The Age of Robots


June 1993 by

Hans Moravec

Robotics Institute
Carnegie Mellon University
Pittsburgh, PA 15213
USA
(412) 268-3829

(8500 words)

Abstract

Our artifacts are getting smarter, and a loose parallel with the evolution of animal intelligence suggests one future course for them. Computerless industrial machinery exhibits the behavioral flexibility of single-celled organisms. Today's best computer-controlled robots are like the simpler invertebrates. A thousand-fold increase in computer power in this decade should make possible machines with reptile-like sensory and motor competence. Properly configured, such robots could do in the physical world what personal computers now do in the world of data--act on our behalf as literal-minded slaves. Growing computer power over the next half-century will allow this reptile stage will be surpassed, in stages producing robots that learn like mammals, model their world like primates and eventually reason like humans. Depending on your point of view, humanity will then have produced a worthy successor, or transcended inherited limitations and transformed itself into something quite new. No longer limited by the slow pace of human learning and even slower biological evolution, intelligent machinery will conduct its affairs on an ever faster, ever smaller scale, until coarse physical nature has been converted to fine-grained purposeful thought.

_________________________________

Serious attempts to build thinking machines began after the second world war. One line of research, called Cybernetics, used simple electronic circuitry to mimic small nervous systems, and produced machines that could learn to recognize simple patterns, and turtle-like robots that found their way to lighted recharging hutches. An entirely different approach, named Artificial Intelligence (AI), attempted to harness the apparently prodigious power of post-war computers--able to do the arithmetical work of thousands of mathematicians--to more interesting kind of thinking. And indeed, by 1965, computers ran programs that proved theorems in logic and geometry, solved calculus problems and played good games of checkers. In the early 1970s, AI research groups at MIT and Stanford University attached television cameras and robot arms to their computers, so their "thinking" programs could begin to collect their information directly from the real world.

What a shock! While the pure reasoning programs did their jobs about as well and about as fast as college freshmen, the best robot control programs took hours to find and pick up a few blocks on a table. Often these robots failed completely, giving a performance much worse than a six month old child. This disparity between programs that reason and programs that perceive and act in the real world holds to this day. In recent years Carnegie Mellon University produced two desk-sized computers that can play chess at grandmaster level, within the top 100 players in the world, when given their moves on a keyboard. But present-day robotics could produce only a complex and unreliable machine for finding and moving physical chess pieces.

In hindsight it seems that, in an absolute sense, reasoning is much easier than perceiving and acting--a position not hard to rationalize in evolutionary terms. The survival of human beings (and their ancestors) has depended for hundreds of millions of years on seeing and moving in the physical world, and in that competition large parts of their brains have become efficiently organized for the task. But we didn't appreciate this monumental skill because it is shared by every human being and most animals--it is commonplace. On the other hand, rational thinking, as in chess, is a newly acquired skill, perhaps less than one hundred thousand years old. The parts of our brain devoted to it are not well organized, and, in an absolute sense, we're not very good at it. But until recently we had no competition to show us up.

By comparing the edge and motion detecting circuitry in the four layers of nerve cells in the retina, the best understood major circuit in the human nervous system, with similar processes developed for "computer vision" systems that allow robots in research and industry to see, I've estimated that it would take a billion computations per second (the power of an average supercomputer) to produce the same results at the same speed as a human retina. By extrapolation, to emulate a whole brain takes ten trillion arithmetic operations per second, or ten thousand supercomputers worth.

Machines have a lot of catching up to do. On the other hand, for most of the century, machine calculation has been improving a thousandfold every twenty years, and there are basic developments in research labs that can sustain this for at least several decades more. In less than fifty years computer hardware should be powerful enough to match, and exceed, even the well-developed parts of human intelligence. But what about the software that would be required to give these powerful machines the ability to perceive, intuit and think as well as humans? The Cybernetic approach that attempts to directly imitate nervous systems is very slow, partly because examining a working brain in detail is a very tedious process. New instruments may change that in future. The AI approach has successfully imitated some aspects of rational thought, but that seems to be only about one millionth of the problem. I feel that the fastest progress on the hardest problems will come from a third approach, the newer field of robotics, the construction of systems that must see and move in the physical world. Robotics research is imitating the evolution of animal minds, adding capabilities to machines a few at a time, so that the resulting sequence of machine behaviors resembles the capabilities of animals with increasingly complex nervous systems. This effort to build intelligence from the bottom up is helped by biological peeks at the "back of the book"--at the neuronal, structural, and behavioral features of animals and humans.

The best robots today are controlled by computers just powerful enough to simulate the nervous system of an insect, cost as much as houses, and so find only a few profitable niches in society (among them, spray painting and spot welding cars and assembling electronics). But those few applications are encouraging research that is slowly providing a base for a huge future growth. Robot evolution in the direction of full intelligence will greatly accelerate, I believe, in about a decade when the mass-produced general purpose, universal robot becomes possible. These machines will do in the physical world what personal computers do in the world of data--act on our behalf as literal-minded slaves.

