Intro Soup Recipe Shelf Sink Chef
Take me higher...

Lagniappe #2

Introduction

I am a brain builder. My dream and ambition in life is to build (i.e. grow/evolve) artificial brains with billions (and upwards) of artificial brain cells, and before I retire (which will be in about 20 years) to see brain building grow into one of the world's largest industries, comparable with construction, electronics, automobiles and oil. I believe I have found a methodology to enable such an industry to develop. If this occurs, I will be very happy, but this is not what this chapter is about. The creation of artificial brains I believe will be achievable within the next few years (at least with hundreds of thousands of artificial neurons whose circuitry is evolved at electronic speeds in special hardware I call Darwin Machines. This period is the short to middle term. What concerns me is the longer term, i.e. well into the 21st century, if I and other brain builders succeed. What then?

I truly believe that the question which will dominate global politics next century will be "Who or what should be dominant species on the planet?" In case this sounds far fetched, consider the following. Imagine a computer with 10 to power 30 components, i.e. a million trillion. This number dwarfs the number of neurons we have in our brains by a factor of 10 to power 18, i.e. a billion human brains would have as many neurons as this computer would have components. I believe we (i.e. humanity) will have such machines within a few decades. The writing is already on the wall. Today, we already have single electron transistors and nano electronics, i.e. molecular scale electronics, which will create a storage capacity of one bit of information per atom. With modern reversible circuitry, it is theoretically possible to perform computer calculations without dissipating heat. Therefore it becomes possible to build 3D circuits without them melting, as would be the case with traditional non reversible electronic circuits if they were in 3D. With adiabatic (heat dissipation less) circuitry it will be possible to build tons of computing material, and if one can build tons, why not build huge asteroid size computers containing 10 to power 40 components and more, by taking the atoms from the asteroid belt. By employing quantum dot technology it will be possible to build "artificial atoms" having the electronic properties one wishes. The circuits could be built with nano scale robots which assemble the building blocks which future Darwin Machines can use to evolve ultra complex circuits (too complex for human engineers to understand). Since these circuits would operate at electronic speeds, i.e. a million times faster than the neural transmissions in our human brains, and given the hugely superior memory capacity, complexity, and number of sensors which could be attached to such devices, it is clear to any informed person that the potential of the computer to generate intelligence is way superior to that of the human brain with its pitiful trillion neurons.

Therefore, I believe that it is only a question of time, before the steady increase in the capacities of computers will begin to generate what I call the "species dominance" debate. At the present time, what I am saying sounds like science fiction to most people and therefore not to be taken seriously or to be concerned about. But, I am a brain builder. I am one of the people who is creating this problem. I'm also an ex theoretical physicist, so I have the consciousness that my ex colleagues a generation earlier, built a device of mass destruction which could have destroyed planetary life as a result of the Cuban Missile Crisis in 1962. I live in Japan, so I went deliberately to Hiroshima to see for myself what horrors scientists can inflict indirectly upon humanity. I was profoundly moved, and each time I give a talk about brain building and the future, I get choked up when referring to what I saw at the Hiroshima museum (e.g. women and children with their skin flaking off in 20 cm strips, molten wine bottles, etc.). This experience made me very conscious of the need for scientists to give thought to the social and political consequences of their work. In my case, with my left hand I am passionately involved with my dream of building artificial brains, and with my right hand, I write articles such as this one, trying to make humanity conscious of the longer term consequences of brain building if it is successful. In the middle term, artificial brains of chimp intelligence would be very useful (so long as they were finally obedient). They could be used to clean the house. Every affluent household would have one, and hence create the market for one of the world's biggest industries. Brain like computers could become personal companions. It is imaginable that such machines could talk, listen, and amuse. It would be possible to have a kind of "relationship" with such machines. But, what worries me is the above numerical comparison between the number of neurons in the human brain and the possibility of 10 to power 30 component machines. I can imagine that the field of brain building will expand as more and more neural circuits can be evolved, as well as their interconnections and control systems, etc. Year by year, we will see increasingly capable devices put onto the market, so that even ordinary people will be able to identify with the question "Where is all this leading?" "How smart should we allow these machines to become?" "The scientists tell us that the machines have the potential to far outstrip human beings and they get obviously smarter every year, so where and when do we draw the line?" "Should we draw the line at all?"

Such questions will become topical and commonplace once the machines truly begin to be based on artificial brain technology that I and other people are developing. I keep thinking that if I am part of the problem, I'd like to be part of the solution (if there is one). One of the things I would like to do, besides build brains, is create an awareness that the "species dominance debate" is coming, so that people have time to think about it before the machines get too smart. I label these ultra intelligent machines "artilects" (as in "artificial intellect"). So I sometimes talk about the "artilect debate", or "artilect question". I believe this debate will determine the broad outlines of 21st century global politics, as I mentioned earlier. To cut a long story short, I suspect strongly that the controversies over the artilect question will be so bitter that warfare, i.e. 21st century style warfare will be inevitable. I am very pessimistic about this question, and it keeps me awake at nights too frequently for the good of my health.

Of course I would like to be more optimistic about the outcome of long term brain building, but no matter from which angle I look at the problem or which of many scenarios I try to think through, I always end up with something really terrible. Of course, its not difficult to come up with a positive scenario, one in which machines and human beings live in harmony (a favorite amongst women I talk to about this), but these scenarios I find to be unrealistic. I hope events prove me wrong, but my gut feeling tells me that the world is in for a most disturbing 21st century in global political terms.

