|
| Beyond Humanity |
|
"Whatever one man is capable of imagining,
other men will prove themselves
capable of realizing."
- JULES VERNE
When Teddy Roosevelt decided to mediate peace between Japan and Russia, scientists knew little about how his brain came to that conclusion. After all, that the brain is the center of all thought had only recently been universally accepted; the long-held belief that the heart is the center of emotions died a hard death. Today, we know a lot more about our brains, but still only a fraction of what there is to know. The goal of figuring out the complexities of the human mind is often thought to be a tough nut to crack. The late scientist-philosopher Heinz Pagels believed it was something we humans "would be coping with for the next centuries." It will take centuries to understand our mind machines only if the knowledge curve grows arithmetically. But it will not.
Growth in knowledge has been exponential; we learned half of what we know about brains in the last decade as our ability to image brains in real time has improved keeping in step with the sophistication of brain scanning computers. For the sake of argument, let's assume we know just two percent of what there is to know about our brains right now. With the knowledge base doubling every ten years or so, we may know most of what can be known in about half a century. The more we know about natural brains, the easier it will be to synthesize them, although we do not necessarily need to understand it all before we can replicate the mind.
Opinions on how hard it will be to make machines think often depend on one's point of view. Some physicists and computer scientists think it will not be so hard. Apply some basic principles to the problem, and away we go. The problem with these fellows is that they work with systems that are rather simple. Atoms, stars, light photons, and bi-digital computers are simple compared to creatures and their brains.
At the other extreme of brain replication reactions are biologists, many of whom love to wax virtual poetry about the extreme complexity of life and how it was built over billions of years of intricate evolution. Of living systems, the human brain is the most complex. Their conclusion is that it will be extremely hard and very long before our crude, inept technology can hope to devise a machine as conscious and capable as a brain. This view is as naïve as is the view that the mind is simple.
The structure and thinking power of the human brain has been the same for tens of thousands, if not hundreds of thousands, of years. The average size was and is the same, 1,500 cubic millimeters, and the shape appears to have stayed the same. Art masterpieces were created on cave walls deep in the pitch black earth tens of thousands of years ago. As far as we know, if you time-snatched a toddler from 100,000 years ago put her into preschool and raised her in modern society, she would grow up to be a typical modern-day person.
People who have studied "primitive" gatherer-hunter societies that used stone tools have found that they are not so primitive. The amount of knowledge crammed into the brain of a hunter is as extensive as it is subtle. Researchers who have accompanied aboriginal hunters on their rounds have been bewildered by how they find, identify, and track prey they may not visually encounter for hours on end. The ability to find hidden plant foods is equally remarkable. It is probable that the capacity of the average individual mind is fully taxed in a gatherer-hunter tribe. As the knowledge contained in increasingly large and technologically advanced societies has risen dramatically, the capacity of individual human minds has not expanded to accommodate the new knowledge. We have been forced to spread out chunks of the new knowledge among a multitude of minds taught over extended periods of education. Humans have also been forced to invent written tablets, books, computers, and the like to store and transmit a knowledge base too large to store even in billions of brains.
The physicists and computer scientists who think consciousness is a simple affair probably are naïve. They may be surprised at how hard it will be to make machines think. The biologists, and philosophers, who think consciousness is an extremely hard affair, are naïve in part because they tend to overstate the problem, but mainly because they understate the power of future technologies. The naturalists may be surprised at how soon the power of technology can converge with the brain. In other words, both sides may be surprised. It looks as if we have a classic case of over-prediction of the near-term (by many computer scientists) and under-prediction of the long-term (by many biologists), which Arthur C. Clarke warned against.
Both sides may be surprised because the brain is a paradox. We have already seen that the human brain is the most complex object in the universe, yet is simple in design. Your brain is the most powerful information processor in the universe. Its circuitry is also pathetically slow. The full complexity of the brain is necessary to achieve its level of cognitive performance only if it is an optimized system, only if it achieves its mental power with the most simple form possible. Because the brain is the product of noncognitive Rube Goldberg evolution, this may not be the case. Indeed, it is hardly likely that the human brain is optimized in terms of simplicity and performance. At every evolutionary stage, the system had to make do with what was already present; a radical transformation to a dramatically new, potentially superior form was impossible. The brain may, therefore, consist of jerry-rigged add-ons getting the job done at a reasonable, but not ultimately efficient, cost. Also consider that the human brain is the first of its kind. It is hardly likely that the first system will be the ideal one. After all, our brains have not been subjected to the competition of other brains of similar performance, forcing a higher level of efficiency.
The simplicity of key parts of the brain suggest that artificial devices of similar performance can remain fairly simple in at least some important parts. If the brain is not optimized in terms of performance versus complexity, it is possible that artificial devices simpler than the brain can match it in performance. Both factors suggest that artificial brains will not have to have the structural complexity of the human brain to be as good or even better.
Consider the brain has some gross deficiencies, such as the sugar-oxidizing power system, which is pathetically weak. The eye's retinas, literally extensions of the brain, are placed behind the nerves and blood vessels that sustain them rather than in front where they would have a better view. Brains upload and store information via sensory systems slowly and sloppily. It takes hours or days to read a novel, years to learn a language or play an instrument well, and an advanced education requires decades of hard study. The learning-retention system is so lousy that we have trouble remembering something as simple as a long-distance phone number. The limitations of the human system will be explored in more detail later in this book.
For the rest of the discussion, we will presume that conscious minds are produced by machines that process information via algorithms. This excludes Penrose quantum effects, but leaves room for alternatives. We will get to the form of computer that may be necessary for thinking later, but first we must consider some minimal requirements for matching the gross thinking performance and physical parameters of the human brain. These are adequate calculating speed and sufficient storage capacity all packed into devices that are small and cheap enough to plunk inside mobile robots.
Of course it is! At the level of basic principles, we are rather like the Soviet physicists in the late 1940s who confidently knew they could build an A-bomb because it had been done already. Likewise, today, we know that small, high-speed, high-capacity computers are workable because we each have one between our ears. As obvious as this is, every few years someone starts wringing their hands, warning that we will soon reach some practical limit to making computers smaller, faster, and more intricate. Do not listen; they've been saying the same thing for 40 years. Every time, the barrier has been bypassed by incremental improvements in the conventional technology or by revolutionary shifts and dramatically new ways of manipulating information. The shift from vacuum tubes (remember those electronic clunkers?) to transistors was a revolutionary transition. The persistent improvement in semiconductors is an example of the incremental, yet swift, transition. There are many reasons to be confident that this combination of revolution and incremental advancement will continue to the point where computers are faster and more sophisticated than organic brains. It may be possible to go well beyond brains and make ultra-computers that work at the level of atoms and molecules.
In contrast to the evolutionary rut our brains are stuck in, it's all too obvious that the power and sophistication of computers is growing exponentially. We will assume that to think as fast as brains, an equivalent machine will have to solve algorithms at approximately the same pace, which we think would be somewhere between 10 to 1,000 teraflops. No matter how brain-like in design, structure, and function we can make a computer, it won't out-power a brain with mere thousands of calculations per second. As for the possibility that speeds well below those of inefficient brains will suffice or the possibility that brain speed has been underestimated, it makes little difference because the speed of computers is climbing so fast. There are further requirements.
In his 1988 book Mind Children, Hans Moravec charted the increase in computer power in our century. It started with early human-operated mechanical calculators in 1900. This combination of man, or usually woman, and machine could do a calculation every dozen seconds or so. Actually, this was not much faster than using a pencil. Things did not improve a whole lot until the 1910 analytical engine (not to be confused with Babbage's nineteenth-century dream machines) did a sizzling one calculation per second, or one "flop" (a "floating operation" is about the same as a calculation, which averages about 8 bits worth of information). As long as people were punching keys, speed didn't pick up much, so we have to go to 1939 when the BTL Model 1, using relays, did five flops. It was the first bi-digital computer and marked a milestone in computer evolution. But it used telephone relays to do its calculations and was limited to the achingly slow speeds possible with gross mechanical action.
Action was fast becoming the name of the game. The next few years saw the first great punctuated leap in computer evolution. Intense wartime pressure to calculate the ballistic tables for all the new kinds of artillery rounds being fired from all the new guns forced the decision to build a "supercomputer" that would use electronic switches to achieve speeds 1,000 times higher than those of physical relays. One year after the war, 50 years ago, the famous ENIAC breathed the word "computer" into the collective consciousness.
|
ENIAC was an unlikely candidate for such a delicate task; a 30-ton behemoth (when big seemed impressive), it was capable of astonishing speeds, more than 7,500 calculations per second (8 kiloflops). Old photographs show this vacuum tube computer taking up a large room with processors connected by thousands of cables. The 18,000 vacuum tubes and supporting electronics sucked up so much electricity (150 kilowatts) that the lights in the building dimmed when ENIAC came online. Eight hours of calculations by a person using an adding machine could be performed by ENIAC in 20 seconds! Reprogramming the beast required replugging most of the cables. The vacuum tubes were constantly bursting and shutting down the bulky wonder of the age. A government study estimated that a dozen of the machines would more than fill the computing needs of the nation! The future of computers was so obscure that only a major battle at IBM got the company going digital in the early '50s.