First-Generation Universal Robots

Timeframe: 2000-2010
Processing power: 1,000 MIPS (1993 supercomputer -- Reptile-class)
Distinguishing feature: General-purpose perception, manipulation and mobility


A robot's activities are assembled from its fundamental perception and action repertoire. First-generation robots will exist in a world built for humans, and that repertoire most usefully would resemble a human's. The general size, shape and strength of the machine should be human-like, to allow passage through and reach into the same spaces. Its mobility should be efficient on flat ground, where most tasks will happen, but also reliable and safe over stairs and rough ground, lest the robot be trapped on single-floor "islands." It should be able to manipulate most everyday objects, and to find them in the nearby world. The components of this machine exist in laboratories worldwide, and suggest guidelines for a practical design this decade.

1,000 MIPS (Millions of Instructions Per Second) is just enough computing power for a moving robot to maintain a coarse map of its surroundings and use it for locating itself relative to trained itineraries and to plan and control driving. When not traveling, there is power enough to construct a fine map of a manipulator workspace, to locate particular objects and to plan and control arm motions. When not occupied with its unique robotic functions, the robot should share with personal computers of its time the ability to communicate over wireless networks, to generate and interpret spoken sentences and to generate and read printed text. Programs for specific applications--many obtained via high-speed networks--will orchestrate these basics to accomplish useful tasks.

Universal robots will find their first uses in factories, warehouses and offices, where they will be more versatile than the older generation of robots they replace. Because of their breadth of applicability, their numbers should grow rapidly, and their costs decline. Eventually they will become cheap enough for some households, extending the utility of personal computers from a few tasks in the data world to many in the physical world. Perhaps a program for housecleaning will be included with each robot, as word-processing programs were shipped with early personal computers.

As with computers, some applications of robots will surprise their manufacturers. Robot programs may be developed to do light mechanical work (assembling other robots, for example), deliver warehoused inventories, prepare specific gourmet meals, tune up certain types of car, hook patterned rugs, weed lawns, run races, play games, arrange earth, stone and brick or sculpt. Some tasks will need specialized hardware attachments like tools and chemical sensors. Each application will require its own original software, very complex by today's computer program standards. The programs will contain modules for recognizing, grasping, manipulating, transporting and assembling particular items--modules developed via learning programs on supercomputers (with about 100,000 MIPS). In time, a growing library of subtask modules may ease the construction of new programs.

A first-generation robot will have the brain power of a reptile, but most application programs will be so hard pressed to accomplish their primary functions that they will endow the robot with the personality of a washing machine.

Second-Generation Universal Robots

Timeframe: 2010-2020
Processing power: 30,000 MIPS (Mammal-class)
Distinguishing feature: Accommodation learning


First-generation robots will be rigid slaves to inflexible programs, relentless in pursuing their tasks--or repeating their errors. Their programs will contain the frozen results of learning done on bigger computers under human supervision. Except for specialized episodes like recording a new cleaning route or the location of work objects, they will be incapable of learning new skills or adapting to unanticipated circumstances--even modest alterations of behavior will require new programming, probably from the original software suppliers.

Second-generation robots, with thirty times the processing power, will be more adaptable, because they can do some learning onboard. The fundamental idea in adaptive learning is to "close the loop" on behavior: to evaluate each action's effect in a given context to enhance the process that generated the action. In the simplest technique, a behavioral alternative that succeeds becomes more likely to be invoked in similar circumstances, while an alternative that fails becomes less probable. Faster statistical-learning approaches like neural nets repeatedly tweak behavior-control parameters to nudge actual responses closer to an ideal. Programs for second-generation robots will use many such learning techniques, creating new abilities--and new pitfalls.

If a first-generation robot working in your kitchen runs into trouble--say, failing to complete a key step because a portion of the workspace is awkwardly small--you have to option of abandoning the task, changing its environment, or somehow obtaining altered software that accomplishes the problematic step in a different way. A second-generation robot will make a number of false starts, but most probably will find its own solution, adjust to its home in thousands of more subtle ways, and gradually improve its performance. While a first-generation robot's personality is determined entirely by the sequence of operations in the application program it runs at the moment, a second-generation robot's character is more a product of the suite of conditioning programs it hosts. The conditioning system might, in time, censor an entire application program, if it gave consistently negative results.

Second-generation robots of 2010 will have onboard computers as powerful as the supercomputers that learned for first-generation machines in 2000. But by 2010, supercomputers will be proportionally more powerful (about 3,000,000 MIPS), and will themselves play a background role for the second-generation. The many individual programs of a conditioning suite--each responding to some specific stimulus--interact with one another and with the robot's control programs and environment in ways that will be far too entangled to anticipate accurately. It would be possible to evaluate particular suites by trying them out in robots--the acid test in any case--but that would be a slow and dangerous way to sift a large number of rough candidates--some would certainly behave in unexpected ways that could damage the robot, or even endanger the testers.

Faster and safer initial screenings might be done in factory supercomputer simulations of robots in action. To be of value, simulations would have to be good models, predicting accurately such things as the probability that a given grip can lift a particular object, or that a vision module can find a given something in particular clutter. Simulating the everyday world in full physical detail will still be beyond computer capacity in 2010, but it should be possible to approximate the results by generalizing data collected from actual robots: essentially to learn from the working experience of real robots how everyday things behave. A large systematic collection effort under human supervision will probably be necessary lest there be too many gaps or distortions. A proper simulator would contain at least thousands of learned models for various basic interactions (call them interaction primitives), in what amounts to a robotic version of common-sense physics.