If so, this is a great pity, because our planet is currently entering a great peaceful era, in which telecommunications are shrinking the planet so much, that a global language and a global government seem to be quite realistic within the next few decades. Stationary orbit satellite TV programs coming down from the sky in their hundreds, and to every global citizen with (dirt cheap) wrist watch receivers/transmitters, thanks to inter-satellite communication, and 90% of them in English, will result in a snowball effect of everyone being able to speak and understand the world language. Ideas will then spread rapidly, resulting in- a cultural homogenization, and hence far greater understanding between peoples, until the peoples become essentially just one giant planetary people. This is the prospect I see for humanity for the next few decades, and I welcome it. Once the superpowers agree, i.e. after China has had its democratic revolution, and Russia is well on the way to a prosperous democracy, the super powers can then impose a world government upon the minor countries, forcing them to obey the decisions of the world government and world court when international political conflicts arise. Trillions of dollars wasted on arms today could then be spent on raising the standard of living of everyone, even Africans.

The question which most divided humanity in the l9th and 20th centuries was that of "Who should own the factories, i.e. the machines?" Some felt that the machines should be in private hands, and others felt that they should be owned by everyone, i.e. the state. In the early industrial revolution, economic exploitation was rampant, and factory owners grew rich by paying subsistence wages to factory workers. When the workers became conscious that the owners were growing rich from the labor of the workers, the latter became bitter and angry. When one feels one is being exploited and that an injustice is being committed at ones expense, one can become murderously angry. As is well known, intellectual currents in Europe in the l9th century began to focus upon solutions to the vicious aspects of early l9th century capitalism. The socialists wanted to collectivize the factories, and share out the wealth. The income generated from sales of the products made by the machines would go to the state, which in turn would pay everyone a wage in accordance to their value to society. This Marxist idea was very seductive to many people, and in the 20th century, many governments were formed based upon these ideals. In the meantime, the workers within the early industrialized countries forced political reforms such as universal (male) suffrage, the right to form labor unions, worker based political parties, progressive taxation, etc., which largely overcame the inhuman face of capitalism. After nearly a century of communistic experimentation, it is now clear that it does not work. Communist regimes are disappearing across the globe. It is clear that the opportunities offered to enterprising, selfish individualists generates more wealth and hence a greater public welfare (provided the above reforms are in place) than a bureaucratically controlled, stultifying centrally planned economies. But, for a while it was really touch and go. It was not clear which of the two major alternatives to social and economic organization would win out. The two regimes, the two ideologies so hated each other, that they were prepared in the limit to exterminate each other in a nuclear holocaust. Fortunately, atomic weapons are so destructive, that the politicians and generals could no longer feel safe "behind the lines" as in conventional warfare, because they would either be killed themselves in the nuclear exchange, or be massacred by their countrymen who survived, in an anti-politician, anti-general witch-hunt. Thank god, it seems that all that is over. We can sleep more peacefully at night, knowing that the "cold war" is over. But is it? I'm not referring to a revival of the old capitalist/communist ideological conflict, but to a new conflict even more bitter, because the stakes are so much higher. The question "Who should own capital?" is trivial compared to the question "Who or what should be dominant species?", because the latter may be concerned with the notion that maybe the rise of the artilects will signal the doom of the human species. What is more important to human beings than our own species survival? The stakes are potentially enormous.

I believe that the peaceful era we now live in will only last a few decades, because the deadly controversy of the recent past, namely between the capitalists and the communists will be replaced by a new dichotomy between the "terrestrialists" and the "cosmists".

Cosmism

I see humanity splitting into two major political groups in the 21st century, namely those (probably the majority) who feel that human beings should remain the dominant species (the "terrestrialists"), and those who feel that the machines should be given the opportunity to become the next form of dominant species, transcending the limitations of human beings (the "cosmists"). The primary motivator of the terrestrialists (or "terras" for short) I believe will be fear - fear of being exterminated by a vastly superior species, and fear of the unknown and the unknowable. The primary motivator of the cosmists I believe will be wonder. Personally I share the views of both groups, but when push comes to shove, would I be a terra or a cosmist? That's a tough one! Let me try to elaborate on these two main views. I begin with the cosmists, because it is they who will be introducing something new to global politics, i.e. the idea that the machines should be allowed to become artilects. Terrestrialism will be a reaction to cosmism. At the moment, there is no debate, because state of the art technology cannot produce artilects, but nanotechnology, and especially nanoelectronics will allow such a possibility. That is why this chapter is relevant to this book on nanotech [Drexler 1992]. Cosmism is a 21st century ideology because it will require 21st century technology to make it possible. The initiative will lie in the hands of the cosmists. What then do the cosmists believe (in other words, what do I think the future group of people in favor of artilect production, will propose in political terms?)