By 1951, UNIVAC was online. Also made of vacuum tubes like the ENIAC, it was less of a programming nightmare because John von Neumann figured out a better way to configure memory systems (your computer uses the same system). UNIVAC could do more than 12 kiloflops. It was this computer that became famous by correctly projecting a landslide win for Eisenhower early on the evening of election day, a projection that was rejected by the network because the human "experts" knew Stevenson was going to make it a close race!
Now electrodigital computing entered the rapid growth era we still enjoy. Four years after UNIVAC appeared, a tenfold leap in calculating speed was made with the respectable 125-kiloflop Whirlwind, the last of the great vacuum tube machines. A switch to smaller, cheaper, and more reliable transistors resulted in IBM's 875-kiloflop 7090 machine in 1959. The growth continued with transistors. In 1961, the Atlas did 3.8 million calculations per second, or 4 megaflops. 1964's CDC 6600 was a 25-megaflop machine. The CDC 7600 of 1969 could run at a maximum speed of about 60 megaflops (note that the average running speed of a computer is significantly less than its optimum running speed).
The 1970s saw another critical leap in computing power. The advent of integrated circuits, or silicon chips, allowed the development of the first supercomputers primarily by Robert Cray. In 1977, the Cray-1 could do an amazing 375 megaflops. The Cray-2 in 1985 could do in the neighborhood of a billion calculations each second, or a gigaflop. Exotic gallium arsenide chips replacing regular silicon in 1990 allowed the Cray to do about 10 gigaflops. Currently, the Cray T90 can perform 60 gigaflops and recently, two Intel machines were joined to run more than 250 gigaflops. Today's supercomputers can accomplish in less than a second what ENIAC took a year to do.
When we graph the higher speeds of 20th century calculating machines, we get - no surprise - an exponential curve with its classic slope. On average, the power of calculating devices increased about twentyfold every decade since 1900. We have plotted the data both arithmetically, to show the spectacular growth of the curve, and exponentially to straighten the same growth line and make it easier to see the details. Now, 20 multiplied by itself nine times with a little added for the last five years is about a trillion, in other words, the power of calculation has grown an astounding trillion times in less than 100 years! Over the last 50 years, computer speed has expanded some ten millionfold, which works out to two dozenfold increases each decade on average. In the last ten years, the semi-conductor industry has hustled to the point that maximum computer power has increased about sixtyfold. The amount of computer speed that can be bought for a dollar is doubling about every 18 months these days.
So far, we have looked at the big mainframes that sit in the centers of large rooms, cost a bundle to make, run up huge electrical bills, need lots of high-tech TLC, and are built to crunch numbers for weather forecasting and aerodynamic studies. What we need for synthetic mind repositories are small, cheap, efficient machines that can take a bounce or two across the floor and be linked up with robots. How have little computers been doing in terms of getting up to speed?
In the middle of the swinging '70s, the first PCs could do a couple of hundred thousand flops, only a minuscule fraction of which is used when you are typing. In the more insipid mid-'80s, about the time it became possible to see a full page-width of text on a cheap PC, home computers could do a megaflop or so. The hot 486 of a few years ago could perform ten megaflops. Today's Pentiums and PowerPCs can beat 100 megaflops. The power of small computers and the chips that run them has gone up about 1,000-fold since they first came into existence, and more than a hundredfold in the last ten years or so. In this respect, PCs have matched the growth rate of mainframes. The PCs have lagged consistently behind the mainframes; it takes about 15 to 20 years for a single common chip to reach speeds once reached only by big machines.
There is one important regard in which the machine has already far surpassed the brain. The telephone relays that worked the first bi-digital computer clicked along only a few times faster than a calculation per second, but even the old vacuum tubes worked about as fast as neurons at a hundred or so cycles per second. Any ordinary modern chip circuit can purr along at many millions of cycles each second.
What about the future? Will past trends continue into the future, or is the growth in computer power about to hit a plateau, forever leaving human-made machines far behind human brains? Certainly, we have a large gap to fill. Today's big supercomputers are something like 100 to 10,000 times slower than a brain, and the laptops are about 50,000 to 5,000,000 times slower than a brain. It is not surprising that no robot can yet do what humans do because they do not have anywhere near the thinking power.
We have discussed how exponential growth is a confusing concept to many people, leading them to over-predict the near future and under-predict the longer-term. Does the exponential growth curve of computer power during this century help us tell where, or whether, we under- or over-predict the course of computer power in the next few decades? The 1968 movie, 2001: A Space Odyssey over-predicted the superintelligent mainframe HAL in the 1990s. We can see why by using the curve. Back in the '60s mainframe computers were a million times less powerful than a brain. The rise of computer power had only just begun to climb above the bottom of the graph. It was impossible for computers to close such an enormous gap in 30 years. That the movie made this mistake is not surprising; no one made any realistic estimates of brain speed at the time, so the enormity of the gap between machine and brain was not understood. Today, we have a better idea of the power level we are reaching for, and growth of computer power has soared until it is nearly vertical. The latest mainframes are about 1,000 times below brain speeds and it will take another 20 years or so to close the gap, with small computers following close behind. The log plot of growing computer power shows how close we are getting to brain speed levels. Looking from the 1990s to three or so decades forward is not at all like looking from the 1960s to thirty years forward. We are much closer to the goal, and the computers are speeding up much faster.
To look at it another way, we have already come a trillion times closer to human brain speed since the turn of the century, and the gap is now only about a thousandfold! Since computers were invented, computer power has grown a thousandfold every 20 years. Human-speed computers should be online in about 20 years. Small hyper-computers suitable for robots can be expected in three dozen years from now. These estimates are in good agreement with calculations by Hans Moravec in Mind Children. In 1988, he predicted that small, cheap computers as fast-thinking as the human brain would be produced by about 2030. AI researchers Maureen Caudill and Daniel Crevier have made similar projections; the trend is too clear and obvious to ignore. We are approaching human brain power so quickly that it has become difficult to over-predict the near future.
The projected speed increase is not speculation just based on lines on a graph. The practical pressure to increase the speed of computers closer to brain levels is intense. Starting around 1990, a number of institutions became frustrated with the slow speed of current supercomputers and began to clamor for "ultra" machines that could do the one trillion calculations per second needed to model the aerodynamics of aircraft in extremely realistic detail or chew through weather data for dramatic improvements in forecasts. In the '80s, such a teraflop machine was something not expected until well into the 21st century. Just a few years ago, folks got a bit over-optimistic and figured a teraflop machine would be up and running by last year. What is happening is that a 1.8-teraflop machine, TFLOP, is being built at Sandia National Laboratories for $46 million. It will be used for simulating exploding nuclear weapons, climates, and advanced materials. It should be online late this year.
But surprise strikes again. In Japan, it was just announced that a specialized research computer, GRAPE-4, can already do a trillion calculations per second!
More ordinary demands are putting the pressure on for small, cheap ultracomputers anyone can afford. As cool as today's computer games may be, they still look like souped-up digital cartoons. Three-dimensional virtual reality is even worse. One of the authors has a fear of heights, and when he recently "walked" off the top of a VR skyscraper and plunged to the ground in 3-D, he was both relieved and miffed to experience not a whiff of vertigo (that the goggles were typical heavy clunkers and the two TV screens were small low-resolution affairs did not improve matters). What we are waiting for is a VR game in which you see the same all-around view as did the dashing Admiral Beatty as he stood on the open bridge of his flagship "Lion," at Jutland in 1916, with dozens of coal smoke-spouting ships spread from horizon to horizon, guns flashing, enormous shells whizzing through the air, and the Admiral's precious battle cruisers blowing up before his eyes. We have already noted that to produce a fully realistic
3-D image that fills the entire visual field requires a computer doing about a trillion calculations per second. Motorola and Intel are about to release chips that perform a few hundred megaflops; they figure they will have chips running at a gigaflop by 2000 or so, and the ways things are going, teraflop chips ought to be going into home computers in the second decade of the next millennium.
A teraflop chip doing heavy VR is pretty impressive, but it is still not at the human level of 10 to 1,000 teraflops. How long will it be until you can shop for a new brain, as it were? The aerodynamic engineers and weather folks have already decided that a teraflop is not going to be good enough. The astronomers who enjoy simulating exploding stars, colliding galaxies, and the end and beginning of the universe sneer at a mere trillion calculations per second. So they are already talking about petaflop machines capable of doing 1,000 trillion calculations every second, which happens to be the high-end estimate of human brain power. Let's figure this out: A petaflop is 10,000 times faster than anything we have online these days, so with the speed of top-of-the-line computers doubling every 18 months or so, it should take 20 to 25 years to build a room-filling petaflop machine. But perhaps we are too pessimistic. The same Japanese group that brought us GRAPE-4 hopes to link 20,000 processors into a $10 million petaflop computer by 2000 or so! As for the cheap and portable, if they continue to lag about 15 years behind the big guys, you should be able to drop in at your local computer store and pick up a petaflop device in, say, 35 to 45 years. If, on the other hand, a 10-teraflop machine will do, they should be available sometime around 2020.