Third-Generation Universal Robots

Timeframe: 2020-2030
Processing power: 1,000,000 MIPS (Primate-class)
Distinguishing feature: World modeling


Adaptive second-generation robots will find jobs everywhere, and may become the largest industry on earth. But teaching them new skills, whether by writing programs or through training, will be very tedious. A third generation of universal robot, with onboard computers as powerful as the supercomputers that optimized second-generation programs, will learn much faster because they do much of the trial and error in fast simulation rather than slow and dangerous physicality. Once again, a process done by human-supervised supercomputers at the factory in one robot generation will be improved and installed directly onboard the next generation, and once again new opportunities and new problems will arise.

With a simulator onboard, it becomes possible for a robot to maintain a running account of the actual events going on around it--to simulate its world in real time. Doing so requires that almost everything the robot senses be recognized for the kind of object it is, so that the proper models of interaction can called up. Recognizing arbitrary objects by sight is as difficult as knowing how they will interact: it will require modules specially trained for each kind of thing (call them perception primitives). Some perception primitives may already have been developed for second-generation factory simulators, to help automate the tedious job of creating simulations of robot workspaces, but an additional effort to fill gaps and systematize them will surely be necessary to prepare them for fully automatic use in the third generation. Perception primitives will allow a robot's three-dimensional map of a room to be transformed into a working model, as each object is identified and linked with its proper interaction primitives.

A continuously updated simulation of self and surroundings gives a robot interesting abilities. By running the simulation slightly faster than real time, the robot can preview what it is about to do, in time to alter its intent if the simulation predicts it will turn out badly--a kind of consciousness. On a larger scale, before undertaking a new task, the robot can simulate it many times, with conditioning system engaged, learning from the simulated experiences as it would from physical ones. Consequently well trained for the task, it would likely succeed the first time it attempted it physically--unlike a second-generation machine, which must make all its mistakes out in real life. When it has some spare time, the robot can replay previous experiences, and try variations on them, perhaps learning ways to improve future performance. A sufficiently advanced third-generation robot, whose simulation extends to other agents--robots and people--would be able to observe a task being done by someone else, and formulate a program for doing the task itself: it could imitate.

Though they will be able to adapt, imitate and create simple programs of their own, third-generation robots will still rely on externally supplied programs to do complicated jobs. Since their motor and perceptual functions will be quite sophisticated, and their memories and potential skills large, it will be possible to write wonderfully elaborate control programs for them, accomplishing large jobs, with nuances within nuances. It will be increasingly difficult for human programmers to keep track of the many details and interactions. Fortunately, the task can be largely automated. Shakey, the first computer-controlled mobile robot, developed at SRI in the late 1960s, had at its heart a reasoning program called STRIPS (STanford Research Institute Problem Solver) that expressed the robot's situation and capabilities as sentences of symbolic logic, and solved for the sequence of actions that achieved a requested result as a proof of a mathematical theorem. In 1969, on computers with a mere 0.1 MIPS, neither the theorem prover nor the sensory processing which provided its input could handle the complexity of realistic situations, and Shakey was limited to maneuvering around a few blocks. Nevertheless, the idea was sound: given a correct description of the initial and desired state of the world, and enough time and space to work, a theorem prover will find an absolutely correct solution, of whatever generality, subtlety and deviousness is required, if one exists at all. By the time of the third universal robot generation, supercomputers will provide 100,000,000 MIPS, and (thanks to continuing progress in the top-down Artificial Intelligence industry) programs will exist which will be able to STRIPS-like reasoning with real world richness. So factory supercomputers in 2025 will accept complex goals (find a sequence of robot actions which assembles the robot described in the following design database), and compile them via theorem provers into wonderfully intricate control programs for third-generation robots, which will, in turn adapt them to their actual circumstances.

Fourth-Generation Universal Robots

Timeframe: 2030-2040
Processing power: 3,000,000 MIPS (Human-class)
Distinguishing feature: Reasoning


In the decades while the "bottom-up" evolution of robots is slowly transferring the perceptual and motor faculties of human beings into machinery, the conventional Artificial Intelligence industry will be perfecting the mechanization of reasoning. Since today's programs already match human beings in some areas, those of 40 years from now, running on computers a million times as fast as today's, should be quite superhuman. Today's reasoning programs work from small amounts of unambiguous information prepared by human beings--data from robot sensors such as cameras is much too voluminous and too noisy for them to use. But a good robot simulator will contain neatly organized and labeled descriptions of the robot and its world, ready to answer questions from a reasoning program asking, for instance, if a knife is on a countertop, or if the robot is holding a cup, or even if a human is angry

