« February 26, 2017 | Main | March 16, 2017 »

Thursday, March 9, 2017

Reading List: The Singularity

Awret, Uziel, ed. The Singularity. Exeter, UK: Imprint Academic, 2016. ISBN 978-1-845409-07-4.
For more than half a century, the prospect of a technological singularity has been part of the intellectual landscape of those envisioning the future. In 1965, in a paper titled “Speculations Concerning the First Ultraintelligent Machine” statistician I. J. Good wrote,

Let an ultra-intelligent machine be defined as a machine that can far surpass all of the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an “intelligence explosion”, and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make.

(The idea of a runaway increase in intelligence had been discussed earlier, notably by Robert A. Heinlein in a 1952 essay titled “Where To?”) Discussion of an intelligence explosion and/or technological singularity was largely confined to science fiction and the more speculatively inclined among those trying to foresee the future, largely because the prerequisite—building machines which were more intelligent than humans—seemed such a distant prospect, especially as the initially optimistic claims of workers in the field of artificial intelligence gave way to disappointment.

Over all those decades, however, the exponential growth in computing power available at constant cost continued. The funny thing about continued exponential growth is that it doesn't matter what fixed level you're aiming for: the exponential will eventually exceed it, and probably a lot sooner than most people expect. By the 1990s, it was clear just how far the growth in computing power and storage had come, and that there were no technological barriers on the horizon likely to impede continued growth for decades to come. People started to draw straight lines on semi-log paper and discovered that, depending upon how you evaluate the computing capacity of the human brain (a complicated and controversial question), the computing power of a machine with a cost comparable to a present-day personal computer would cross the human brain threshold sometime in the twenty-first century. There seemed to be a limited number of alternative outcomes.

  1. Progress in computing comes to a halt before reaching parity with human brain power, due to technological limits, economics (inability to afford the new technologies required, or lack of applications to fund the intermediate steps), or intervention by authority (for example, regulation motivated by a desire to avoid the risks and displacement due to super-human intelligence).
  2. Computing continues to advance, but we find that the human brain is either far more complicated than we believed it to be, or that something is going on in there which cannot be modelled or simulated by a deterministic computational process. The goal of human-level artificial intelligence recedes into the distant future.
  3. Blooie! Human level machine intelligence is achieved, successive generations of machine intelligences run away to approach the physical limits of computation, and before long machine intelligence exceeds that of humans to the degree humans surpass the intelligence of mice (or maybe insects).

Now, the thing about this is that many people will dismiss such speculation as science fiction having nothing to do with the “real world” they inhabit. But there's no more conservative form of forecasting than observing a trend which has been in existence for a long time (in the case of growth in computing power, more than a century, spanning multiple generations of very different hardware and technologies), and continuing to extrapolate it into the future and then ask, “What happens then?” When you go through this exercise and an answer pops out which seems to indicate that within the lives of many people now living, an event completely unprecedented in the history of our species—the emergence of an intelligence which far surpasses that of humans—might happen, the prospects and consequences bear some serious consideration.

The present book, based upon two special issues of the Journal of Consciousness Studies, attempts to examine the probability, nature, and consequences of a singularity from a variety of intellectual disciplines and viewpoints. The volume begins with an essay by philosopher David Chalmers originally published in 2010: “The Singularity: a Philosophical Analysis”, which attempts to trace various paths to a singularity and evaluate their probability. Chalmers does not attempt to estimate the time at which a singularity may occur—he argues that if it happens any time within the next few centuries, it will be an epochal event in human history which is worth thinking about today. Chalmers contends that the argument for artificial intelligence (AI) is robust because there appear to be multiple paths by which we could get there, and hence AI does not depend upon a fragile chain of technological assumptions which might break at any point in the future. We could, for example, continue to increase the performance and storage capacity of our computers, to such an extent that the “deep learning” techniques already used in computing applications, combined with access to a vast amount of digital data on the Internet, may cross the line of human intelligence. Or, we may continue our progress in reverse-engineering the microstructure of the human brain and apply our ever-growing computing power to emulating it at a low level (this scenario is discussed in detail in Robin Hanson's The Age of Em [September 2016]). Or, since human intelligence was produced by the process of evolution, we might set our supercomputers to simulate evolution itself (which we're already doing to some extent with genetic algorithms) in order to evolve super-human artificial intelligence (not only would computer-simulated evolution run much faster than biological evolution, it would not be random, but rather directed toward desired results, much like selective breeding of plants or livestock).