We challenge skeptics to come up with a list of people who, 30 years ago, foresaw 100-megaflop chips in wee little computers costing a few thousand dollars for sale in this century. There would have been people happy to explain in boring detail why today's laptops were impossible to build. If we assume computer power will remain below human brain levels well into the next century, the projected growth curve takes a sharp turn from vertical toward horizontal for the first time in decades, just when our knowledge of how to make computers ever tinier and faster is reaching new levels. This may not be impossible, but it is far from a safe prediction. It is more likely that a sudden breakthrough could produce hypercomputers earlier than we predicted.
We have estimated the future growth of computer power by extrapolating past trends into the future. This is a powerful method, but not the only reason to believe hypercomputers will be in labs and homes very soon. Some of the researchers have set up an online database intended to pull together the varied disciplines speeding up the development of such machines. Our predictions are not extreme; rather, they are mainstream. We have extrapolated long-term past trends a few decades forward and found that they are in line with the current plans of the computer industry.
A brain-like machine with lots of speed requires ready access to enough stored memories to have a sense of identity at the human level. Considering the data density of visual memories, it is not likely that such a system can get away with less than what a brain needs, about 1,000 trillion bits.
Computer memory capacity improves at about the same pace as processing speeds. From almost nothing to fantastic in a few decades, exponential growth of computing power should reach human brain levels in a few more decades. We are more interested in the random access memory contained inside the computer rather than the data stored on peripherals. ENIAC could store about 1,000 bits, UNIVAC 40,000, a mid-'60s mainframe a few tens of millions of bits. The first Cray stored 300 million bits, and the Sandia beast will probably hold a few hundred billion bits. At this rate, the memories of the most powerful machines should reach the human range around 2015. As usual, there is a two-decade lag between the memory levels of top-end machines and those you can afford.
Let's not, however, ignore the peripherals. One advantage of an artificial brain will be the ability to link its information processor directly to precise digital memories, greatly expanding the available database. No more accessing external data through the senses. So far, memory storage has been hampered by its two-dimensional nature, first as rust on tapes and then as rust on disks. Magnetic storage is early-20th-century technology, and its continued use is something of a low-tech embarrassment. To achieve brain-level memory storage capacity, it will be necessary to do what the brain does: store information in three dimensions, preferably in a holographic manner. The first step along these lines may be read and write CDs storing data in a single disk on multiple levels. A library of books could reside on such a disk. But disks need to be spun at high rpms, a crude affair when things ought to be done electrically with no moving parts. The next step may be installing light-sensitive crystals that use little lasers to code molecules storing and retrieving data holographically. Because the data is stored in three dimensions, terabytes of information could be stored in a space the size of a child's wooden block. Even bigger libraries can be stuffed into a crystal the size of a D-battery and, look Ma, no moving parts! Better still, the data can be up- and downloaded thousands of times faster than with spinning disks.
The exponential increase in the calculating speed and memory capacity of computing machines has been paralleled for an equally dramatic decline in cost. ENIAC cost a few million in today's dollars and could do about a billion calculations over two years, so each dollar bought about 500 calculations. Not bad, but these days a $2,000 machine will do a few thousand trillion calculations if you push it to factor primes for a couple of years. So you get about a trillion calculations for your buck! Quite a deal.
It costs a few thousand dollars to raise and support a person in the third world and a few hundred thousand to raise and support a Ph.D. candidate. For 20 years, a Ph.D. employee plus computer will do about ten million trillion calculations, so a dollar buys somewhere around ten trillion calculations. Computers doing calculations at such speeds and at a similar cost should be available in three decades, plus or minus a number of years. After that milestone has been passed, the cost of doing calculations should quickly drop below human standards.
Computers are doing what they are doing because, in the larger scheme of things, it is relatively cheap and easy to improve them, and those who do the improving get a lot of money for the effort. Computers are small, and relatively small groups can tinker with them in relatively small labs. Big discoveries and innovations can be worked up by only a few people with a modest amount of money. Huge rocket boosters needed to go into space or the enormous tunnel machines and big reactors needed by particle and atomic physicists aren't important in our "small-is-beautiful" domain.
With computers, smaller is better. What it all comes down to - the stuff about speed, memory, size, and cost - is information density. The higher the density, the smaller the machine can be. The brain has a very high information density; all that speed and memory packed into just 1.5 kilograms. In the outer cortex, there are a few hundred million neurons per square inch, and the density of synapses in the cortex is about ten times higher. The latest commercial chips have a few million circuits per square inch. Because a modern chip is a thousandfold short of the information density of a brain, even TFLOP will be hundreds of times bigger than a brain, although its overall power will be less. Much better than ENIAC, though, which had one vacuum tube for every few square inches!
As information density increases, cost goes down. It is cheaper to manufacture something the size of a chip than a room full of vacuum tubes. The smaller a computer, the cheaper the machine is to assemble, so costs spiral downward. The physical needs of high-speed computing are also driving down the size of fast machines. The smaller the machine, the shorter the distance between circuits, so the less time it takes for them to communicate, so the faster the machine is. For example, one plan for making the first 10-teraflop computer envisions a hybrid technology combining protein molecules and semiconductors in a package one-fiftieth the size of a Cray mainframe using 100 times less power. The advantage of building the fastest computers as small as possible is one reason that the time lag between mainframes and portable computers of the same power continues to decrease.
However, there is a cost to making computers smaller. A single vacuum tube was simple and cheap. A high-end chip of smaller dimensions is more intricate and expensive. This is because a chip is orders of magnitude more intricate than any tube, and it takes lots of high-tech machinery and know-how to make the chip. Designing a new high-end chip today can cost a billion dollars, and building the factory to make them can cost another billion or two. Such expenses have driven smaller chipmakers out and renewed predictions that we are about to hit a plateau in chip development.
What we are really seeing is a consolidation of the industry comparable to the shift from hundreds of carmakers in the first decades of this century to the Big Three. A completely redesigned new car model costs a billion or so to develop and produce; even a sports stadium can cost half a billion. Development of a new airliner costs so many billions that it threatens the finances of even a big company such as Boeing. A new airport or subway costs multibillions, and eight miles of superhighway runs a billion dollars these days. One reason that new chip factories cost a billion is because they are big, and they are big because it is efficient to be so and because demand for chips is so high. The chip market these days amounts to hundreds of billions of dollars. There is lots of money to be made in chips, and to make the big bucks the chips have to be as fast as, if not faster than, anything else. People are more than willing to put down hard cash in exchange for more computational power, and the investment/return ratio is very favorable, unlike space flight, which is just a bottomless money pit. So competition is fierce, and money is skimmed off the big cash inflow to make better chips. Although the cost of these chips is substantial and they carry so much more information than the vacuum tubes of old, the deal is almost infinitely better.
With the post-cold war Department of Defense in decline, deficit-busting in vogue, and corporate downsizing all the rage, concern has been expressed about an investment decline in new, radical computer technologies. The huge monetary investment already sunk into silicon chips acts to dampen any major diversion to alternative, potentially more effective, technologies. The long-awaited teraflop ultracomputer will be made of assemblies of commercial chips rather than some standards, upsetting new and exotic design. The tendency away from exotic computer concepts is more fertile ground for predictions that computer performance is about to hit a plateau, and it is possible, if not likely, that a modest decline in the growth of computer technologies will occur. However, this will be part of a regular cycle, not a cause for panic. After all, science continued to progress rapidly during the Great Depression, and things are certainly better than that today. Economies are growing rather than contracting, and pressures and incentives to improve computers remain high. Intense global competition is spurring computer evolution; no one can afford to be left behind. As China comes online as a modern economic power, it can be expected to boost computer research in the next century. And lest we forget, probably 95 out of 100 computer scientists and engineers who have ever lived are still tinkering today.
Silicon semiconductors have been pushed far beyond where many thought they could be pushed just ten years ago, and they still have some good years left. But melted sand appears to have limits. How small can it be fashioned and still perform effectively as switches and gates? For a while, it seemed gallium arsenide, that stuff Cray liked to make supercomputer chips out of, might give silicon some competition, but this tricky material has proved a commercial bust. Silicon-germanium combines compatibility with traditional silicon technologies with superior micro-performance and may displace silicon, at least for a while.
