The Age of Spiritual Machines: When Computers Exceed Human Intelligence

by Ray Kurzweil

  20 See the discussion on these music composition programs in the section The Creative Machine in chapter 8, “1999.”

  21 See W S. Sarle, ed., “Neural Network Frequently Asked Questions,” . This web site has numerous resources on past and current research on neural nets. G. E. Hinton’s “How Neural Networks Learn from Experience,” in the September 1992 issue of Scientific American (144-151), also provides a good introduction to neural networks.

  22 Researchers at the Productivity from Information Technology (PROFIT) Initiative at MIT have studied the effectiveness of neural networks in understanding handwriting.

  The PROFIT Initiative is based at MIT’s Sloan School of Management. The mission of the initiative is to study how the private and public sectors use information technology. Abstracts of working papers on this and other research on neural networks and data mining can be found at .

  23 “Miros, Inc. is located in Wellesley, Massachusetts, and specializes in providing face recognition software. Miros’ products include TrueFace PC, the first face recognition solution for computer, network and data security; and TrueFace GateWatch, a complete hardware/software security solution that allows or denies access to buildings and rooms by automatically recognizing a person’s face taken by a video camera.” From Miros Company Information at .

  24 For more information on BrainMaker’s aptitude to diagnose illnesses, and to predict the Standard and Poor 500 for LBS Management, see California Scientific’s home page at .

  25 The reset time stated here is an estimated average for neural connection calculations.

  For example, Vadim Gerasimov estimates the peak firing frequency of neurons (which significantly exceeds the average rate) to be 250-2,000 Hz (0.5-4 ms intervals) in “Information Processing in the Human Body” at . The firing time is affected by a number of variables, including, for example, the level and duration of a sound, as discussed in Jos. J. Eggermont, “Firing Rate and Firing Synchrony Distinguish Dynamic from Steady State Sound,” NeuroReport 8, issue 12, 2709-2713.

  26 Hugo de Garis maintains a web site on his research for ATR’s Brain Builder Group at .

  27 For an intriguing account of this research, read Carver Mead, Analog VSLI and Neural Systems (Reading, MA: Addison-Wesley, 1989), 257-278. Synaptics is briefly highlighted in Carol Levin, “Here’s Looking at You,” PC Magazine (December 20, 1994): 31. Carver Mead’s web site also provides detailed information on this research at the “Physics of Computation-Carver Mead’s Group” at .

  28 The SETI (Search for Extraterrestrial Intelligence) Institute conducts research on other signs of life in the Universe, its primary goal being the search for extraterrestrial intelligence. The institute is a nonprofit research organization, funded by government agencies, private foundations, and individuals, which in turn provides funding for several dozen projects. For more information, see the SETI Institute web site, .

  29 The author is dictating portions of this book to his computer through the continuous speech recognition program called Voice Xpress Plus from the dictation division of Lernout & Hauspie (formerly Kurzweil Applied Intelligence). See note 9 on Voice Xpress Plus in chapter 2 for more information.

  30 To find out more on State Street Global Advisor’s purchase in a majority stake in Advanced Investment Technology, read Frank Byrt, “State Street Global Invests in Artificial Intelligence.” Dow Jones Newswires, October 29, 1997. The genetic algorithm system used by the AIT Vision mutual fund is described in S. Mahfoud and G. Mani, “Financial Forecasting Using Genetic Algorithms.” Applied Artificial Intelligence 10 (1996): 543-565. The AIT Vision mutual fund opened at the beginning of 1996 and has publicly available performance numbers. In its first full calendar year (1996), the mutual fund increased 27.2 percent in net asset value, compared to 21.2 percent for its benchmark, the Russell 3000 index.

  It should be noted that outperforming its benchmark index does not in itself prove a superior level of decision making. The algorithm may have been making higher-risk investments (on average) than the average in the index.

  31 There are many online resources on evolutionary computation and evolutionary and genetic algorithms. One of the best is “The Hitchhiker’s Guide to Evolutionary Computation: A List of Frequently Asked Questions (FAQ),” edited by Jörg Heitkötter and David Beasley at : This guide includes everything from a glossary to links to various research groups.

