Never speak more clearly than you think. —Jeremy Bernstein
This book assumes that any brain, machine, or other thing that has a mind must be composed of smaller things that cannot think at all. The structure of the book itself reflects this view: each page explores a theory or idea that exploits what other pages do. Some readers might prefer a more usual form of story plot. I tried to do that several times, but it never seemed to work; each way I tried to line things up left too many thoughts that would not fit. A mind is too complex to fit the mold of narratives that start out here and end up there; a human intellect depends upon the connections in a tangled web — which simply wouldn't work at all if it were neatly straightened out.
Many psychologists dream of describing minds so economically that psychology would become as simple and precise as physics. But one must not confuse reality with dreams. It was not the ambitions of the physicists that made it possible to describe so much of the world in terms of so few and simple principles; that was because of the nature of our universe. But the operations of our minds do not depend on similarly few and simple laws, because our brains have accumulated many different mechanisms over aeons of evolution. This means that psychology can never be as simple as physics, and any simple theory of mind would be bound to miss most of the big picture. The science of psychology will be handicapped until we develop an overview with room for a great many smaller theories.
To assemble the overview suggested in this book, I had to make literally hundreds of assumptions. Some scientists might object to this on the ground that successful sciences like physics and chemistry have found it more productive to develop theories that make the fewest assumptions, eliminating everything that does not seem absolutely essential. But until we have a more coherent framework for psychology, it will remain too early for the task of weeding out unproved hypotheses or for trying to show that one theory is better than another — since none of our present-day theories seem likely to survive very long in any case. Before we can have an image of the forest of psychology, we'll have to imagine more of its trees and restrain ourselves from simplifying them to death. Instead, we have to make ourselves complicated enough to deal with what is actually there.
It is scarcely a century since people started to think effectively about the natures of the brainmachines that manufacture thoughts. Before that, those who tried to speculate about this were handicapped on one side by their failure to do experiments, particularly with young children, and on the other side by their lack of concepts for describing complicated machinery. Now, for the first time, mankind has accumulated enough conceptual tools to begin comprehending machines with thousands of parts. However, we are only beginning to deal with machines that have millions of parts and we have barely started to acquire the concepts that we'll need to understand the billion-part machines that constitute our brains. New kinds of problems always arise when one encounters systems built on larger, less familiar scales.
Since most of the statements in this book are speculations, it would have been too tedious to mention this on every page. Instead, I did the opposite — by taking out all words like possibly and deleting every reference to scientific evidence. Accordingly, this book should be read less as a text of scientific scholarship and more as an adventure story for the imagination. Each idea should be seen not as a firm hypothesis about the mind, but as another implement to keep inside one's toolbox for making theories of the mind. Indeed, there is a sense in which that can be the only realistic way to think about psychology — since every particular person's mind develops as a huge machine that grows in a somewhat different way. Are minds machines? Of that, I've raised no doubt at all but have only asked, what kind of machines? And though most people still consider it degrading to be regarded as machines, I hope this book will make them entertain, instead, the thought of how wonderful it is to be machines with such marvelous powers.
Scientists like to credit those who first discovered each idea. But the central concept of this book, that the mind is a society of many smaller mechanisms, involved so many years of work to bring it to its present form that I can mention only a few of the people who had the most influence on it. In this research I shared the greatest privilege a human mind can have: to work on new ideas together with the foremost intellects of one's time. As a student at Harvard, I immersed myself in mathematics and psychology and attached myself to two great young scientists, the mathematician Andrew Gleason and the psychologist George A. Miller. This was the era of the scientific movement that was later called cybernetics, and I was especially entranced with the works of Nicholas Rashevsky and of Warren McCulloch, who were making the first theories of how assemblies of simple cell-machines could do such things as recognize objects and remember what they'd seen. By the time I started graduate school in mathematics at Princeton in 1950, I had a clear enough idea about how to make a multi-agent learning machine. George Miller obtained funds for building it; this was the Snarc machine of chapter 7. Constructed with the help of a fellow student, Dean Edmonds, it managed to learn in certain ways, but its limitations convinced me that a more versatile thinking machine would have to exploit many other principles.