Ever tinier circuits bring us to another modern technology threatening to run out of post-modern steam: circuit-making via etching. Chipmakers use light to etch the circuit pattern onto silicon chips. Problem error rates are high, and avoiding the smallest dust contamination is a chronic nightmare (which is why chips are made in clean rooms). Basic physical limitations soon will make it impossible for photo lithography to pack yet more circuits onto a given area of chip. A given color of light can etch points and lines only as small as its wavelength, and even smaller blue wavelengths are much bigger than atoms and most molecules. There is talk of using X-rays to etch the chips because the shorter wavelengths will cut finer than visible light. Even if this can be made to work, and the low cutting power of X-rays casts doubt on this, the dust contamination problems will be worse than they are with light etching. A very different technology proposes to stamp out chips rather than etch them. The process avoids the contamination problem and should cut costs, but the technique has not been converted to commercial use.
|
Currently, silicon semiconductors can be etched with circuit lines about one micron in width, about 100 times thinner than a hair. It is believed that the technology can be pushed until lines are one-tenth of a micron in diameter. Just how impressive this is becomes obvious when it is considered that neurons are about 20 microns across and the axons that connect them are one-tenth to one micron in diameter. Conventional semiconducting technology will soon be able to produce some components as small as biology, but not the nanoscale structures and operations that go on inside neurons.
Eventually, the conventional semiconductor industry will run out of silicon steam. The swift increase in computer speed is about to reach a practical barrier, eventually reaching a plateau much as the rapid evolution of airliners came to a sudden halt 30 years ago. This analogy is inappropriate. Airliners are unnatural machines with no close, living equivalent showing the way to dramatic improvements. Computers will have to become more like brains if we are to exploit their full potential. Rather than stopping the great computational increase, the need for ever-higher speeds and smaller circuitry is likely to force a dramatic switch away from the top-down process of etching silicon. Replacing it with bottom-up nanotechnologies will grow a suitable material (diamond, or harder-strength-designed materials, is a strong contender) into extremely small information processors. If and when this happens, silicon chips and the semiconductor industry making them will go the way of vacuum tubes. As important as the semiconductor industry has been to cyber evolution it is not likely to form the substrate for the Extraordinary Future. Such a switch should not be surprising; it will be an example of technological punctuated equilibrium.
|
It is possible that the smooth slope increase in computer power will become irregular as new ways of boosting computer data density come online, only to be bumped off by better technologies. Such a situation might slow down the overall increase, but it also opens up the possibility of sudden jumps to extremely powerful compact machines in the near future.
Consider one possibility for such a jump, the electron spin computer. An ESC would exploit quantum mechanics to bring order to the random spin of electrons and turn the tiny particles into on-off switches. One of the marvelous things about ESCs is that the smaller they are, the better they work! ESCs would be made of conductive metals rather than semiconductors, and elements may be as little as one-hundredth of a micron across, or the width of a hundred atoms. A trillion transistors could be crowded onto one chip. The information densities of quantum computers would surpass those of brains many times over.
So far, it has taken small computers years to catch up with the big boys. This may change if the most sophisticated computers are nanobuilt machines in which switches are made from a few molecules. It may be necessary to make computers small to keep the wiring as short as possible and thereby minimize the data transmission time between circuits. Of course, a hypercomputer the size of a bread box will always be cheaper to make than one the size of a room, so cost savings may also drive the big squeeze. When the high-end computers bow out, the size-related performance lag will be eliminated. This may mean that the advent of small petaflop devices will be pushed forward.
Basic technologies for making computers capable of rivaling human brains in speed, capacity, circuit miniaturization, cost, and size will be online in 20 to 50 years. It is improbable that it will take much longer - well over half a century to a few centuries - to reach these targets because the economic forces pushing such capacity are strong and because there are no apparent barriers to such information density. However, the development of computers remains subject to unpredictable evolutionary forces, including surprising leaps in technology. Our estimates of the growth of speed, capacity, and complexity are conservative. They assume the complexity of computers will continue with the same top-down development manner practiced today. Techniques for evolving and growing computers from the bottom up may spring up as if out of nowhere, giving us computers of extreme calculating power and memory capacity. It is hard to rule out a sudden leap to human levels before the curve suggests. What this all means is that some of the minimum requirements for making mind machines will probably be met in the near future.
Just because small computers as fast-working as the human brain is may be available around 2030 does not necessarily mean it will be possible to run a conscious mind on one of them! They may just be faster and fancier versions of the number crunchers we already have. Or maybe not.
The belief that small computers will match the human brain in gross speed and capacity within a few decades is not especially controversial. Where opinions diverge is whether these systems will be as capable as the human brain; especially whether they will be conscious to any degree.
The field of AI has always been contentious; the idea of AC is more so. It's true that people have been ostracized at certain institutions for their ideas. Promising areas of research have been suppressed, such as the near-disappearance of neural networks, for more than a decade. Yet when people are outraged at the other guy's idea or skepticism of an idea, a call to neurons is spread far and wide leading to increased interest and creative output by motivated individuals and institutions. A general consensus could be wrong. It's better to ensure a diversity of competing ideas, increasing the possibility that someone may be on the right path. A clear consensus has emerged that classic, strong AI will not succeed. Programming computers to be as smart as we are is considered a nightmare of tedious complexity. Beyond this minimal consensus, just about anything goes, and the argument follows two main lines of thought.
The evolutionary AI folks trust that as the gross computational power of computers rises and exceeds human levels and as software programs evolve, exploiting the full capacity of these machines, the equivalent of human intelligence will be at hand, if not met. Hans Moravec is one leading proponent of this view. AI should be developed in a manner analogous to human flight. Human flight has a strong top-down component in the use of fixed wings rather than flapping wings. If the combination of mimicking and diverging from nature worked for getting into the air, why not for getting to high intelligence?
Getting high-level intelligence out of evolutionary AI is one thing, but getting AC may be another. Some AI researchers think smart hypercomputers will mimic human intelligence, but will not be aware they are doing so. This view assumes that even the highest levels of intelligent action are not dependent on being conscious. Others subscribe to the argument that consciousness inevitably emerges from high levels of cognition. In the latter case, the form of the mind machine is irrelevant except that it must be powerful enough to run the mind on.
If the evolutionary AI argument were correct, there should be a close correlation between robotic computer speed and the capabilities it produces. The simple action-reaction behavior of single-cell creatures has been mimicked by simple robots with low power chips (which discredits the assertion by Stuart Hameroff that the supposedly complex behavior of paramecium shows that they must have a quantum supercomputer running in their microtubules). These days, modern computer chips have computational speeds similar to those of insect brains, yet no one has built a robot as agile and intelligent as a fly.
Consider the ordinary housefly, a tiny self-guided biomechanical marvel causing robot builders and nanotech enthusiasts to turn green with envy. The fly has complex oscillating wings driven by miniature muscular motors and rapidly navigates in four-dimensional space (i.e., it flies in three dimensions while accounting for temporal/spatial changes in its surroundings) while interpreting stereo signals from two large, compound eyes consisting of more than 80 lenses. It does this swiftly enough to detect and avoid imminent danger (try hitting a fly with your hand) and to invert in flight so as to land upside down on a ceiling. It uses sophisticated chemical detectors to locate food that is processed in a miniature digestive factory. And the fly, of course, mates and reproduces. This behavior is guided by a speck of a brain and decentralized control centers consisting of only a few hundred thousand neurons, a value below the number of transistors in a late-model PC. The fly brain runs on about one hundred-thousandth of a watt, compared to the dozen or more watts burned by Pentium-grade chips.
By implication, bi-digital chips are missing something involved in replicating biointelligence. Penrose and Hameroff have suggested that bugs tap into even quantum mechanics. However, before getting carried away with the chip versus bug brain disparity, it should be realized that insects are not doing that much. Insect behavior is stereotypically limited. It is doubtful that conscious thought is involved in their behavior. It can be argued that the navigation ability of robovans and cruise missiles is approaching that of insects. When the human pilot of a F-117 Stealth attack jet becomes disoriented and the plane starts tumbling through the dark sky, the pilot pushes a panic button and the computer gets the craft re-oriented, a mean aerodynamic stunt. The decentralized walking action of real insects is being replicated to a certain extent by Brooks' robotic insects. No one has yet put big bucks into software packages needed to make a robotic insect. Instead, insects, mass parallel-processing analog-digital systems optimized for bug behavior, and chips, are built to do very different things. A Pentium chip is optimized for video games, word processing, and other nonbug-like tasks. The miniaturization and reproductive capacity of insects is due to their being nanosystems beyond the capacity of current microsystems, and their energy efficiency is attributable to analog versus digital information processing.
The way to high intelligence may prove to be a jerry-rigged compromise that includes both conventional programming and AL. Only time will tell. What we will do for now is set the goals high and assume that to think like a brain, computers will have to be built and work very much like a brain. What, then, will be needed to progress from number crunchers to thinking machines?
A better knowledge of how brains work requires better computers that work more like brains! It is no accident that computer intelligence and our knowledge of brain function are growing exponentially at the same time. The remote scanning systems being used to examine brain function in real time are possible only because of high-power computers. Brain research is fairly cheap; a university hospital or lab can afford the machinery needed to do cutting-edge brain research. At the same time, researchers are using computer-driven scanners to peer into the brains that they are using other computers to mimic and model the workings of the brain. The investigative power of the combination of remote scanning and computer modeling cannot be exaggerated. It helps force neuroscientists to propose rigorously testable hypotheses that can be checked out in simplified form on a neural network such as a computer. Neuroscience is fast going from a field of speculative opinions to productive hypotheses based on confirmed facts and tested via real experiments.
