"The Artilect War"

Second Version, 2001

Prof. Dr. Hugo de Garis

Head, Brain Builder Group,
Computer Science Department,
Utah State University (USU),
Old Main 423, Logan,
Utah, 84322-4205, USA.
tel: + 1 435 797 0959,
fax: + 1 435 797 3265,
cellphone: +1 435 512 1826,
degaris@cs.usu.edu
http://www.cs.usu.edu/~degaris




Prof. Dr. Hugo de Garis


===================================


Contents


Chapter 1. Introduction

Chapter 2. Who is this de Garis?

Chapter 3. Artilect Enabling Technologies

Chapter 4. The Cosmists

Chapter 5. The Terrans

Chapter 6. The Artilect War

Chapter 7. The Artilect Era

Chapter 8. Questions

Chapter 9. Brief Summary

Glossary

 =========================================



Chapter 1 INTRODUCTION

My name is Professor Hugo de Garis. I'm the head of a research group which designs and builds "artificial brains", a field that I have largely pioneered. But I'm more than just a researcher and scientist - I'm also a social critic with a political and ethical conscience. I am very worried that in the second half of our new century, the consequences of the kind of work that I do may have such a negative impact upon humanity that I truly fear for the future.

You may ask, "Well, if you are so concerned about the negative impact of your work on humanity, why don't you just stop it and do something else?" The truth is, I feel that I'm constructing something that may become rather godlike in future decades (although I probably won't live to see it). The prospect of building godlike creatures fills me with a sense of religious awe that goes to the very depth of my soul and motivates me powerfully to continue, despite the possible horrible negative consequences.

I feel quite "schizophrenic" about this. On the one hand I really want to build these artificial brains and to make them as smart as they can be. I see this as a magnificent goal for humanity to pursue, and I will be discussing this at length in this book. On the other hand, I am terrified at how bleak are some of the scenarios that may ensue if brain building becomes "too successful", meaning that the artificial brains end up becoming a lot more intelligent than the biological brains we carry around in our skulls. I will be discussing this too at length in this book.

Let me be more specific. As a professional brain building researcher and former theoretical physicist, I am in a position to see more clearly than most, the potential of 21st century technologies to generate "massively intelligent" machines. By "massively intelligent" I mean the creation of artificial brains which may end up being smarter than human brains by not just a factor of two or even ten times, but by a factor of trillions of trillions of trillions of times, i.e. truly godlike. Since such gargantuan numbers may sound more science fiction like to you than any possible future science, the next chapter of this book will explain to you the basic principles of those 21st century technologies that I believe will allow humanity, if it chooses, to build these godlike machines. I will try to persuade you that it is not science fiction, and that strong reasons exist to compel humanity to believe in these astronomically large numbers. I will present these technologies in as simple and as clear a way as I can, so that you do not need to be a "rocket scientist" (as the Americans say, i.e. someone very smart) to understand them. The basic ideas can be understood by almost anyone who is prepared to give their study a little effort.

Now, once you have read the next chapter which introduces to you all these fabulous 21st century technologies that will permit the building of godlike massively intelligent machines, a host of ethical, philosophical, and political questions will probably occur to you. The prospect of humanity building these godlike machines raises vast and hugely important questions. The majority of this book is devoted to the discussion of such questions. I don't pretend to have all the answers, but I will do my best.

One of the great technological economic trends of our recent history has been that of "Moore's law", which states that the computational capacities (e.g. electronic component densities, electronic signal processing speeds, etc) of integrated circuits or "chips", have been doubling every year or two. This trend has remained valid since Gordon Moore, one of the founders of the Intel microprocessor manufacturing company, first formulated it in 1965. If you keep multiplying a number by 2 many times over, you will soon end up with a huge number. For example, 2 times 2 times 2 times 2 ... (ten times) equals 1024. If you do it 20 times you get 1048576, i.e. over a million. If you do it 30 times, you get a billion, by 40 times you get a trillion, etc. Moore's law has remained valid for the past few decades, so that the size of the doublings recently has become truly massive. I speak of "massive Moore doublings".

Moore's law is a consequence of the shrinking of the size of electronic circuits on chips, so that the distance that electrons (the elementary particles whose flow in an electronic circuit is what constitutes the electronic current) have to travel between two electronic components, for example two transistors, is reduced. According to Einstein, the fastest speed at which anything can move is that of the speed of light (about 300,000 kms/sec) and this is a constant of nature that electronic currents have to respect. If one shortens the distance between two electronic components, then an electronic signal between them (i.e. the flow of electrons between them) has less distance to travel, and hence takes less time to traverse that distance (at the constant speed of light).

A huge amount of effort over the past few decades has been devoted by the chip manufacturing companies into making electronic circuits smaller, and hence denser, so that they function faster. The faster a microprocessor chip functions, the more economically attractive it is. If you are the CEO of a chip manufacturing company and your competitor down the road in California's "Silicon Valley" brings a rival chip onto the market that is 30% faster than yours and 6 months ahead of you, then your company will probably go out of business. The market share of the rival company will increase significantly, because everyone wants a faster computer. Hence for decades, electronic circuitry has become smaller and hence faster.

For how much longer can Moore's law remain valid? If it does so until 2020, then the size of the electronic components in mass memory chips for example, will be such that it will be possible to store a single bit of information (a "bit" is a "binary digit", a 0 or a 1, that computers use to represent numbers and symbols to perform their calculations) on a single atom. So how many atoms (and hence how many stored bits) are there in a human sized object, such as an apple? The answer is astonishing - a trillion trillion atoms (bits), i.e. a 1 followed by 24 zeros, or a million million million million.

Are you beginning to get an inkling for why I believe that massively intelligent machines could become trillions of trillions of times smarter than we are later this century?

Not only is it likely that 21st century technology will be storing a bit of information on a single atom, it will be using a new kind of computing called "quantum computing", which is radically different from the garden variety or "classical computing" that humanity used in the 20th century. The following chapter will attempt to give a brief outline of the principles of quantum computing since it is likely that that technology will form the basis of the computers of the near and longer term future.

The essential feature of quantum computing can however be mentioned here. It is as follows. If one uses a string of N bits (called a "register" in computer science, e.g. 001011101111010) in some form of computing operation (it doesn't matter for the moment what the operation is) it will take a certain amount of time using "classical computing". However in the same amount of time, using "quantum computing" techniques, one can often perform 2N such operations. (2N means 2 multiplied by 2 multiplied by 2 ... (N times)). As N becomes large, 2N becomes astronomically large. The potential of quantum computing is thus hugely superior to classical computing. Since Moore's law is likely to take us to the atomic scale where the laws of physics called "quantum mechanics" need to be applied, humanity will be forced to compute quantum mechanically, hence the enormous theoretical and experimental effort in the past few years to understand and build "quantum computers".

Quantum computing still has many conceptual and practical problems which need to be solved before quantum computers are sold to the public. But progress is being made every month, so personally I believe that it is only a question of time before we have functional quantum computers.

Now, start putting one bit per atom memory storage capacities together with quantum computing and the combination is truly explosive. 21st century computers could have potential computing capacities truly trillions of trillions of trillions ... of times above those of current classical computing capacities.

I hope you have followed me so far.

At this point in the argument, you may be racing ahead of me a little and object that I seem to be assuming implicitly that massive memory capacities and astronomical computational capacities are sufficient to generate massively intelligent machines, and that nothing else is needed. I have been accused by some of my colleagues of this, so let me state my personal opinion on this question.

There are people (for example, Sir Roger Penrose, of black hole theory fame, and arch rival of the wheel-chaired British cosmologist Stephen Hawking) who claim that there is more to producing an intelligent conscious machine than just massive computational abilities. I am open to this objection. Perhaps such critics are right. If so, then their objections do not change my basic thesis much, since I feel that it is only a question of time before science understands how nature builds us, i.e. before science understands the "embryogenic" process, used in building an embryo and then a baby, consisting of trillions of cells, from a single fertilized egg cell.

We have the existence proof of ourselves, who are both intelligent and conscious, that it is possible for nature to assemble molecules in an appropriate way to build us. When a pregnant woman eats, some of the molecules in her food are rearranged, and then self assemble into a large molecular structure consisting of trillions of trillions of atoms which becomes her baby. The baby is a self assembled collection of molecules that gets built to become a functional three dimensional creature that is intelligent and conscious.

Nature, i.e. evolution, has found a way to do this, therefore it can be done. If science wants to build an intelligent conscious machine, then one obvious strategy is to copy nature's approach as closely as possible. Sooner or later, science will end up with an artificial life form that functions in the same way as a human being.

Common sense says that it would be easier to build an artificial brain if science had a far better knowledge of how our own biological brains work. Unfortunately, contemporary neuroscience's understanding of how our brains work is still painfully inadequate. Despite huge efforts of neuroscientists over the past century or more to understand the basic principles of the functioning of the human brain, very little is known at the micro-neural circuit level as to just how a highly interconnected neural circuit does what it does. Science just does not yet have the tools to adequately explore such structures.

However, as technology becomes capable of building smaller and smaller devices (moving down from the micro-meter level to the nano-meter level (i.e. from a millionth of a meter (the size of bacteria) to a billionth of a meter (the size of molecules)) it will become possible to build molecular scale robots that can be used to explore how the brain functions.

Science's knowledge of how the biological brain works is inadequate because the tools we have at our disposal today are inadequate, but with molecular scale tools (called "nanotech" or "nanotechnology") neuroscientists will have a powerful new set of techniques with which to explore the brain. Progress in our understanding of how the brain functions should then be rapid.

Brain builders like me will then jump on such newly established neuro-scientific principles and incorporate them rapidly into our artificial brain architectures.

Hopefully in time, so much will become known about how our own brains function, that a kind of "intelligence theory" will arise, which will be able to explain on the basis of neuronal circuitry (a neuron is a brain cell) why Einstein's brain for example, was so much smarter than most other people's brains. Once such an intelligence theory exists, it may be possible for neuro-engineers like myself to take a more engineering approach to brain building. We will not have to remain such "slaves to neuroscience". We will be able to take an alternative route to producing intelligent machines (although admittedly initially based on neuro-scientific principles).

So with the new neuro-scientific knowledge that nanotech tools will provide, and the computational miracles that quantum computing and one bit per atom storage allow, brain builders like me will probably have all the ingredients we need to start building truly intelligent and conscious machines.

At this point, a host of questions start arising, and I will spend most of this book trying to answer a lot of them. Lets jump into the future for a moment and try to imagine how the above technological developments will impact on ordinary peoples lives.

Pretty soon, it will be possible to buy artificially brained robots that perform useful tasks around the house. If the price of such robots can be made affordable, then the demand for them will be huge. I believe in time that the world economy will be based upon brain based computers. Such devices will be so useful and so popular that everyone on the planet will want to own them. As the technologies and the economics improve, the global market for such devices will only increase to the point that most of the planet's politics will be tied up in supporting it. Not only will the commercial sector be heavily involved in the production of ever smarter and ever more useful robots and artificial brain based devices, but so too of course will the military forces of the world.

It is unlikely in the next few decades that the planet will have formed a truly global state, with a global police force to defend its global laws. I feel there will be a growing political rivalry over the next half century between the United States and China to be the world's most powerful nation. This rivalry will ensure that the ministers of defense of both countries cannot afford to allow the other country to develop more intelligent soldier robots and other artificial brain based defense systems than their own. Hence national governments will be heavily involved in pushing the development of military based artificial brain research, that will only spill over in time to the commercial sector, as has been the pattern for over a century.

Thus the rise of artificial brain based robotics and related fields, seems unstoppable. There will be so much military and commercial momentum behind it that it is difficult to imagine how it could be stopped, unless some mass political movement is formed to block its development.