  Another helpful online resource is the web site for the Santa Fe Institute. The institute’s web site can be accessed at .

  For an offline introduction to genetic algorithms, read John Holland’s article “Genetic Algorithms,” Scientific American 267, no. 1 (1992): 66-72. As mentioned in note 22 in chapter 1, Holland and his colleagues at the University of Michigan developed genetic algorithms in the 1970s.

  For more information on the use of genetic algorithm technology to manage the development and manufacturing of Volvo trucks, read Srikumar S. Rao, “Evolution at Warp Speed,” Forbes 161, no. 1 (January 12, 1998): 82-83.

  See also note 22 on complexity in chapter 1.

  32 See “Information Processing in the Human Body,” by Vadim Gerasimov, at .

  33 See “Information Processing in the Human Body,” by Vadim Gerasimov, at .

  34 I founded Kurzweil Applied Intelligence (Kurzweil AI) in 1982. The company is now a subsidiary of Lernout & Hauspie Speech Products (L&H), an international leader in the development of speech and language technologies and related applications and products. For more information about these speech recognition products, see .

  CHAPTER 5: CONTEXT AND KNOWLEDGE

  1 Victor L. Yu, Lawrence M. Fagan, S. M. Wraith, William Clancey, A. Carlisle Scott,

  John Hannigan, Robert Blum, Bruce Buchanan, and Stanley Cohen, “Antimicrobial Selection by Computer: A Blinded Evaluation by Infectious Disease Experts,” Journal of the American Medical Association 242, no. 12 (1979): 1279-1282.

  2 For an introduction to the development of expert systems and their use in various companies, read: Edward Feigenbaum, Pamela McCorduck, and Penny Nii, The Rise of the Expert Company (Reading, MA: Addison-Wesley, 1983).

  3 William Martin, Kenneth Church, and Ramesh Patil, “Preliminary Analysis of a Breadth-First Parsing Algorithm: Theoretical and Experiential Results.” MIT Laboratory for Computer Science, Cambridge MA, 1981. In this document, Church cites the synthetic sentence:

  “It was the number of products of products of products of products of products of products of products of products?” as having 1,430 syntactically correct interpretations.

  He cites the following sentence:

  “What number of products of products of products of products of products of products of products of products was the number of products of products of products of products of products of products of products of products?” as having 1,430 X 1,430 = 2,044,900 interpretations.

  4 These and other theoretical aspects of computational linguistics are covered in Mary D. Harris, Introduction to Natural Language Processing (Reston, VA: Reston Publishing Co., 1985).

  CHAPTER 6: BUILDING NEW BRAINS ...

  1 Hans Moravec is likely to make this argument in his 1998 book Robot: Mere Machine to Transcendent Mind (Oxford University Press; not yet available as of this writing).

  2 One hundred fifty million calculations per second for a 1998 personal computer doubling twenty-seven times by the year 2025 (this assumes doubling both the number of components, and the speed of each component every two years) equals about 20 million billion calculations per second. In 1998, it takes multiple ca
lculations on a conventional personal computer to simulate a neural-connection calculation. However, computers by 2020 will be optimized for the neural-connection calculation (and other highly repetitive calculations needed to simulate neuron functions). Note that neural-connection calculations are simpler and more regular than the general-purpose calculations of a personal computer.

  3 Five billion bits per $1,000 in 1998 will be doubled seventeen times by 2023, which is about a million billion bits for $1,000 in 2023.

  4 NEC’s goals to build a supercomputer with a maximum performance of more than 32 teraflops is chronicled in “NEC Begins Designing World’s Fastest Computer,” News-bytes News Network, January 21, 1998, located online at .

  In 1998, IBM was one of four companies chosen to participate in PathForward, an initiative from the Department of Energy to develop supercomputers for the twenty-first century. Other companies involved in the project are Digital Equipment Corporation; Sun Microsystems, Inc.; and Silicon Graphics/Cray Computer Systems (SGI/Cray). PathForward is part of the Accelerated Strategic Computing Initiative (ASCI). For more information on this initiative, see .