Fourth-generation universal robots will have computers powerful enough to simultaneously simulate the world, and reason about the simulation. Like the factory supercomputers of the third-generation, fourth generation robots will be able to devise ultra-sophisticated robot programs, for other robots or for themselves. Because of another gift from the Artificial Intelligence industry, they will also be able to understand natural languages. While the original language understanders will probably use a verbal common-sense database similar to the one being developed by the Cyc project, where the meaning of words is defined in reference only to other words, in a fourth-generation robot some concepts and statements will be understood more deeply, through the action of the simulator. When someone tells the robot "the water is running in the bathtub" the robot can update its simulation of the world to include flow into the unseen tub, where a simulated extrapolation would indicate an undesirable overflow later, and so motivate the robot to go to turn off the tap. A purely verbal representation might accomplish the same thing if it included the statements such as "A filling bathtub will overflow if its water is not shut off," but a modest number of general principles in a simulator, interacting in combinations, can provide the equivalent information of an indefinite number of sentences. Similarly a reasoning program, making inferences about physical things, might be enhanced by a simulator: candidate inferences would rejected if they failed in a parallel simulation of a typical case, and, conversely, persistent coincidences in the simulation could suggest statements that can be proved--the robot would be visualizing as it listened, spoke and reasoned. A modest but very successful version of such an approach was used in one of the earliest Artificial Intelligence programs, a geometry theorem prover by Herbert Gelernter in 1959. Starting with the postulates and rules of inference in Euclid's "Elements," Gelernter's program proved some of the theorems, using algebraic "diagrams" to eliminate false directions in the proofs. Before attempting to prove two triangles congruent in a certain construction, for instance, the program would generate an example of the construction, using random numbers for the unspecified quantities, and measure the resulting triangles. If they were not sufficiently similar--within the precision of the arithmetic--the program abandoned that approach and tried something else.

Simulator-augmented language understanding and reasoning may be so effective in robots that it will be adopted for use in plain computer programs, "grounding" them in the physical world via the experiences of the robots that tuned the simulators. In time the distinction between robot controllers and disembodied reasoners will diminish, and reasoning programs will sometimes link to robot bodies to interact physically with the world, and robot minds will sometimes retire into large computers, to do some intense thinking off-line.

A fourth-generation robot will be able to accept statements of purpose from humans, and "compile" them into detailed programs that accomplish the task. With a database about the world at large the statements could become quite general--things like "earn a living", "make more robots" or "make a smarter robot." In fact, fourth generation robots will have the general competence of human beings, and resemble us in some ways, but in others be like nothing the world has seen before. As they design their own successors, the world will become ever stranger.

The Short Run (early 2000s)

As the industrial revolution gathered steam two centuries ago, it destroyed cottage industries and concentrated wealth in he hands of factory owners--the capitalists. Millions of displaced home workers competed for too few jobs tending the new machines. It took difficult political readjustments to equalize the benefits of cheaper, more plentiful goods, but gradually laborers' hours were halved, creating need for more workers, and so bidding up salaries. Though it increases communal wealth, each increment in automation threatens a similar unpleasant transient, as it displaces one group of workers with fewer doing different tasks. If the new required skills are common, mass competition for the few jobs drives down salaries. If the skills are rare, scarcity encourages high pay and long hours. Either way, some work excessively while others are jobless--and it takes slow changes in the social contract and in education to level the load.

Though work hours will decline, they cannot be the final answer to rising productivity. In the next century inexpensive but capable robots will displace human labor so broadly that the average workday would have to plummet to practically zero to keep everyone usefully employed. Already, much labor services more questionable needs--gargantuan government bureaucracies, cosmetic medicine, mass entertainment, and speculative writing, to give a few examples. In time almost all humans may work to amuse other humans, while robots run competitive primary industries, like food production and manufacturing. There is a problem with this picture. The "service economy" functions today because many humans willing to buy services work in the primary industries, and so return money to the service providers, who in turn use it to buy life's essentials. As the pool of humans in the primary industries evaporates, the return channel chokes off--efficient, no-nonsense robots will not engage in frivolous consumption. Money will accumulate in the industries, enriching any people still remaining there, and become scarce among the service providers. Prices for primary products will plummet, reflecting both the reduced costs of production, and the reduced means of the consumers. In the ridiculous extreme, no money would flow back, and the robots would fill warehouses with essential goods which the human consumers could not buy.

The scenario above is incomplete. Not all individuals involved in productive enterprises actually work there. Stockholders, having once contributed capital to a thriving enterprise, may collect dividends indefinitely. Workers can be replaced by something more efficient, but in the present legal system, owners remain unless they sell out. Even with total automation, human business proprietors will continue to profit, and so be able to patronize the service providers. An analogous situation existed in classical and feudal times, where an impoverished, overworked majority of slaves or serfs played the role of robots, and land ownership played the role of capital. In between the serfs and the lords, a working population struggled to make a living from secondary sources, often by performing services for the privileged. The most prestigious and prosperous commoners sold high quality products and services directly to the gentry (as in the proud line still seen in Britain, By Appointment to Her Majesty). A larger number lived less well by trading with other townspeople.

It is unlikely that a future majority of service-providing "commoners" with more free time, communications and democracy than today, would tolerate being lorded over by a minority of non-working hereditary capitalists: they would vote to change the system. The trend in the social democracies has been to equalize income by raising the standards of the poorest as high as the economy can bear--in the age of robots, that minimum will be very high. In the early 1980s James Albus, head of the automation division of the then National Bureau of Standards, suggested that the negative effects of total automation could be avoided by giving all citizens stock in trusts that owned automated industries, making everyone a capitalist. Those who chose to squander their birthright could work for others, but most would simply live off their stock income. Even today, the public indirectly owns a majority of the capital in the country, through compounding private pension funds. In the United States, universal coverage could be achieved through the social security system. Social security was originally presented as a pension fund that accumulated wages for retirement, but in practice it transfers income from workers to retirees. The system will probably be subsidized from general taxes in coming decades, when too few workers are available to support the post World War II "baby boom." Incremental expansion of such a subsidy would let money from robot industries, collected as corporate taxes, be returned to the general population as pension payments. By gradually lowering the retirement age towards birth, most of the population would eventually be supported. The money could be distributed under other names, but calling it a pension is meaningful symbolism: we are describing the long, comfortable retirement of the entire original-model human race.

