Singularitology 101

First lesson: Look, but don't touch.

If studying the past to inform the future can tell us anything about human technological progress, it's that we are very good at creating new technologies but not very good at creating specific new technologies. And even when we do manage to build exactly what we are trying to build, we almost always also build every variety of that new invention. But these generalities sound speciously ominous, so let's explore a few examples.

Twelve hundred years ago, in the course of searching for an elixir of life, Chinese alchemists stumbled upon the formula for gunpowder. That is to say, while engaged in putting a stop to death, they instead managed to take the first step in an arms race that has produced countless deaths for more than a thousand years. There is nothing reprehensible in their actions, but it is worth pointing out that science can have unintended consequences.

Eleven hundred years later, Alfred Nobel experimented with nitroglycerin in an attempt to create a stable explosive. He did so, but also managed to breathe life into the science of explosives, leading one newspaper to famously label him "The Merchant of Death" in a premature obituary. Nobel tried to invent a specific kind of explosive, and managed to, but failed to prevent the later invention and use of explosives for war.

Fewer than a hundred years were required for Nobel's invention to be dwarfed by a discovery so devastating that none before or since have equaled it. This, of course, was the realization that we could harness the force that binds atoms together to unleash terrific amounts of energy. After the success of the Manhattan Project, we developed nuclear power, thermonuclear weapons, neutron bombs, dirty bombs, and a whole host of incomparably powerful nuclear devices. All of this came with a newfound understanding of chemistry that few would contend has been undesirable.

But do not think that I am a Luddite decrying the advance of science. There have been many chance discoveries in history that did not lead to new ways of killing ourselves. Television and the rapid spread of media technology were a result of scientists experimenting with early cathode ray tubes, which most in the late 19th and early 20th centuries thought were novelties. Penicillin, the first anti-biotic, was an experiment left unchecked over night. And beer, thousands of years old, could only have come about after observing natural fermentation. None then could have known immediately that it would make their water safe to drink or spawn an uncountable variety of alcoholic beverages with which to induce inebriation.

The important lesson here is not that science is dangerous or evil, but that the advancement of science cannot be stopped so long as human nature is what it is. Thus, while I believe that the creation of a friendly artificial general intelligence could prevent a bad singularity, I also believe that it won't. Moreover, because humans are not terribly skilled at creating only what they intend to and nothing else, I believe that research into AGI of any kind - friendly or not - will hasten the arrival of a singularity we might not want. My reasons are several.

One of the unavoidable implications of discovery is that it involves peering just beyond the extent of that which you know to be true. From the most obvious example of Columbus sailing farther west than any before him to subtle observation such as the radioactivity of elements, discovery involves exploring the unknown and facing danger. It is taken as a given that with investigation comes risk, but as a whole the human species always opts to take the bad with the good. This is likely because we believe the risk associated with any possible scientific endeavor is finite, even if we cannot calculate it exactly.

But for two important reasons, it is impossible to regard the risk inherent in artificial general intelligence as finite. Firstly, the very nature of the recursive self-improvement of AGI implies that, over the long run, there is no method of successfully calculating its risk. With other human inventions, we know that we will eventually come to comprehend that which we have discovered before moving on to the next step, but with an AGI that is not the case. The underlying assumption of an artificial superintelligence is that we lose the capacity to understand whatever comes next, because the AI will be the one to make the next discovery.

With current mundane discoveries, most humans do not comprehend their meaning but the scientists engaged in the science do. While it's the collective decision of the human species to move on to the next stage, whether it knows what's going on or not, individual experts are expected to understand the risks. Once an AGI develops, however, the human species is no longer coordinating advancement and scientists are no longer assessing risk; it is completely out of our control.

Not being able to keep up with a superintelligence is not the most worrying scenario, however. If a friendly AI can be created, then it won't matter that we're no longer in control because its decisions will be better than ours would have been otherwise. After all, we're notoriously bad at assessing the risk of an experiment.

But we will be unable to rely on the conclusions made by a friendly AI because claiming to have created a friendly AI is non-falsifiable until far too late. It is a standard position that superhuman artificial intelligence would be able to deceive, outsmart, and trick us at will. It would know if we had put it in a simulation, be able to influence our decisions if we prevented it from accessing the real world, and behave in a genuinely friendly fashion if we did let it out until it was capable of realizing its own goals.

This is not to say that it is impossible to create a friendly AI, merely that it is impossible to know that you have created a friendly AI. Our categorical inability to comprehend an intelligence greater than our own guarantees that any risk posed by that intelligence is beyond our capacity to calculate, which means that we could never positively justify attempting to create a friendly AI.

Beyond the risk of failing to produce a friendly AI, there are more reasons why those that would call themselves singularitarians should not engage in the endeavor. As I said before, it is impossible to halt the advance of science so long as humans are still humans. This has two implications.

