The Age of Spiritual Machines: When Computers Exceed Human Intelligence

by Ray Kurzweil

  The above paradigm is called an evolutionary (sometimes called genetic) algorithm.31 The system designers don’t directly program a solution; they let one emerge through an iterative process of simulated competition and improvement. Recall that evolution is smart but slow, so to enhance its intelligence we retain its discernment while greatly speeding up its ponderous pace. The computer is fast enough to simulate thousands of generations in a matter of hours or days or weeks. But we have only to go through this iterative process one time. Once we have let this simulated evolution run its course, we can apply the evolved and highly refined rules to real problems in a rapid fashion.

  Like neural nets, evolutionary algorithms are a way of harnessing the subtle but profound patterns that exist in chaotic data. The critical resource required is a source of many examples of the problem to be solved. With regard to the financial world, there is certainly no lack of chaotic information—every second of trading is available online.

  Evolutionary algorithms are adept at handling problems with too many variables to compute precise analytic solutions. The design of a jet engine, for example, involves more than one hundred variables and requires satisfying dozens of constraints. Evolutionary algorithms used by researchers at General Electric were able to come up with engine designs that met the constraints more precisely than conventional methods.

  Evolutionary algorithms, part of the field of chaos or complexity theory, are increasingly used to solve otherwise intractable business problems. General Motors applied an evolutionary algorithm to coordinate the painting of its cars, which reduced expensive color changeovers (in which a painting booth is put out of commission to change paint color) by 50 percent. Volvo uses them to plan the intricate schedules for manufacturing the Volvo 770 truck cab. Cemex, a $3 billion cement company, uses a similar approach to determining its complex delivery logistics. This approach is increasingly supplanting more analytic methods throughout industry.

  This paradigm is also adept at recognizing patterns. Contemporary genetic algorithms that recognize fingerprints, faces, and hand-printed characters reportedly outperform neural net approaches. It is also a reasonable way to write computer software, particularly software that needs to find delicate balances for competing resources. One well-known example is Microsoft’s Windows95, which contains software to balance system resources that was evolved rather than explicitly written by human programmers.

  With evolutionary algorithms, you have to be careful what you ask for. John Koza describes an evolutionary program that was asked to solve a problem involving the stacking of blocks. The program evolved a solution that perfectly fit all of the problem constraints, except that it involved 2,319 block movements, far more than was practical. Apparently, the program designers had neglected to specify that minimizing the number of block movements was desirable. Koza commented that “genetic programming gave us exactly what we asked for; no more and no less.”

  Self-Organization

  Neural nets and evolutionary algorithms are considered self-organizing “emergent” methods because the results are not predictable and indeed are often surprising to the human designers of these systems. The process that such self-organizing programs go through in solving a problem is also often unpredictable. For example, a neural net or evolutionary algorithm may go through hundreds of iterations making apparently little progress, and then suddenly—as if the process had a flash of inspiration—things click and a solution quickly emerges.

  Increasingly, we will be building our intelligent machines by breaking complex problems (such as understanding human language) into smaller subtasks, each with its own self-organizing program. Such layered emergent systems will have softer edges in the boundaries of their expertise and will display greater flexibility in dealing with the inherent ambiguity of the real world.

  The Holographic Nature of Human Memory

  The holy grail in the field of knowledge acquisition is to automate the learning process, to let machines go out into the world (or, for starters, out onto the Web) and gather knowledge on their own. This is essentially what the “chaos theory” methods—neural nets, evolutionary algorithms and their mathematical cousins—permit. Once these methods have converged on an optimal solution, the patterns of neural connection strengths or evolved digital chromosomes represent a form of knowledge to be stored for future use.

  Such knowledge is, however, difficult to interpret. The knowledge embedded in a software neural net that has been trained to recognize human faces consists of a network topology and a pattern of neural connection strengths. It does a great job of recognizing Sally’s face, but there is nothing explicit that explains that she is recognizable because of her deep-set eyes and narrow, upturned nose. We can train a neural net to recognize good middle-game chess moves, but it will likewise be unable to explain its reasoning.