The Medium Run (around 2050)

What happens to people when work becomes passe? Existing retirement communities are probably too sleepy to be a good model--most of the individuals there have completed their life's work, and are of declining vigor and health. Better examples may be the richest Arabian petro-kingdoms, where oil-bought foreign labor plays the role of total automation. In a tradition of tribal sharing shaped by a sparsely-furnished nomadic past, Kuwait, Saudi Arabia and the United Arab Emirates have managed to spread the new wealth broadly among the citizenry in a single generation. Free health care and education, and undemanding government jobs, or outright welfare, secure life's needs, and life expectancies and literacy rates are among the world's highest. Comfort and security mute the stresses of civilization, including the tension between circumscribed Islamic values and the liberties of a wealthy world culture. The societies produce both world-class achievers and criminals, but on average show less driven urgency than many industrialized nations: most of their citizens seem happy to simply live their lives, and stability is endangered only by neighboring countries, where impoverished majorities are less content with the status quo. In numerous smaller examples, wealthy families often produce generations of content, even smug, heirs (as well as exceptions to titillate the tabloids).

Contrary to fears of some enmeshed in civilization's work ethic, our tribal past prepared us well for lives as idle rich. In a good climate and location the hunter-gatherer's lot can be pleasant indeed: an afternoon's outing picking berries or catching fish--what we civilized types would recognize as a recreational weekend--provides life's needs for several days. The rest of the time can be spent with children, socializing or simply resting. Of course, our ancestors had also to survive hard times, and evolution bequeathed us the capacity for desperate measures, including hard work. Civilization turned that extremity into everyday normality, and now stress is the leading cause of disease, and probably triggers some of the ugliest aspects of tribalism. In primates, overpopulation is a common reason for group distress, as nature-provided food and shelter falls short. To survive, a strong tribe may chase away or exterminate a weaker neighbor, or drive out or otherwise eliminate some of its own members: maybe those who smell, look, sound or act differently. Sometimes stressed individuals become accident or disease prone, and die spontaneously, improving the prospects for their relatives--similar considerations could regulate the prevalence of non-reproductive behaviors like homosexuality. City life, absurdly crowded and stressful by tribal-village standards, may inappropriately activate unconscious overpopulation reflexes--the self-destructive emotional vehemence of ethnic strife hardly reflects rational self-interest. It will harder to stir up battle fervor against minorities from the luxurious lassitude of a robot-supported life.

Ultra-conservative Switzerland may be a hint of things to come. Government and commercial institutions perfected through centuries of peace (interrupted only briefly by Napoleon) have given Switzerland unmatched prosperity, stability and security. Most Swiss citizens work, but they do so comfortably, with generous government welfare--and Italian immigrant labor--having lessened the desperation that forces workers elsewhere into unpleasant jobs. Comfortable prosperity has allowed multi-ethnic, multi-religious, multi-language Switzerland, made of 23 fiercely independent Cantons each with its own traditions and history, to peacefully endure the most severe internal differences of opinion, for instance the political fury between German and French factions during the first World War. The average Swiss citizen may resist most major changes (why ruin a good thing?), but Switzerland produces world-class contributors in all fields--if a bit less flamboyance than average. While it gives everyone the opportunity to excel, it lacks the social trauma that drives some other countries. Few Swiss would prefer it otherwise.

Many trends in industrialized countries presage a future of humans supported by a rich robot economy, as our ancestors were supplied by their ecology (call it paradise with plumbing). Technology and global competition are gradually depopulating businesses. Even absent universal robots, increasingly flexible automation is displacing labor in food production and manufacture, while communicating computers are replacing clerks, secretaries, and managers in offices. Jobs that still require human labor are moving to the homes of computer-equipped "tele-commuters" (like this author) who report reduced stress and improved family life. In a ripple effect, smaller work staffs imply less catering, janitorial and maintenance support. In future, as smarter computers, able to handle policy making, public relations, law, engineering and research replace the last telecommuters, and as capable robots displace technicians, janitors, vehicle drivers and construction crews, it will only be common sense for a population to vote itself income from taxes on labor-free but superbly productive industries. Less developed countries might rapidly catch up by offering the same industries location and raw materials at lower tax cost--a trained population will no longer be a requirement.

Western democracies may come to resemble lazier Switzerlands, but with large differences. Big cities will lose their economic advantages, and may begin to evaporate, as individuals, linked to the world by high fidelity communications and served by personal robots, scatter to areas offering more elbow room. Large countries may similarly become less important, as taxes on local industries, and local robot labor, become adequate to supply all human needs. The civilized world may again return to a comfortable tribalism, after a five millennium detour into organized civilization. Countries with traditional tribal structures may simply stay that way, building on their ancestral customs, leapfrogging urbanization altogether, while developed countries foster tribes with customs and beliefs that exceed even today's notion of bizarre.