Computers are starting to become more like brains and as they do, they allow a greater understanding of brains, which can be put back into making the computers more brain-like. The evolutionary convergence between brain and computer is inevitable and speeding up. As computers get better, brain scans are more detailed, and modeling becomes more sophisticated and realistic. In turn, a better understanding of how the brain works is applied to computers, making them more powerful and cognitive and making it easier to do brain research, which allows the computers to become more brain-like and powerful, and so on.
It is universally agreed that the new mind machines will have to duplicate the brain by being complex parallel processors running many algorithms at the same time. If for no other reason, this will be necessary to achieve the high calculating and memory access speeds required to form a practical intelligence. It is probably physically impossible to squeeze many trillions of calculations through just one or a few processors. The difficulty of reaching very high speeds through one processor is already forcing a switch away from the rather simple linear machines we have lived with for half a century toward more complex multiprocessor systems. For example, TFLOP will use more than 9,000 processors using new Intel P6 chips (the successor to the Pentium) to achieve its 2 teraflops.
TFLOP is actually simple compared to some supercomputers that have already been built. Even recent Crays have had only a few processors working in parallel; the high speeds are achieved by extremely high-speed special chips. However, about ten years ago, thinking machines began to build massively parallel-processing connection machines that used first thousands and then tens of thousands of relatively common chips to achieve total speeds matching those of more conventional supercomputers. Software problems and post-cold war cutbacks have helped to cripple Thinking Machines, but some of the latest computers of this type have a few hundred thousand parallel processors working at once. Computer complexity has, therefore, leaped forward a few hundred thousandfold in only 20 years. Does the failure of companies like that of thinking machines spell stagnation for computer complexity? Perhaps for a while, but only for awhile. There are many problems that naturally flow through parallel systems better than through linear machines, even when the overall speed of the machines is about the same. The airflow of aerodynamics and air masses of weather systems fall into this category. This, combined with the constraints imposed by physics, is likely to force still more parallel processing into computers and the development of the software that will control the machines.
It is, therefore not surprising that the Brain Build Group in Japan plans to build a neural network they call an "artificial brain" with one billion neuron-like circuits by 2001. This machine will not operate at human levels, but perhaps HAL is not so far off after all.
Along with parallel processing, just about everyone agrees that an AC computer will have to be a self-learning neural network storing holographic memories by strengthening and weakening connections. This means the processor and a large portion of the memory will be intimately unified. Whether synaptic-like structures for memory storage are critical is not certain, but they couldn't hurt. The result will be systems with human-like pattern recognition and redundancy.
Here, we come to an important unknown: High-performance computers that employ parallel processing and neural networking are inherently more complex than conventional computers. How much complexity does it take to do all that a brain does? Certainly, we do not have to build a computer as complex as a brain to match the gross calculating speed of a brain. Remember that brains have more than a hundred billion neurons and trillions of interconnections because the neurons signal each other so slowly. A computer with circuits working at a million cycles a second would need only millions of neural circuits to match the total speed of a brain. But would a machine as fast and as data-dense as a brain, but much less complex than a brain, think as well as a brain? What if high-level consciousness can be achieved only by trillions of bits of information flowing in parallel between circuits that are generating standing waves of only a few hundred cycles per second? This is one of the major questions of AC at this time.
Another uncertain complexity issue is whether conscious machines will need to be special-purpose devices with various functions split into subunits as per brains, or whether general-purpose machines will do.
Previously, we discussed how neurons send fairly complex analog-digital signals, rather than simple digital signals used by most computers. Also discussed were the neuron-mimicking analog chips that produce signals virtually indistinguishable from the original item. It is possible, many would say probable, that conscious thought is dependent on analog-digital signals and the fuzzy logic they produce, and that purely digital-only systems will always be reflexively intelligent in the manner of insects. Others disagree.
Even if the latter are correct, analog-digital circuits have the advantage of being much more energy-efficient and flexible than digital circuits, which is one reason why brains are many times more energy-efficient than digital computers. Although digital processing is simpler than analog, it is more energy-expensive because it takes work to convert information from its initial analog form into large digital codes and then back again. If an AC machine the size of a brain used purely digital chips, it would not only sop up a large amount of energy, but it would also run hot, perhaps hot enough to melt. We conclude that thinking machines will probably need to be fuzzy-thinking analog-digital machines.
If the Penrose-Hameroff quantum hypothesis of consciousness is correct, fitting the synthetic neurons with synthetic quantum sensitive parts should produce quantum AC.
Existence may prove dull for smart computers unless they have feelings as they apply to both sensory awareness and emotions. Some AI researchers who think consciousness will emerge in step with increasing power and intelligence suggest the same will be true of emotions, as well. Moravec, for instance, tells the story of the cognitive household robot that gets locked out of the home and is running out of electricity. Its "anxiety" rises as its power runs down, and it tries to decide whether to enter an unlocked home to tap its power - something it knows is bad. Conversely, we can imagine its "relief" and "pleasure" when its owner unexpectedly returns and lets the robot get its juice.
The contrary view is that feelings will prove one of the most difficult brain functions to replicate in view of the fact that animal feelings are highly dependent on wet chemistry. A neural-chemical reaction is just another form of information processing, one that the brain uses because it is convenient for water-rich organic materials. If cyberbrains are dry, electrical waves may be used to stimulate and shift moods. Alternatively, if emotions require chemical agents, cyberbrains may be partly wet, emotional machines! In any case, emotions may not be critical to early AC, although one can never be sure about these things.
At some point, the convergence between human brains and brain-like computers will go so far that the latter should start to do what the former does, generate conscious creative thought and foresight. This convergence in form and especially function is as close to inevitable as one can get without being there, but the details will remain murky until the event occurs. As previously mentioned, if AC does not emerge without prompting as computer intelligence rises, a more deliberate effort may prove necessary. AC machines may mimic brains with precise synchronous communications between circuits, generating transient standing mind waves. Perhaps information will need to be looped through a central processor, analogous to the critical part of the thalamus, before special neuron-like circuits communicating at a common frequency can become aware of the information. Alternatively, the mechanism for generating synthetic conscious thought may prove dramatically different from ours.
In any case, there is little doubt that AC will be a form of after-the-fact virtual consciousness of the sort we practice. This will allow the cyberminds to put together a coherent picture of the world before they are aware of it. It will also allow them to simulate, practice, and think through complex action repeatedly before the robot tries to do something difficult, just like we do. Of course, many actions, such as the details of walking and forming words, will be done subconsciously.
Obviously, if one is trying to emulate brains that are so intricate and complex because they evolve as they grow, molecule by molecule it makes sense to build equally capable cybersystems the same way. Indeed, from the bottom up may be the only way to create AC and the computers to run it on.
Two current bottom-up technologies are artificial life and nanotechnology. The first involves the immense knowledge base needed for high-level cognitive intelligence. AL also produces the complex debugged genetic algorithms that code for the design and basic growth of intelligent computers. Nanotech provides the means by which the AL codes are translated into the hardware of vast numbers of hyper-tiny circuits and wires needed to make a fake brain. For example, some of the researchers doing the preliminary work on petaflop computers are investigating how to use bioengineered bacteria to manufacture molecular transistors.
We can envision self-learning computers that evolve increasingly complex hardware in parallel with their increasing intelligence. They may be placed in mobile robots so they can learn about the real world. In this regard, evolving robots will build up a broad knowledge base by having a richness of experiences. Will this work? Insects, frogs, birds, cats, and humans develop working mind machines all the time.
Where the synthetic systems will differ from us will be in their ability to quickly build broad, deep knowledge bases by uploading from other AI expert systems. No need to go to school to get a degree.
|
We have already explained that the 3 lb. human brain runs at 15 watts with one or two hundred billion neurons firing a slow 100 cycles per second over trillions of interconnections to perform 10 trillion to 1,000 trillion calculations per second. The existence of the brain itself makes us confident that it is physically possible to build machines with similar capacities. We conservatively project a 21st century cybercomputer with one billion to ten billion circuits working in parallel via a trillion connections in a small package weighing 1 lb. The circuits are analog-digital switches that mimic neurons, and they form a self-learning neural network storing memories in holographic form while processing information at both subconscious and conscious levels. The analog-
digital circuits are energy efficient, like neurons, but because the new systems are not organic, there is the potential to give them new and improved power and cooling features. The circuitry can be run many times faster than human circuitry. For example, we are not limited to sugar power; we can plug the cybercomputer into an outlet or a high-tech battery, so we up the power to 100 watts. The system will run hotter than our brains, but the device is made of diamond or some other material with high heat-resistance and we attach cooling fans. The biggest difference is that the nanocircuits run at a technologically respectable one million to ten million cycles per second and the wiring zips electrons millions of times faster than nerves. Running at a normal 10 percent of total capacity per second calculating speed of our new cyberpride and joy is a petaflop.
