different "organizational levels". (p. 9) Of all the ideas presented in the book, this last one has the greatest promise for applicability beyond GOFAI.

Chapter 1, "The Saga of the Modern Mind", is a condensed bit of intellectual history. Haugeland introduces the philosophical children of the Copernican revolution-Hobbes, Descartes, and Hume-and the ways they grappled with understanding the world of the mental with the ideas that had proven so effective in the physical sciences. We soon encounter the "paradox of mechanical reason": if reason is the meaningful manipulation of symbols, and meanings are not physical entities, then how can machines manipulate them? (p. 39)

Chapter 2 serves as an extended definition of "Automatic Formal Systems", that is, computers. This material is the most challenging in the text, but the important concepts (formal games, digital systems, medium independence, etc.), are well-described, except for finite playability. The students I tutored through this work found it impossible to determine just what point was being made, and so did I.

How does one assign meanings-connections to the "real", outside world-to the symbols that a computer manipulates? This question is taken up in Chapter 3, "Semantics"-and answered, it seems, by sleight-of-hand. Haugeland gives to this the name "the formalist's motto": "if you take care of the syntax, the semantics will take care of itself". (p. 106) Neither I nor my students found this simple resolution at all satisfying. In every example of a formal game that the author presents, whatever semantic interpretation it has is provided from outside the system.

Chapter 4, "Computer Architecture", charts the milestones of computing. It begins with the analytical engine, and lauds Babbage's single-handed invention of programming without noting, however, that a human mind does not resemble the tabula rasa of a computer's memory bank. Moving quickly to the twentieth century, we get insightful descriptions of Turing machines, von Neumann machines (which turn out to be the kind of computer we are accustomed to), the mind-bending tree-structured LISP machines, and Newell's pragmatic production machines.

Chapter 5, "Real Machines", might be better titled "Real Problems". Haugeland presents some of the brick walls that AI research has run into. These can be grouped into the phenomenon of the combinatorial explosion: in order to interact with the real world in a manner that demonstrates "common sense", an AI must have access to an impossibly large store of information (while accessing what it needs in due time), and be able to consider an equally impossibly large set of potential courses of action. (p. 178) Methods to restrict what the AI has to consider, such as the focus on "micro-worlds", result in a system with no sense. Haugeland acknowledges these problems, and offers nothing but hope in scientific and technological progress to answer them.

Chapter 6, "Real People", develops means by which the sense that humans exhibit, and machines are far from realizing. Dennett's intentional stances and Grice's conversational implicatures are intelligent-if partial-characterizations of perspicuous reasoning. They are, however, frustratingly slippery for computer programmers, so it's not surprising that Haugeland, with some exasperation, groups them together under the "nonasininity canon": "An enduring system makes sense to the extent that, as understood, it isn't making [a rear] of itself." (p. 219) I feel that, if a reader has followed the author this far, then he or she deserves better than this.

Yet Haugeland and his colleagues are bound to feel frustration. Computers are electromechanical in nature, while humans are neurochemical. Computers can engage in numerical calculation with speed and precision, while most people find mathematics to be their most difficult school subject. Computers are tools that we devised to assist us. Human behavior was forged in the four-billion cauldron of evolution, and psychologists have barely begun to sort out the seething stew of vestigial loves, hates, and motivations that shape our behavior. And honest cognitive science will admit that humans and supercomputers are each masters of two separate, very different worlds. At the end, Haugeland finally admits this possibility-without contemplating the alternatives to the computation theory of might that this possibility demands.

Rating:
Summary: Don't judge this book by its cover...
Review: Don't judge this book by its cover-or at least by its title. Haugeland's Artificial Intelligence: The Very Idea does not adequately serve as a general introduction to the conceptual underpinnings and philosophical background of the quest to create an artificial mind. Rather, it focuses on one specific approach to how natural and man-made thought works: "thinking...essentially is rational manipulation of mental symbols." (p. 4) Haugeland plows forward with this as his core assumption, barely noting that some AI researchers see thought from a very different perspective (for example, the connectionists) and others find the whole enterprise fraught with theoretical difficulty (such as Dreyfus).

So Haugeland's story is that of a particular theory of mind that held predominance for several decades (what the author himself dubs "good, old-fashioned artificial intelligence" or "GOFAI", p. 112) but is now gradually being superceded. His introduction to this story concludes with a description of the Turing test and a justification for its use, and a brief statement of the efficacy of describing a system in different-even contradictory-ways through different "organizational levels". (p. 9) Of all the ideas presented in the book, this last one has the greatest promise for applicability beyond GOFAI.

