Why the Singularity is Nearer

Why super-human intelligence might arrive within the next five years DRAFT, June 8, 2008 In his well-known books on the topic, Ray Kurzweil outlines a vision in which greater-than-human intelligence arrives somewhere in the 2020s or 2030s. Following Hans Moravec, he constructs graphs so that the processing power of a $1000 machine is compared against the estimated processing power in human intelligence (or the estimated processing power to reproduce human intelligence), and a time of cross-over is computed. These graphs are actually one of my favorite parts of the works of Moravec and Kurzweil; I had for a while carried a copy of Moravec's graphs around in my laptop bag, because I liked them so much. Perhaps a better approach to estimating the arrival date of super-human intelligence, however, is to look at Moravec's metaphor of a rising tide (or water level) of computational ability, starting in the lowlands of arithmetic calculations, and heading ever higher for such peaks as natural language processing. It seems clear that the existence of a computer with super-human processing abilities does not immediately result in a general super-human intelligence, for both theoretical and empirical reasons. It has been noted that very expensive machines, such as supercomputers, would probably not be used to perform activities such as grocery shopping or movie reviewing. Indeed, the current crop of supercomputers, some close to petaflops (and hence likely also exceeding human abilities) in power, are not used for such tasks. But is the $1000 target the correct target? Granted, household computers have typically been on the order of that price, probably due to some function of usefulness and longevity, and prices of household computers seem to generally be dropping. But whether household computers are the correct sort to look at for predicting the arrival of super-human intelligence is unclear. Also unclear is whether more capable (and hence more useful) computers might not be more desired even at home, and hence fetch higher prices even there. An alternative place to look would be in the workplace. Even traditional customer-service roles such as grocery checkout are starting to be replaced now, to some degree, by computers, but of course no one would argue that the power of such computers as are used for these tasks now approaches that of general human intelligence. On the other hand, such computers are not the most powerful used in business, so that a comparison might be made between what business is willing to buy and what the computer market can provide. As a very rough estimate, the amount a company might be willing to pay for a computer of roughly human intelligence might be based on the salary of the human. If a computer is estimated to last three years, and to work about three times as long per day as a human, then an upper bound of about nine times the yearly salary of a human worker might be reasonably allocated toward a computer capable of performing that person's job. Whether the computer replacement would need anything approaching general human intelligence to perform the job would depend on the job, but this level could be used as a reasonable bound. Calculating with this bound, and a salary of $50,000, we might guess that,subject to various caveats, business would probably be willing to pay about $450,000 for a computer of roughly human capabilities. Granted, computers have various overheads and might not always be fully utilized, but these caveats seem to apply also to human workers. Whether a computer of a given capability were programmable to perform some specific task is also unclear, but given the variety of jobs currently performed by humans, it seems likely that at least some jobs will be easily programmed for, and that the business of replacing a variety of human jobs with computers will probably lead to improvement in the general technologies for doing so. According to top500.org, the last-place supercomputer is currently estimated to perform at about 6 teraflops, which might be estimated as a factor of ten to twenty away from human-level processing capabilities. If this computer were purchased for on the order of a million dollars (with a factor of ten error in price shifting date estimates by about three years), then business should be expected to show interest in replacing a wide variety of jobs by computers sooner than four to eight years from now. Given the variety in the sorts of jobs that could possibly be replaced, some jobs might disappear in a year, some in five years, but at least most long-term positions should be amenable to computer replacement within eight years. It is interesting to consider what will probably occur after such replacement starts. If the trends in diminishing price for processing power continue to hold, it can be estimated that each year will bring something like a halving in the level of salary at which computer replacements will start to be feasible. If one year sees replacement of $50,000 per year jobs, then the next would likely see replacements in $25,000 per year jobs, and the year after, replacements of $12,500 per year jobs. Soon, computers would be found in almost all areas of the workforce, commensurate with a human population of hundreds of millions or billions. At this point, it might be said that general human-level intelligence in computers is not only achievable, but is, in fact, commonplace, and that the singularity had long since passed. If we use the replacement of $50,000 per year jobs as a threshold for the singularity, then we get the earlier-noted estimate of some date no later than the next four to eight years, which is considerably sooner than the 2020s or 2030s estimated in various works. The difference in approach taken here is simply looking at costs associated with salary for employment stretching over three years or more, rather than looking at the costs associated with the purchase of a typical modern household workstation. This same observation can be applied to other methods of calculation as well, perhaps more reliable than the one outlined above. To take another approach, start with the earlier-noted 2020 or 2030 estimates for $1000 computers with human-level processing capabilities, and assume that the price trends in computer technology remain roughly the same. To be conservative (as we are subtracting time), assume a relatively short period of price halving, on the order of a year, and we see that $450,000 computers with human-level abilities would probably arrive sometime between 2011 and 2021, which is considerably sooner than most estimates for human-level intelligence arising. To look in the other direction, it seems that replacement or partial replacement of more highly paid jobs should occur even sooner. News stories of robotically assisted surgeries are fairly common now, and, although humans currently direct the robots in their actions, this need not always be so.* For certain surgeries, autonomy on the part of the robot might be especially preferred, just as some autonomy for the Martian rovers was desirable. With increasing autonomy, more aspects of the surgeon's job are replaced, and hence funds might be allocated more toward better autonomous robots than toward increasing the population of human surgeons. (As a point of reference, note that expert systems are already considered to in some cases be superior to humans in the area of diagnosis, so that this sort of higher-level decisionmaking is already not the exclusive domain of humans.) I once overheard two classmates of mine talking, when studying many years ago, about the virtues of a physics degree; the conversation was rather short, as one asked the other what he would do with a physics degree after computers had become better at doing physics than humans are. Indeed, it might be equally asked what would be done with a medical degree once computers are better than humans at medicine, and what might be done with a computer science degree once computers become the better programmers. At least in the case of computer programmers, with degree options falling within the bound of four years, it might be expected that some sort of employment might be obtained just prior to the replacement of human jobs with computer. But what of starting a medical education now, which in the United States at least can stretch on for a decade or so? Does that make sense? Even if uploading of human brains into computers becomes possible, or integration of the two, would it make more sense to learn such skills by attending a lengthy and expensive degree program rather than by eventually transferring information over electronic interface? The above hints at the importance of the question of whether an eventual merging or transfer of computer and human abilities is ever possible. If so, as in the case of uploading or neural interfaces, it might mean the acquisition, perhaps into a different place from our current brains, of new abilities and experiences that humans have never before had, so that humans and computers become merged. If this does not happen, or is so difficult or expensive that it only happens rarely, then perhaps humans will live very differently from how we live today; we would no longer be on the forefront of development, of technology, or of decision-making, but those roles would rather be left to our non-human descendents. Such, then, might be the importance, if a rather selfish one, of working now on methods of creating high-bandwidth links between human brains and computers. But whether we succeed in this or not, it seems that we can expect super-human general intelligence to arrive in the very near future. * In April, 2008, a supercomputer in Austin, Texas performed laser surgery on a dog in Houston without intervention by a human surgeon. (Updated June 9, 2008) (This work is (C) 2008 David Hilvert <dhilvert@gmail.com>, and may be distributed under the terms of the Creative Commons Attribution Share-Alike License version 3.0. For specific terms, see http://creativecommons.org/licenses/by-sa/3.0/ )