The minds produced by the machines are as self-aware and conscious as those of humans. They all have no trouble passing any Turing test thrown at them; some may grab a human by the proverbial collars and insist on their consciousness. On one hand, pattern recognition performance and the robust knowledge base of the system rivals that of humans. On the other hand, the digital component of the systems is better developed than in humans, so memory retention and mathematical abilities are superior. Whether the rational capacity of these first mind machines will exceed that of humans is uncertain. Also open is whether their emotional capacity will fall short of match or exceed the human condition.
So far, we have assumed that artificial consciousness will be achieved from the ground up by building an entirely new means of producing minds that may or may not mimic the animal version. There is another way to AC one that is not so alien: piggyback artificial thought on the natural system. The idea is simple enough. Gradually replace original brain parts with new ones that do the same thing until the entire brain has been replaced.
This concept has been used as a basic thought experiment to show the viability of AC and mind transfer. The concept is gradually being put into practice as a means of correcting brain defects. Initial experiments center around hooking up minicams to the visual cortex and replacing defective retinas with new ones. So far, these crude affairs only allow the blind to see patterns of dots and letters, but the technology will improve. The synthetic analog-digital retinas currently residing in research labs will form the basis for high-resolution eye replacements. Of course, this is to allow the blind to see, but what will happen when it inevitably becomes possible to replace healthy eyes with superior artificial replacements?
|
Even more intriguing are the people at the University of Southern California who have developed the chip they say mimics the function of part of the hippocampus. The chip is about the size of a floppy disk, too large to be useful to us, but just as the earth spins on its axis, the chip will become ever smaller. The aim is to replace the hippocampus when it becomes too damaged to function properly. A few labs are looking into how failing neurons can be replaced one on one with new analog-digital circuits via nanotechnologies. This is intended to correct such ills as spinal injuries and Alzheimer's.
The main point is that AC may never be developed from scratch, but captured from the human system!
There are important differences between the two alternative means for making new minds. Because evolution is such a complex affair, it is not possible for two lines of evolution to produce identical results. If AC is evolved in an entirely new manner, its evolution will only roughly mimic, not precisely match, the evolution of human consciousness. The new minds will be very different from ours to the point that they will be aliens. Although they will retain aspects of humanity within them because we built them, we will have more in common with apes.
If AC is derived from modified brains, the new minds will not start out alien, but will soon become so. The difference between modified and alien minds has important implications for mind transfer, which we will address in more detail later.
Whether the first synthetic petaflop AC computers are made from scratch or from modified brains, they may run at only human speeds. But why stop there? We can be confident that we can do better than the brain. Why? Because big brains are too big! The long connections between neurons in different parts of the brain slow things down. The slow neurons and nerve transmission rates also can be improved on immensely. Instead of being aware of reality a substantial fraction of a second after it happens, the lag will be cut to a tiny fraction of a second. As nanotechnology improves the size of the neuron-mimicking circuits quickly decreases, while their number and speed increases. Because the intelligent computers are evolving, there is nothing limiting them to human levels of thinking speed, memory capacity, structural complexity, and overall mental powers. The machines were at one point equal to humans, then edge a little above human levels, then a little higher, and still higher. Power quickly soars far above human levels. Rather than taking a minute each to simulate a complex action a dozen times before doing it, the cybermind runs through 100 simulations in a few seconds. Because the cybermind knows the intimate workings and details of its circuitry, it can control and modify the evolution of the circuitry better than a human, so the cybermind does not get into mental ruts as easily as we do.
|
The cybermind has additional digital memory storage and retains enormous amounts of information with a nonhuman degree of precision. In addition, the system can tap directly into the international computer network and upload whole libraries-worth of information in a few minutes. The human brain barrier is not merely broken, it is smashed.
As impressive as these superbrains may seem, they are possible because they represent only moderate increases over the speed, power, complexity, and information density of biobrains that clearly can be improved on. In this sense, they will be early and crude devices that may seem challenging to us, but in terms of size and power do not go far beyond the brains we take for granted. Some think we can do much better. Although brains grow at the level of molecules, the neuronal subsystems that process information are each made of multitudes of molecules.
A computer with information densities far better than brains would operate at the level of a few molecules and atoms. Folks are already looking into how atomic-level computers might be made, while others are questioning whether atomic-level computers can be made. Atoms and molecules are ideal building materials, they are perfect and uniform in structure and they can articulate in precise and perfectly predictable joints. A simple atomic switch has already been constructed. It is not clear whether machines will work at such amazingly small scales and do anything important. For one thing, the amount of energy running through such tiny devices might melt them (then again, energy-efficient analog signaling may keep them sufficiently cool). Beyond atomic computers is the concept of subatomic computers, which would use the energy states of individual electrons to store and process information. Electron computers would be half a million times more information-dense than today's computers and thousands of times more compact than brains. We have no example of an atomic- or subatomic-level computer in existence showing they will work (unless the Penrose-Hameroff hypothesis is correct after all), and the failure of evolution to produce one may be telling us something. Then again, life has not produced 500-ton flying objects, either.
|
We cannot judge at this time whether the tiniest atomic or even subatomic computers will ever be possible. We can say with a good deal of confidence that hypercomputers that match and exceed the performance of the human brain at least in terms of sheer power and even in complexity are inevitable. It is even possible that the first AC machines will be nanotech-based, atomic-level devices, in which case they will surpass brains from the get-go!
The biggest question is whether artificial consciousness will prove a hard nut to crack, requiring special circuitry based on an intimate knowledge of brain function, or if it will emerge on its own as computers evolve intelligence. As tempting as it may be to say the former, the latter cannot be ruled out. We just do not know.
Just because it is possible to do something, does not mean it will be done. There has to be a reason to do it, and the results must be worth the investment. We could already have colonies on the moon and be on the way to Mars if we wanted to spend hundreds of billions of dollars for no obvious gain. The forces propelling the development of smart machines are much more compelling. Whether they are digital, analog-digital, or quantum mechanical in their operations, there are strong pressures to make mobile machines think and do what we do.
As much as they may protest to the contrary, industry and commerce really do not like human labor, physical and intellectual. Consider things from the viewpoint of owners and investors: Even a nonskilled worker requires up to two decades for the potential employee to grow and be educated to a minimal level; skilled workers can take three decades to get online. These are costs industry must subsidize through the elevated wages they pay parents and skilled employees. Once the person is working, they want to do so for only one-third to one-half of a day with one or two days a week off as well as paid vacations. Even when at work, down time due to lunch and coffee breaks, restroom visits, personal phone calls, gossiping, computer games, daydreaming, drowsiness, and so forth cuts down further on productive time. It does not help that humans resent controls meant to keep them as productive as possible, but costly efforts must be made to keep employee moral at sufficiently high levels to maximize productivity. Laying off large numbers of workers when they become too expensive or their skills obsolete adversely impacts employee moral. This is when human workers tend to get into ruts, which makes retraining them difficult and costly. Every worker is a potential thief, whether it be a few items from the candy aisle or a few hundred million from the brokerage firm. At the same time, even honest human workers want enough money to live pleasant lives outside of their jobs including large structures to live in, vehicular conveyances, and assorted and sundry hobbies, pets, and entertainment. All this assumes the worker remains healthy. If not, at best productive time is lost; at worst, the business must help defray medical costs either directly or indirectly. After just three or four decades of working for only one-quarter of each year (for a total at-work time of about ten years), the retired worker actually expects to continue to be supported until death! If the human employees are not happy with the arrangements, they may go join a union that restricts the freedom of the owner/investors and may even hold a strike that hampers or even ceases operations. If the employees are really mad, they may engage in a little sabotage or go so far as to destroy the facilities and attack the owners.
No wonder businesses want smart robots and computers that will work and think like humans, only more efficiently, without requiring long education, around the clock, without vacation pay and benefits, always be honest, and never go on strike.
Industry also wants smart robots that can go where no human should be and do things people cannot do, such as working deep in nuclear facilities, or in mines where ore digging and transporting machines are well on the way to replacing human miners. Instead of exposing human divers to the dangers of the deep, and their employers to the costs associated with their too-frequent deaths, why not use underwater robots? How about electrical high wire robots? The list is almost endless.
So many leading vulcanologists have died studying the objects of their research that the survivors are pressing for ambulatory robots such as Dante II that can descend into natural hells on earth. Oceanography is made for robots. Bob Ballard, who discovered the wrecks of the Titanic and Bismarck, has become a leading proponent of deep seabots, perhaps in part because of some close calls he experienced while squeezed into tiny submersibles a few miles down. Even when things go well, humans can spend only a few hours each year in deep waters. To fully explore the depths will require a host of smart swimming seabots. NASA knows that astronauts cannot devote enough extracurricular time to building large structures in orbit or on alien bodies at a reasonable cost. Nor is it yet practical to send large numbers of human explorers to other planets. NASA is, therefore, a leading funder of research into robotics.