  5 By harnessing the accelerating improvement in both density of components and speed of components, computer power will double every twelve months, or a factor of one thousand every ten years. Based on the projection of $1,000 of computing being equal to the estimated processing power of the human brain (20 million billion calculations per second) by the year 2020, we get a projection of $1,000 of computing being equal to a million human brains in 2040, a billion human brains in 2050, and a trillion human brains in 2060.

  6 By 2099, $1,000 of computing will equal 1024 times the processing power of the human brain. Based on an estimate of 10 billion persons, that is 1014 times the processing power of all human brains. Thus one penny of computing will equal 109 (one billion) times the processing power of all human brains.

  7 In the Punctuated Equilibrium theories, evolution is seen to progress in sudden leaps followed by periods of relative stability. Interestingly, we often see similar behavior in the performance of evolutionary algorithms (see chapter 4).

  8 Dean Takahashi, “Small Firms Jockeying for Position in 3D Chip Market,” Knight-Ridder /Tribune News Service, September 21, 1994, p. 0921K4365.

  9 The entire February 1998 issue of Computer (vol. 31, no. 2) explores the status of optical computing and optical storage methods.

  Sunny Bains writes of companies using optical computing for fingerprint recognition and other applications in “Small, Hybrid Digital/Electronic Optical Correlators Ready to Power Commercial Products: Optical Computing Comes into Focus.” EE Times, January 26, 1998, issue 990. This article is online at .

  10 For a nontechnical introduction to DNA computing, read Vincent Kiernan, “DNA-Based Computers Could Race Past Supercomputers, Researchers Predict,” in the Chronicle of Higher Education (November 28, 1997). Kiernan discusses the research of Dr. Robert Corn from the University of Wisconsin as well as the research of Dr. Leonard Adleman. The article can be accessed online at .

  Research at the University of Wisconsin can be accessed online at .

  Leonard Adleman’s “Molecular Computation of Solutions to Combinatorial Problems” from the November 11, 1994, issue of Science (vol. 266, p. 1021) provides a technical overview of his design of DNA programming for computers.

  11 Lambertus Hesselink’s research is reported by Phillip F. Schewe and Ben Stein in Physics News Update (no. 219; March 28, 1995). The description is available online at .

  12 For information on nanotubes and buckyballs, read Janet Rae-Dupree’s article “Nanotechnology Could Be Foundation for Next Mechanical Revolution,” Knight-Ridder/ Tribune News Service, December 17, 1997, p. 1217K1133.

  13 Dr. Sumio Iijima’s research on nanotubes is summarized in the following article at the NEC site, .

  14 The research of Isaac Chuang and Neil Gershenfeld is reported in “Cue the Qubits: Quantum Computing,” The Economist 342, no. 8005 (February 22, 1997): 91-92; and in an article by Dan Vergano, “Brewing a Quantum Computer in a Coffee Cup,” Science News 151, no. 3 (January 18, 1997): 37. More technical details and a list of Chuang and Gershenfeld’s publications can be found at the Physics and Media Group/MIT Media Lab and at the Los Alamos National Laboratory .

  Other groups working on quantum computation include the Information Mechanics group at MIT’s Lab for Computer Science and the Quantum Computation Group at IBM .

  15 “Student Cracks Encryption Code,” USA Today Tech Report, September 2, 1997.

  16 Mark Buchanan, “Light’s Spooky Connections Set Distance Record,” New Scientist, June 28, 1997.

  17 Roger Penrose, The Emperor’s New Mind (New York: Penguin USA, 1990).

  18 To understand the concept of tunneling, it is important to understand how transistors on an integrated circuit chip work. An integrated chip is engraved with circuits comprised of thousands or millions of transistors, which electronic devices use to control the flow of electricity. Transistors are made up of a small block of a semiconductor, a material that acts as both an insulator and a conductor of electricity. The first transistors were comprised of germanium and were later replaced with silicon.