How might such a movement get off the ground? It's not too difficult to imagine what might happen. Imagine in about a decade from now that millions of people have already bought household cleaning robots, sex robots, teaching machines, babysitter robots, companionship robots, friendship robots, etc, and that these brain based machines talk quite well and understand human speech to a reasonable extent. A few years later what happens? Not surprisingly, the models of that earlier year are now seen by their owners to be rather old fashioned and not as attractive as the latest models. The latest models will be more "intelligent" because their speech is of higher quality. They will understand more and give better, more appropriate answers. Their behavioral repertoire will be richer. In short, they will make the earlier models look quite inferior.

So what does everyone do? Of course, they will scrap their old robots and buy new ones, or have their old ones updated with better artificial neural circuitry. In a further few years, the same process will repeat itself, in a fashion similar to the way buyers of personal computers behaved in the 1980s and 1990s.

However, some of the more reflective buyers may start noticing that their household machines and robots are becoming smarter and smarter every machine generation, so that the IQ gap between human beings and robots keeps getting smaller. Once the robots start getting really quite smart, suddenly millions of robot owners will start asking themselves some awkward questions -

"Just how smart could these artificially brained robots become?"

"Could they become as smart as human beings?"

"If that's possible, is that a good thing?"

"Might not the robots then be smart enough to be a threat to humanity?"

"Could the robots become smarter than humans?"

"If so, how much smarter?"

"Should humanity allow these robots to become smarter than human beings?"

"If they become a lot smarter than human beings, might they decide that humans are a pest, a cancer on the surface of the planet, and decide to wipe us out?"

"Should humanity take the risk, that that might happen?"

"Should a limit be placed on the robot's AIQ (Artificial Intelligence Quotient), so that the robots are smart enough to be useful to human beings, but not too smart so as to be threatening?"

"Will it be possible to stop the rise of robot AIQ?"

"Will it be politically, militarily, economically possible to stop the robots becoming smarter every year?"

"There are lots of people who see the creation of massively intelligent machines as the destiny of the human species. These people will not like any limits being placed on AIQ levels. Won't this create conflict amongst human beings?"

You may be able to think of other such questions relating to the rise of artificial intelligence and the creation of artificial brains with ever-greater capabilities.

How do I see humanity facing up to the challenge of the rise of smart machines? My personal scenario that I find the most plausible I will now present to you. However, before doing so, I would like to introduce a new term that I will use from now on throughout this book, as it is a useful shorthand for the term "godlike massively intelligent machine". The new term is "artilect", which is a shortened version of "artificial intellect". The term "artilect" features in the very title of this book "The Artilect War", so it is probably the most important concept and term in this book.

I believe that the 21st century will be dominated by the question as to whether humanity should or should not build artilects, i.e. machines of godlike intelligence, trillions of trillions of times above the human level. I see humanity splitting into two major political groups, which in time will become increasingly bitterly opposed, as the artilect issue becomes more real and less science fiction like.

The human group in favor of building artilects, I label the "Cosmists", based on the word "cosmos" (the universe), which reflects their perspective on the question. To the Cosmists, building artilects will be like a religion; the destiny of the human species; something truly magnificent and worthy of worship; something to dedicate one's life and energy to help achieve. To the Cosmists, not building the artilects, not creating the next higher form of evolution, thus freezing the state of evolution at the puny human level, would be a "cosmic tragedy". The Cosmists will be bitterly opposed to any attempt to stop the rise of the 21st century artilect.

The second human group, opposed to the building of artilects, I label the "Terrans", based on the word "terra" (the earth) which reflects their inward looking, non cosmic, perspective. The Terrans, I strongly suspect, will argue that allowing the Cosmists to build their artilects (in a highly advanced form) implies accepting the risk, that one day, the artilects might decide, for whatever reason, that the human species is a pest. Since the artilects would be so vastly superior to human beings in intelligence, it would be easy for the artilects to exterminate the human species if they so decided.

But you may argue that if the artilects truly become very smart, they would realize that human beings gave birth to them, that we are their parents. Therefore the artilects would respect us and treat us well. This may be what happens, but the point is, I argue, is that you could not be certain that the artilects would treat humanity with the level of respect that we would like.

Don't forget, the artilects have the potential of becoming trillions of trillions ... of times smarter than we are, so there is always the possibility that they could become so smart that human beings would appear to them to be so inferior that we would simply not be worth worrying about. Whether humanity survives or not, might be a matter of supreme indifference to them.

It is not exaggerating to say that there is quite a close analogy between an artilect trying to communicate with a human being, and a human being trying to communicate with a rock.

To make another analogy, consider your feelings towards a mosquito as it lands on the skin of your forearm. When you swat it, do you stop to consider that the creature you just killed is a miracle of nano-technological engineering, that scientists of the 20th century had absolutely no way of building. The mosquito consists of billions of cells, each of which can be looked upon as a kind of molecular city, where a molecule in a cell is equivalent to a person in a city. The comparative scale of molecule to cell is about the same as person to city.

Despite the fact that the mosquitoes, which took billions of years to evolve, are extremely complex and miraculous creatures, we human beings don't give a damn about them, and swat them because from our perspective they are a pest. We have similar attitudes towards killing ants when we walk on them during a stroll through the forest, or when flushing spiders down the plug hole.

Who is to say that the artilects might not have similar attitudes towards human beings, and then wipe us out. With their gargantuan "artilectual" intelligence, it would be as easy as pie for them to do so.

The critical word in the artilect debate to the Terrans is "risk". The Terrans will argue that humanity should never take the risk that the artilects, in an advanced form, might decide to wipe out the human species. The only certain way that the risk remains zero, is that the artilects are never built in the first place.

When push comes to shove, if the Terrans see the Cosmists are truly serious about building artilects in an advanced state, then to preserve the survival of the human species, the Terrans will exterminate the Cosmists. Killing a few hundred or a few million Cosmists will be considered justifiable by the Terrans for the sake of preserving the survival of the whole human species, i.e. billions of people.

Such a sacrifice would be deemed reasonable by the Terrans. To make a historical analogy - when Stalin's troops were pushing west at the end of WW2, to capture Berlin and destroy Hitler's Nazi regime that murdered 20 million Russians, they were losing about 100,000 Russian soldiers killed or injured for every major east European city captured from the Nazis. To Stalin, such a sacrifice was considered justifiable for the greater good of ridding the Russian people of the horror of mass murdering Nazism.

You may now ask - "Would anyone in their right mind genuinely choose, when push comes to shove, to be a Cosmist, and truly risk the annihilation of the human species?"

I think that in the future, millions of people will answer yes to this most fundamental of questions. I think that as more people become fully conscious of what the artilects could become, many of these people will end up choosing in favor of their creation. This book will devote a whole chapter to arguments in favor of building artilects when it presents the Cosmist case.

These people, these "Cosmists", will place a higher priority on the creation of godlike, immortal, go anywhere, do anything creatures (where one artilect is "worth" a trillion trillion human beings) than denying the risk of the extermination of the human species at the hands of the artilects.

Let me spell this out, so that there is no doubt about the stance of the Cosmists. A Cosmist, by definition, is someone in favor of building artilects. The artilects, if they are built, may later find humans so inferior and such a pest, that they may decide, for whatever reason, to wipe us out.

Therefore the Cosmist is prepared to accept the risk that the human species is wiped out. If humanity is wiped out, that means your grandchildren will be wiped out, my grandchildren will be wiped out. It would be the worst calamity in human history, because there would be no more history, because there would be no more humans. Humanity would thus join the long list of over 99% of species that have ever existed on the earth, which have already become extinct.

Thus to the Terrans, the Cosmists are monsters incarnate, far worse than the regimes of Hitler, Stalin, Mao, the Japs, or any other regime that murdered tens of millions of people in the 20th century, because the scale of the monstrosity would be far larger. This time we are not talking about deca-mega mass murder, we are talking about the potential annihilation of the whole human species, billions of people.

But to the Cosmists, the survival or not of the human species, on an insignificant planet, circling a star that is one of about 200 billion in our galaxy, in a known universe of a comparable number of galaxies (also in the billions), and with probably as many universes in the "multiverse" (according to several recent cosmological theories) is a matter of miniscule importance. I have labeled the Cosmists Cosmists for a reason. Their perspective is cosmic. They will look at the "big picture" - meaning that the annihilation of one ultra-primitive, biological, non-artilectual species (i.e. human beings) on one insignificant little planet, is unimportant in comparison with the creation of artilect gods.

There will be two chapters later in the book presenting the Terran and the Cosmist cases, one for each viewpoint. There are very powerful arguments on both sides, which I believe will only make the inevitable conflict between Terranism and Cosmism all the more bitter as the artilect debate heats up in the coming decades.

What makes me particularly gloomy about the potential bitterness of this coming conflict is how evenly people's opinions are split along the Terran/Cosmist divide. For example, I often invite audiences to whom I present the Cosmist/Terran/Artilect scenario in public lectures, to vote on whether they would be Terran or Cosmist. I find that the voting is not what I first expected it would be (namely about 10% Cosmist, 90% Terran) but rather 50/50, 60/40, 40/60. This issue truly divides people.

What makes me even gloomier is that the artilect issue (i.e. should artilects be built or not) will heat up in the 21st century to such an extent, that it is almost certain it will lead to a major war between the Terrans and the Cosmists in the second half of this new century. This conflict will take place with 21st century weaponry. If one extrapolates up the graph of the number of major deaths in major wars from the beginning of the 19th century (e.g. the napoleonic wars) to the end of the 21st century, one arrives at the depressing figure of billions, what I call "gigadeath".

But the population of the earth is only several billion people, so we arrive at the tragic conclusion that to avoid the risk of the total annihilation of the human species by the artilects, humanity goes to war against itself and kills itself off (or almost).

This "Artilect War" as I call it, will be the most passionate in history, because the stake has never been so high, namely the survival of the whole human race. It will be waged with 21st century weapons and hence the casualty figures will be of 21st century grandeur.

The sad thing about this gloomy scenario is that despite considerable effort on my part, I have been unable to find a way out of this mess. I lie awake in bed trying to find a realistic scenario that could avoid "gigadeath". I have not succeeded, which makes me feel most pessimistic. In fact I am so pessimistic that I am glad to be alive today. At least I will die peacefully in my bed. However I fear for my grandchildren. They will see the horror of it and very probably they will be destroyed by it.

I will die within about 30-40 years, given my age, but that is not enough time I believe, for the artilect scenario to unfold. I believe it will take longer than that to obtain the necessary knowledge to build massively intelligent artificial brains or artilects. However, what I will see in my lifetime, and obviously this book is aimed at producing just that, is a vociferous debate over the artilect issue.

There are a growing number of researchers and professors like myself who are starting to see the writing on the wall, and who are claiming publicly in media appearances and books that the 21st century will see the rise of massive artificial intelligence. I am the only one so far who is saying that this rise of massive AI will probably lead to a major war, the "artilect war".

Thus the issue is really starting to hit the world media, and countries such as the US, the UK and France are leading the pack. In fact I believe that within only a few years, the issue will have passed from one that is confined largely to academic audiences, to a wider general public, with representatives from such fields as politics, religion, defense etc, with each field contributing its views from its own perspective.

The "artilect debate" will seem like science fiction, and set too far into the future, for most people to worry about, but as the machines start getting smarter and smarter every year, it will take on an intensity that will become truly frightening.

So what is my position on all this? Why am I writing this book?

Deep down, I'm a Cosmist. I think it would be a cosmic tragedy if humanity chooses never to build artilects. To illustrate my views on why I'm a Cosmist in my heart, I like to tell a little story.

Imagine you are an ET (an extra terrestrial) with godlike technological powers and you come to the earth 3 billion years ago. You observe the life forms at that time on earth and notice that they are still at the primitive bacterial single-celled stage. In a sweep of your magical technological wand, you fiddle all the DNA in all the bacteria of the planet so that (for the sake of the argument) it will never be possible in the future for these bacteria to evolve into multi-celled creatures. Hence, there will never be any plants, no animals, no human beings, no Einstein, no Beethoven's 9th. Is that a tragedy? Once multi-celled creatures did evolve on the earth, zillions of bacteria were eaten by them. The evolutionary rise of multi-celled creatures on the earth was no picnic for the bacteria.