Then there is the military. Faced with a public that does not think its young men, and now women, should die in combat, the military wants a lot of good robowarriors. With a budget almost a large as the combined militaries of the rest of the world, even the post-cold war Defense Department remains on the cutting edge of robotics and AI.
Understanding brain function, replacing brain parts with new ones, using remote scanning technologies, and cracking entire genetic codes all depend on dramatic advances in computers, AI, and AL. The nightmare of diagnostic procedures and the craving for cost cutting is spurring the development of sophisticated AI expert systems. Even robot surgeons are being introduced to perform special tasks such as prepping bones for artificial joints that clumsy human hands do poorly. Now being researched is a special little robot that will explore the entire colon for polyps and other things that don't belong there. The ultimate dream is a microscopic nanorobot that will correct defects from clogged arteries to cancer with no muss and no fuss.
You want smart machines. That's right, you. Once upon a time, middle class families could afford in-house help, and much of the driving was handled by horsepower with enough horse sense built in not to run off the road or run into other conveyances or pedestrians. It is one of those ironies of modern life that automobiles are so dumb that you, the driver, must guide them along every inch of every road mile, or you will suffer the consequences. After you bring back the good old days with your street-smart robocar, after a few years, you will want a personal robot to clean the house (without taking anything from it), do the dishes and put them away, watch the kids (without looking into someone's records for past charges of child abuse), be available any time of the day, and not require filling out any Social Security forms. You also want a smart house that provides its own security, detects and puts out fires, turns the lights on and off, raises and lowers the shades, and cooks the meals. Even if you don't want them, others do.
For your computing pleasure, you would like a whole set of intelligent machines. Most obvious would be the general-purpose machine that listens and responds to you, is smart enough to read your miserable excuse for handwriting, and uses intelligent agents to help determine your wants and needs, bringing a little order to your chaotic postmodern life. Also useful would be other little intelligences inserted unobtrusively into common devices such as phones, ovens, thermostats, calculators, remote controls, lawn mowers, saws and other powered hardware, and just about anything else that can be improved with a little AI.
The plain fact is that lots of people want smart systems that do just about everything humans do. Today's computers actually do only a little of what we would like them to do because they are still weak. A must for making machines really useful and smart is speed, speed, and more speed along with more and more memory inside smaller and smaller devices. Premiere among the wished-for smart devices are robots that can move at least as fast and adeptly as people, identify, pick up, and manipulate objects as readily as we do and use them as intelligently as we do, all by voice command. The amount of money to be made by those who design and produce such mobile thinking machines will dwarf the automobile and aircraft industries combined. Even downsizing, bankruptcies, and associated cuts in research budgets sweeping through industry and the military have only whittled down, not stopped, leading-edge research in robotics.
The attempt to make robots mimic what we do will compel scientists and engineers to resort to increasingly brain-like evolving analog instruments. At the same time, the pressures of medical research encourage computer scientists to use their machines to better understand the workings of our cranial computers. The result is a fast-rising learning curve. Some of the wealth generated by the commercial success of increasingly smart machines will be used to make them smarter in a classic capitalist feedback loop. This pattern may enter a new phase if and when nanotechnology comes online and pumps the evolution of AC machines to new levels of sophistication. Computer science and neuroscience are in the midst of a mutually beneficial evolutionary feedback loop. The ultimate result is obvious: Computers are becoming more and more like brains.
At some point, all these forces probably will come together and force a shift away from the simple calculating machines we are familiar with to more sophisticated and subtle systems that combine the best aspects of computers and brains. Conventional computers may remain common because of technological inertia and because they can continue to perform simpler tasks. However, the leading edge of cybertech will be the brain-like machines.
The limited success of today's robots is not a good guide to future developments. In 1896, no one had an engine that was light and powerful enough to propel a manned aerial machine, and the need for a dynamic flight control system was not well understood. In 1996, no one has computers small and powerful enough to match human-level performance, and much has to be learned about how brains work.
Computers can run rings around humans when it comes to doing math and are beating the human champion players of complex board games. Computers are coming close to passing Turing's test (in line with Turing's 1950 prediction that it would take 50 years to do so). The gross calculating speed and memory capacity of computers are growing exponentially and have come so far that human levels will probably be reached in a few decades. The miniaturization and information density of computers are approaching brain-like levels at the same burning pace. The complexity of the machines is rising swiftly as parallel processing is introduced and expanded. A symbiotic interaction between neuro- and computer sciences is facilitating the exponential increase in our understanding of brain function, which is then being used to make computers more brain-like. The knowledge has been used to produce neural networking computers that learn on their own and do a modest level of innovative thinking. Some of the networks are made of analog-digital circuits that closely mimic the function of real neurons. The functions of small parts of the brain and retina have been reproduced with sophisticated chips. Research into replacing brain parts with synthetic systems is under way. Computers are beginning to understand and use language; others are learning to recognize faces. AI expert systems are becoming a normal part of business and government. Robots are learning to navigate dangerous terrain. Some are learning to walk and do flips. Robots are exploring the oceans and space. A few robots can understand and carry out fairly complex requests in a flexible manner. Mobile robots are taking their first tentative steps into the commercial realm where static robots already rule many assembly lines. Some weapons are as smart as insects. AL computers are evolving complexity from simplicity. Whatever it is that computers and robots do, they are getting better and they are doing new things all the time. Although the rate of progress may seem slow, it is shockingly swift by geological and even historical standards. At the same time, researchers are plumbing the depths of the human mind machine, and smaller-thinking researchers are laying the foundations for building new mind machines atom by atom, molecule by molecule, and intend to out-do life at what it does best.
What has already been done with AI, AL, and AC is astonishing, and the rate of progress is equally so. It is odd that this fact is not better recognized, perhaps events are moving too fast for people's perceptions. It is also important that a critical defect barring the way to human-level cognition in machines has yet to arise, and as far as can be conjectured, there is no such barrier. Because the effort to produce evolving neural networking analog-digital robots built on the nanoscale has just begun, it is tempting to say that AI/AC is still in its infancy. That was certainly true until the '80s, but a case can be made that things have already progressed beyond that point. The effort to endow machines with human-level cognition has passed into its early childhood and gone from taking baby steps to making the first running strides, which should soon be followed by adolescent leaps early in the next century, and then Olympic pole vaults soon after.
MIT's Rodney Brooks, the builder of those little roboinsects, has launched a project under the direction of Lynn Stein to "evolve" during the next five years an android with the physical skills, cognitive abilities, and, hopefully, the consciousness of a two-year-old child. The upper body of Cog has already been built and its computers are starting to learn about the world. It is said to be unnerving to be tracked by Cog's twin eyes, and people in the lab have set up a group to consider the ethical problems associated with a budding consciousness. These goals are unlikely to be reached because even a child's brain is far more powerful and complex than any computer will be in the next ten years.
Cog does not represent the only attempt to construct a working android. At Waseda University in Japan, WABOT-2 is an articulating torso plus head that can recognize music written or played and then play the same tune on a piano to near-concert standards. WABOT-2 is meant to be a stage toward fully functional androids that interact with humans, even on the emotional level.
Then there is Steve Thaler, a private inventor working on his Creativity Machine. This neural networking computer system has already "written" thousands of original tunes, designed soft drinks, discovered unique minerals that may match diamonds in hardness, and has been hired by a high-tech company to search for high-temperature superconductors. Thaler also has his current-generation CM working on the next-generation CM. Thaler hopes that his Creativity Machines will evolve until they become conscious and do so rapidly enough so he can leave his aging body behind and live forever in cyberspace.
Although the Cog, WABOT, and Creativity Machine programs appear over-ambitious at this time, they are worth pursuing because much will be learned. Besides, maybe we will be surprised and Brooks will make naysayers of his critics with the first highly capable robot. It is more likely that Cog will prove a sophisticated insect. Perhaps a better model for a demonstration robot would be a robofly, a small, autonomous aerial machine that can reliably and flexibly carry out tasks by voice command. Georgia Tech holds a yearly competition to see if student teams can get a flying robot to pick up a small metal disc in the center of one raised ring, carry it over a barrier, and deposit the disc in another ring. A simple enough task it would seem, but no radio controls are allowed. Dozens of machines, balloons, tail sitters, and model helicopters have shown up. Some do not get off the ground, others drift off in the breeze, but most prove too sensitive to slight perturbations to complete the job before they tip over and grind spinning blades to bits and pieces. (The robot competition increases as their ability to do "human" things of creativity and innovation increases.) So far, no aerial robot has won the competition in a convincing manner. In the last competition, one chopperbot succeeded in picking up the disc with a magnet and carrying it to the other ring, but it could not drop the disc. The machines get better every competition cycle, and one will someday complete the task. Then the task will be made a little harder, a craft will do that version, the task will get still harder, and before we know it, agile robots will be on hand.