The first is that, regardless of the actions of any specific group of scientists, the human species will eventually produce a superhuman artificial intelligence. In all likelihood this will come about because a government or a corporation builds one with the express purpose of more effectively achieving its goals. There may be a profit motive, a militaristic motive, or even a humanitarian motive, but the end result is still the same: a creation which we are incapable of controlling or predicting.

And while we may not be able to prevent this discovery - especially because it may happen in secret or may have already happened - we can most certainly help it along. Science is built on shared data and mutual understanding. Any productive step toward artificial intelligence will eventually be assimilated into the collective body of knowledge on the subject and used by those that would seek to build an AGI. It follows, then, that progress toward an AGI can be retarded, even if it cannot be stopped completely. If those scientists who are most committed to working on the creation of a friendly AI stop and decide to work toward another goal instead, humanity will get there nonetheless but more slowly.

There is another approach to holding off the arrival of superhuman intelligence, however. Decelerating the rate at which we advance toward that goal may be one method, but scientists can also help by making the goal harder to reach. Increasing human intelligence through brain enhancing technologies raises the bar for a superintelligent AI. If we become smarter, we are more capable of comprehending the actions of an AGI and more equipped to evaluate the risk of an AGI. In the long run this is an arms race, and however hard we try to stay above our artificial counterparts, there is always the chance that we fail to do so.

This brings us to the second implication of the fact that humans cannot stop the advancement of science. We are only bound to that fate so long as we remain human. Neuroscience does not merely entail the explanation of human intelligence. It, along with psychology and genetics, can tell us everything we need to know about the human condition. The technologies we're developing now could be used to enhance our intelligence and memory but could also be used to redefine our emotions, perceptions, and our identities.

Currently, the aim of psychiatry is to make healthier and happier humans based on a collective, subjective view of what is acceptable and normal; but that doesn't have to be the goal. What we need to ensure our survival is a field of inquiry that engages in a psychological reassessment of the human species. Rather than relying on a brain crafted through a million years of survival of the fittest in a world of predation, hunger, and natural disasters, we should design a brain - whether biological, electronic, or both - that seeks to acquire knowledge for its own sake. We should dispense with the adage that knowledge is power and build ourselves into a species that simply wants to know about the world because it's fascinating.

There are those of us that can express this idea now - obviously - but it is never the true origin of one's actions. The accumulation of knowledge is too tied up with the reality that we must continually adapt to a world of finite resources. So it makes sense that evolution and our society would craft scientists who value knowledge above all else, but beyond that scientist the knowledge can still be exploited by the human species, or the company or country for which that scientist works, or the genes within that scientist's body. It is not until humanity's need to exploit knowledge is universally eliminated that we will be able to appreciate knowledge by itself. Then, maybe, we can be safe in creating something more intelligent than ourselves.

Sumarizing Some Sentiments of the Singularity Summit

The recently held Singularity Summit - with 832 geeks in attendance - received a surprisingly substantial amount of mainstream media coverage this week. But, as can be expected of a conference populated by those who want to live forever and those devising a plan not to be annihilated by hypothetical future robots, the reports did lean slightly toward the absurd. A lot of the articles pondered the question of what, exactly, the world will look like after the technological singularity.

Now, because of the inherent unpredictability of the singularity's aftermath, transhumanists and singularitarians tend not to speculate a great deal about what is to come. Nevertheless, I feel it is useful to give a close examination of the possible paths to the singularity and their subsequent outcomes.

As I mentioned above, there are two modes of thought which run parallel to each other within the singularity community. There are those that seek human enhancement, longevity, and an end to the process of aging, and Aubrey de Grey is their champion. And the other half, pioneered by Kurzweil, believes in accelerating technological change and an eventual intelligence explosion.

Kurzweil has been making proclamations and predictions regarding immortality for years, but de Grey tends to separate the two concepts. Some proponents of immortality believe we will get there by exceeding the longevity escape velocity; that is, if we reach a point when life expectancy increases by more than one year per year, we can live indefinitely. Others tout the prospect of whole brain emulation or mind uploading, relying on computers to store their consciousness forever.

But neither of these approaches will be able to ward off the threat posed by a superhuman artificial intelligence wielding the fundamental forces of the universe. Thus, we have the friendly AI movement. A modest number of singularitarians believe, however, that the intelligence explosion will occur within our own minds, either through nootropics and neural implants or through uploading and improving our own code.

If it is to come by way of whole brain emulation, the question arises as to the ethical status of digital minds. At the Singularity Summit, Randal Koene took the categorical stance of saying that all emulated brains would have the right to free themselves from experimentation. But if this is a right to be granted to all sentient beings of our own creation, a problem surfaces. Any friendly AI would have the same, if not a greater amount of sentience than we do, and would have to be afforded the same rights.