I hope you see the analogy. If we build artilects and billions of human beings are wiped out as a result, what will be the equivalent of Beethoven's 9th that the artilects will produce with their godlike intellects? As human beings, we are too dumb to know. We are just too inferior to be capable of recognizing such things. It would be like asking a mouse to study Einstein's General Theory of Relativity. It just couldn't do it, because it doesn't have the necessary neural circuitry to allow it, nor do most humans, for that matter.

But, you may ask, if I'm a Cosmist at heart, why am I writing this book? The answer is that I'm not 100% Cosmist. If I were, I would be quietly getting on with my brain building work and not trying to raise the alarm on the artilect issue to the general public. Part of me is also Terran. On my death bed I would be proud to be considered the "father of the artificial brain", but if history condemns me as being the "father of gigadeath", then that prospect truly horrifies me. My second wife's mother was gassed by the Nazis at Auschwitz. I know to some extent what genocide means at an emotional level, and have had to live with its consequences for years.

I'm writing this book to raise the alarm, because I think humanity should be given the choice to stop the Cosmists before they get too advanced in their work, if that is what most human beings choose. So should I stop my brain building work now? No. I don't think so. I believe that producing near human-level artificial intelligence is a very difficult problem that will take decades to solve. Over the next 30 to 40 years, it is likely that the AIQ of robots will become high enough to be very useful to humanity. They will perform so many of the boring, dirty and dangerous tasks. Humanity will be liberated from such work, and hence have more time to pursue more rewarding tasks. The robots can do most of the work allowing human beings to do more fun things.

It would be premature to stop the research on artificial brains now. However, once these artificial brains really do start becoming smart and threaten to become a lot smarter and perhaps very quickly (a scenario called "the singularity") then humanity should be ready to take a decision on whether to proceed or not. Making an informed decision on an issue that concerns the survival of the whole species is something so important that the necessary discussion on the artilect issue should begin earlier. There should be enough time for all the issue's intricacies to be thrashed out before the artilect age is imminent.

So publicly I'm Terran. I'm trying to raise the alarm. Privately I'm Cosmist. Hence I feel quite schizophrenic, as I mentioned in the very first page of this book. I feel so torn on the issue, so ambivalent. I believe that similar feelings will be felt by billions of people in the future as the artilect debate really takes hold. From the Terran viewpoint, to be a Cosmist is to be a "speciecidal monster" (a species killer). A Cosmist accepts the risk of seeing the human species wiped out by the artilects. This is inherent in the nature of the situation. The decision whether to build artilects has a binary answer - we can build them or not. The decision to build them is the decision to accept the risk that they will wipe us out.

On the other hand, not to build them is the decision not to build gods, a kind of "deicide" (god killing). From the Cosmist viewpoint, Terrans are "deicidal monsters".

In passing, I should mention that there are some people who feel that the whole Cosmist/Terran conflict can be avoided by having human beings themselves become artilects by adding components to their heads etc to become "cyborgs" (cybernetic organisms, i.e. part human, part machine). Personally I find such arguments naïve, since they would only work if the whole of humanity made the transition from human to artilect at the same rate, which obviously is not going to happen.

There is more potential computing capacity in a grain of sugar than there is in the human brain by a factor of trillions. Incorporating such a grain into the human brain would simply make the human cyborg an "artilect in human disguise" as seen from the perspective of a Terran. The Terrans would hate the cyborgs with as much venom as they would the artilects and would be motivated to destroy both. Having a human exterior would not make the cyborgs any less threatening to the Terrans.

Let me try to express this Terran revulsion against the cyborgs in an even more graphic way that may have a stronger appeal to women than to men. Take the case of a young mother who has just given birth. She decides to convert her baby into a cyborg, by adding the "grain of sugar" to her baby's brain, thus transforming her baby into a human faced artilect. Her "baby" will now spend only about a trillionth of its mental capacity thinking human thoughts, and the rest of its brain capacity (i.e. 99.9999999999% of it) will be used for thinking artilect thoughts (whatever they are). In effect, the mother has "killed" her baby because it is no longer human. It is an "artilect in human disguise" and totally alien to her.

Thus to me, the cyborg option will not avoid the Cosmist/Terran conflict. If anything, it will probably only worsen it, because it will increase the level of paranoia of the Terrans when they cannot distinguish easily a cyborg from a human at a distance.

For about 10 years I sat on the fence, presenting my ideas in a "on the one hand, on the other hand" kind of way, presenting the two cases, one in favor of the Terrans, and the other in favor of the Cosmists. After some years, my friends began to accuse me of being a hypocrite. "Hugo, you expect humanity to choose between being Terran or Cosmist, but you don't do the same yourself". "Fair enough", I thought, so I chose. In my heart I'm a Cosmist, and I'll try to present the many arguments and feelings in favor of building artilects in the chapter on the Cosmist viewpoint. This chapter tries to justify why I and other Cosmists feel so passionately about building artilects, that we are prepared to run the terrible risk of the extermination of the human species.

In the chapter on the Terran viewpoint, I will present the case why the Terrans feel that building artilects would be a total disaster.

Later on in this book, I will try to paint a picture as to how I see the conflict brewing and what the possible outcome might be.

This introductory chapter has given you an overview of what the "artilect war" is about. The later chapters will provide greater detail on the ideas sketched out so far.

I hope this book will make you think. It is written to help make you conscious of an issue that I believe will dominate the global politics of the 21st century, that will color and define the age, namely, the question of "species dominance", "Should humanity build artilects or not?" This question I believe will divide humanity more bitterly in the 21st century than the question which divided humanity so bitterly in the 20th, namely, "Who should own capital?" The bitterly opposed answers to that question led to the Capitalist/Communist dichotomy. The question which will dominate 21st century global politics will be "Who or what should be dominant species, artilects or human beings?"

I end this chapter with a little slogan that expresses rather pithily, the essence of the artilect debate.

"Do we build gods, or do we build our potential exterminators?"

=================================

Chapter 2 Who is this de Garis?

Who is this de Garis, who makes such outrageous claims - that machines will become trillions of trillions of trillions of times smarter than human beings by the end of the century - that there will be a major war over the issue of species dominance, and that as a result billions of human beings will die? Is he a mad man? Is he a science fiction writer? Does he deserve to be listened to, or can humanity afford simply to ignore him?

This chapter is about who I am. It is partly autobiographical, partly a description of my work, and attempts generally to paint a portrait of me the person, so that readers may be in a better position to judge the credibility of my ideas by knowing something about the person who wrote them.

This chapter will be divided into 3 main sections. The first gives a brief autobiography, the second is a longer description of my current work and the third is a presentation of my future work goals and dreams.

2.1 Autobiography

I was born in Sydney, Australia in 1947, making me a middle aged man at the time of writing. I've been divorced, widowered and will probably marry a third time in the near future. I have two children by the first wife. By temperament I am a passionate intellectual, with over 6000 books in my private library. I am scientist, a research professor, a social activist, a writer, and a social critic.

As an adolescent, growing up in Australia, I felt that my passionate intellectual values were not valued by Australia's phlegmatic anti-intellectual brawn-based culture. During the time of the Sydney Olympic games, a BBC journalist said of Australians that they would rather win a gold medal than a Nobel prize. By the time I was 23 and had finished my basic undergraduate degrees, in applied mathematics and theoretical physics, I wanted to leave the country for ever. I had been conscripted to fight in the Vietnam war, which simply made me hate my own government. "How dare they force me to risk my life to defend their ideology!" I found an antiwar psychiatrist who spoke with me for all of 2 minutes and wrote out a recommendation to the conscription medical board that I had a "severe personality disorder". I failed the medical and soon after took a boat for England.

The first day in London I felt overwhelmed by the feeling that I had set foot in an unquestionably superior culture. That night on BBC TV watching a debate, I was struck at its intellectual quality. I felt a great weight lift off my shoulders. I had found my home, a culture that valued my values.

A few years later, I was browsing a world atlas that I had bought for my first wife, an Australian whom I had met on the 5-week boat trip from Australia to England. The idea occurred to me that I could live in a cosmopolitan city like Brussels and hence benefit from the intellectual stimulus of several superior cultures. All I would need to do would be to learn a few languages and then move there. That's what I did. I got fluent in French, German and Dutch and absorbed those cultures into my personality. I loved it. My wife however, did not. After 5 years living in England where she was fairly happy - her mother was English - and a further 6 in Belgium, she longed to get back to her native Australia. This conflict of interests broke us up. She has since remarried and lives in Australia. After the breakup I lived in Brussels and married a French speaking Belgian woman, my second wife.

I got bored doing computer work for industry and decided at age 40 to return to university and get a Ph.D. in artificial intelligence (AI) and become a researcher. This I did at the University of Brussels (ULB). Early in 1992, I and my second wife left Europe to live in Japan. I had gotten a postdoc fellowship to do AI in Tsukuba in Japan. I and millions of others believed at the time that by the year 2000, Japan would be the world's dominant economic power, overtaking the US. It was not to happen. I spent 8 years in Japan, working towards building the world's first artificial brain.

I hated Japan. It was too feudal, too fascist, too repressive of individualism, too intellectually sterile, too socially backward for me to tolerate it for very long. I stayed as long as I did because at least Japan was paying for the construction of a remarkable new type of computer that I will talk more about in the next section.

I got a new job doing the same kind of work at a research lab in Brussels. I returned alone, because my second wife had recently died of lung cancer. Yes, the idiot smoked heavily in her younger years. Young women take note! During my stay in Japan I became increasingly friendly with a Japanese woman who may become my third wife. We Iphone (internet phone) each other every day, while she pays off her bank debt in Japan. Then we will live together to see how compatible we are. If we pass this "time test", we will marry.

I'm not just a research scientist. I'm also a social critic. I get very upset if I feel an injustice is being committed against me and fight hard against it. I am a "masculist", i.e. a men's libber, fighting for the liberation of men, largely from the traditional financial parasitism of women who expect to live off the money of their husbands. I find this a form of slavery. I have contempt for such women, whom I labeled "fluffies". I coined a whole vernacular of masculist terms, packaged the ideas and presented them to the European media. After a decade of feminism of the 70s, the media were very happy to hear from the men in the 80s. I was quite successful, and got on the media over 150 times in half a dozen countries.

I mention this masculist period of mine, because I think it gives some insight to what I am doing now with a new ideology that I call cosmism. The masculist period gave me the training and the knowledge that I was capable of pushing a new ideology to the media. I found in practice that I could do it quite well.

As an example of some of the masculist terms I coined, take the word "masculist" itself. It is the obvious equivalent of the word feminist. I spoke of FIPs, i.e. financially independent persons, i.e. women who had careers and pulled their financial weight more or less equally with men and took it for granted that they had a moral obligation to share the burden with the men of their lives in earning the family living. Traditional men who accepted unquestioningly the traditional male role of paying for fluffies, I labeled "robots". I hit the media warning young fluffies, that if they wanted to get a man, they would have to get a career. I tried to persuade robots that they would be better off with a FIP. FIPs are cheaper to divorce. They share the financial burden. They free up men from their traditional financial cages, and are generally more fun and sexier to be with than boring fluffie housewives. I argued that if large numbers of robots switched to relationships with FIPs, the fluffies would rot on the shelf and would be forced to convert themselves into FIPs. A fluffie can only survive if some robot is prepared to be parasited upon. The media lapped it up.


2.2 My Work

In this section I will describe at some length the work that I have done over the years, with emphasis on what I have been doing during the past decade, since it is most relevant to the theme of this book.