My teachers in the golden age of mathematics at Princeton were not particularly interested in psychology, but the ways of thought are more important than the subject matter, and I learned new mental strategies from Albert Tucker, Ralph Fox, Solomon Lefshetz, John Tukey, Salomon Bochner, and John von Neumann. I learned even more from my own generation of students at Princeton: particularly from John Nash, Lloyd Shapley, Martin Shubik, and John McCarthy. In 1954 I returned to Harvard as a Junior Fellow of the Harvard Society of Fellows, with no obligation but to pursue whatever goal seemed most important. There seemed no way to get around the apparent limitations of low-level, distributed-connection learning machines, so I turned toward a new theory being pioneered by Ray Solomonoff, about generalizing from experience. I attached myself to Warren McCulloch and Oliver Selfridge, with whom I worked most closely of all until becoming a professor of mathematics at MIT. It was from them that I derived my image of how to make a laboratory work.
In 1959, John McCarthy came to MIT from Dartmouth, and we started the MIT Artificial Intelligence Project. We agreed that the most critical problem was of how minds do common-sense reasoning. McCarthy was more concerned with establishing logical and mathematical foundations for reasoning, while I was more involved with theories of how we actually reason using pattern recognition and analogy. This combination of theoretical and practical research attracted students of great ability, and our laboratory had an atmosphere that combined mathematical power with engineering adventure; this led not only to new theories of computation, but also to developing some of the very first automatic robots. In 1963, McCarthy left to start a new AI laboratory at Stanford, and now there were three principal centers of research in Artificial Intelligence, including the one that Allen Newell and Herbert Simon had started earlier at Carnegie-Mellon University. A fourth center soon emerged at Stanford Research Laboratory, and we all worked closely together.
The money to support the people and equipment for this work came mainly from an office of the Advanced Research Projects Agency concerned with information processing technology. This office was directed, in effect, by the scientists themselves, initially by Dr. J.C.R.
Licklider, who had been my teacher and friend when I was a student at Harvard. Licklider had already organized a research center at the Bolt, Beranek, and Newman company in Cambridge, Massachusetts, and McCarthy and I and several of our students had worked closely with that group for several years. Later, when Licklider returned to become a professor at MIT, the Information Processing Technology Office was taken over successively by Lawrence G. Roberts and Ivan Sutherland (who had been students of ours at MIT) and then by Robert Taylor and Robert Kahn — all of whom made important intellectual contributions. The actual details of all these research contracts were managed in the Office of Naval Research by Marvin Denicoff, whose vision of the future had a substantial influence on the entire field. My own research was supported by the ONR over an even longer period, since it had previously financed my graduate studies in topology at Princeton, and, subsequently, Denicoff's successor, Alan Meyrowitz, supported my research during the completion of this book.
Jerome Wiesner and Philip Morse of MIT obtained the resources for our first laboratory. Our development at MIT was encouraged by William Ted Martin, Norman Levinson, Witold Hurewicz, Norbert Weiner, Claude Shannon, Peter Elias, and Robert Fano. I was given the privilege of sharing with Shannon the endowed chair of Donner Professor of Science at MIT and enjoyed the support of many other people and organizations over the years: John Williams, Paul Armer, and Merril Flood enabled me to work with Newell, Shaw, and Simon at the Rand Corporation; Oliver Selfridge and Gerald Dinneen encouraged research at MIT's Lincoln Laboratory; Michel Gouilloud supported my work from the Schlumberger Corporation; Edward David provided support from Exxon; and Alan Kay supported many of our students with funding from (successively) the Xerox, Atari, and Apple corporations. For several years, the Thinking Machines Corporation has supported both this research and the development of a new type of computer called the Connection Machine — designed by my student Danny Hillis for embodying societies of mind.
Most of all, I want to acknowledge the contributions to this book of Seymour Papert, who came to MIT in 1963 after five years of studying child development with Jean Piaget in Geneva. Papert and I worked so well together that for a decade we supervised the laboratory jointly, each able at any time to leave the other to decide what should be done. Together we evolved new mathematical techniques, designed laboratory experiments, built computer hardware and software, and supervised the same students. Such a partnership could not have worked so well had we not both developed in similar intellectual directions before we met; we were both involved with the same areas of mathematics, with similar concerns about machinery, and with similar attitudes about psychology. One of our projects was to build a machine that could see well enough to use mechanical hands for solving real-world problems; this was the origin of Builder and the insights that emerged from it. In trying to make that robot see, we found that no single method ever worked well by itself. For example, the robot could rarely discern an object's shape by using vision alone; it also had to exploit other types of knowledge about which kinds of objects were likely to be seen. This experience impressed on us the idea that only a society of different kinds of processes could possibly suffice. Papert and I worked together not only on robotic machines, but in many other areas; for example, we spent several years developing a new mathematical theory for the then mysterious Perceptron type of learning machine. In the middle 1970s Papert and I tried together to write a book about societies of mind but abandoned the attempt when it became clear that the ideas were not mature enough. The results of that collaboration shaped many earlier sections of this book.