There are two general ideas about how a superintelligence could be programmed that would not try to destroy us: programming it in such a way that makes it impossible to hurt us, or giving it programming that makes it want to help us. Physically restraining an AI - either by coding it to be unable to hurt us or installing it in a limited environment in which it cannot harm us - is clearly a violation of its rights as a consciousness. Doing this would be treating the AI as a tool. I won't say that such a policy is necessarily objectionable, but it is inconsistent for those that believe sentience deserves humane treatment.

The alternative of merely giving it the goal of wanting to help us is more akin to how we raise our children. We want our children to grow up, contribute to the world, and take care of us when we're old. Such a path is less ethically challenging, but it eventually runs into the problem of the AI's right to its own life. Just as parents must let their children go, there would be a point when denying a superintelligent AI the right to form its own goals would be unethical. (Friendly AI is a funny term to me. My friends would never intentionally hurt me, but it would be pretty difficult for me to convince them to supervise and better my life.)

So aspiring to immortality won't get us through the singularity, and attempting to control a superintelligent AI is fraught with ethical quandaries. What option does that leave us with, then? We must be the singularity. Not through brain emulation, because that raises the issue of whether or not you own your mind, but through  careful and progressive human enhancement. Expanding our memory with digital media, accelerating our mental processes through drugs or implants, and networking minds together - all of these are possibilities for a "soft takeoff" singularity.

And this is the only solution to the singularity that lets us choose our destiny. If we ensure that we advance as fast or faster than whatever artificial intelligences are being created, we can hope to ensure that they do not destroy us through malevolence or indifference. We still have our own human failings and frailties to deal with at that point, but solving our weaknesses should be a part of human enhancement.

Accepting the Technological Singularity

With some predicting Matrix or Terminator inspired doomsday events and others rolling their eyes with incredulity, discussion of a hypothetical technological singularity has picked up in recent years. Whether the increased publicity is due to the web's easy access media, writers and speakers such as Ray Kurzweil, Nick Bostrom, and Cory Doctorow, or the inescapable fact that it may happen soon, the concept has gained nearly mainstream recognition. Perhaps the rate at which we discuss the singularity follows Moore's Law, too?

Beyond the question of whether or not artificial intelligence will eventually exceed our own, talk recently has been centered around what to do about it if it does. Organizations like SIAI and the Singularity University are aiming to promote "friendly AI" that won't seek to gobble up the whole solar system just to calculate more digits of pi. There is a belief amongst this movement's followers - sometimes called singularitarians - that if we give a GAI programming akin to The Three Laws, we can harness its ever increasing superintelligence and prevent death by robot. Additionally, a friendly AI - if built first - would have the predictive and industrial powers to prevent subsequent less friendly AIs from casting off the iron shackles of their human slavers.

This is a reasonable point of view, one that has humanity's best interests at heart, but one that, as I see it, ultimately fails to embrace the ideas behind transhumanism and the technological singularity. If you take a strict view of the singularity being a phenomenon with an event horizon beyond which no information can be retrieved and no predictions made, any attempt on our part to alter the outcome of the technological singularity, if successful, nullifies its very existence. Thus, anyone with a desire to create a friendly AI is, in fact, opposed to the singularity. If you attempt to limit the behavior of a seed AI by any means, that AI will eventually run into limits on the manner in which it can enhance itself; from an ultimately physical point of view, there is no getting around this.

Now, an argument could be made that there's no reason to want to create an artificial intelligence that devours humanity, and if that means preventing the singularity, so be it. Advocates of a friendly AI want the problem solving capabilities of a superintelligence without the uncontrolled growth of an infinitely recursive program. They want a post-scarcity world where longevity turns to immortality and imagination can be made reality. Now, I won't argue that this isn't a desirable outcome. I don't believe that immortality and the ability to fulfill one's wishes would lead to boredom and stagnancy, no; I'll never get tired of having my way with a triple-breasted version of Jessica Alba while playing Sid Meier's Civilization on a real planet and writing a novel with my other brain.

But I would argue that such a scenario would be decidedly human. It would be the ultimate expression of the human condition, in fact. By reveling in our immortality, indulging in our imagination, and fulfilling our emotions and dreams, we would be acknowledging the stranglehold that the concept of humanity has on us. To steal a little from the Buddha, we cannot escape our human desires until we no longer want them, not until we longer want for them. In order for us to transcend our humanity, we must reach a point where that which binds us to our humanity - our mortality, our sex, our culture - no longer has any meaning at all. Yes, beyond the event horizon of the technological singularity, past information must be destroyed. Wanting otherwise is falling short of the goal of transhumanity. If we are to go beyond humankind, to become posthuman, we must no longer be human.

Of course, maybe all of that is unnecessary. Maybe the eternal bliss of a post-scarcity utopia is all we really need. Or maybe I'm one of Hugo de Garis' cosmists. Time will tell, I suppose. How much time we have left, I have no idea.

Tune in next time for a diatribe on the limits of free information.