After arriving in the UK, and spending a year in London, with its awful air pollution in the early 1970s, I had constant catarrh and decided to move to beautiful and academic Cambridge. I became a free lance mathematics tutor to the undergraduates of some half dozen Cambridge University colleges. The students would come in pairs to my apartment and be helped with the problems they were having with the math questions given to them by their lecturers.

I loved Cambridge, its green, its beauty, and especially its intellectuality. My first wife finished her philosophy degree at a London college and wanted to get pregnant. I wanted to move to Brussels. I got a job in a large Dutch electronics/computer firm and then moved later to Brussels, working in the computer field, but became increasingly frustrated and bored. I missed the intellectual life of Cambridge and its academic lifestyle. After I split up with my first wife, and worked for a few years again with computers in industry, I began a PhD in artificial intelligence and artificial life at the University of Brussels.

I began to evolve neural networks using a form of software simulated Darwinism, called Genetic Algorithms (GAs). I started this work in about 1989 and began publishing a steady stream of scientific research articles. I had 20 published by the time I had finished my Ph.D.

A neural network can be envisaged as a 3D array of brain cells - neurons - interconnected by branch like fibers called axons and dendrites. In an axon, a signal originating from a neuron travels away from it. In a dendrite, the signal is sent to the neuron. When an axon connects with a dendrite or another neuron, the junction is called a synapse.

In a real biological brain, each neuron or brain cell can have tens of thousands of synapses, that is it can be influenced by signals arriving from tens of thousands of other neurons. Those neural signals arriving at the neuron at the same time get reinforced or "weighted" and then summed. If the total signal strength is above the threshold firing value of the neuron, then it will fire, i.e. it will send pulses of electricity down its axon at a frequency proportional to how much greater the summed value is above its threshold value. The axon pulses then travel down to their synapses to influence further neurons.

This kind of biological neural network can be simulated in software. Typically the number of neurons simulated in a single network in the 80s and 90s was tens to hundreds. For my Ph.D. work, I was using at most 16 neurons per network. This contrasts so sharply to my present work, which deals with nearly 100 million artificial neurons.

A genetic algorithm (GA) uses a software simulated form of Darwinian evolution to optimize the performance of whatever is being evolved. For example, take my application of GAs to the evolution of neural networks. I simulated the behavior of a neural net in the following way. The first problem was how to represent the neural net itself. I took 16 neurons and had them all connect to themselves and all other neurons, so that there was a total of 16*16 = 256 connections. The incoming signal strengths, represented by ordinary decimal pointed numbers, e.g. 10.47, were multiplied by a weighting factor, e.g. 0.537, and then summed. As an illustration of this idea, imagine a very simple network of only 2 neurons, hence 4 connections. Neuron 1 sends a signal to itself at connection or synapse C11 and to neuron 2, at connection or synapse C12. Neuron 2 sends a signal to itself at connection or synapse C22 and to neuron 1 at connection or synapse C21. Assume that the signal strengths at a given moment are S1 and S (e.g. 10.54 and 7.48).

Each connection Cij or synapse possesses a corresponding weighting factor wij which is used to multiply the signal strength of the signal coming through the synapse. So the sum of the signal strengths arriving at neuron 2 would be (w12*S1 + w22*S2). Similarly for neuron 1. There will be 4 of these weights. Assume that the value of each weight lies between -1 and +1. Thus each weight can be represented as a fractional binary number with say 8 bits (binary digits, 0s or 1s). 4 such numbers can be represented by 4*8 = 32 bits which can be laid out in a row of length 32 bits. With 16 neurons, I had a row or bitstring as it is called, of 16*16*8 = 2048 bits to represent the 16*16 weights of my neural network that I was evolving.

If I knew the 2048 bit values (0s or 1s) I could calculate all the 256 weight values, and hence construct a fully interconnected neural net from them. The reverse process was also possible. If one knew the values of all the weights, and the values of the initial incoming signals from outside the net, one could calculate the signal strength of each neuron as it fired. If one knew how each neuron fired, one knew how the whole neural network fired, or behaved. One could extract the signals from some of the neurons and use them as control signals to control some process, e.g. the angles of legs of a robot to make it walk.

To explain how to use a GA in this application, imagine generating 100 random bitstrings of length 2048 bits each. From each bitstring one can construct its corresponding neural net. To each net one applies the same initiating input signals to kick-start the signaling of the network. One extracts the output signals and uses them, for example, to make some stick legs walk by controlling the angles of the 4 lines that make up the stick legs. One then measures how far the legs walk in a given time.

Those bitstrings that generate neural nets that generate longer distance walking survive into the next generation. Those that generate shorter distance walking are killed off, Darwinian style, "survival of the fittest". The fitter bitstrings, i.e. those with higher performance scores or "fitnesses", have copies made of them, called their "children" or offspring. The children and their parents are then "mutated", meaning that at low probability, each bit may be flipped (a 0 to a 1, or a 1 to a 0) and/or crossed over. There are various ways to cross over, but one simple form is to take two parent bitstrings or "chromosomes" as they are usually called, cut them both at the same position, and then swap components. This is the equivalent of sex, which is basically only mixing of genes from two parents to form the offspring.

The fitter parents have more offspring. Each generation of the GA has a fixed population size, e.g. 100. Most mutations and crossovers cause the chromosomes to have lower fitnesses, so they get weeded out of the population. Occasionally, a mutation or crossover actually improves the fitness of a chromosome by a small amount, so that in time it squeezes its parents and other inferior chromosomes out of the population. By looping through this cycle hundreds of times, it is possible to evolve a neural network or whatever one is trying to evolve, that performs quite well.

For my Ph.D. at the University of Brussels I was evolving neural networks that gave time dependent output signals. As far as I know that was the first time that anyone had done such a thing. Previously, a few people had applied GAs to neural network evolution, but the applications were static, i.e. the signals being extracted did not change with time. This struck me as being unnecessarily restrictive. The GA should be able to handle time dependent outputs. Once I had this insight, I started to evolve a neural net that made some stick legs walk. It worked. It required a few tricks to get it to evolve, but it did work.

That initial discovery, that it was possible to evolve neural network dynamics (as distinct from statics) opened up a whole new world for me, and created a new research branch called "evolutionary neural systems". I began to wonder what I would do next. The thought occurred to me that if I could evolve one behavior with one neural net, I could evolve a different behavior with a second neural net, i.e. one with a different set of weights. The weight set determines the dynamics of the output signals.

I became more ambitious. Instead of playing with simple sticklegs confined to a 2D plane, I would evolve behaviors for a 3D simulated quadruped creature that I called "Lizzy". If I could evolve one behavior successfully, then I could evolve a whole library of behaviors, with one neural net per behavior. I could probably then switch behaviors by having Lizzy at first walk and then turn. To achieve a smooth behavioral transition, all that was necessary was to switch off the inputs to the "walk straight" behavior generating network (or module as I started calling them), and input the outputs of the walk module to the turn module. Simulation experiments showed that the motion transition was smooth. Great. I now knew I could get a quadruped creature like Lizzy to display a whole library of behaviors.

The question arose as to when one would want to switch behaviors. Perhaps such decisions might arise from stimuli from the environment. I started to see if I could evolve detector modules, e.g. signal strength detectors, frequency detectors, signal strength difference detectors, etc. Yes, it was possible. The next logical step was to attempt to evolve decision type modules, e.g. of the type - "if the strength of the 1st input signal is greater than S1, and the strength of the 2nd input signal is less than S2, then switch on action An", i.e. a stimulus signal would be sent to the module that executes action An.

Putting all 3 kinds of modules together, i.e. behavior generating or behavioral modules, detector modules and decision modules, it seemed to me that it would be possible to start making artificial nervous systems. If there were a lot of such modules, then I thought it would be fair to call such a collection, an "artificial brain". It was at this stage that I started to become very ambitious. I began to see myself as the future pioneer of artificial brains, as Mr. Brain Builder.

But there were problems. The computer I was using in the late 80s and early 90s was hopelessly slow for the task I had in mind. By the time I was playing with a dozen evolved modules, the simulation speed of Lizzy on the computer screen was becoming noticeably slow. Every time I added another module's weights, the simulation speed slowed further. It became obvious to me that this was not the way to go. How to get around this problem?

By this stage I had finished my Ph.D. and was now postdocing in Japan, in 1992. In the summer of that year I was in the US talking with an electronic engineer acquaintance of mine at one of the universities that I am associated with, namely George Mason University in Virginia. I was asking this acquaintance how it might be possible to use electronics to speed up the evolution of my modules. After about an hour's discussion, he mentioned something called an FPGA (a field programmable gate array). Not being an electronic engineer, I had never heard of such a thing. "What's an FPGA?" I asked. He told me that it was a special kind of chip that was programmable, i.e. one could send in a bit string that would instruct the chip how to wire itself up or configure itself, to use the technical term.

I suddenly got very excited. A vision flashed before my eyes. Since I had spent the past few years using GAs to evolve neural nets, my immediate inclination was to imagine the configuring bit string as a GA chromosome, so the idea that it might be possible to evolve hardware directly in the chip suddenly looked plausible. I began to grill my acquaintance. Can the configuring bit string be sent in an unlimited number of times? He thought for a moment, and replied that if the chip were based on RAM, i.e. computer memory, then like ordinary RAM in any computer, the programmable chip could be reprogrammed as often as one likes.

I felt overjoyed. It meant that it might be possible to send in random bit strings that would configure or wire up the programmable chip in a random way, generating a complex random circuit. If there was another circuit, programmed by a human being to measure the performance of the randomly programmed chip, then it might be possible to perform a GA directly in hardware at hardware speeds.

I was so excited by this vision, that as soon as I got back to my Japanese research group, I gave a seminar on my idea and launched the research field of "evolvable hardware". I wrote papers on this idea, preached it to colleagues, gave talks on it at conferences, etc. The research field of Evolvable Hardware, or just EH, is now an established research field, with its own conferences every year in the US, Europe and Japan, plus its own academic journals. I feel I am the father of this field, and use its basic ideas in my daily work.

The following year, 1993, I moved to a research lab in Kyoto, Japan where I began work on building an artificial brain. I was convinced after my discovery of the possibility of evolvable hardware, that I had found a tool that would make the building of an artificial brain practical.

I started writing papers announcing that I intended to build an artificial brain with a billion artificial neurons by the year 2001, which happens to be the year that I am in as I write this. In 1993, to make such an announcement invited disbelief, because at the time, most neural net researchers were dealing with tens to hundreds of neurons, as I had been in earlier years. To hear someone suddenly announce that he was going to use a billion, sounded ludicrous. I was laughed at, ridiculed.

But, I was convinced that my vision was sound. If one could build a special kind of computer based on the principles of evolvable hardware, then its electronic evolution speeds would make brain building practical. I did the math and reasoned that a billion neuron artificial brain by 2001 would be just about doable. I had a contract with my Japanese lab for 7-8 years, so I thought I had the time to be ambitious.

My first task was to choose some kind of medium in which to grow and evolve neural nets. I chose to use cellular automata (CA). Each cell of a cellular automaton can be likened to a square on a chess board, but with two differences. One is that the chess board has an unlimited number of squares. The other is that the squares are not confined to be only black or white but can be any of a finite set of colors. Each square can change its color into any other of the set only at the tick of a clock. The color that a particular square changes into depends on its current color, and the colors of its 4 immediate, touching neighbor squares. For example, if the North square is red, the East square is yellow, the South square is blue, the West square is green and the central square in question is brown, then at the next clock tick, the central square changes its color to purple.

By appropriately choosing thousands of such rules, it was possible for me to make these cellular automata cells behave like a neural network that grows and evolves. For example I could grow pathways 3 cells wide, in which I would send growth cells that moved down the middle of the path. When a growth signal hit the end of the growing path it would make the path extend, turn left, turn right, split etc, depending on the color of the growth signal. By mutating the sequence of these growth signals that were sent down the middle of the CA pathways, I was able to evolve the CA based neural net.