A few years ago, Hans Moravec tentatively estimated it will take 40 to 50 years to make robots as intelligent, capable, and potentially as conscious as humans. This projection is based on expectations of small cheap computers becoming as powerful as brains. Not surprisingly, this prediction is controversial; most simply refuse to believe that humans can be matched so soon. Roger Penrose, for instance, hints that it will take a long time to build his quantum mind machines, but he is a Platonic mathematician who thinks arithmetically, rather than an evolutionary biologist who thinks exponentially.
The common belief that we are far from achieving human-level performance in computers is a naive illusion due to the exponential growth curve of science and technology. Because today's computers have only a fraction of a percentage of the power and complexity needed to make a human-equivalent robot, it seems natural and obvious to assume it will be a long time before it happens. Not necessarily. Computer power and sophistication has come so far and is growing so fast that the gap may well be covered in a few decades.
We, too, have a bone to pick with Moravec's prediction, but for different reasons. Moravec presumes top-down programming will be involved in making robots increasingly smart via evolutionary AI, although there will be a strong self-learning component, as well. As part of the process of teaching robots to do better and better, Moravec outlines a series of ten-year stages to human-level performance. It starts with not-so-intelligent but useful mobile robots by or soon after the turn of the century, followed by a series of major moves toward human cognition every ten years or so. In this view, there will be a period of decades in which mobile robots are capable and fairly intelligent, but still well below human norms.
Moravec's outline of a fairly smooth and linear build-up to human-equivalent robots may be as naïve as the naysayers' belief that it will be a long time before we get to the ultimate goal. Look at it this way: Many people will go along with the idea that in the not-too-distant future, it will be possible to make a mobile robot with a mind that is the mental equivalent of, say, a dog. They will then go on to argue it will be long after that before a human-equivalent system is up and operating. They could hardly be more wrong because they forget the implications of exponential growth leading to over-prediction of the near future and under-prediction of the distant future. They also fail to fully appreciate their pet dog's abilities. Remember, dog-like brains first evolved only a few tens of millions of years ago, in the geological perspective, it was only a few moments before human brains showed up. Dog brains are powerful, complex, fast-running, and intelligent systems in their own right. They probably have at least a tenth of the gross calculating power of a human brain. The gap between a man and his dog is much less than that between a dog and its flea. Once we have achieved dog-level robotic performance, the projected growth rate in computer power means that machines with human-level performance will arrive only four years later! Even if we assume a time lag due to the difficulty of programming hypercomputers with human-level knowledge, the gap should be measured in years rather than decades. By the same token, MIT's Cog, if it works as well as Brooks says, will also exhibit a large fraction of adult human intelligence, and the jump to adult levels would be short and modest.
To squeeze dog- or child-level mental performance out of a robot is going to require AC-capable computers almost as powerful as those that can support a human-equivalent mind. If we assume that the power and complexity of small computers rises smoothly to human levels in 45 years, robots with dog-equivalent minds will not be running until 35 to 40 years from now. Until then, we will have to make do with insect-level robots, then lizard-equivalent systems followed by ostrich-ferret-level cyberminds, and so on until the dog-level is reached. When it becomes clear that cybersystems as smart as dogs are up and running, there will not be a long way to go before machines will be our equals. It is more like one to ten years.
Further complicating the jump to human-equivalent intelligence is the precocial intelligence of AI computers. Take language. As mammalian intelligence developed over time, complex language did not appear until the later stages of human evolution when brains were nearly as large or as large as they are now (just which of our hominid ancestors had sophisticated language is uncertain; some suggest all species of Homo sapiens did while others assert that even Neanderthals could not speak fluently). Although computers will be far less capable than dogs, much less humans, for the next decade or two it is probable that some of the systems will be able to understand and use language much better than any dog and approach the human level. The same machines will have access to other AI expert programs that will give them far more human-like knowledge than any animal. The implications of precocial intelligence in computers for the timing of the onset of human-level intelligence are hard to sort out amidst a subject already deeply perplexing. It hints that as soon as machines can support a human level of intelligence, they will be able to do so. It also hints that the final leap to human-equivalent intelligence may be very big and very fast.
|
If it takes nothing more than computers with enough speed and learning, all we have to do is wait for the computers to get fast enough and the robots experienced enough to be as smart and aware as we are. If, instead, the recipe includes a large portion of AL, mixed with nanotech and laced with a detailed knowledge and imitation of brain structure, the meal might be delayed. However, the exponential increase of our understanding of cognition, human and cyber, suggests it will not take centuries to make special thinking machines. We cannot say it will take any longer than the "easy" approach because of the potential for dramatic leaps in AL and nanotechnology. Even if Penrose is right and nonquantum machines cannot think, the need to build quantum mind machines will, at most, put off the day of reckoning, but they may not even do that if it is discovered that quantum computations are the quickest way to high-powered machines!
So, you are asking, just how long is it going to take for the Extraordinary Future to bring us to the CyberOz of human-equivalent AC? Unlike those who search for hidden numbers in the Bible or arcane predictions in the works of Nostradamus, we who speculate cannot conjure up some magical date. Perhaps it means we will not embarrass ourselves too much.
Our tentative timing sequence to human-level artificial cognition presumes a combination of AL and nanotechnology is the way to the goal, although some conventional programming and microtechnology cannot be ruled out. This bottom-up process may progress smoothly and exponentially. However, it may prove subject to sudden dramatic jumps to new levels as hard work leads to the unexpected breakthroughs we call serendipity. We will also assume there will be no global catastrophes, economic, political, or environmental. If there are, things will be speeded up or delayed to a greater or lesser extent.
Although the thinking power of computers is still a small fraction of our brains', nothing much will happen and all will seem secure to those who do not favor cyberevolution. In this case, mobile robots may not become as capable as soon as Moravec suggests they will. The attempt to make machines smart will remain a study in frustration as progress remains modest, with no quick breakthrough to human-level performance. There will be failures, and naysayers will pan thinking machines just as they dismissed flying machines. They will fail to appreciate that this is a normal part of the early trial-and-error stage of exponential evolution and that the mistakes are helping outline the path to success as Darwinian selection winds its way through multiple approaches to the problem. Never forget how few in the 1890s thought people would fly early in the 20th century. The knowledge base will expand toward a critical breakout point. A major station on the way to AC will be the development of nanotechnology, giving us the ability to develop computers as intricate and complex as brains. The techniques for doing so will become understood as neuroscience details the function of the brain.
Although major breakthroughs toward cognitive machines of great complexity will be made perhaps in only one lab, it is more likely that they will be made in a number of research centers at the same time. The power and sophistication of nanocomputers growing and evolving on their own as well as uploading masses of information in short order may not double every year or two, but every month or two. The great evolutionary increase will be decoupling itself from slow-moving human intelligence as it has already abandoned the still slower DNA. The fast-improving nanocomputers will also exhibit increasing signs of profound intelligence. With thinking power and a self-acquired knowledge base increasing exponentially on a subyearly basis, the dog, monkey, ape, and human stages may be accomplished in less than a year! If this happens, events will then come with sudden unanticipated swiftness almost before people realize it. Suddenly and with great surprise, they will be among us.
Alternatively, AC is not achieved entirely through new machines, but by altering and upgrading brains with cybertech. Again, the early stages may appear frustratingly slow, but when nanotech allows new brain parts to be grown in place, rather than inserted via surgery, progress to completely artificial brains can be expected to be swift.
When all is said and done, we cannot tell you exactly how and when we will develop artificial consciousness; a marvelous characteristic about evolution is its very ability to surprise. If truly conscious, cognitive robots thinking as fast as humans, even human children, are around in ten years, we will be shocked. If it happens 20 years from now, we will be very surprised. By 30 to 50 years from now, our surprise should be mild indeed. If not by 2095, we would be back to being shocked. Do not bet on the Big Show arriving within your Baby Boomer lifetime. But with DNA computers suddenly appearing out of nowhere and technoevolution on an exponential curve, do not rule it out, either.
The timing, method, and details are inevitably murky, but there are technological and scientific trends much too powerful to ignore. Make no mistake: In a few decades, there will be information-processing machines as powerful, and then more powerful than a brain. For that matter, sometime early in the next century, there will be a computer whose calculating speed equals that of all those working now! It is hardly conceivable that this will not change our lives in ways far more dramatic than anything that has gone before. So, while the human brain is the most complex device in the universe, it won't be for long. Even boosting the power of brains via technology is not likely to allow oxygen-burning organic systems to compete with artificial systems powered by enhanced systems and means. Do people really believe that as the power of the computer soars to levels far beyond those achieved by nature that no one will figure out how to turn them from mere "computers" into true thinking machines? The rate of progress and complexity of the world a hundred years from now promises to be orders of magnitude greater than we know, and such swift complexity may prove beyond human abilities. Those who believe human minds are going to remain supreme and in control in the brave new world of the next century may prove marvelously naïve.
© 1996-2003 Charles River Media. All Rights Reserved