dmo@turkey.philips.com (Dan Offutt) (06/21/89)
Suppose that AI-based design systems that can think a million times as fast as a human designer become possible, inexpensive, and numerous. What changes would this imply in the rate of technological advance? It seems clear that there will be *some* increase in the rate of technological advance. But the increase will be much less than proportional to the hardware speedup obtained. Million-times-faster designers cannot bring in one year the designs that unspeeded designers would bring in a million years. One reason, briefly, is that a speedup in conscious design cannot serve as a substitute for real-world testing of design realizations. Real-world testing takes time, cannot be speeded up without substantial risk, and produces empirical data about design performance that cannot be obtained in any other way and which is a critical ingredient in subsequent design efforts. For example, the testing of a particular make and model of automobile is performed by consumers who drive the automobiles through precisely the environment to which it must be fit, if the design is to be a success. This testing process produces a steady stream of feedback to designers: consumer complaints about performance and asthetics, manner of failure in accidents, repair rates and types, median useful lifetime and so forth. This information is invaluable in uncovering *in-principle-unpredictable* design flaws. Testing cannot be speeded up: One must wait patiently for consumers to slowly generate design-performance data as they go about their everyday driving activities. The objection may be raised that design flaws are predictable by simulation or simplified mock-ups of real environments. But many design flaws are still unpredictable because design failure can be a function, in part, of almost anything the environment to which the design must be fit. And complete information about such environments is never available to the designer, simulation programmer, or mock-up builder. These remarks apply to designs in general, and nanomachine designs in particular. Nanomachines are likely to be more complex than present-day machines (holding size constant). In general, the more complex the machine, the more difficult it will be to predict its interaction with the environment to which it must be fit. Consequently, collecting performance data during the testing phase will be at least as important for nanomachines as it is for today's machines. Thus the time required for testing nanomachines will limit the rate of nanotechnological progress to much less than might be suspected, given the availability of a million-fold speedup in the speed of AI-based design programs. These observations apply to the distributed nanomachines called active shields. If a prototypical active shield is not tested in the real-world then many or most of its design flaws will not be identified. If it is tested in the real world under the actual noxious conditions it is supposed to protect against, then it is already too late to be of help. If it is tested in a scaled-down sealed ecosystem (a sealed greenhouse, for example), then any characteristic of the complete environoment not present in that scaled-down enviroment is a potential source of design failure. Simulations are even more unsatisfactory. It follows that active shields are less likely to be ready in time to protect against the replicating nanomchines that will inevitably be unleashed into the environment. Caveats: For a given design, some types of quality feedback will come earlier and some later. The lifetime of some automobiles is ten years. Certain facts about such automobiles are discovered only during the tenth year, and not earlier. One may invest more or less effort in acquiring information about the environment to which one's design must be fit. There is the issue of which types of feedback to seek out. The building of a model of this environment is a resource consuming task itself. There is the issue of how much small errors or incompletenesses in the designer's information about the target environement affect the success of a design. Personally, I suspect that seemingly insignificant details about an artifact's environment can often have a very large impact upon success, especially if those details have a long period of time over which to act. The strictly-internal interactions among the components of a design can be complex enough to make the success of a design unpredictable even when the environment is both simple and fully understood. Consider the failure of Apollo 13. Empty space is a very simple environment. A Saturn-V rocket is fairly complex. There is the question of whether the design realization can be neatly distinguished from its environment. Design affects choice of environment since different artifacts will be sorted into different niches. Sports cars are sorted into different niches in the economy than passenger sedans. Sports cars end up in accidents more frequently than passenger sedans. Dan Offutt dmo@philabs.philips.com
dmocsny@uceng.uc.edu (daniel mocsny) (06/22/89)
In article <Jun.20.23.27.17.1989.28085@athos.rutgers.edu>, dmo@turkey.philips.com (Dan Offutt) writes: > Suppose that AI-based design systems that can think a million times as > fast as a human designer become possible, inexpensive, and numerous. > What changes would this imply in the rate of technological advance? It would change everything in ways we can hardly imagine at present. But we can amuse ourselves by speculating, and arguing ;-) > ... the increase will be much less than > proportional to the hardware speedup obtained. Million-times-faster > designers cannot bring in one year the designs that unspeeded > designers would bring in a million years. One reason, briefly, is > that a speedup in conscious design cannot serve as a substitute for > real-world testing of design realizations. Real-world testing takes > time, cannot be speeded up without substantial risk, and produces > empirical data about design performance that cannot be obtained in > any other way and which is a critical ingredient in subsequent > design efforts. OK, but hold on a second! Think about all the data consumers generate every day that vendors have no choice but to ignore because (1) they can't handle the data volume (2) no communication systems are in place to make gathering the data easy and (3) the data is unavailable for political reasons (e.g., trade secrets, inter- and intra-corporate rivalry). If we grant your original premise, that mechanical super-intelligence is cheap and ubiquitous (and further assume that humans will be able to stay on top of it!), then vendors will have *vastly* increased ability to gather data and use it. Similarly, consumers will have a vastly increased ability to record and report complaints. Even if real-world data doesn't get generated any faster, if we simply start using a vastly larger portion of the data now going to waste, product improvements will speed up drastically. Think of all the millions of consumers out there using all of those products. How much time passes now before major design flaws filter back to the vendors and are corrected? Too much. Similarly, enormous amounts of data are already available for every major product category one cares to name. I suggest that most of what a present-day vendor will learn from a