This process occurred in two phases. The first was the growth phase. After a few hundred clock ticks, the growth would saturate. No more 3-cell wide CA trails or paths could be grown. These trails were the axons and dendrites of the neural net. Once the growth phase was completed, i.e. the growth instruction cells had cleared themselves from the network, the grown neural net could then be used for the subsequent signaling phase. Input signals could be applied, which would propagate over the network. These signals behaved like the signals in the neural networks that I had evolved in earlier years. They could be extracted at output points and used to control processes whose fitness or performance quality could be measured. The fitness of the performance became the fitness of the network, which in turn was grown from a sequence of growth instructions, i.e. a random string of 6 different integers (whole numbers).

What I had done was marry neural nets with cellular automata. This had not been done before as far as I know. The reason for doing this was that I thought CAs would be a suitable medium in which to have billions of CA cells, more than enough for a billion neurons. It seems to me to be practical. The workstations (i.e. computers a bit more powerful than PCs) of the time would have a gigabyte (a billion bytes) of RAM memory in them. RAM is cheap, so since I could store the state or color of one CA cell in a single byte (8 bits) of RAM, and my workstation could have a gigabyte of RAM, that would allow me to store the colors of a billion CA cells, a billion! That's a lot, more than enough in which to put an artificial brain which a huge number of neurons. Space would not be a problem. The technology of the time would allow it. It would be practical.

It took me about a year to write all the rules (NorthEastSouthWestCenter type rules) to show that a 2D version of a CA based neural net would work, that it would evolve. I had to hand code (with software productivity tools to help me) about 11,000 such rules to get it to work, but work it did. I successfully evolved oscillator circuits, signal strength detection circuits, line motion detector circuits etc. It was time to move on to a 3D version which would have quite a different topology. In 2D, circuits have to collide. They cannot go past each other. Whereas in 3D, CA trails can pass each other using the 3rd dimension. The dynamics and evolvability of 3D circuits would be much richer than the 2D case.

I got the 3D version to work but only after another 2 years, and roughly 60,000 rules. By this stage I was feeling quite miserable in Japan. My immediate group boss had a policy of having only one person per project, which made me terribly lonely and intellectually sterile. I had noone to really talk with. After exerting some pressure I finally got a young German M.Sc. level student to help me for the year 1996.

I explained to him that the 3D version was pretty well finished, and that I was becoming increasingly disillusioned with the particular CA model that I had been using. I explained to him my dream of growing and evolving CA based neural circuits directly in electronics, at electronic speeds. I felt that together we would need to simplify the CA model, so that it would be possible to fit it all into the electronics of the time, i.e. 1996. He listened to my list of desiderata and then disappeared for 2 weeks. He returned with a new, much simplified neural net model that kept the essential features of my old model, but added features that simplified it to such an extent that indeed the new model could be put directly into electronics. This new model was called CoDi, and was conceived by Felix Gers.

At about this time in the second half of 96, I was contacted by a Russian/American electronic engineer by the name of Dr. Michael Korkin, who lived in Boulder, Colorado. He had found my papers interesting and wanted to collaborate. I sent him details of Gers's new model and asked him if he thought he could implement it in hardware using special FPGAs that were then on the market. He said he thought he could. My Japanese boss approved the financing of the idea and a close collaboration between Mike Korkin and myself then started. Unfortunately I lost Gers only after one year. He went to do a Ph.D. in Switzerland. My Japanese boss reverted to his old policy of one person per project and I became more miserable than ever. I was also becoming increasingly fed up with Japan, with its repression of individuality, its insularity, its social backwardness, its status as an unrepentant criminal nation after killing some 30 million people in the war and not feeling any guilt about it. The Japanese government keeps knowledge of its massive crimes hidden from the general population. I was starting to want to leave, but couldn't, because the new machine had just been approved for construction. How I survived 8 years in Japan seems a miracle to me now. After so many years in the same lab with my growing contempt for Japanese culture, it was only a question of time before the management there would be happy to see me go. Mitigating this was the fact that I was putting their lab on the world map with all the world wide media publicity I was getting. I was claiming to be the guy who was going to be the first on the planet to build an artificial brain. I must have been a real thorn to them. What do you do with a researcher who makes the lab famous, yet truly despises your culture?

Relations with my group manager became increasingly strained, especially after I discovered he employed a policy of putting his name on academic journal papers written by his subordinates to which he made absolutely no intellectual contribution. He asked me to put his name on one of my journal articles. I refused, and told him that in the west that would be considered disgusting, an abuse of power, and corrupt. After that, relations soured fast. I was allowed to stay on until the end of the year 1999, and then I would have to leave. Since two thirds of the division of some 100 researchers would have to leave at the same time, I felt only half fired. The Japanese economy had performed so badly during the 90s, "the lost decade", that the whole research division was considered too blue sky, too fringey to be funded in times of economic scarcity.

I got another job. Ironically it was in Brussels, and to do the same work as I had done at my Japanese lab.

During the years 1996 and mid 1999, Mike Korkin was working away solidly in the US on constructing the special piece of hardware that would fulfil my dream of building artificial brains. It was slow going for him. He had only a limited budget from my Japanese group boss. He could afford only one full time assistant plus a few part timers on limited term contracts.

During the course of his work, the US company making the FPGA chips that the machine was based on, decided to take them off the market. Mike then had to fight the company to get the remaining chips. This caused many months of delay. The chips were finally obtained, but were untested. Thus he had to test them himself, without the thorough testing software that the company would have - more delays.

It was not until mid 2000 that the machines that I called CAM-Brain Machines (CBMs) were finally debugged sufficiently that true evolution experiments could begin. CAM stands for Cellular Automata Machine, because the original work was to put an artificial brain inside cellular automata.

The first CBM was delivered to my Japanese lab in early 1999, but it still contained bugs. With untested chips and small manpower, work progressed slowly. But all was not gloomy. Other people became interested in the CBM. As I write in early 2001, there are 4 such machines in the world. The first remains at my old Kyoto lab in Japan. The second was bought by a Belgian speech-processing lab and later transferred to a bioinformatics company also in Belgium. The third was bought by my new Brussels lab, and the fourth is owned by Mike Korkin himself. Thus with 2 of the 4 machines in Belgium, Belgium is in a sense the world leader in this field. In 2000, I managed to get a million dollar grant from the Brussels government to build an artificial brain to control a small robot, giving it hundreds of behaviors. As you can see, my current work is really only a glorious extension of my old Ph.D. thesis work.

Just what can the CBM do? I believe it is truly a miraculous machine, that in time, once people appreciate its significance, will take its place in the history of computing. It implements the CoDi CA based neural net model directly in electronics. It evolves a neural net in a few seconds, i.e. it performs a complete run of a genetic algorithm, i.e. tens of thousands of neural circuit module growths and fitness measurements. It can change the color of CA cells at the phenomenal rate of about 130 billion a second. It can handle nearly 100 million artificial neurons. It has the processing capacity of about 10,000 PCs, so is definitely a supercomputer but costs only $500,000.

The CBM has two main roles. The first is to evolve individual neural circuit modules, or just modules, I call them. A neural net is grown/evolved inside a 3D CA space of 24*24*24 CA cells or little cubes. About 1000 neurons can fit inside this space. Branch-like axons and dendrites grow randomly inside this space. A programmed FPGA is used to measure the quality of the neural signaling of the network that is grown. The basic ideas are similar to what I was working on before 1996. Once a module is evolved, it is downloaded into a portion of a gigabyte of RAM memory. 64000 of such modules can be evolved one at a time, each with its own fitness definition (i.e. task or function) as specified by human "evolutionary engineers" (EEs) and downloaded into the RAM. Later, "brain architects" (BAs) interconnect by hand the downloaded modules to form their humanly specified artificial brain architectures to perform the tasks that they want.

2.3 Future Tasks and Dreams

At the present time I and a small team of full time collaborators at my present lab in Brussels have recently started using the CBM to evolve individual modules, mainly for pattern recognition tasks. For example, we can evolve a module capable of detecting whether a line of input stimulus moves up or down an input face. We are testing the level of "evolvability" of the neural net model we have implemented in the CBM. Of course this model is constrained by the state-of-the-art hardware that it is implemented in.

The modules do not always evolve the way we want or even at all sometimes. What I call "evolutionary engineering" is a black art. There is no theory to guide evolutionary engineers (EEs) on how to improve evolvabilities, a concept fundamental to this field. At the present time, we are getting a feel for what the CBM can do, its strengths and limitations. Noone has done this kind of thing before, so we are struggling in the dark. There are no signposts. This is research. Every step of the way is new and may often blunder into an unanticipated problem. But, we are making headway, even if at a slower pace than I had estimated way back in 1993 when I started this project.

If those people who had laughed at my preposterous assertion that I would build an artificial brain with a billion neurons by 2001 were able to see the CBM in 1993 as it exists in 2001, they would not have laughed. Admittedly the machine cannot handle a billion neurons. The actual figure is 75 million, but that's only one order of magnitude off. That's not bad. Admittedly also, the task of architecting the artificial brain, a huge task, still lies ahead of us, and will take several years. There is still a lot of work to do, and I still suffer from critics. With all the delays, whether for commercial, intellectual, managerial, or personal reasons, I still do not have an artificial brain to show off to people. Some journalists are starting to get impatient, and are wondering when I will deliver.

During the next year or so, if all goes ahead as planned, my team needs to complete its evolvability studies, evolving one module at a time. If the evolvability levels are not sufficient, we may have to change the fitness definitions we use in the CBM. We may also have to change the neural net model implemented in the reprogrammable FPGAs. Once that stage is over, the next will be to start building multi-module systems, with 10s of modules, then 100s, then 1000s, up to 64,000, to build an artificial brain aimed at controlling the behavior of robots. We intend to show off a robot with many behaviors controlled by an artificial brain. One will not need to have a Ph.D. to understand what is going on, as is the case with the CBM, but just by simple observation of the robot, one should be able to see that "there is a brain behind it".

In parallel with all this work, which should take several more years, is the need to start serious thinking about the next generation of brain building machine, that I call the BM2 (brain building machine, 2nd generation). I have started collaborating with another American colleague who has had some revolutionary ideas for the next generation of electronics that self configure. He has estimated that with a budget of a few million dollars, it should be possible to build a next generation machine within about 4 years which should be about 1000 times more performant compared with the CBM.

In fact, it is my ambition to continue trying to build a new generation brain building machine and its corresponding brain every 4-5 years. I'm now in my mid 50s, so if I choose to retire in my 70s, that gives me about 20 years, or 4 more generations. In 20 years, if Moore's law continues to be valid that long, it will give humanity the ability to put one bit of information on one atom. Once that happens it will be possible to build what I call "Avogadro Machines", i.e. machines with a trillion trillion components. Avogadro's number is the number of molecules in an object of human scale, e.g. an apple in one's hand.

If the second generation brain building machine can be funded and can be built within the next 4-5 years, then it will be possible to make the next generation brain more similar to the biological brain. The neural net model it implements can be more sophisticated and closer in its behaviors to those of biological neurons.

Within a mere 20 years, i.e. my own working lifetime, humanity, hence I and other brain builders, will have the technologies and the tools to build ever more performant artificial brains.

Is it any wonder then, that someone who is as politically and socially critical as I am, is beginning to feel alarmed at the rapid progress that brain building can be expected to make in the coming 20 years. What will our artificial brains be doing for humanity in 20 years? I would say it is highly likely they will be in our homes, cleaning them, babysitting our kids, talking with us, giving us infinite information from knowledge banks all over the planet. We will be having sex with them, be educated by them, be entertained by them, made to laugh by them etc. The brain building industry 20 years from now I estimate will be worth about a trillion dollars a year worldwide. By 2005 I hope and expect that if my own group can "prove concept" within the next year or so that brain building is doable, then a new "brain building" research field will have been established.