To answer this question, I will try to appeal to your gut, to your emotional instincts. When was the last time that you were out in the country, far from the city lights on a moonless night, and you slept out under the stars? Maybe never, but if you have, and you kept staring up at the Milky Way, did you not feel overwhelmed by the triviality of human existence? You may have been preoccupied by the dominant issues of your life, "Will Sue marry me?", "Will I get that promotion?", "Is my growth cancerous?", etc., etc. Answer? "Who cares?" There are a trillion trillion stars, probably most of them (at least second generation stars) have planetary systems, and some of them are billions of years older than our solar system. Hence there are probably civilizations out there in their zillions. Who cares if you might die next year, or if you will have to sell the house because you cant keep up the mortgage payments after being fired, and that your wife left you and took the kids, and that your drinking made you lose all your friends, and that you have a heart problem because you have been eating and smoking way too much the past few years. Who cares? Your-- little life is of no significance when compared to what you see when looking up at the stars. There are bigger things, higher things, issues of cosmic significance. Big questions like - "Why is there matter?", "What's the point of existence, if any?" "Are there ETIs (extra terrestrial intelligences) out there?" "What do they do?" "Have they discovered the answers to the really big questions of universal i.e. cosmic significance?"

The cosmists are concerned with these cosmic issues. Cosmism is a "scientist's religion", i.e. a religion in the sense that traditional religions energize and direct the lives of groups of people. In the case of cosmism, billions of people will be involved, because it concerns the destiny of the human species. Traditional religions and modern sects are not scientific. Some profess to be, but mostly they are just superstitious dogmas, based on ideas that scientific knowledge treats with contempt. Few critically minded, scientifically educated people are religious. But cosmism is essentially a scientific vision or enterprise. By its nature, it is very "high tech". Its beliefs are quite compatible with the latest in scientific knowledge, and therefore is not easily dismissed by cynical scientists who treat such two thousand year old superstitions such as the idea of life after death, or resurrections, or virgin births, or the like (in western cultures) with derision. The religious impulse is probably in all of us. It is one of the few cultural universals, such as the incest taboo. Virtually all cultures (some 5000 to lO,OOO of them on this planet) have invented their own gods.

Personally speaking, it would be nice to believe in something that gave me a sense of direction and even a sense of wonder or excitement, but was also compatible with my scientific knowledge and cynical questioning mind. Cosmism would be a serious candidate for such a "religion" or ideology. Perhaps the word religion is more suitable than the word ideology, because when one asks what the difference is between an ideology and a religion (because the two have many common characteristics), the immediately obvious answer is that a religion usually is concerned with beliefs in some kind of super being with power over human beings, whom often human beings can influence to their own benefit. Cosmists could argue that the artilects would be the new gods, perhaps more capable than human beings of answering the big questions, and exploring and understanding the utterly mysterious cosmos. Hence he word "cosmism". The horizons of the cosmists are cosmic. The horizons of the terrestrialists are largely terrestrial and human centered

In the middle term, i.e. several decades from now (now being 1995), the development of artificial brains will probably be largely beneficial, and therefore will be pursued with vigor So I can feel guilt free for a time. I have a Japanese friend to whom I joke, "If I succeed in building brains and later, a Cosmist-Terrestrialist war breaks out, there will be people who will say to you, you should have killed me when I was younger." You mean like now?" she jokes back. I don't think the artilect debate will really get going until the machines really begin to show signs of brain like abilities and a large repertoire of behaviors. One of my own goals for example is to see a "NASA moonshot" type research and development project undertaken by the Japanese government to build artificial brains, using the skills of thousands of people. I certainly don't want to manage such an enterprise, but I would very much like to see such a project get off the ground. I believe it is coming. Building a truly capable artificial brain will be an enormous undertaking. The payoff will also be enormous. Industry will be able to take the results and put them into many "smart" products. Researchers such as myself will have been propounding the pros and cons of cosmism for many years, and will probably have been largely ignored, because of its science fiction aspect. But, once millions of people, year by year, see with their own eyes the growing intelligence of their own machines, they will start thinking, and in time a great debate will be generated. In the meantime, the machines will continue to get smarter and smarter. Sooner or later, some kind of incident will shake humanity from its slumber. For example, household robots might begin to talk back to their owners, or frighten their owners with their million fold faster thought processes. When confronted "flesh to metal" with an alien artifact from an industrial company which keeps upgrading its products every year, it is to be expected that millions of people will be forced to confront the "artilect issue". Its only a question of time. Maybe the early versions of artilects, i.e. of quasi human intelligence will begin to make decisions that human beings do not agree with nor understand. The artilects may attempt to explain but not be understood. If they are truly smart, they will be able to avoid human objections. The issue of whether to keep the artilects as docile and obedient slaves the whole time is critical. Asimov's "3 laws of robotics" are relevant here as an attempt to formulate policy regarding the characteristics of future artilects. However, I doubt whether artilectual intelligence and total obedience are compatible, in which case, Asimov's ideas are worthless for advanced artilects.

In the appendix, I will present my work on brain building, showing how it is possible to build artificial brains (in an elementary form at the moment, with about 100,000 neurons) without understanding how they function. They are too complex, but it is possible nevertheless to build them in a Darwinian fashion using a methodology I call "evolutionary engineering". To build an artificial brain of a billion neurons, (my goal for the early 21st century) will not be possible with traditional engineering "blueprinting" techniques. Traditionally, engineers make a plan how to build something, implying they understand how it functions, and then they build it. However, with evolutionary engineering, a concept that I have helped pioneer, it is possible to build something successfully without understanding fully how it functions. One simply uses some form of linear set of instructions to specify (at a low level) how some system is to be built or how it should function. One employs a population of such linear instructions each a little different from the rest. Those that by chance produce a system which functions well according to some human criterion, survive and have more offspring in the next generation. By making blind random changes in the instructions, it is possible to generate superior systems by chance. These superio- instruction chains then begin to dominate the population, until the next favorable mutation etc. Nature has employed this mutate and select technique for billions of years and has produced us as a result, and probably zillions of other ETIs around the cosmos.