Regardless of the path or paths taken, the outcomes will be one of the three discussed above: either a singularity or no singularity. Assume, arguendo, that the singularity occurs, whether before 2050 as some optimists project or many decades later. What will it be like? Will it be good or bad? Chalmers writes,

I take it for granted that there are potential good and bad aspects to an intelligence explosion. For example, ending disease and poverty would be good. Destroying all sentient life would be bad. The subjugation of humans by machines would be at least subjectively bad.

…well, at least in the eyes of the humans. If there is a singularity in our future, how might we act to maximise the good consequences and avoid the bad outcomes? Can we design our intellectual successors (and bear in mind that we will design only the first generation: each subsequent generation will be designed by the machines which preceded it) to share human values and morality? Can we ensure they are “friendly” to humans and not malevolent (or, perhaps, indifferent, just as humans do not take into account the consequences for ant colonies and bacteria living in the soil upon which buildings are constructed?) And just what are “human values and morality” and “friendly behaviour” anyway, given that we have been slaughtering one another for millennia in disputes over such issues? Can we impose safeguards to prevent the artificial intelligence from “escaping” into the world? What is the likelihood we could prevent such a super-being from persuading us to let it loose, given that it thinks thousands or millions of times faster than we, has access to all of human written knowledge, and the ability to model and simulate the effects of its arguments? Is turning off an AI murder, or terminating the simulation of an AI society genocide? Is it moral to confine an AI to what amounts to a sensory deprivation chamber, or in what amounts to solitary confinement, or to deceive it about the nature of the world outside its computing environment?

What will become of humans in a post-singularity world? Given that our species is the only survivor of genus Homo, history is not encouraging, and the gap between human intelligence and that of post-singularity AIs is likely to be orders of magnitude greater than that between modern humans and the great apes. Will these super-intelligent AIs have consciousness and self-awareness, or will they be philosophical zombies: able to mimic the behaviour of a conscious being but devoid of any internal sentience? What does that even mean, and how can you be sure other humans you encounter aren't zombies? Are you really all that sure about yourself? Are the qualia of machines not constrained?

Perhaps the human destiny is to merge with our mind children, either by enhancing human cognition, senses, and memory through implants in our brain, or by uploading our biological brains into a different computing substrate entirely, whether by emulation at a low level (for example, simulating neuron by neuron at the level of synapses and neurotransmitters), or at a higher, functional level based upon an understanding of the operation of the brain gleaned by analysis by AIs. If you upload your brain into a computer, is the upload conscious? Is it you? Consider the following thought experiment: replace each biological neuron of your brain, one by one, with a machine replacement which interacts with its neighbours precisely as the original meat neuron did. Do you cease to be you when one neuron is replaced? When a hundred are replaced? A billion? Half of your brain? The whole thing? Does your consciousness slowly fade into zombie existence as the biological fraction of your brain declines toward zero? If so, what is magic about biology, anyway? Isn't arguing that there's something about the biological substrate which uniquely endows it with consciousness as improbable as the discredited theory of vitalism, which contended that living things had properties which could not be explained by physics and chemistry?