If we have all this within 20 years, where will humanity be in 50 years, in a 100? Given the exponential progress in the accumulation of our knowledge of brain science, all of which can be immediately incorporated into neuro-engineering the moment it is discovered, I feel that the initial positive feelings about artificial brains will later turn sour and develop into fear.

I am attempting to become the father of the artificial brain. I am already the father of evolvable hardware and of evolutionary engineering, which are the enabling technologies of this new field. If I were a traditionally minded engineer or scientist, I would probably be quite content to get on with my work and not worry about its longer term social consequences, but I'm not like that. I'm a very political animal, and I'm very worried. My rather unusual combination of being a scientist/engineer and at the same time a social critic and media person makes me an appropriate person I believe to raise the alarm on the artilect issue.

I'm hoping that my credibility or otherwise as a professional brain builder will aid my attempts to raise the alarm on the rise of the 21st century artilect. However, the two need not be connected. Even if I fail to build an artificial brain, others will succeed. For me to succeed with each brain-building-machine-generation, and the building of its corresponding brain, I will need to raise more money, hire more people as the scale of the enterprise keeps increasing. I will need to become like Goddard, the US rocket pioneer, or Werner von Braun, who put an American on the moon. Both these men started with toy rockets, but had a vision. In the 20s Goddard's first contraptions were not much better than 2m tall ancient Chinese style rockets. 20 years later both he and von Braun were heavily subsidized by their respective governments to build highly sophisticated rockets capable of travelling great distances. In the late 60s von Brawn played a major part in an organization that put Armstrong on the moon.

I have similar dreams. I dream of national projects paying billions of dollars to build artificial brains. I have talked of the J-Brain Project (Japan's national brain building project), the A-Brain Project (America's), the E-Brain and C-Brain Projects (Europe's and China's). Within 20 years, and in possession of Avogadro machines, there will be so much work to be done in building a brain with not billions of components but trillions of trillions of components, that a huge team of people will be needed. That's my longer term dream, 20 years from now.

After that, once I have retired, I hope I will be able to play the role of the wise old man who advises younger minds on where the whole brain builder effort ought to be headed. As this book shows, I am not optimistic about the future survival of humanity when faced with machines that become ever smarter at exponential rates.

My ultimate goal is to see humanity, or at least of portion of it, go Cosmist and to do it successfully by building truly godlike artilects that tower above our puny human intellectual, and other, abilities. That is my true goal. I won't live to see it unfortunately. True artilects won't be built within the 30-40 years I have left. I will not live to see the ultimate fruits of my work. This is a source of great frustration and disappointment to me, but there is one consolation. At least I will probably die peacefully in my bed of old age. As this book shows, I fear for my grand children who I believe are likely to be destroyed in a gigadeath war over the issue of species dominance late this century.

So, dear reader, you have now heard the more technical side of my story, a description of my life's work. Does knowing this make you feel that my political opinions concerning a possible artilect war are more credible? Should I tell you that I am not just a Ph.D. but a guest or adjunct professor in China, Japan, and the US. I am also a Davos Science Fellow, the only one in Japan, so I get to go to the Davos World Economic Forum every 4 years to entertain the billionaires. I'm in the Guinness book of world records (p126, 2001) for the CBM. I was a guest editor of a special issue of an academic journal on "Evolutionary Neural Systems", which is usually an honor reserved for the person who is considered to be the best in the world in a given specialty by the editor in chief of the journal concerned.

If a lot of people consider what I am trying to do to be crankish, then I hope it is clear that at least I am a competent crank. The point of this chapter is to try to convince you that the author of this book, the coiner of the terms artilect, cosmist, terran, gigadeath, etc is worthy of being listened to. Whether I have succeeded is for you to judge.

=====================================

Chapter 3. Artilect Enabling Technologies

Some years ago, when I was trying to get an earlier draft of this book published in the US, I received an email from an American literary agent, saying that my manuscript, which she had read, was "quite well written", but "fantastical", making it a very hard sell for publishers?. Since then I have found in practice, that the greatest obstacle I have to face in trying to persuade people to accept these ideas, is their seemingly "science fiction" like character.

Most people, when confronted with such concepts as "massively intelligent machines" with artificial intelligence levels trillions of trillions of times about the human level, or of an "artilect war" killing billions of people ("gigadeath"), or of asteroid sized computers, etc, not surprisingly, their immediate reaction is one of incredulity. They will often laugh at the preposterousness of these ideas. Even many of my colleagues (especially the non physicists) do not take a lot of these ideas seriously. For example, I have been trying for the past year or so to persuade the most eminent "applied ethics" professor on the planet, Professor Peter Singer of Princeton University, USA, to take up these ideas. I'm trying to persuade him write a book on the topic of "Artilect Ethics", which would deal with the huge moral and ethical issues concerned with the possible construction of artilects this century. His reply was illuminating. I quote him from one of his emails to me. "To be blunt, I am not sure how to place you between the "total flake" and "genius ahead of his time" views of your ideas". This is from someone with a very open mind.

So you see my most pressing problem for the moment is one of credibility. How to persuade people that these ideas are not just a piece of non-serious "science fiction", but are very probable "future science" ideas. Admittedly, the persuasion task has become easier recently as the world media increasingly takes up the message. I am constantly on the world media lately, (TV, newspapers, magazines, radio, web, etc) in the major countries, (mainly the US, UK, France, Australia, Poland, etc).

Despite the growing credibility, there is still a long way to go, so it is essential in this book for me to try to persuade you that these ideas are worthy of serious consideration, and that they should not be dismissed out of hand.

This chapter is devoted to trying to persuade you that it will be possible to build artilects this century. The fabulous technologies that will be developed in the next 100 years will be so capable and so fantastic, that they will force the issue as to whether artilects should be built this century or not.

Once you have read this chapter, I hope you will be left with the strong impression that the artilect's potential intelligence is truly gargantuan. Artilects will have the ability to surpass human intelligence levels by many orders of magnitude, not just ten times smarter, or a thousand times or even a million times smarter, but by trillions, quadrillions, quintillions, truly zillions (using the generic term) of times smarter. (If something is 10 times larger than something else, it is said to be an order of magnitude larger. If it is 100 times larger, it is two orders of magnitude larger, etc).

This chapter will try to persuade you that these numbers are not exaggerated. There are very good reasons, based on the new technologies, that we can expect to be developed this century, to motivate us to believe that artilect building is a realistic proposition within the next 100 years.

This chapter will be the most complicated of the book, since it will be discussing scientific ideas and technologies that are new or do not yet exist. I will try to make this chapter as easy to understand for the general, non-scientific reader as I can.

As I wrote in the introductory chapter, one of my life goals, besides building artificial brains, is to raise the alarm on the "Artilect Issue", or if you prefer to call it, the "Species Dominance Issue", or the "Cosmist-Terran Conflict". There are several ways to label the same basic problem that is coming.

This issue is far too important to be confined to intellectual discussions amongst a bunch of "nerdy scientists". In time, it will concern everyone, because if Cosmists are serious in their "threat" to build artilects, everyone will be affected, one way or another. One does not initiate a great public debate by confining ones worries to the scientific specialists, a tiny proportion of humanity, less than 1%.

An effective beginning to getting people to talk about the artilect issue is to write a book. A book will help the journalists become familiar with the problem, and they in turn will write about it for the greater reading public. Similarly with the TV and radio journalists, who can present these ideas to an even wider audience, because unfortunately, only about a half of the population reads books.

Probably the most effective way to get the message across would be to have Hollywood make a block buster movie on the theme. I hope this will come. There are already significant nibbles from filmmakers in various countries at the time this book was written.

Before launching into details of these new or yet to be developed technologies, I ought to say a little about the category of readers who could most benefit from this chapter, which I think is the most difficult of the book.

To fully appreciate this chapter, it will be helpful if readers have at least a high school education in science, but this is not essential. I need to talk about some very "high-tech" technologies and even technologies that don't yet exist, so I will have to go into some level of detail. I hope that few readers will be put off.

If you are, I suggest you just read as much of this chapter as you can follow, without too much effort, then skip to the next chapter, which discusses the many points of view of the Cosmists. However, if you do decide to skip this chapter, I suggest at least you accept its main conclusion, which is (to labor a point) that this century's technologies will enable the building of artilects, which could become zillions of times smarter than human beings.


Moore's Law

I begin the introduction of the artilect enabling technologies of this chapter with the phenomenon known in the electronics world as "Moore's Law" that I talked about briefly in the introductory chapter. This time however, the concept will be treated in greater detail. Gordon Moore is a person, still alive in the early 21st century, who was one of the co-founders of the "Intel" microprocessor company, in Silicon Valley, California, USA. In the mid 1960s, he noticed that integrated circuits were increasing their speed and density (i.e. the number of transistors crammed onto the surface of a silicon chip) by a factor of two every year or so. This doubling rate has remained more or less true for the past 30 years and many people believe that it will continue right down to the molecular scale.

What is the point of trying to make electronic components smaller and more densely packed? If two electronic components have to signal each other, and given the finite speed of light (i.e. the maximum speed with which electronic components can send messages to each other) then the closer these components are to each other, the faster they can influence each other. Also, the smaller the size of the components, the greater is the number of them that can be crammed into a given surface area. Hence the chip can deliver greater performance because it has more components to do more things.

The microchip industry is thus under constant pressure to scale down, to make its transistors smaller, its circuits smaller. If a company falls behind in this frenetic race, it will lose sales and go out of business. If the rival company down the road is six months ahead in its development cycle, and thus releases a new batch of products ahead of your company, you are at a great disadvantage. New generations of chips and the computers based on them come out every year or two. We are getting used to this now. We know that if we wait six months or a year, we will be able to buy a better, more performant computer.

Moore's Law is probably one of the most important technological and economic phenomena of our times. It is fueling the digital revolution, which is now driving our global economy. So many jobs and such a large proportion of the GNPs (Gross National Products) of many nations are now tied up with the electronics-computer- telecommunications industries that if Moore's Law were ever to stop, humanity would be in for a real shock. However, there is a problem.

As the size of electronic components, particularly transistors, gets smaller and smaller, a scale is eventually reached which is so small that a different set of physical laws, which governs their behaviors, begins to apply.

If Moore's Law can continue right down to the molecular scale, i.e. if the size of electronic components can reach that of molecules and still be functional, then new laws of physics must be applied. The old "classical mechanics" discovered by Newton in the 17th century is no longer appropriate, and must be replaced by the newer 20th century based "quantum mechanics".

Quantum mechanical laws govern the behavior of atoms and molecules (and even smaller scales). For example, as the line widths of wires connecting electronic components on the silicon surface of a chip are reduced below about 0.1 micron (a micron is a millionth of a meter, about the size of a bacterium), quantum mechanical phenomena begin to appear. These phenomena make themselves felt with such a strength that the usual transmission of electrons (i.e. electric current) down the wire at larger scales, is severely disturbed.

There are many other similar reasons why researchers in electronics are worried today. They know that they must shift away from conventional electronic principles into quantum mechanical principles if electronics is to continue its incredible "Moore doublings" phenomenon. Instead of looking upon these quantum effects as a disturbance of conventional electronics, a growing number of electronics researchers are accepting the inevitable, and have started to think of new electronic and computing techniques which embrace the quantum phenomena as their functioning principles.

If Moore's Law continues unstopped until 2020 or thereabouts, it will be possible to store one bit of information (a zero or a one, a "0" or a "1") on a single atom. An excited atom (in which an electron circling the nucleus of the atom has a higher energy than usual) could be interpreted as storing a "1", and an unexcited atom as storing a "0". The two different states "0" or "1" would correspond to the two different energy levels of the atoms.

The enormous significance of this scaling down to the atomic level, is the huge number of potential electronic components one could then have in a given volume. It was the Italian chemist Avogadro in the 19th century who first estimated the number of molecules in an object of human scale, such as an apple. The number is so large that it is almost impossible for the human mind to conceive.