Evolutionary engineering is probably the only effective means scientists and engineers will have to build artificial brains. The complexities of artificial brains, with their huge numbers of neurons and circuits, will be monstrous, and quite beyond human understanding. However, I routinely evolve neural circuit modules and their interconnections which perform as I want. I don't understand how they function, but they do. I can imagine several decades from now an army of "brain developer" engineers with an army of Darwin Machines who will build artificial brains with thousands if not millions of neural modules, and not understand how it all functions.

It is this inherent ignorance in the structure and function of evolutionary engineered artilects that will be the basis of the terrestrialists' intellectual objections. The terras will say that the artilects are inherently non understandable, and hence their behaviors will be inherently unpredictable. They are simply too complex. As their intelligence increases, their actions will become increasingly threatening because their greater intellects will allow them to do more threatening things to human beings. Human beings could never be sure what attitudes the artilects would take towards human beings. As humans, we routinely kill animals for food, sport or just for the hell of it. We could never be sure that advanced artilects would treat human beings the same way.

Hence to be safe, it is quite possible that global legislation will eventually be passed which will ban the development of artilects beyond a certain intelligence level. This legislation will probably result from a growing public outcry concerning the growing menace of the increasingly artilectual nature of the machines and the increasing power they will exert over human beings. The creation of this legislation itself will be the result of enormous debate. Probably the cosmist minority will not agree to it and may go underground. For the cosmists, it will be extremely frustrating to stop artilectual development. The natural curiosity of the scientist will be blocked by such a ban, and be anathema to those scientists who are cosmists. Confronted with a ban on artilectual development will truly separate the sheep from the goats, i.e. the terras from the cosmists. The more strident cosmists will be most frustrated and will probably begin to organize their own political institutions. In a global setting, there will probably no longer be a geographical correlation with ideology. In a global culture; ideological views will probably be more based on personality-differences rather- -than cultural differences. In a globally homogenized culture, such differences will be minimal.

So what are the cosmists to do in face of a global ban on their future activities? One idea is to form their own state, somewhat similar to the Jews and Israel or the Mormons and Utah. The cosmists could buy large tracts of land in cheap areas of the planet, and then select those people as its citizens who want to build artilects. Those cosmists who have serious second thoughts and who are public about it could be excluded from the colony, and could return to the terrestrialist majority. But would the terras allow such a cosmist colony to be established on the earth in the first place? One can imagine the terras arguing "What if the cosmists succeed in their artilectual experiments?" The artilects would then be on the earth, and hence be a threat not only to the cosmists, but more importantly, to the terras. This to the terras would be too dangerous to be tolerable, and hence cosmist colonies would be banned under threat of total annihilation. In the limit, if the existence of secret cosmist micro colonies is discovered, they would be exterminated by military means (i.e. they would be "nuked"). The cosmists would not even be arrested or convicted. It is quite imaginable that the fear of the terras of the possible success of the cosmists would be so great, and the threat to humanity so large, that easy justification for the vaporization of the secret cosmist colonies would be found.

In the light of such terrestrialist opposition, the cosmists, driven by a religious fervor to promote the "big picture" - the "cosmic issues", will be contemptuous of the terras, and will push hard for the opportunity to be given the freedom to develop their artilects. One other possibility is that using 21st century technology, it would be possible to rocket many cosmists off the earth so that they can build their artilects elsewhere. But, the earth might be considered a very valuable place by the artilects, which may want to return. The hostility of the terras towards artilects, would make the human beings a threat to the artilects. Who knows what the artilects might do to such an inferior species as the terras. Anticipating the possibility of such a return by the artilects would motivate the terras to ban any such cosmist extra planetary exodus. It is still too dangerous. Thus the cosmist dream of escaping into the depths of outer space, of traveling to other stars to create their cosmist colonies would be blocked. The frustration of the cosmists can only increase. In time, it is possible that the cosmists will realize that the only way that they will be able to do what they want will be by force of military arms.

I can imagine cosmist strategists planning the creation of distant cosmist colonies, so far from the earth, that most "earthians" ("terrians"), will not be so concerned with the threat of a cosmist artilectual success. The cosmists could pursue a type of evasion strategy, initially in a small way, moving further and further away from the earth, whenever there is a threat from the terras. In time, the cosmists may be able to establish themselves in sufficient strength to be able to defend themselves from a terrestrialist attack.

The scenario I am painting is one of continuous vigilance on the part of the terras, and a corresponding desire on the part of the cosmists to escape from the constraints imposed by the terras on the freedom of the cosmists to develop artilects. I see a continuous war between these two groups, now geographically separated, but this time, not on a planetary scale but on a 21st century interstellar scale. The cosmos is huge, so there should be space enough for terras and cosmists to exist quite independently of each other. The big problem is reaching such a state. The terras could always argue that if the artilects exist, their superior intelligence would allow them to overtake the earth is they so choose, no matter how far away they are. If they were able to get so far away, they would be able to get back to the earth, a wonderful piece of real estate.

Now, to return to whether I am a cosmist or a terra. Well, I guess I'm a cosmist, with enough terra in me, not to want to be swatted like a mosquito by an artilect, but nevertheless, I want to see artilects built, preferably far from the earth. I think it would be tragic on a cosmic scale if the evolution of nature, which has gone from elementary particles to intelligent creatures like ourselves, were to stop at the human level, considering that the potential of the computer is so hugely greater. This opinion of mine I can imagine being shared by many people, so there will be no shortage of cosmist sympathizers. However the hard nosed terrestrialist politicians and strategists will not allow such sympathy to take concrete form. It is too dangerous.