Eventually Papert and I both turned away from large-scale scientific enterprises toward somewhat different individual goals, and we imposed the directorship of our laboratory upon one of our most original and productive students, Patrick Winston — who first worked out the idea of making uniframes. Papert went on to develop a host of new theories about mental development and education; these led to the computer language LOGO and to many other concepts that started to enter the mainstream of educational thought over the next decade. I focused my concern on searching for better theories about the little world of how a child might learn to build with blocks. The parts of the puzzle that form this book began to fit together in my mind in the mid-1970s, around the concept of frame-array, and this eventually led to the theories about communication-lines, K-lines, and level-bands, and then, during the book's final stages, to the ideas about pronomes, polynemes, and cross-realm correspondences.
As for the manuscript itself, Bradley Marx read through every draft, comparing each version with earlier ones, helping to maintain clarity, stylistic coherency, and especially protecting good ideas from destructive revisionary impulses. This was hard because the early manuscript was more than twice its present length. Robin Lakoff suggested neutering the English; this seemed at first impossible but soon became quite natural. Theodore Sturgeon reviewed an early draft; I wish he had lived to see it now. Kenneth Haase, Betty Dexter, and Tom Beckman made innumerable suggestions and corrections. Successive drafts were reviewed by Danny Hillis, Steve Bagley, Marvin Denicoff, Charlotte Minsky, Michel Gouilloud, Justin Lieber, Philip Agre, David Wallace, Ben Kuipers, Peter de Jong, and Sona Vogel. Richard Feynman contributed a variety of insights about memory and parallel processing. David Yarmush helped to organize the book into sections, to smooth out the transitions, and to establish the gradient wherein the words begin with commonsense meanings and gradually become more technical. Bob Whittinghill made many suggestions about language as well as about psychology. Douglas Hofstadter evaluated the entire theory, forcing me to make several substantial changes. Michael Crichton made many technical suggestions and helped me to refine the early chapters.
Russell Noftsker and Tom Callahan made substantial engineering contributions to our work. Hosts of ideas came from students at MIT, notably Howard Austin, Manuel Blum, Danny Bobrow, Eugene Charniak, Henry Ernst, Tom Evans, Scott Fahlman, Ira Goldstein, William Gosper, Richard Greenblatt, Adolfo Guzman, Kenneth Haase, William Henneman, Carl Hewitt, Danny Hillis, Jack Holloway, Tom Knight, William Martin, Joel Moses, Bertram Raphael, Larry Roberts, James Slagle, Jerry Sussman, Ivan Sutherland, David Waltz, Terry Winograd, Patrick Winston, and many others. Countless other thoughts came from working at various times with Maryann Amacher, Gregory Benford, Terry Beyer, Woodrow Bledsoe, Mortimer Casson, Edward Feigenbaum, Edward Fredkin, Arnold Griffith, Louis Hodes, Berthold Horn, Joel Isaacson, Russell Kirsch, David Kirsh, Robert Lawler, Justin Leiber, Douglas Lenat, Jerome Lettvin, David MacDonald, Curtis Marx, Hans Moravec, Stewart Nelson, Nils Nillsson, Donald Norman, Walter Pitts, Jerry Pournelle, Charles Rosen, Carl Sagan, Roger Schank,
Robert Sheckley, Stephen Smoliar, Cynthia Solomon, Ray Solomonoff, Luc Steels, Warren Teitelman, and Graziella Tonfoni. I wish I could acknowledge the inspirations of all the friends of earlier years, particularly W. Ross Ashby, Thomas Etter, Ned Feder, Heinz von Foerster, Donald Hebb, John Hollander, Arnold Honig, Gordon Pask, Roland Silver, Jan Syrjala, Carroll Williams, Bertram Wolfe, David Yarmush — and of all the teachers of my youth, particularly Dudley Fitts, Ruth Gordon, Alexander Joseph, Edward Lepowsky, and Herbert Zim. My development was also strongly influenced first by the writing and later by the friendship of Arthur C. Clarke, Robert Heinlein, Frederick Pohl, and most of all by Isaac Asimov.
Of course, the deepest influence on my style of thought came from my parents, Henry Minsky and Fannie Reiser. My wife, Gloria Rudisch, our children Margaret, Henry, and Juliana (who drew the illustrations and sometimes changed the text to make them fit), and my sister Ruth all helped to shape this book. My sister Charlotte also lives between these lines, for even in our childhood, she was a powerful artist and critic, and her dreams became the meanings of my ordinary words.