Avogadro's Number is 6.023 times 1023, i.e. nearly a trillion trillion (a 1 followed by 24 zeros). That number is a hundred trillion times larger than the number of human beings alive on the earth at the beginning of the 21st century.

Molecular scale electronics holds the promise of truly huge computational capacities, and all this perhaps by the 2020s. When I talk about an artilect having a potential artificial intelligence of trillions of trillions of times the human level, part of that assumption is based upon the enormous computational capacities that we will have in a mere few decades, which a future artilect could possess.


Reversible Computing

The above idea of having trillions of trillions of electronic components inside a small volume (say that of an apple, or a few cubic centimeters) contains an implicit assumption, and that is that the electronic circuits contained in such a volume would be distributed throughout that space. They would be three dimensional circuits (3D). But today's electronic circuits are all 2D, imprinted on the surface of silicon chips. Why is this? Why doesn't modern electronics take advantage of the far greater storage capacities of 3D circuits?

The answer has to do with the problem of heat dissipation. The following paragraphs will explain.

For the past few decades, theoretical physicists have been asking themselves some fundamental questions about the ultimate limits of the physics of computing. This branch of physics goes under the label of "phys-comp" (physics of computation). One of the questions that has been asked in the phys-comp field is "What is the minimum amount of energy or heat that must be dissipated to perform an elementary computational step?"

If you put your hand over your PC, or if you have your laptop on your lap as I do now as I type this, you will be fully conscious that your computer is generating heat. Computing inevitably generates heat it seems, or does it?

In the 1960s, a researcher named Landauer discovered that what was generating the heat in computers was the process of "resetting" memory registers (a register is a linear storage chain of 0s and 1s), i.e. wiping out their contents and resetting them to 0s. He discovered that the heat was generated when information was "wiped out" or destroyed.

To be a bit more technical, to wipe out the contents of a register implies increasing its order, making it less random. In physics, the concept of "entropy" is used to measure how disordered a physical system is. For example, ice has a lower entropy than water, because it is more ordered, less chaotic.

One of the basic laws of physics, known as the "Second Law of Thermodynamics" is that entropy never decreases in a closed system (one where energy cant get in or out). So if the contents of a register are wiped out, its entropy, its measure of chaos, decreases, so where does the rest of the entropy go, given that the total cannot decrease? The answer is in the form of heat to the surrounding environment of the computing component.

Today's computers generate heat because we are using thermodynamically irreversible processes (i.e. we can't reverse the effects at a later time). We generate heat every time we destroy information, i.e. wipe out bits (resetting to 0s). Landauer thought that this was inevitable, because when he looked at how the computers of his time all functioned, he saw that they were full of "AND gates", and the like.

An AND gate is an elementary piece of electronic circuitry which has two input signal lines (A and B), and one output line. If both input lines are set at a high voltage (i.e. have a "1" on their line) then the output line will become a "1", i.e. if both input line A AND input line B are set at "1", then the output line becomes a "1". In any other case (i.e. A=0, B=0; A=0, B=1; A=1, B=0) the output line becomes a "0".

Since there are two input lines containing a total of 2 bits of information in an AND gate, and only one output line containing 1 bit of information, of necessity, the AND gate destroys information. (If you are told in which state a system is in, that can have two possible states, then you are given 1 bit of information. For example, take the question "On which side of the road do the Japanese drive?" When you are told "On the left-hand side", you have been given 1 bit of information).

Every time two bits go through the AND gate, only one bit comes out. The AND gate is irreversible, i.e. you cannot always deduce from the output what the input was. For example, if the output was a 1, then you know the inputs were both 1, but if the output was a 0, you don't know if the inputs were, (0,0), or (0,1) or (1,0). For a gate to be reversible (i.e. you can deduce what the inputs were from its outputs and vice versa), common sense says that you have to have the same number of input lines as output lines.

People began to dream up reversible elementary circuits (or "gates") with an equal number of input and output lines. (A "gate" is an elementary electronic circuit that performs some basic operation, e.g. an AND gate, an OR gate, a NOT gate, etc). One famous such gate was called the "Fredkin Gate", which had 3 inputs and 3 outputs. The Fredkin gate is reversible, so no bits of information are destroyed. It is also "computationally universal", i.e. by feeding the outputs of Fredkin gates to the inputs of other Fredkin gates, larger circuits of these gates could be built up that could perform any of the functions that computers need to perform.

Since the individual gates of the computer were reversible, the computer itself could be made reversible. In other words, one could input the initial bits into the left-hand side of the computer, and these would be processed by the Fredkin gates in the computer design. The resulting output (the answer) would appear exiting from the gates at the right-hand side of the computer.

You can make a copy of the answer (which might generate a little bit of heat) and then send the answer back into the computer from right to left. Since all the gates of the computer are reversible, you will end up with what you started with at the left-hand side. You have performed a reversible computation. No bits have been lost and hence no heat has been generated. Nevertheless, you have the answer you want, because you made a copy of it half way through the computational process, i.e. before you "reversed" the direction of processing.

Reversible computing may take twice as long as traditional computing, because you have to send the result backwards through the same circuit (or an identical copy), but at least there's no heat generated.

What is the significance of this! Why am I spending so much time and energy explaining such things? Because I believe the theoretical discovery of reversible, heatless computing in the 1970s was one of the greatest scientific discoveries of the twentieth century and is of great relevance to the main ideas of this book.

Since this is such a strong statement and will probably be treated with some skepticism by many people, particularly some of my research colleagues, let me try to justify why I have this opinion.

A few years ago, some phys-comp theorists were wondering, "If Moore's Law extends right down to the molecular scale, how hot would molecular scale circuits become if one continues to employ conventional irreversible, bit destroying, information processing techniques?" The answer was shocking.

Not only would such highly dense circuits melt with the heat, they would become so hot they would explode. It became clear that molecular scale circuits, if ever they are to be built, would have to abandon the traditional irreversible style of computing, and start using the new reversible style.

Only recently have researchers started thinking seriously about reversible computer designs. The laptop and palmtop computer industries are very interested in reversible computing, because it might help them with their "battery lifetime" problem.

If their computers could use electronic circuits that were more reversible, the circuits would consume less battery energy, because they would generate less wasteful heat. Hence the battery would drain more slowly and have a longer life. Consumers will be more likely to buy laptop computers which have batteries that last longer. Wouldn't it be nice to have a single laptop battery that lasted for a full transatlantic flight, for example.

So, it is inevitable that reversible computing has to happen. As Moore's Law continues to bite, pressure will increase on computer designers to use the reversible paradigm. It is only a question of time.

But, if we start taking the concept of heatless computing seriously, we can begin to play with some revolutionary ideas. For example, why are today's electronic circuits two-dimensional? Why do we talk of 2D "chips" (i.e. slices) of silicon, rather than 3D "blocks"? Well, because of heat. If we made 3D blocks of silicon with today's level of density of electronic components, there would be so much heat, the blocks would melt. Also, how would we build them and debug them once they were built? We do not have the techniques yet to do such things. We don't even bother trying to build 3D circuits because we know it would be a waste of time, due to the heat dissipation problem.

But, with reversible heatless circuits, we have the luxury to build large 3D circuitry, with in principle, no limit to size. We could make circuits the size of a cubic centimeter, or a cubic meter, or the size of a room, or a house, or a building, or a city, or even a large asteroid hundreds of kilometers across. (An asteroid is a huge boulder of metal or rock that orbits the sun at a radius between those of Mars and Jupiter. There are thousands of very large asteroids in the "asteroid belt").

In theory we could make computers the size of moons or planets, but the gravitational effects might prove to be problematic.

You are now probably beginning to suspect why I think reversible computing is so terribly important. Ask yourself how many bits of information you could store in an asteroid. The answer is about 1040, i.e. a "1" followed by 40 zeros, i.e. ten thousand trillion trillion trillion atoms and hence bits.

Also ask yourself how many brain cells (neurons) we have in our heads. The answer is of the order of 1010, i.e. tens of billions. If we could accurately simulate in a computer the behavior of one biological neuron using a trillion bits (and that may be overkill) we would still have 1018 (18 = 40 - 10 - 12, i.e. a million trillion) human brain equivalents in one asteroid.

I suggest you really study these numbers. They are the writing on the wall for me. What this means is that, sooner or later, humanity will be able to create vast computing capacities, enormously eclipsing human brain levels. It is therefore only a question of time before humanity has to choose whether to fully exploit such enormous computing potential or not.

Let me spell out a bit more explicitly just how phenomenal such an asteroid sized computer might be, and what it could do. Firstly, it could "think" a million times faster than do our brains. The neurons in our skulls communicate with each other at maximum hundreds of meters a second. Electronic signaling speeds, as in computers or an artilect, would be a million times faster, i.e. close to the speed of light, which is 300 million meters a second.

Even if these artilects had the same intelligence levels as humans, they could do in seconds what takes us years. Instead of getting a Ph.D. in 4 years, it would take such an artilect only 4*50*5*8*60*60/1,000,000 = 30 seconds. But an artilect has far more than just one human brain equivalent. It has zillions of times more. So if it could distribute its thinking over all its asteroid brain, then it could do what we do in 4 years in picoseconds or less. (A picosecond is a trillionth of a second).

Artilects would be so fast in their thinking that our human pace of thought would seem as slow to them as humans trying to communicate with rocks. Over millions of years, rocks change their shape, which might be interpreted as conveying a message, but humans don't have the patience (nor the life span) to wait.

There is a strong case to be made by the Cosmists that advanced artilects would be totally bored by humans with our glacial thinking speeds, and simply ignore us. They would invent whole histories within themselves in the time it would take us to utter one word.

But the artilects need not be limited to human intelligence levels. It is not difficult to make out an argument saying that one ought to be able to extrapolate the trend in human IQ levels as we discover the neurobiological structures and functions that make one human being smarter than another. In time we should be able to look at an ordinary person's brain and Einstein's brain and notice neuro-physiological features that correlate with higher intelligence.

We could then plot a graph depicting IQ on the vertical axis, and the neuro-physiological features that correlate with high IQ (e.g. number of connections per neuron in certain regions of the brain, etc) on the horizontal axis, and then just extend the trend. We may see the development of an "Intelligence Theory", as we learn more about how the human brain works and understand just what it is that makes humans more intelligent than other animals. It may become clear to us, that if we simply increase certain parameters in the design of artificial brains, we may be able to increase the level of intelligent behavior in the robots that these artificial brains control.

So, asteroid sized artilects need not be limited to architectures which generate human level intelligence. Artilects could not only think faster, with hugely more components, but in qualitatively superior ways as well.

Their huge surface areas, would allow them to attach huge numbers of external sensors to themselves, including use of the full range of electromagnetic wavelengths from gamma rays to radio waves. They could communicate with other asteroid artilects across the asteroid belt and deeper into space.

Such asteroid sized artilects using nanotech based principles are probably the logical extreme of human technological imagination (unless we can create something called "femtotech", which I will discuss a bit later.

Before asteroid sized artilects are built, earlier versions will certainly be much smaller, more on a human scale, but even at this smaller scale, when talking about one bit per atom, we will still have many technological problems to solve in order to make such computers.


Nanotechnology : Molecular Scale Engineering

This brings me to the need for "nanotech", as distinct from "femtotech", that I mentioned just above. "Nanotech" is an abbreviated form of "nanometer scale technology", i.e. molecular scale engineering. Nanotech builds things at the scale of a nanometer, which is one billionth of a meter, about the size of molecules. "Femtotech" is an abbreviated form of "femtometer scale technology". A femtometer is a quadrillionth of a meter, i.e. a thousandth of a trillionth of a meter, which is about the scale of a proton or neutron inside the nucleus of an atom. Femtotech would be nuclear or even quark scale engineering. Quarks are "elementary particles" which combine to constitute protons, neutrons and other such particles.