I therefore see no peaceful way out of this terrible dilemma. This is how 21st century warfare will be generated I believe. It will be nuclear in form and total, in the sense that the terras will want to completely eliminate the cosmists. The cosmists on the other hand may have some sympathy for the views of the terras, because they too are human, but the difference is that they are prepared to run the risk of the artilect threat because they see the "rise of the artilect" as their primary raison d'etre.

Discussion

Of course I can't predict the future, and my views are inevitably idiosyncratic. In this section, I will try to anticipate objections to the opinions above. Firstly, why should there be a threat from the artilects at all? Would it not be possible to build artilects that are wonderfully useful to human beings but which are safe, and sit quietly in a corner, being totally harmless? Of course, it will be possible to build such machines, and maybe they will be highly intelligent. But one questions whether artificial intelligence can be placed into a stationary box. Is not motion needed to interact with the world? How else is learning about the environment possible. One can imagine a computer lying in the corner with eyes, but how would it check its visual hypotheses about the world if it could not touch the objects it sees? I suspect strongly that motion is essential to the generation of intelligence. Mobile robots will be the vehicles of artificial intelligence according to this opinion. But, mobile robots are potentially dangerous. Using their mobility, they could kill people if they decided to. A stationary computer with no limbs to manipulate its world is harmless, unless it could move indirectly, by persuading human beings to do things for it. The development of pre-human intelligence in artilects will almost certainly take place in mobile robots.

Isn't a peaceful coexistence between artilect and humans possible? Look at humans and trees. Each have their own niche and have lived together for millions of years. Could not the artilects and humans live peacefully together by simply ignoring each other? The humans would not understand how the artilects pass their time. The artilects would probably find the humans totally uninteresting because the humans are such slow thinkers, i.e. a million times slower, and with massively inferior memory capacities. This scenario is possible I suppose, but its great weakness is its uncertainty. Human beings would have no way of knowing what the artilects think about human beings. Artilects are not human. From the point of view of human beings, artilects are utterly alien, and hence capable of generating all the -fear reactions that evolution has built into us over millions of years. The unknown was often life threatening and was avoided or killed, e.g. saber toothed tigers, or snakes, etc. Since we don't know what the artilects will be like, we have to be prudent. Politicians have a saying, "Hope for the best, prepare for the worst". The worst case artilect scenario is annihilation of the human species by a superior artilect species. Anticipating-this possibility and given the ultimately high stakes, the option of allowing the building of artilects (from the terrestrialists point of view) is not to be tolerated.

Is it possible that artilect production will be inevitable, and that they cannot be stopped at all? As mentioned earlier, I believe that brain building will end up one of the world's top industries. Tremendous capital investment will be devoted to the creation of early artilects, before they become a potential threat. The problems will arise once the artilects begin to approach human intelligence levels. With millions of peoples' economic lives tied up in the creation and use of low level artilects, how will it be possible to turn off the tap? One can imagine, "Oh, just this one extra feature on our company's new artilect!". It will be difficult to overcome this kind of incrementalism. In practice, it is likely that brain building will proceed rapidly, until human intelligence levels are reached or approached. If no really negative events occur up to this point, then artilect development will probably be uninterrupted. What happens when artilects reach human level? One can imagine responsible people pushing for super human artilectual intelligence for various reasons, e.g. "Oh we need a super smart computer to control economic policy for us. The economy is so complex, that no human economist can follow it, but a super smart computer might be able to. Therefore we should have one". Physicists might argue that their theories are too complex to be easily understood. Perhaps an artilect could figure out how the world functions and explain it to the human physicists. Probably the strongest argument in favor of wanting to build artilects is the sheer fascination of being able to do such a thing. Hence there is a real possibility that artilects of superior intelligence to humans will be built until some kind of negative experience occurs, and some kind of threat is felt. It is also possible by that stage that human dependence on these artilects will be so strong that it may be nearly impossible to stop using them. This is why I suspect some kind of crisis involving artilects will be behind the mass rejection of them on the part of the majority of human beings.

The above two scenarios which I call the "coexistence" and the "creep" scenarios respectively, can be supplemented by a "explosion" scenario. I doubt very much whether we will have true artilects within my lifetime (probably another 40 years). So we have plenty of time to indulge in the "artilect debate". But the danger is at the last moment, when some kind of scientific breakthrough is made, and suddenly there are artilects. It is possible that the world will be caught off guard by some development in a research lab somewhere, that suddenly passes some threshold, and from then on, human beings begin to share the planet with a superior species.

Conclusion

I obviously don't have all the answers, or even all the questions concerning the "artilect debate" or the "dominant species" debate, etc. I hope this chapter has made you think about the issues of genuine artilectual success. The potential of the computer to generate intelligence is massively superior to that generated by our own puny nervous systems. Therefore its only a question of time before the artilect debate gets going. Hopefully the ideas in this essay have given you some food for thought. Hopefully it will serve as a catalyst in a much larger debate. Finally, I see war. I am quite pessimistic about the long term outcome of the artilect issue. Both the terra and cosmist positions are powerful and will be defended with passion. Unfortunately its a zero sum game. I hope I'm wrong about all this, but my gut tells me "No, there will be real conflict. The artilect issue has all the prerequisites to fuel a global ideological war, even an interstellar; war. Postscript

How to Build an Artificial Brain

This section is devoted to a description of my work. The idea is to show how it is possible to build artificial brains in an elementary form, but nevertheless containing hundreds of thousands to millions of artificial neurons. This is very much work still in progress, so this section will quickly date.