Now let's consider another kind of uploading. Instead of incremental replacement of the brain, suppose an anæsthetised human's brain is destructively scanned, perhaps by molecular-scale robots, and its structure transferred to a computer, which will then emulate it precisely as the incrementally replaced brain in the previous example. When the process is done, the original brain is a puddle of goo and the human is dead, but the computer emulation now has all of the memories, life experience, and ability to interact as its progenitor. But is it the same person? Did the consciousness and perception of identity somehow transfer from the brain to the computer? Or will the computer emulation mourn its now departed biological precursor, as it contemplates its own immortality? What if the scanning process isn't destructive? When it's done, BioDave wakes up and makes the acquaintance of DigiDave, who shares his entire life up to the point of uploading. Certainly the two must be considered distinct individuals, as are identical twins whose histories diverged in the womb, right? Does DigiDave have rights in the property of BioDave? “Dave's not here”? Wait—we're both here! Now what?

Or, what about somebody today who, in the sure and certain hope of the Resurrection to eternal life opts to have their brain cryonically preserved moments after clinical death is pronounced. After the singularity, the decedent's brain is scanned (in this case it's irrelevant whether or not the scan is destructive), and uploaded to a computer, which starts to run an emulation of it. Will the person's identity and consciousness be preserved, or will it be a new person with the same memories and life experiences? Will it matter?

Deep questions, these. The book presents Chalmers' paper as a “target essay”, and then invites contributors in twenty-six chapters to discuss the issues raised. A concluding essay by Chalmers replies to the essays and defends his arguments against objections to them by their authors. The essays, and their authors, are all over the map. One author strikes this reader as a confidence man and another a crackpot—and these are two of the more interesting contributions to the volume. Nine chapters are by academic philosophers, and are mostly what you might expect: word games masquerading as profound thought, with an admixture of ad hominem argument, including one chapter which descends into Freudian pseudo-scientific analysis of Chalmers' motives and says that he “never leaps to conclusions; he oozes to conclusions”.

Perhaps these are questions philosophers are ill-suited to ponder. Unlike questions of the nature of knowledge, how to live a good life, the origins of morality, and all of the other diffuse gruel about which philosophers have been arguing since societies became sufficiently wealthy to indulge in them, without any notable resolution in more than two millennia, the issues posed by a singularity have answers. Either the singularity will occur or it won't. If it does, it will either result in the extinction of the human species (or its reduction to irrelevance), or it won't. AIs, if and when they come into existence, will either be conscious, self-aware, and endowed with free will, or they won't. They will either share the values and morality of their progenitors or they won't. It will either be possible for humans to upload their brains to a digital substrate, or it won't. These uploads will either be conscious, or they'll be zombies. If they're conscious, they'll either continue the identity and life experience of the pre-upload humans, or they won't. These are objective questions which can be settled by experiment. You get the sense that philosophers dislike experiments—they're a risk to job security disputing questions their ancestors have been puzzling over at least since Athens.

Some authors dispute the probability of a singularity and argue that the complexity of the human brain has been vastly underestimated. Others contend there is a distinction between computational power and the ability to design, and consequently exponential growth in computing may not produce the ability to design super-intelligence. Still another chapter dismisses the evolutionary argument through evidence that the scope and time scale of terrestrial evolution is computationally intractable into the distant future even if computing power continues to grow at the rate of the last century. There is even a case made that the feasibility of a singularity makes the probability that we're living, not in a top-level physical universe, but in a simulation run by post-singularity super-intelligences, overwhelming, and that they may be motivated to turn off our simulation before we reach our own singularity, which may threaten them.

This is all very much a mixed bag. There are a multitude of Big Questions, but very few Big Answers among the 438 pages of philosopher word salad. I find my reaction similar to that of David Hume, who wrote in 1748:

If we take in our hand any volume of divinity or school metaphysics, for instance, let us ask, Does it contain any abstract reasoning containing quantity or number? No. Does it contain any experimental reasoning concerning matter of fact and existence? No. Commit it then to the flames, for it can contain nothing but sophistry and illusion.

I don't burn books (it's некультурный and expensive when you read them on an iPad), but you'll probably learn as much pondering the questions posed here on your own and in discussions with friends as from the scholarly contributions in these essays. The copy editing is mediocre, with some eminent authors stumbling over the humble apostrophe. The Kindle edition cites cross-references by page number, which are useless since the electronic edition does not include page numbers. There is no index.

Posted at 01:22 Permalink