  Transistors work by holding a pattern of electric charge, allowing that pattern of charge to change millions of times every second. Tunneling refers to the ability of electrons (small particles that circle around the nucleus of an atom) to move or. “tunnel” through the silicon. Electrons are said to tunnel through the barrier as a result of the quantum uncertainty as to which side of the barrier they are actually on.

  19 Knowledge chunks would be greater than the number of distinct words because words are used in more than one way and with more than one meaning. Each different word meaning or usage is often referred to as a word “sense.” It is likely that Shakespeare used more than 100,000 word senses.

  20 Quoted from Douglas R. Hofstadter, Gödel, Escher, Bach: An Eternal Golden Braid (New York: Basic Books, 1979).

  21 Michael Winerip, “Schizophrenia’s Most Zealous Foe,” New York Sunday Times, February 22, 1998.

  22 The goal of the Visible Human Project is to create highly detailed, three-dimensional views of the male and female human body. The project is collecting transverse CT, MRI, and cryosection images. The web site is located at .

  23 Researchers Mark Hübener, Doron Shoham, Amiram Grinvald, and Tobias Bonhoeffer published their experiments on optical imaging in “Spatial Relationships among Three Columnar Systems in Cat Area 17,” Journal of Neuroscience 17 (1997): 9270-9284.

  More information on this and other brain-imaging research is located at the Weizmann Institute’s web site and at Amiram Grinvald’s web site .

  24 The work of Dr. Benebid and other researchers is summarized in an online article, “Neural Prosthetics Come of Age as Research Continues,” by Robert Finn, The Scientist 11, no. 19 (September 29, 1997): 13, 16. This article may be found at .

  25 From an April 1998 phone interview by the author with Dr. Trosch.

  26 Dr. Rizzo’s research is also reviewed in Finn’s article, “Neural Prosthetics Come of Age as Research Continues.”

  27 To read more about the “neuron transistor,” visit the web site of the Membrane and Neurophysics Department at the Max Planck Institute for Bioche
mistry .

  28 Robert Finn, “Neural Prosthetics Come of Age as Research Continues.”

  29 Carver Mead’s research is described at .

  30 W B. Yeats, “Sailing to Byzantium,” from Selected Poems and Two Plays of William Butler Yeats, edited by M. L. Rosenthal (New York: Macmillan, 1966).

  CHAPTER 7: ... AND BODIES

  1 Herbert Dreyfus is well known for his critique of artificial intelligence in his book What Computers Can’t Do: The Limits of Artificial Intelligence (New York: Harper and Row, 1979). Other theorists who may be considered to support the mind-beyond-machine perspective include J. R. Lucas and John Searle. See J. R. Lucas’s “Minds, Ma . chines and Gödel,” Philosophy 36 (1961): 120-124; and John Searle’s “Mind, Brains, and Programs,” The Behavioral and Brain Sciences 3 (1980): 417-424. Also, see Searle’s more recent book The Rediscovery of the Mind (Cambridge, MA: MIT Press, 1992).

  2 “Researchers led by Dr. Clifford Steer at the University of Minnesota Medical School report in the current Nature Medicine that they have eliminated the need for viruses by harnessing the body’s own genetic repair processes. In a landmark proof-of-concept experiment, the Minnesota team permanently altered a blood-clotting gene in 40 percent of the liver cells in a group of rats. The researchers started by splicing their DNA patch into a slip of RNA. Then they encased the hybrid molecule in a protective coating, laced it with sugars that seek out liver cells, and injected it into lab rats. True to plan, the hybrid molecules zeroed in on the targeted gene and lined up alongside it. An enzyme in the rats’ own liver cells did the rest: Whenever it spotted a mismatched DNA, it simply removed the offending DNA and stitched in a replacement. Now the trick is to show that it will work with other tissues—and other species.” From “DNA Therapy: The New, Virus-Free Way to Make Genetic Repairs.” Time, March 16, 1998.