I begin with some fundamental assumptions, some of which have already been mentioned in the text. Firstly, I feel it is vital to use the techniques of evolutionary engineering. I work in a research group called the "Evolutionary Systems" department at the ATR labs in Kyoto, which is based on this idea. Brain building is far too complex for low level design. In broad outline, one can design artificial structures, but the fine circuit details are left to the whims of evolution. I also collaborate closely with the new "Evolutionary Technologies (ET)" department at NTT, the world's biggest company. So there is plenty of scope for future brain building (at least in Japan). I also want this evolution to occur quickly, i.e. at electronic speeds if possible. Other research groups are using mechanical methods to evolve neural circuits, but such an approach is impossibly slow if one aims to build a billion neuron artificial brain (which I hope to do within the next decade or less).

The vehicle for growing/evolving neural circuits in large numbers that I have chosen, is cellular automata or CAs [Codd 1968]. CAs are like the squares on a multicolored chessboard, where the squares change their color at the same instant (at the tick of a clock). The color that a square (in 2D, or a cube in 3D) will change into at the next clock tick, depends on its present color and the colors (or "states" as they are usually called) of its neighbors. Rules governing the state change take the form as shown in Fig. 1. It is possible to hand code thousands of these rules so that cellular automata behave like neural networks that grow and evolve.

How this is done will be explained shortly. Right now I'm trying to explain the main points. The states (i.e. integers) can be stored relatively cheaply in gigabytes of RAM memory chips, so too the state transition rules. Hence with today's technology it is possible to grow CA based neural networks containing millions of artificial neurons. I have already put 100,000 such neurons in the CA machine I have on my desk. When you are talking about that number of neurons, it is no longer "neural networks" that you are talking about, but the building of ''artificial brains", or "brain building". The bottle neck in the whole system is the processor which updates the states (using a lookup table which contains the rules). Using a conventional workstation it is too slow to be interesting. But state-of-the-art CAMs (cellular automata machines), e.g. MIT's "CAM-8" can update 200,000,000 CA cells per second [Toffoli & Margolus 1987, 1990]. By collaborating closely with NTT's Evolutionary Technologies department, I intend building a super-CAM which is thousands of times faster. Using CAMs, one can grow/evolve neural circuits at a reasonable speed, fast enough to be interesting. The CAM in effect is the Darwin Machine, i.e. one which evolves its own structure. (Strictly speaking, the CAM-8 does not change its hardware architecture, but the architecture of CA based neural circuits does change and evolve). The CAM-8 on my desk ($40,000) contains 32 megabytes of RAM, which is enough (at a pinch) to place 50-100,000 initial neurons, ready to grow their axons and dendrites under evolutionary control. Fig. 2 shows a 3 CA cell wide CA trail. CA rules are hand coded which cause the cells in the middle to move down the trail. When a signal cell hits the end of the trail, it makes the trail turn left or right, or split, depending on the nature of the signal cell. For example, if we want the signal cell in state S at the right of Fig. 2 to move one cell 1 to the right, the cell to its right (in state 1) would need to change to state 5. Hence the state transition rule (using the C.T.R.B.L.C'. convention) would be 1.2.2.2.5.5. Figs. 3 to 6 show the state changes of CA cells which cause the trail to extend, turn or split.

It is the sequence of these signal cells that gets evolved. Initially, one generates a random string of integers consisting of the states of the possible signal cells. These integers are sandwiched with pairs of follower and trailer cells. The resulting triplets of cells move down the middle, like spaghetti through a drinking straw. At every third clock cycle a triplet is fed into the CA trails of a CA based artificial neuron. These trails grow into axons and dendrites. When axons and dendrites collide, they form synapses, as shown in Fig. 7. Thus, the sequence of the signal cells determines the pattern of growth of the resulting CA circuit. If a turn left signal is mutated by chance into a turn right signal, a different circuit is obtained. This growth phase is the first of a two-phase process. The second phase consists of using the circuit just grown as a neural circuit. Starting at the neurons, neural signal triplets (of constant value, i.e. state) are fed in, which then propagate around the CA network, making it behave like a neural network. The signal strength of a cell in an axon remains as it is when it is created, but once a signal enters a dendrite, its strength drops off linearly with distance. Hence the strength of he signal when it arrives at the receiving neurons depends on the distance of the dendral trail, but this distance is evolvable. This evolutionary control of the signal strengths according to distance is equivalent to the weighting of signals in artificial neural networks. Positive valued neural signals are colored pinkish, negative values neural signals are colored bluish. Neural signals sum their strengths when they collide. Synapses are either excitatory (neural signals don't change sign when crossing the synapse) or inhibitory (they do change sign). An excitatory neuron has only excitatory synapses. An inhibitory neuron has only inhibitory synapses.