  The same is true for human memory. There is no little data structure in our brains that records the nature of a chair as a horizontal platform with multiple vertical posts and an optional vertical backrest. Instead, our many thousands of experiences with chairs are diffusely represented in our own neural nets. We are unable to recall every experience we have had with a chair but each encounter has left its impression on the pattern of neuron-connection strengths reflecting our knowledge of chairs. Similarly, there is no specific location in our brain in which a friend’s face is stored. It is remembered as a distributed pattern of synaptic strengths.

  Although we do not yet understand the precise mechanisms responsible for human memory—and the design is likely to vary from region to region of the brain—we do know that for most human memory, the information is distributed throughout the particular brain region. If you have ever played with a visual hologram, you will appreciate the benefits of a distributed method of storing and organizing information. A hologram is a piece of film containing an interference pattern caused by the interaction of two sets of light waves. One wave front comes from a scene illuminated by a laser light. The other comes directly from the same laser. If we illuminate the hologram, it re-creates a wave front of light that is identical to the light waves that came from the original objects. The impression is that we are viewing the original three-dimensional scene. Unlike an ordinary picture, if a hologram is cut in half, we do not end up with half the picture, but still have the entire picture, only at half the resolution. We can say that the entire picture exists at every point, albeit at zero resolution. If you scratch a hologram, it has virtually no effect because the resolution is insignificantly reduced. No scratches are visible in the reconstructed three-dimensional image that a scratched hologram produces. The implication is that a hologram degrades gracefully.

  The same holds true for human memory. We lose thousands of nerve cells every hour, but it has virtually no effect because of the highly distributed nature of all of our mental processes.32 None of our individual brain cells is all that important—there is no Chief Executive Officer neuron.

  Another implication of storing a memory as a distributed pattern is that we have little or no understanding of how we perform most of our recognition tasks and skills. When playing baseball, we sense that we should step back when the ball goes over our field of view, but most of us are unable to articulate this implicit rule that is diffusely encoded in our fly-ball-catching neural net.

  There is one brain organ that is optimized for understanding and articulating logical processes, and that is the outer layer of the brain, called the cerebral cortex. Unlike the rest of the brain, this relatively recent evolutionary development is rather flat, only about one eighth of an inch thick, and includes a mere 8 million neurons.33 This elaborately folded organ provides us with what little competence we do possess for understanding what we do and how we do it.

  There is current debate on the methods used by the brain for long-term retention of memory. Whereas our recent sense impressions and currently active recognition abilities and skills appear to be encoded in a distributed pattern of synaptic strengths
, our longer-term memories may be chemically encoded in either the ribonucleic acid (RNA) or in peptides, chemicals similar to hormones. Even if there is chemical encoding of long-term memories, they nonetheless appear to share the essential holographic attributes of our other mental processes.

  In addition to the difficulty of understanding and explaining memories and insights that are represented only as distributed patterns (which is true for both human and machine), another challenge is providing the requisite experiences from which to learn. For humans, this is the mission of our educational institutions. For machines, creating the right learning environment is also a major challenge. For example, in our work at Kurzweil Applied Intelligence (now part of Lernout & Hauspie Speech Products) in developing computer-based speech recognition, we do allow the systems to learn about speech and language patterns on their own, but we need to provide them with many thousands of hours of recorded human speech and millions of words of written text from which to discover their own insights.34 Providing for a neural net’s education is usually the most strenuous engineering task required.

  I FIND IT FITTING THAT THE DAUGHTER OF ONE OF THE GREATEST ROMANTIC POETS WAS THE FIRST COMPUTER PROGRAMMER.

  Yes, and she was also one of the first to speculate on the ability of a computer to actually create art. She was certainly the first to do so with some real technology in mind.