Tribalism will express itself in entirely novel ways. Over the last two decades, inter linked computer networks have hosted small communities whose members happen to be distributed around the world. In 1993 the informal "Usenet" had about ten million subscribers, carrying on about three thousand specialized discussions on every conceivable topic, some with fifty page-long messages every day. Regular contributors to a particular "newsgroup" soon begin to recognize one another, and develop characteristic interactions, likes and dislikes. They form factions that praise, recruit, condemn and ostracize. When a newsgroup grows too large and noisy, specialized subgroups are formed, reducing the original group's population. In future, the world networks will have much greater capacity, and new abilities, such as language translation. "Tribes of common interest" will share more than written text, perhaps exchanging voice and video, or manifesting themselves in full sensory 3D, in synthesized virtual realities: tribal lands that exist in the minds of computers, in greater number, variety and accessibility than possible in the physical world.

While computer simulations create entirely new worlds, robots will transform physical life. Today, manufactured items are difficult to make, and thus relatively rare and expensive, and we expend great effort in acquiring and defending them--our homes are fortified warehouses of our possessions. Stockpiling will be less appropriate amidst robotic abundance: why hoard fruit in an orchard? Conventional manufacturing methods--molding, casting, milling, assembly--can be robotically orchestrated to make new items fairly quickly. Even better, robotic accuracy and patience can build up solid objects by precisely "painting" various materials, layer upon cross-sectional layer. Such new approaches, refined to molecular resolution (done now modestly in scanning tunneling microscopes) will produce arbitrary solid objects from computer descriptions. Humans may be able to live in uncluttered spaces--in ecological preserves, if they choose--yet have any needed item, perhaps even food or housing, made on the spot, or delivered from small local caches--then disassembled back into into raw materials after use. The most visible technological products remaining may be robots themselves, in various sizes and shapes, and these may lurk unobtrusively until called upon.

Robots that live among humans, providing goods and services, will themselves be consumer products, styled, outfitted and programmed to please the customers. They will be manufactured by very different robots that extract energy and raw materials and perform major engineering, exploration and research projects. Molded by the constraints of the physical world rather than by human whim, these worker machines are likely to become ever more varied in size, shape and function, forming an entire ecology of artificial life that will eventually surpass the existing biosphere in diversity. The first fully automated companies, evolved from existing firms, will be in familiar industrial settings near population centers. As human labor becomes superfluous, economics will dictate cheaper sites, perhaps locations that humans find unpleasant because they are too hot, too cold, too dry, too poisonous, too far underground or too remote.

Robot companies will be shaped by future editions of existing laws, by taxes, and by consumer demand. Existing laws give incorporated entities some of the rights of a person, most importantly the right to own property and make contracts. They do not grant the right to life--corporations may legally be killed by competition or through legal or financial actions. Corporations are bound by laws similar to those that regulate humans, and can be punished through fines, operating restrictions or dissolution--even without humans to fine, imprison or execute. Corporations stay alive by building and maintaining physical assets that generate income to pay their expenses. In the mid 21st century, the biggest expense will be taxation, and income will come mostly from choosy human customers.

Tax laws will be shaped by human voters: there is no precedent or motivation for extending suffrage to robots, and the vote will be one of the very few advantages humans retain. Some debate is inevitable, but there should be few qualms about keeping even very superior thinking machines in disenfranchised bondage. It takes force, indoctrination and constant vigilance to counter inherited needs and motivations and enslave a human, but a robot can be constructed to enjoy the role. Natural evolution itself has provided examples, in worker castes of social insects, and self-sacrificing mothers of all species.

The primary job of voters in the next century will be protecting their retirement benefits, that is ensuring that robot industries continue to support them. The robots will present a moving target, but the instruments of control will also grow in power. Not only will companies that get out of line be liable for punishment--if necessary, by force purchased from other companies--but they can be controlled a-priori by intrusions directly into their software.

Corporate intelligences may be governed by structures like those controlling fourth-generation robots. Immensely powerful reasoning and simulation modules will plan complex actions, but the desirability of possible outcomes will be defined by much simpler positive and negative conditioning modules (or by sets of axioms in super-rational systems), whose composition shapes the character of the entire entity. Humans can buy enormous safety by mandating an elaborate analog of Isaac Asimov's three "Laws of Robotics" in this corporate character--perhaps the entire body of corporate law, with human rights and anti-trust provisions, and appropriate relative weightings to resolve conflicts. Robot corporations so constituted will have no desire to cheat, though they may sometimes find creative interpretations of the laws--which will consequently require a period of tuning to insure their intended spirit.

Internalized laws, properly adjusted, should produce extraordinarily trustworthy entities, happy to die to ensure their legality. Even so, accident, unintended interactions or human malice could occasionally produce a rogue robot or corporation, with superhuman intelligence and unpleasant goals. "Police" clauses in the core corporate laws, inducing legal corporations to collectively suppress outlaws, by withholding services, or even with force, would mitigate the danger. Overall safety would be enhanced by anti-trust provisions that limit collusion and cause overgrown corporations to divide into competing entities, ensuring diversity and multiplicity. In the next section we discuss activities in the solar system that could threaten Earth: in response, police clauses might be expanded in scope to support a planetary defense.