It is possible to ensure that signals collide synchronously (i.e. no phase lags) by making the CA trails follow a 6 cell wide grid, i.e. 2 triplets. This "gridding" is a result of making each growth (e.g. turn left, turn right, split etc.) extend the trail by the same number of cells, i.e. two. See figs. 3 to 6 which shows this to be true. When a dendrite hits an axon in the growth phase, there is no doubt about which way the later neural signal is to travel. It simply moves from the axon into the dendrite and on towards the receiving neuron (which grew the dendrite in the first place), but when an axon hits a dendrite, the later neural signal does not know which way the receiving neuron is situated. Should the neural signal turn one way or the other? To overcome this problem, a marker signal is sent down the dendrite CA trail in the growth phase which changes or marks the first of the two synapse cells it comes next to. The second of the ttwoo synapse cells is unchanged. Since the first synapse cell lies closer to the receiving neuron, the later neural signaling cell knows which way to turn, i.e. toward the marked synapse cell. There are many tricks like this which I needed to make the system work. Another is the need for delay cells. The growth signals are sent out for a given number of clock cycles, but these signals take time to circulate throughout- the network. If the marker cells were placed immediately after the growth cells, it is possible that new synapses will be formed after the marker cells have passed. So delay signals (dummy signals which don't do anything except take up time) are used for enough clock cycles, so that the above risk is removed. So the sequence of signal types is :- growth signals, delay signals, marker signals, and finally neural signals (in the second phase).

Once the neural signals are circulating around the network, they can be tapped and used to control some process, whatever it is, e.g. controlling the angles of legs on an artificial creature, to make it walk or turn etc., or to control the actions of an artificial retina in an artificial eye, or whatever neural networks can be used to control. Once one has a system which can be controlled in this way, one measures the quality of the performance of this control (called the "fitness"). By having a whole population of evolving neural circuits, those which have superior fitnesses, have more offspring in the next generation (in an algorithm called the "genetic algorithm" [Goldberg 1989]). By mutating the growth instructions randomly with a small probability, it is possible to generate a string of growth instructions which by chance generates a slightly superior circuit than its rivals. In this way, it is possible to evolve increasingly superior neural circuits.

This is the essence of the idea. I already have a lot of experience in evolving neural network modules [de Garis 1990, 91, 92, 93, 95]. For my thesis work, I evolved modules (not using CAs or CAMs) to control the motions of an artificial quadruped which could detect frequencies, signal strengths, signal strength differences, and walk straight, turn left, turn right, peck at food, and mate. Every module I tried to evolve did so successfully. But each time I added a module to the system, the simulation speed slowed. I was thus limited to a small number of modules. I had at the time a dream of evolving a thousand modules to give my creature the behavioral repertoire of a kitten. It was this failure that motivated me to find a way to evolve neural circuits at electronic speeds and in huge numbers. Hence this research project, which I call the "CAM-Brain Project", for obvious reasons.

At the time of writing (April 1995), I have been working just over 2 years on this project, and will continue until 2001 (Japanese research projects tend to be long term, a real plus for researchers). I began with a 2D simulation, which worked well, but cost me 18 months work, most of which was spent hand coding 11,000 CA state transition rules, to make the system work. I then used it in a few evolutionary experiments, just to satisfy myself that it did do what I hoped it would do. I did not spend a lot of time on 2D experiments, because I didn't take the 2D version very seriously. It was just a vehicle to teach me how to do it, so that I could take the techniques learned to the more serious 3D version, which-r-m still busy doing. So far I've hand coded over 36,000 3D CA rules, and expect to reach about 60,000. Since each rule in 3D is rotated 24 ways (6 faces of a cube, and 4 rotations once the face is chosen), that means that the system has over a million rules in its (hashed) lookup table.

The few 2D experiments which I did undertake were as follows. The first and simplest was to maximize the number of synapses formed for 4 neurons on the computer screen. This worked fine, (over 100 synapses). The next was to see if a constant valued output signal could be generated (by tapping a signal at an arbitrary fixed point). If a circuit which generates such a simple behavior could not be evolved, then the whole project would be shown to be pointless, but it evolved perfectly. The next step up in difficulty, was to evolve a circuit which generated a sinusoidal output with a desired arbitrary amplitude and frequency. This evolved slowly but evolved. The next step up was to evolve a motion velocity vector detector. A square grid of "detector" neurons had a simulated "line of white light" pass over it in various directions. Two arbitrary neurons were chosen whose tapped outputs gave the velocity vector of the white line (i.e. moving up, across from the left, etc.) This also evolved slowly, but evolved.

I was impatient to move on to the 3D version, so I did not pursue further 2D experiments. The problem with the 2D version is that collisions are too easy. If a trail is heading towards another which lies across its path, a collision is inevitable, but in 3D, most trails simply pass each other, thus enabling long trails to be grown, which is essential if I want to grow artificial brains.

The 3D version is a conceptually straightforward extension of the 2D version, although will involve a much larger rule set. I have written software productivity tools to accelerate the generation of the many rules necessary. In the 2D version a CA trail is 3 CA cells wide. In the 3D version, a trail has a cross section of 5 cells, in the form of a cross. With the 24 rotations possible for these 5 cells, there is no leftness or rightness, so I gave each of the four outer cells in 3D a different color, so instead of talking about East, West, North South, Up, Down, I use Red, Blue, Green, Brown, etc. The 2D version has 6 growth instruction types, i.e. extend the trail, turn left, turn right, T split, split left forward, split right forward. The 3D version has 15 growth instruction types, etc. extend, turn red, split forward-red, split red-green etc. Synapsing was a bit more complex; because of the S cell cross section, but posed no real difficulties. The problem is just the sheer number of ways that dendrites and axons can collide, which results in many rules. Both the 2D and 3D versions use 3 kinds of CA trails, one for the dendrites, one for the excitatory axons (which form excitatory synapses) and one for the inhibitory axons (which form inhibitory synapses). These 3 types of trails need 3 sets of CA states.