  TECHNOLOGY THAT NEVER WORKED.

  Unfortunately, that’s true.

  WITH REGARD TO TECHNOLOGY, YOU SAID THAT WAR IS A TRUE FATHER OF INVENTION—A LOT OF TECHNOLOGIES DID GET PERFECTED IN A HURRY DURING THE FIRST AND SECOND WORLD WARS.

  Including the computer. And that changed the course of the European theater in World War II.

  SO IS THAT A SILVER LINING AMID ALL THE SLAUGHTER?

  The Luddites wouldn’t see it that way But you could say that, at least if you welcome the rapid advance of technology.

  THE LUDDITES, I’VE HEARD THEM.

  Yes, they were the first organized movement to oppose the mechanized technology of the Industrial Revolution. It seemed apparent to these English weavers that, with the new machines enabling one worker to produce as much output as a dozen or more workers without machines, employment would soon be enjoyed only by a small elite. But things didn’t work out that way. Rather than produce the same amount of stuff with a much smaller workforce, the demand for clothing increased along with the supply. The growing middle class was no longer satisfied owning just one or two shirts. And the common man and woman could now own well-made clothes for the first time. New industries sprung up to design, manufacture, and support the new machines, creating employment of a more sophisticated kind. So the resulting prosperity, along with a bit of repression by the English authorities, extinguished the Luddite movement.

  AREN’T THE LUDDITES STILL AROUND ?

  The movement has lived on as a symbol of opposition to machines. To date, it remains somewhat unfashionable because of widespread recognition of the benefits of automation. Nonetheless, it lingers not far below the surface and will come back with a vengeance in the early twenty-first century.

  THEY HAVE A POINT, DON’T THEY?

  Sure, but a reflexive opposition to technology is not very fruitful in today’s world. It is important, however, to recognize that technology is power. We have to apply our human values to its use.

  THAT REMINDS ME OF LAO-TZU’S “KNOWLEDGE IS POWER.”

  Yes, technology and knowledge are very similar—technology can be expressed as knowledge. And technology clearly constitutes power over otherwise chaotic forces. Since war is a struggle for power, it is not surprising that technology and war are linked.

  With regard to the value of technology, think about the early technology of fire. Is fire a good thing?

  IT’S GREAT IF YOU WANT TO TOAST SOME MARSHMALLOWS.

  Indeed, but it’s not so great if you scorch your hand, or burn down the forest.

  I THOUGHT YOU WERE AN OPTIMIST?

  I have been accused of that, and my optimism probably accounts for my overall faith in humanity’s ability to control the forces we are unleashing.

  FAITH? YOU’RE SAYING WE JUST HAVE TO BELIEVE IN THE POSITIVE SIDE OF TECHNOLOGY?

  I think it would be better if we made the constructive use of technology a goal rather than a belief.

  SOUNDS LIKE THE TECHNOLOGY ENTHUSIASTS AND THE LUDDITES AGREE ON ONE THING—TECHNOLOGY CAN BE BOTH HELPFUL AND HARMFUL.

  That’s fair; it’s a rather delicate balance.

  IT MAY NOT STAY SO DELICATE IF THERE’S A MAJOR MISHAP.

  Yes, that could make pessimists of us all.

  NOW, THESE PARADIGMS FOR INTELLIGENCE—ARE THEY REALLY SO SIMPLE?

  Yes and no. My point about simplicity is that we can go quite far in capturing intelligence with simple approaches. Our bodies and brains were designed using a simple paradigm—evolution—and a few billion years. Of course, when we engineers get done implementing these simple methods in our computer programs, we do manage to make them complicated again. But that’s just our lack of elegance.

  The real complexity comes in when these self-organizing methods meet the chaos of the real world. If we want to build truly intelligent machines that will ultimately display our human ability to frame matters in a great variety of contexts, then we do need to build in some knowledge of the world’s complications.