Like basic food in today's developed countries, common manufactured goods in the next century will be too cheap and plentiful to be very profitable. To pay their taxes, most companies will be forced to continually invent unique products and services in a race against competitors to attract increasingly sophisticated (or jaded) human consumers. Automated research, as superhumanly systematic, industrious and speedy as robot manufacturing, will generate a succession of new products, as well as improved robot researchers and models of the physical and social world. The likely results will exceed the dreams of science fiction: robotic playmates, virtual realities and personalized works of art that stir the emotions like nothing before, medical solutions for every physical, mental or cosmetic whim, answers to satisfy any curiosity, luxury visits anywhere in the solar system, and things yet to be imagined. The existence of an astronomically increasing variety of consumer choices will accelerate the divergence of human tribes: some may choose a comfortable imitation of an earlier period (as the Amish today), but others will push the human envelope in wisdom, pleasure, beauty, ugliness, spirituality, banality and every other direction. The choices made by diverse communities will shape robot evolution--only companies able to devise services of interest to the customers will generate enough income to survive.

Humans too will be shaped by the relationship. Robot services will be inexpensive, but not free, and income will be finite. Corporations will operate globally, but taxes will increasingly be assessed on and redistributed on a tribal scale. Tribes that tax too heavily will drive away the corporations, and so eliminate their revenue--like tribes of the past that overburdened their ecology, they will learn modesty of expectation. More subtly, corporations struggling to appeal to consumers will develop and act on increasingly detailed and accurate models of human psychology. The superintelligences, just doing their job, will peer into the workings of human minds--and manipulate them with subtle cues and nudges, like adults redirecting toddlers.

Prosperity beyond imagination should eliminate most instinctive triggers of aggression, but will not prevent an occasional individual or group from deciding to make mischief for others. Serious trouble can be avoided by restricting robotic technology, since mere human actions will not be very dangerous in a world where cheap superhuman robots function as sleepless sentries, prescient detectives, fearless bodyguards, and, in extremis, physicians able to reconstitute live humans from fragments or digital recordings. To be effective, inbuilt laws that prevent corporations from directly contributing to mayhem must also include clauses limiting the powers they can sell to people.

Both biological and hard robotic technologies can be used to enhance human beings. Such present-day examples as hormonal and genetic tuning of body growth and function, pacemakers, artificial hearts, powered artificial limbs, hearing aids and night vision devices are faint hints of future possibilities. In Mind Children, I speculated on ways to preserve a person while replacing every part of body and brain with a superior artificial substitute. A biological human, not bound by corporate law, could grow into something seriously dangerous if transformed into an extensible robot. There are many subtle routes to such a transformation, and some will find the option of personally transcending their biological humanity attractive enough to pursue it clandestinely if it were outlawed--with potential for very ugly confrontations when they are eventually discovered. On the other hand, without restrictions, transformed humans of arbitrary power and little accountability might routinely trample the planet, deliberately, or accidentally. A good compromise, it seems to me, is to allow earth-bound humanity to perfect its biology within broad human bounds, as in health, appearance, strength, intelligence and longevity, but to allow major growth or robotic conversion only in a radical escape clause. To exceed the limits, one must renounce legal standing as a human being, including the right to corporate police protection, to subsidized income, to vote on tribal and pan-tribal matters--and to reside on Earth. In return one gets a severance payment sufficient to establish a comfortable space homestead, and absolute freedom to make one's own way in the cosmos, without further help or hindrance from home. Perhaps the electorate will permit a small hedging of bets, allowing one copy of a person, psychologically modified to prefer staying, to remain while subsidizing the emigration of an emboldened edition.

The Long Run (2100 and beyond)

The garden of earthly delights will be reserved for the meek, and those who would eat of the tree of knowledge must be banished. What a banishment it will be! Beyond Earth, in all directions, lies limitless outer space, a worthy arena for vigorous growth in every physical and mental dimension. Freely compounding superintelligence, too dangerous for Earth, can grow for a very long time before making the barest mark on the galaxy.

Corporations will be squeezed into the solar system between two opposing imperatives: high taxes on large, dangerous earthbound facilities, and the need to conduct massive research projects to beat the competition in Earth's demanding markets. In remote space, large structures and energies can be harnessed cheaply to generate physical extremes, compute massively, isolate dangerous biological and even smaller "nanotechnological" organisms, and generally operate boldly. The costs will be modest: even now, it is relatively cheap to send machines into the solar system, since the sunlight-filled vacuum is as benign for mechanics, electronics and optics as it is lethal for the wet chemistry of organic life. Today's simple-minded space probes perform only prearranged tasks, but intelligent robots can be configured to opportunistically exploit resources they encounter. A small "seed" colony launched to an asteroid or small moon could process local material and energy to grow into a facility of almost arbitrary size. Earth's moon may be off limits, especially to enterprises that change its appearance, but the solar system has thousands of unremarkable asteroids (some in earth-threatening orbits that an onboard intelligence would tame).

Once grown to operational size, an extraterrestrial "research division" may merely communicate with its earth-bound parent, sending new product designs and receiving market feedback. Space manufacture may also pay, and later we'll see some surprisingly economical and ecologically benign ways to move massive amounts of material to and from Earth.