The first people to think about the possibility of building things at the nanometer level were presenting their ideas in the 1950s. In the 1990s, these ideas had become well accepted and regular monthly progress was being made in this domain. The essential idea is that atoms can be placed into exact position to build molecular scale machines, e.g. tiny molecular scale robots which pick up atoms and position them with great precision.

When one begins to imagine the kinds of things that could be done with molecular scale machines, the field begins to sound truly science fiction like, yet the possibilities exist. Many scientists feel that given the current rate of research progress in the field of nanotech, it will probably be well established by about the year 2020. This is about the same time that it will be possible to store one bit of information on a single atom, according to Moore's law.

Consider some of the more fantastic things we could do with a fully-fledged nanotech. Imagine tiny robots sent into the blood stream of human beings, which are programmed to detect cancer cells. They would travel throughout the body, detect the cancerous cells, kill them, and then self-destruct or be flushed out over time. A similar story could hold for "immortality generating" robots, which could repair aged cells and restore them to a state like those of young children. With a regular dose of such "fountain of youth" robots, people could become immortal.

Each of the cells in our bodies contains a DNA program which explicitly or implicitly causes the cell containing it to die. This DNA program takes the form of a molecular structure that can be reprogrammed by a molecular scale robot, a nano device. Hence nanotech offers humanity the prospect of immortality. If that happens, we will need a new politics to decide who lives forever, who dies, and who reproduces.

Another favorite nanotech idea is to have one's head or one's whole body frozen soon after death on the assumption that in a century or so, it will be technologically possible to restore the damage to the dead brain and make it come alive again. Nanomachines, the theory goes, would be able to enter the dead tissues and repair them.

There are already hundreds of people who have paid for their bodies or brains to be frozen for an indefinite period. They establish a monetary fund whose interest pays the cost of the apparatus and materials to keep the body frozen.

Molecular scale robots (nano scale robots, or "nanots") could build copies of themselves. They could reproduce and hence grow exponentially in numbers, 1, 2, 4, 8, 16, 32, 64, etc. After 20 such doublings, the numbers are into the millions. If ways could be found to make these nanots cooperate to build human scale products, then conventional economics would be revolutionized. It would cost almost nothing to make huge numbers of nanots which then build the product. The price of the product would then be merely the cost of the raw materials. Goods could become amazingly cheap, effectively costing nothing.

The cost of designing the first such self reproducing nanot, capable of cooperating with others of its kind to build specific products, would be amortized over the many purchases of the same nanot design all over the planet, and hence would cost almost nothing. The whole concept of economic scarcity would need to be reconsidered.

Construction materials could be made many times stronger, because today's materials still contain cracks, faults, etc which weaken their strengths. Nanotech could assemble these materials with atomic precision with no faults, no cracks, no blemishes, and hence they would be much stronger. This would probably mean that we could construct buildings that would be kilometers high if we wanted to. The structural skeletons of the buildings would be strong enough to withstand the stresses and strains generated by strong winds. Diamond like materials could be built with amazing strength.

As I see it, there are at least two major paradigms conceivable when discussing how nanotech could build human scale products. One is to imagine zillions of self-reproducing nanots which once reproduced, would combine to build the product.

How would such a mammoth task be coordinated? One would need to think of the manufacturing process like a molecular scale city with a huge infrastructure to make it all happen. One could imagine nanots each doing their tiny thing on conveyor belts, assembling their few atoms at this point, at that point, and passing on the result, further down the line, where other nanots do something different. It would be Henry Ford at the nanoscale.

With zillions of nanots doing the same thing simultaneously (in parallel, as computer people say) it would be conceivable to imagine a human scale device being built. Molecular modules could be built from atoms, and these modules used as components to build larger macro-modules, which in turn become components of macro-macro-modules etc, until a human scale product is built. To make such a construction system work, the enormous molecular infrastructure needed may or may not prove to be very practical.


Artificial Embryology

Personally, I prefer a nanotechnology based on the method nature has used for billions of years to build its life forms, i.e. an "embryological approach". In the embryological process, one starts with a fertilized cell which divides and divides until some cells (depending upon their position in the embryo) begin to differentiate. Their intercellular environment sends them chemical signals which are used to switch on and switch off certain portions of their DNA, which in turn, results in different proteins being built, which perform different tasks. These different proteins then change the nature of the differentiating cells. Eventually, the mass of differentiating cells creates a living three dimensional biological creature.

Evolution has created a growth mechanism that takes a linear one dimensional coded string of chemical instructions (usually called DNA) and translates it into a three dimensional functioning living creature. The study of how this miracle of nanoscale engineering occurs is called "embryology", or "development". The machines which instruct the differentiating cells how to switch on and switch off genes at the appropriate time in the growth process are of molecular scale. A biological cell can be viewed as a molecular scale city, with millions of molecular inhabitants all organized into one functioning whole.

I would like to see the creation of a new branch of science that I call "Artificial Embryology", which would aim to mimic the same process that nature employs to grow dinosaurs or gnats from single fertilized eggs. Scientists and engineers would need to understand how nature does it in far greater detail than is known at the beginning of the century. But, as the molecular biologists are discovering all there is to know, more or less, about certain single-celled bacteria, many of these scientists are changing specialties towards studying how multi-cellular creatures are built. Embryology is now a hot research topic, so we can expect a steady flow of discoveries in this domain over the coming decades.

Eventually, I expect to see the creation of what I call "Embryofacture" (embryological manufacture), i.e. using artificial embryological techniques to manufacture human scale products from the nano scale. Instead of needing a complex nanoscale infrastructure using nanots as described earlier, one would need a complex timing control system which decides when particular genes in the DNA (or its humanly designed equivalent) switch on and off when stimulated by certain molecular signals in their inter and intra cellular environment.

Designing such a complex control system top-down from scratch will probably be beyond the abilities of human scientists, so a more likely approach will be to use an "evolutionary engineering" approach. The mapping between an artificial "DNA" sequence of molecular based growth instructions and the final 3D product, whether living or not, is probably impossible to predict due to its complexity, so probably the only method remaining is the one nature uses to learn how to "embryofacture" its creatures, namely evolution.


Evolutionary Engineering

An evolutionary engineering approach to embryofacture might work in the following way. One begins with a zillion random molecular "artificial DNA" strings, which translate themselves into blob-like 3D molecular structures. Predesigned molecular scale nanots then move in and measure how closely the actual blob resembles the shape or function of the microproduct that is desired. Those blobs that get a higher score, will see their corresponding artificial DNAs survive and have more copies (children) made of them in the next generation.

The less functional blobs are killed off, Darwinian style, thus generating a kind of "survival of the fittest" strategy. The child-DNAs are then "mutated" slightly (i.e. the chemical instructions contained in the artificial DNA are modified somewhat). Occasionally, a mutated child-DNA will create a 'fitter" (more performant) blob than its parent. By cycling through this loop many times, a desired artificial DNA is formed which grows the desired shape or function of its blob. The result is a molecular product which performs some useful function.


Self Assembly

However, evolving single components is not enough. These components then need to be complementary in shape so that they can "self assemble", i.e. fit together like jigsaw-puzzle pieces to form a greater functioning whole. Viruses form this way. Portions of DNA (genes) code for the construction of viral components. Once they are built, they click together to form whole viruses.

So the component parts need to have lock and key shape complementarities. They need to have the capacity to self assemble, simply by bumping into each other (as occurs frequently in the chaotic motion at the molecular scale).

This notion of self-assembly is very important when it comes to building an asteroid sized artilect, or even one of human size. A human sized artilect (or a human sized anything) contains trillions of trillions of molecules. To build a human sized artilect would require that all the atoms of that artilect, all trillions of trillions of them, be placed with atomic precision at just the right places. Such an artilect I believe, would have to build itself through an embryological process. It would have to embryofacture itself. So how would such an artilect be built and designed in the first place?

Initially the first (very primitive) artilects would need to be built by evolutionary engineers (people like me). Perhaps they should be called "embryofacturers" or "embryological engineers". At first, the evolved 3D molecular structures could be assembled piece by piece into a larger 3D structure. Later, more sophisticated artilects could be built which perform their own evolution (perhaps within their own bodies) and make their own decisions at electronic speeds.

Of course, human beings would need to abandon all hope of fully understanding how these evolving, "Darwinian artilects" would develop. Their artilectual structure and functioning would be so complex and change so fast, that full human understanding of it all, would be totally impractical.

Such human ignorance will later prove to be powerful ideological fuel to the Terrans, who will argue that the very nature of artilect construction (i.e. Darwinian, self-assembling embryofacture) makes artilect behavior inherently unpredictable and hence potentially very dangerous for human beings. This point will be discussed again at length in Ch.4, which presents the case of the Terrans.


Putting the Technologies together

So let me try to summarize a bit here. After all, the point of this chapter has been to introduce those technologies which will enable the construction of artilects this century.

So far, the vision presented in this chapter is that of an asteroid sized artilect containing 1040 atoms or bits of information, using nanotech based, self-assembling, embryofactured, heatless, reversible, 3D, computer circuitry, thinking at least a million times faster (and probably a lot faster) than humans. It will contain a huge number of sensors attached to its surface, with enormous memory capacities etc. But there's more.


Quantum Computing Artilects

These artilects will be using molecular and atomic sized components, so these components will be subject to the laws of quantum mechanics. Recently the new field of "quantum computing" has become popular, as theoretical and experimental physicists compete with each other to dream up new ways to "quantum compute" and to implement these ideas in real hardware.

I hesitate to describe what quantum computing is to the general public. It is very counter-intuitive and difficult to grasp. If the following few paragraphs sound like gobbledygook to you, then just flip to the next topic. In a sense no one really understands quantum theory. It seems like a bunch of mathematical recipes that give good numerical answers to problems, but seems totally unintuitive conceptually.

Atoms behave in the weirdest ways, quite unlike what human beings are accustomed to at our scale of things. Quantum mechanics IS truly weird and abstract. It is a branch of mathematical physics which gives the probabilities of certain measurement results when atomic scale systems interact with human scale measurement devices. In classical mechanics, the state of a physical system is distinct, i.e. it has given values, e.g. its velocity at a given moment is V, its position is X, its kinetic energy is K, etc. In quantum mechanics, things are more abstract.

The state of a quantum system is represented by an abstract mathematical sum of numbers, where each number is associated with a measurement result if a measurement is performed. This summing and linear weighting of states is called a "superposition", and is the conceptual heart of quantum mechanics. Don't fret too much if you don't understand this. It is not essential to the understanding of this chapter.

It is this superposition that is the great feature of quantum computing. The superposition evolves over time, in a sense performing many calculations at once, whereas a classical computer can only do one thing at a time.

In classical computing, the state of a register (a storage chain of bits) is a definite string of 0s and 1s (e.g. 0011011101001). In a quantum computing register, the state is a superposition of a huge number of possible classical register states. For example, if there are N bits in the register, then there are 2N possible different classical register states (e.g. if N = 3, there are 8 different classical states, 000, 001, 010, 011, 100, 101, 110, 111). If N is large, then 2N is huge.

The enormous advantage of quantum computing is that this huge number of classical states gets treated as though it is just one (superimposed) state, one quantum state that the quantum system can handle. In order to perform a calculation with classical computing it is often necessary to test each classical register state one at a time, for all possible states. This is a slow business, and as N increases, the number of tests rises exponentially (i.e. like 2, 4, 8, 16, 32, 64 etc).

With quantum computing however, only one test needs to be done, because in a sense, all possible classical states are blended in together in the quantum register state. Quantum computing is potentially incredibly more efficient that classical computing. It is therefore not surprising that many physicists around the world are now racing each other to see who can build the next most performant quantum computer.

Since the artilect will be built with atomic scale components, it will need to function as a quantum computer. Since quantum computers are more efficient than clas