  OKAY, LET’S GET PRACTICAL FOR A MOMENT. THESE EVOLUTION-BASED INVESTMENT PROGRAMS, ARE THEY REALLY BETTER THAN PEOPLE? I MEAN, SHOULD I GET RID OF MY STOCKBROKER, NOT THAT I HAVE A HUGE FORTUNE OR ANYTHING?

  As of this writing, this is a controversial question. The security brokers and analysts obviously don’t think so. There are several large funds today that use genetic algorithms and related mathematical techniques that appear to be outperforming more traditional funds. Analysts estimate that in 1998, the investment decisions for 5 percent of stock investments, and a higher percentage of money invested in derivative markets, are made by this type of program, with these percentages rapidly increasing. The controversy won’t last because it will become apparent before long that leaving such decisions to mere human decision making is a mistake.

  The advantages of computer intelligence in each field will become increasingly clear as time goes on, and as Moore’s screw continues to turn. It will become apparent over the next several years that these computer techniques can spot extremely subtle arbitrage opportunities that human analysts would perceive much more slowly, if ever.

  IF EVERYONE STARTS INVESTING THIS WAY, ISN’T THAT GOING TO RUIN THE ADVANTAGE?

  Sure, but that doesn’t mean we’ll go back to unassisted human decision making. Not all genetic algorithms are created equal. The more sophisticated the model, the more up to date the information being analyzed, and the more powerful the computers doing the analysis, the better the decisions will be. For example, it will be important to rerun the evolutionary analysis each day to take advantage of the most recent trends, trends that will be influenced by the fact that everyone else is also using evolutionary and other adaptive algorithms. After that, we’ll need to run the analysis every hour, and then every minute, as the responsiveness of the markets speeds up. The challenge here is that evolutionary algorithms take a while to run because we have to simulate thousands or millions of generations of evolution. So there’s room for competition here.

  THESE EVOLUTIONARY PROGRAMS ARE TRYING TO PREDICT WHAT HUMAN INVESTORS ARE GOING TO DO. WHAT HAPPENS WHEN MOST OF THE INVESTING IS DONE BY THE EVOLUTIONARY PROGRAMS? WHAT ARE THEY PREDICTING THEN?

  Good question—there will still be a market, so I guess they will be trying to out-predict each other.

  OKAY, WELL MAYBE MY STOCKBROKER WILL START TO USE THESE TECHNIQUES HERSELF. I’LL GIVE HER A CALL. BUT MY STOCKBROKER DOES HAVE SOMETHING THOSE COMPUTERIZED EVOLUTIONS DON’T HAVE, NAMELY THOSE DISTRIBUTED SYNAPTIC STRENGTHS YOU TALKED ABOUT.

  Actually, computerized investment programs are using both evolutionary algorithms and neural
nets; but the computerized neural nets are not nearly as flexible as the human variety just yet.

  THIS NOTION THAT WE DON’T REALLY UNDERSTAND HOW WE RECOGNIZE THINGS BECAUSE MY PATTERN-RECOGNITION STUFF IS DISTRIBUTED ACROSS A REGION OF MY BRAIN ...

  Yes.

  WELL, IT DOES SEEM TO EXPLAIN A FEW THINGS. LIKE WHEN I JUST SEEM TO KNOW WHERE MY KEYS ARE EVEN THOUGH I DON’T REMEMBER HAVING PUT THEM THERE. OR THAT ARCHETYPAL OLD WOMAN WHO CAN TELL WHEN A STORM IS COMING, BUT CAN’T REALLY EXPLAIN HOW SHE KNOWS.

  That’s actually a good example of the strength of human pattern recognition. That old woman has a neural net that is triggered by a certain combination of other perceptions—animal movements, wind patterns, sky color, atmospheric changes, and so on. Her storm-detector neural net fires and she senses a storm, but she could never explain what triggered her feeling of an impending storm.