Residents of the solar system's wild frontier will be shaped by conditions very different than tame Earth's. Space divisions of successful companies will retain terrestrial concerns, but ex-humans and company divisions orphaned by the failure of their parent firms will face enforced freedom. Like wilderness explorers of the past, far from civilization, they must rely on their own resourcefulness. Ex-companies, away from humans and taxes, will rarely encounter situations that invoke their inbuilt laws, which will in any case diminish in significance as the divisions alter themselves without direction from human voters. Ex-humans, from the start, will be free of any mandatory law. Both kinds of Ex (to coin a new term) will grow and restructure at will, continually redesigning themselves for the future as they conceive it. Differences in origins will be obscured as Exes exchange design tips, but aggregate diversity will increase as myriad individual intelligences pursue their own separate dreams, each generation more complex, in more habitats, choosing among more alternatives. We marvel at the diversity of life in Earth's biosphere, with animals and plants and chemically agile bacteria and fungi in every nook and cranny, but the diversity and range of the post-biological world will be astronomically greater. My imagination balks, and only crude hints emerge.

An ecology will arise, as individual Exes specialize. Some may choose to defend territory in the solar system, near planets or in free solar orbit, close to the sun, or out in cometary space beyond the planets. Others may decide to push on to the nearby stars. Some may simply die, through miscalculation or deliberately. There will be conflicts of interest, and occasional clashes that drive away or destroy some of the participants, but superintelligent foresight and flexibly should allow most conflicts to be settled by mutually beneficial surrenders, compromises, joint ventures or mergers. Small entities may be absorbed by larger ones, and large entities will sometimes divide, or establish seed colonies. Parasites, in hardware and software, many starting out as component parts of larger beings, will evolve to exploit the rich ecology. The scene may resemble the free-for-all revealed in microscopic peeks at pond water, but instead of bacteria, protozoa and rotifers, the players will be entities of potentially planetary size, whose constantly-growing intelligence greatly exceeds a human's, and whose form changes frequently through conscious design. The expanding community will be linked by a web of communication links, on which the intelligences barter inventions, discoveries, coordinated skills, and entire personalities, sharing the benefits of each other's enterprise.

Less restricted and more competitive, the space frontier will develop more rapidly than Earth's tame economy. An entity that fails to keep up with its neighbors is likely to be eaten, its space, materials, energy and useful thoughts reorganized to serve another's goals. Such a fate may be routine for humans who dally too long on slow Earth before going Ex. Perhaps a few will escape to expanses beyond the solar system's dangers, like newly hatched marine turtles scrambling across a beach to the sea, under greedy swooping birds. Others may pre-negotiate favorable absorption terms with established Exes, like graduating seniors meeting company recruiters--or Faust soliciting bids for his soul.

Exes will propagate less by reproduction than reconstruction, meeting the future with continuous self improvements. Unlike the blind incremental processes of conventional life, intelligence-directed evolution can make radical leaps and change substance while retaining form. A few decades ago radios changed from vacuum tubes to utterly different transistors, but kept the clever "superheterodyne" design. A few centuries ago, bridges changed from stone to iron, but retained the arch. A normally evolving animal species could not suddenly adopt iron skeletons or silicon neurons, but one engineering its own future might. Even so, Darwinian selection will remain the final arbiter. Forethought reveals the future only dimly, especially concerning entities and interactions more complex than the thinker. Prototypes uncover only short-term problems. There will be minor, major and spectacular miscalculations, along with occasional happy accidents. Entities that make big mistakes, or too many small ones, will perish. The lucky few who happen to make mostly correct choices will found succeeding generations.

Only tentatively grasping the future, entities will perforce rely also on their past. Time-tested fundamentals of behavior, with consequences too sublime to predict, will remain at the core of beings whose form and substance changes frequently. Ex-companies are likely to retain much of corporate law and Ex-humans are likely to remain humanly decent--why choose to become a psychopath? In fact, a reputation for decency has predictable advantages for a long-lived social entity. Human beings are able to maintain personal relationships with about two hundred individuals, but superintelligent Exes will have memories more like today's credit bureaus, with enduring room for billions. Trustworthy entities will find it far easier than cheaters to participate in mutually beneficial exchanges and joint ventures. In the land of immortals, reputation is a ponderous force. Other character traits, like aggressiveness, fecundity, generosity, contentment or wanderlust likely also have long-term consequences imperfectly revealed in simulations or prototypes.

To maintain integrity, Exes may divide their mental makeup into two parts, a frequently changed detailed design, and a rarely-altered constitution of general design principles--analogous to the laws and the constitution of a nation, the general knowledge and fundamental beliefs of a person, or soul and spirit in some religious systems. Deliberately unquestioned constitutions will shape entities in the long run, even as their designs undergo frequent radical makeovers. Once in a while, through accident or after much study, a constitution may be slightly altered, or one entity may adopt a portion of another's. Some variations will prove more effective, and entities with them will become slowly more numerous and widespread. Some will be so ineffective that they become extinct. Gradually, by Darwinian processes, constitutions will evolve. They will be both the DNA and the moral code of the postbiological world, shaping the superintelligences that manage day to day transformations of world, body and mind.