product-testing program must already be available in principle. Here we can divide the data into essence and accident. The accidents are all those things you should have already known (for example, so many automobiles have been sold that by now the general outline of consumer preference should not be any great surprise), whereas the essence is whatever really is new about the product and heretofore untested. With massive increases in data-gathering and -handling power, vendors will be able to greatly increase their efficiency in designing products that work the first time. But we will observe even more fundamental changes. For example, if we had super-intelligent machines, we would probably not use them to build better automobiles. Instead, we would no longer need the present levels of automobile use, because the existence of such machines would imply the existence of communication technology fast and transparent enough to make most of our present physical travel a waste of time. When the jet engine appeared, nobody tried to mount one on a horse. New technological capabilities do not always help you do better what you are already doing. Instead, they often push you into doing entirely new things. Dan Mocsny Snail: Internet: dmocsny@uceng.UC.EDU Dept. of Chemical Engng. M.L. 171 513/751-6824 (home) University of Cincinnati 513/556-2007 (lab) Cincinnati, Ohio 45221-0171
macleod@drivax.UUCP (MacLeod) (06/22/89)
In article <Jun.20.23.27.17.1989.28085@athos.rutgers.edu> dmo@turkey.philips.com (Dan Offutt) writes: >These remarks apply to designs in general, and nanomachine designs in >particular. Nanomachines are likely to be more complex than >present-day machines (holding size constant). In general, the more >complex the machine, the more difficult it will be to predict its >interaction with the environment to which it must be fit. I would sleep better if all engineers, of every discipline, read a slender volume called "Systemantics" by a medical doctor named John Gall. It is a short, humerous series of essays exploring a number of empirically derived axioms about system behavior. Like its predecessor, "The Peter Principle", it is actually profound truth wrapped in humor. Gall shows that as system complexity grows the possibilities - and likelihood - of anomalous behavior increases, presumably as some function of the number of machine states. The larger the system, the more it tends to impede its own functioning. The examples Gall cites as climax designs often perversely generate exactly the problem they were originally designed to surmount - the classic example is the mammoth VAB at Cape Canaveral. Built to protect Saturn V components from the weather, it generates its own rain internally. Michael Sloan MacLeod (amdahl!drivax!macleod) [This is often called the "law of unintended effect" and applies to almost any complex system, not just engineered mechanisms. Indeed, it applies less to engineered machines than to most other complex systems. The VAB really does protect rockets from the strong winds that are common on the Florida coastline. However, purchasing departments and their regulations typically cause organizations to spend twice as much for what they buy. Expanded legal liability for manufacturers and doctors cause talented people to leave the field, and safety oriented products and medicines to be withdrawn. The more complex something is, the greater the chance its design and production will be done by committee and bureaucracy. This is the major reason for the more-than-linear decrease in reliability and effectiveness with size. There is some reason to hope that for engineered machines, AI systems will have their biggest impact simply by letting bigger projects be handled by a single individual. Furthermore, I'll wager that the first corporation to replace its *management* with a computer program will wipe up the competition in no time flat. Of course, as I have noted here before, there are some dangers inherent with trying the same thing with the government... --JoSH]
alan@oz.nm.paradyne.com (Alan Lovejoy) (06/24/89)
In article <Jun.20.23.27.17.1989.28085@athos.rutgers.edu> dmo@turkey.philips.com (Dan Offutt) writes: >... Million-times-faster >designers cannot bring in one year the designs that unspeeded >designers would bring in a million years. One reason, briefly, is >that a speedup in conscious design cannot serve as a substitute for >real-world testing of design realizations. Real-world testing takes >time, cannot be speeded up without substantial risk, and produces >empirical data about design performance that cannot be obtained in >any other way and which is a critical ingredient in subsequent >design efforts. Ok. You make a good point. But I have three quibbles: 1) So what? The "million times speed-up" is a VERY conservative estimate of the possible increase in computing speeds due to nanotechnology, quantum circuits and other as-yet-unknown advances. What does it matter if the effective speed-up in technical progress which is practically obtainable is only 10**5? The thrust of Drexler's argument is not mortally wounded simply by a one or two order of magnitude overestimation. 2) The problems which you rightly identify as having a significant dampening effect on the rate of progress which is practically achievable do not apply to the same degree to ultra-intelligent full-spectrum AIs. Your arguments hit with full force only in the case of idiot savant machines, not in the case of intelligences which transcend our own in all ways. Such intelligences will be able to find elegant solutions to many problems that we find hard, intractable or even do not see at all. They are really beyond our ability to predict or understand. We are like salamanders trying to envision the problem-sovling skills of homo sapiens. 3) All the problems which you have listed that hinder the development of active shields also work against the designers of gray goo. The goo has to both defend itself against the shield and try to accomplish its main objective of destroying the enviroment. The shield must defend itself against the goo and attempt to destroy and/or incapacitate the goo. Both may masquerade as the other. Both will attempt to interfere with the other's communications. Assemblers and disassemblers will mindlessly follow whatever program they receive, regardless of its origin. The shield may be able to shut the goo down just by jamming the goo's communications, destroying its energy supplies and immobilizing ALL molecules in the affected area, including those needed or used by the goo, in such a way that nothing of importance (e.g., human bodies) is injured beyond the ability of nanomachines to repair. It is VERY difficult to win a game of chess against an opponent who is determined to achieve a draw, unless the player who wants to win is much better than the player who wants merely not to lose. The shield "wins" merely by making it impossible for nanomachinery to function, or at least by making it impossible for nanocomputers to send commands to assemblers and disassemblers. Alan Lovejoy; alan@pdn; 813-530-2211; AT&T Paradyne: 8550 Ulmerton, Largo, FL. Disclaimer: I do not speak for AT&T Paradyne. They do not speak for me. ______________________________Down with Li Peng!________________________________ Motto: If nanomachines will be able to reconstruct you, YOU AREN'T DEAD YET.