The 3D version will probably need about 1000 states, and hence 10 bits per cell, so the left hand side of a CA state transition rule for 3D will be 70 bits long where the neighbor cells are positioned at Top, South, East, North, West, Bottom). This will be a challenge when the time comes to build a new CAM to update my 3D rules at speed.

Speaking of CAMs, I obtained an MIT CAM-8 machine several months ago, and now face the challenge of porting my 3D rules to it. The disadvantage of the CAM-8 is that it is only a 16 bit machine. It is basically a lookup table machine which outputs 16 bits which become an address for the next cycle. The art is to generate the table of 16 bit words. This can be done, but at times it is very tedious and requires many tricks. I will work with the CAM-8 until I have nothing more to learn from it, and will then (with help from my colleagues in the Evolutionary Technologies department at NTT) design and build a superCAM which will perform thousands of times faster than MIT's CAM-8 and which will overcome its weaknesses - for example, the CAM-8 updates every cell in the space, instead of only those that are filled or their neighbors. Thus I . a large space, with only a small number of initially filled cells, most of the time is wasted in updating empty cells - not very bright.

Further down the line, I will be collaborating with the Quantum Effect Devices group at NTT, who are at present developing -nano scale electronic devices. One of these researchers hopes to use his device to build cellular automata machines. If he succeeds, then I would like to take his technology and build a 'nanoCAM", i.e. one which uses molecular scale electronics and hence will have great speed and memory capacities. Since a close collaboration between my research lab (Brain Builder Group, Evolutionary Systems dept. at ATR) and NTT (Evolutionary Technologies dept.) exists, so long as the research ideas are valid, there is plenty of opportunity to see them developed into successful prototypes and later products (which is very much the Japanese way.)

Where will I be in a few years, research wise? Obviously I will have finished the 3D simulation and have my 60,000+ rules. This will be ported one way or another to the CAM-8. Once this happens, I can take advantage of the great processing speed of the CAM-8, i.e. 200,000,000 CA cell updates per second. Since the CAM-8 has 32 megabytes of memory (or 128 megabytes if I change the 4 megabit memory chips for 16 megabit chips) I have enough space to put about 50,000 to 100,000 neurons. I will then need to find a way to feed in growth instructions to so many neurons. Once all these problems have been solved, I can begin to evolve large neural systems, and a new research field will be born, namely "brain building". Just how I will be able to evolve so many neurons is an open question. I don't yet know how I will do this. I'm a researcher and no one has done this before. I guess I will not simply evolve a 100,000 neuron blob all at once. I might, just to see if it works, but a blob is a blob.

Later, I will need to evolve in incremental steps, i.e. first evolve a small system which performs certain functions, then add extra modules of new neurons, and evolve their intra and inter modular connections so that they perform some extra functions. Just how I will do this I cannot tell you yet. It is a topic of future research, but at least I will have a fabulous new tool to enable me to begin investigating such issues. I believe that once the tool exists, the theory will follow rapidly. So much of research is not theory dominated but tool dominated, e.g. molecular biology and the electron microscope etc.

I hope this brief description of my work gives you some idea of why I believe humanity will have artificial brains next century, and hence why I believe we should start thinking about the longer term consequences of brain building work. Of course its early days yet, but for me at least, the writing is already on the wall, and it frightens me.

References

[Codd 1968] E.F. Codd, CellularAutomata, Academic Press, NY, 1968

[de Garis 1990] Hugo de Garis, Genetic Programming: Modular Evolution for Darwin Machines, ICNN-90-WASH-DC, (Int. Joint Conf. on Neural Networks), January 1990, Washington DC, USA

[de Garis 1991] Hugo de Garis, Genetic Programming, Ch.8 in book Neural and Intelligent Systems Integration, ed. Branko Soucek, Wiley, NY, 1991

[de Garis 1992] Hugo de Garis, Ch.14 in book Dynamic, Genetic, and Chaotic Programming, ed. Branko Soucek and the IRIS Group, Wiley, NY, 1992

[de Garis 1993] Hugo de Garis, Evolvable Hardware: Genetic Programming of a Darwin Machine in Artificial Neural Nets and Genetic Algorithms, R.F. Albrecht, C.R. Reeves, N.C. Steele (eds.), Springer Verlag, NY, 1993

[de Garis 1995] Hugo de Garis, Brain Building: The Evolutionary Engineering of Artificial Nervous Systems, Wiley, in prep., 1995

[Drexler 1992] K.E. Drexler, Nanosystems: Molecular Machinery, Manufacturing and Computation, Wiley, NY, 1992

[Goldberg 1989] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-Wesley, Reading, MA, 1989

[Toffoli & Margolus 1987, 1990] T. Toffoli & N. Margolus, CellularAutomata Machines, MIT Press, Cambridge, MA, 1987; and CellularAutomata Machines, in Lattice Gas Methods for Partial Differential Equations, SFI SISOC, eds. Doolen et al, AddisonWesley, 1990.

Copyright & # 169; 1996, Nanothinc. All Rights Reserved.
Comments and Queries to webmaster@nanothinc.com.

Take me higher...
Intro Soup Recipe Shelf Sink Chef