Chapter 1, "The Saga of the Modern Mind", is a condensed bit of intellectual history. Haugeland introduces the philosophical children of the Copernican revolution-Hobbes, Descartes, and Hume-and the ways they grappled with understanding the world of the mental with the ideas that had proven so effective in the physical sciences. We soon encounter the "paradox of mechanical reason": if reason is the meaningful manipulation of symbols, and meanings are not physical entities, then how can machines manipulate them? (p. 39)

Chapter 2 serves as an extended definition of "Automatic Formal Systems", that is, computers. This material is the most challenging in the text, but the important concepts (formal games, digital systems, medium independence, etc.), are well-described, except for finite playability. The students I tutored through this work found it impossible to determine just what point was being made, and so did I.

How does one assign meanings-connections to the "real", outside world-to the symbols that a computer manipulates? This question is taken up in Chapter 3, "Semantics"-and answered, it seems, by sleight-of-hand. Haugeland gives to this the name "the formalist's motto": "if you take care of the syntax, the semantics will take care of itself". (p. 106) Neither I nor my students found this simple resolution at all satisfying. In every example of a formal game that the author presents, whatever semantic interpretation it has is provided from outside the system.

Chapter 4, "Computer Architecture", charts the milestones of computing. It begins with the analytical engine, and lauds Babbage's single-handed invention of programming without noting, however, that a human mind does not resemble the tabula rasa of a computer's memory bank. Moving quickly to the twentieth century, we get insightful descriptions of Turing machines, von Neumann machines (which turn out to be the kind of computer we are accustomed to), the mind-bending tree-structured LISP machines, and Newell's pragmatic production machines.

Chapter 5, "Real Machines", might be better titled "Real Problems". Haugeland presents some of the brick walls that AI research has run into. These can be grouped into the phenomenon of the combinatorial explosion: in order to interact with the real world in a manner that demonstrates "common sense", an AI must have access to an impossibly large store of information (while accessing what it needs in due time), and be able to consider an equally impossibly large set of potential courses of action. (p. 178) Methods to restrict what the AI has to consider, such as the focus on "micro-worlds", result in a system with no sense. Haugeland acknowledges these problems, and offers nothing but hope in scientific and technological progress to answer them.

Chapter 6, "Real People", develops means by which the sense that humans exhibit, and machines are far from realizing. Dennett's intentional stances and Grice's conversational implicatures are intelligent-if partial-characterizations of perspicuous reasoning. They are, however, frustratingly slippery for computer programmers, so it's not surprising that Haugeland, with some exasperation, groups them together under the "nonasininity canon": "An enduring system makes sense to the extent that, as understood, it isn't making [a rear] of itself." (p. 219) I feel that, if a reader has followed the author this far, then he or she deserves better than this.

Yet Haugeland and his colleagues are bound to feel frustration. Computers are electromechanical in nature, while humans are neurochemical. Computers can engage in numerical calculation with speed and precision, while most people find mathematics to be their most difficult school subject. Computers are tools that we devised to assist us. Human behavior was forged in the four-billion cauldron of evolution, and psychologists have barely begun to sort out the seething stew of vestigial loves, hates, and motivations that shape our behavior. And honest cognitive science will admit that humans and supercomputers are each masters of two separate, very different worlds. At the end, Haugeland finally admits this possibility-without contemplating the alternatives to the computation theory of might that this possibility demands.

Rating:
Summary: A great exposition of the fundamentals and more.
Review: This is a great exposition of the fundamental notions involved in the philosophy of AI. While at first look may appear like a good undergraduate read, it is, in fact, quite subtle and deep in most of the material it touches. Great scholarship.

Rating:
Summary: THE VERY BEST ON CLASSICAL AI
Review: This is the very best book on classical AI. However, there's a catch, as classical AI has many pitfalls, such as the frame problem or the symbol grounding problem. But there are ways to overcome these pitfalls, and if you want to see what's really hot in AI today you should check out Douglas Hofstadter's Fluid Concepts and Creative Analogies.

Rating:
Summary: THE VERY BEST ON CLASSICAL AI
Review: This is the very best book on classical AI. However, there's a catch, as classical AI has many pitfalls, such as the frame problem or the symbol grounding problem. But there are ways to overcome these pitfalls, and if you want to see what's really hot in AI today you should check out Douglas Hofstadter's Fluid Concepts and Creative Analogies.