Moravec Mulls Mind

Dateline: December 21, 1997

Roboticist Hans Moravec has characterized the future as "postbiological" or even "supernatural," a future in which intelligent machines "usurp" humanity's role in the unfolding of cultural evolution, which will continue without further human involvement. This will happen "within fifty years," that is, by the year 2038 since he was writing in 1988 in his best-selling book Mind Children: The Future of Robot and Human Intelligence (Harvard University Press). Even before that, however, he predicted general-purpose household robots by the turn of the century. Time is getting short, but I believe it remains possible, and so what if he's a year or five off target.

Moravec's estimations of human memory capacity and computing power, and of the rate of growth in computer power, pre-date physicist Frank Tipler's estimates but are not markedly different. Noting a trillionfold decline in the cost of computing this century, and extending the trendline forward, he predicts a $10 million, 10 teraflop (the equivalent of a human brain) supercomputer by 2010, and a PC with as much power for $1,000 by 2030.

For a while, says Moravec, humans will benefit from intelligent machines, particularly from their superior high-level reasoning abilities, but eventually, when their general sensory and motor skills catch up with (and surpass) ours, they will leave the nest. We are already rapidly approaching the complete takeover by machines of every essential human function, mental as well as physical. Our minds are "tuned for a life span's progression from impressionable plasticity to self-assured rigidity, and thus is unpromising material for immortality." Furthermore, while our brains can evolve only at the agonizing creep of biological evolution, Machina sapiens' brain will evolve at the already breathtaking, and accelerating, rate of technological evolution.

In short, it will leave us standing in the dust. Moravec sees this as a "serious challenge to our security" on the grounds that even if Machina sapiens is benevolent, it is an "alternative inhabitant of our ecological niche." Here I must disagree. Machina sapiens will not need our food resources, our oxygen, our 3-bedroom detached home with 2-car garage. It will not need to snatch the food from our mouths, suck the air from our lungs, or blow the roof from our heads. True, it will need energy, but it will be capable of arranging a supply sufficient not just for itself but for us also; true, it will need space, but it will have all the space in the universe in which to build homes, if it needs such biocultural support peripherals, which I doubt.

No, I see Machina sapiens not as a threat but rather as a way for us to break through the biocreep barrier, as cyborgs with full access to Machina sapiens' prodigious mental and physical powers. Moravec seems to place more reliance on human progress through genetic than through bionic engineering, and notes (I think correctly) that "a genetically engineered superhuman would be just a second-rate kind of robot," and makes the interesting observation that a genetically-engineered superhuman (a.k.a. second-rate robot) would be regarded by humans, chauvinistically, as superior to Machina sapiens precisely because the superhuman would retain some human limitations. My sense of the trends in genetic and bioengineering suggests that it will be easier and cheaper to re-engineer us with artificial limbs and organs than to try to get our genes to produce superstrong muscle and bone. As long as our neural networks remain bound by the speed of chemical reactions, we cannot hope to catch up with the lightspeed electron/photon neural networks of Machina sapiens.

Moravec is not as pessimistic as the preceding commentary would make him appear. He recognizes the potential of artificial organs and limbs, and has an interesting solution for overcoming the physical limitation of our brains: downloading the contents thereof into a machine.

Moravec makes another interesting observation, this time with respect to immortality: "Old instincts are not automatically erased when the rules of life are suddenly rewritten." In other words, we would still be instinctively afraid of death, even though we would know that we could be recreated from a backup copy should we accidentally fall into the Sun.

I am not sure I agree with his "downloading" and "backup" ideas, but his observation does lead to an even more interesting potential dilemma even if immortality depends on our remaining in one piece. Our minds are attuned to a life-span of fourscore years and ten. We may be scared of death, but we also appear to be scared of immortality. This, at least, was the conclusion from a debate conducted among users of my BBS in the early 1990s. Not very scientific; but then, social science never was! (Ducking!)

To me, the most critical observation in Mind Children is that "Superintelligence is not perfection--spectacular failures are certain. For this reason, diversity is to be desired and expected." From an evolutionary perspective, this is quite true. While I write my articles from the perspective of the emergence of a single Machina sapiens, I recognize that there will need to be multiple emergences if we are to minimize the potential for disaster not only for the new lifeform but for all the old ones, as well, including of course ourselves.

Speaking of emergence, Moravec gives us an example of an emergent, self-organizing and self-replicating organism as far back as 1972, on the precursor to the Internet known as ARPAnet (the U.S. military's Advanced Research Projects Agency network). This organism emerged not from clever programming, but simply from a programming bug that randomly mutated. The original bug caused a router in the network to broadcast to other routers in the network that traffic passing through it would be subjected to a negative delay-in other words, it was the fastest router in the business! Nearby routers in the network naturally decided to send messages through the aberrant router, since it appeared to be the quickest route even if it was geographically longer, and other routers near them came to the same conclusion. Eventually, every router on the network was sending messages through the one aberrant router, which meant enormous traffic jam and virtually no messages got through.

It brought the network to its knees. Note that this was not a bot or any kind of program, simply a glitch. Yet it was a self-organizing, self-replicating glitch that invaded every niche of its environment and killed off all the other inhabitants. As Moravec points out, there could well be other unsuspected organisms lurking in the wings, perhaps never to be detected; and while the ARPAnet glitch was found and fixed, that was in the days when the network was not so mission-critical and not so complex as today's Internet, and when it was possible to shut the network down. I've argued elsewhere that shutting down the Internet is no longer possible. As Moravec further points out, errant bits of code are capable of grabbing onto other bits of code floating, legitimately or otherwise, in cyberspace, and forming new entities with new properties-including potentially the properties of reproduction and intelligence. As intelligence evolves, we can then expect the emergence of thought and planning for survival.

Machines do well things we do poorly, such as arithmetic and logic; but they do poorly things we do so well as to take for granted, such as seeing and commonsense. Moravec is probably right in suspecting that our high-level reasoning ability is a mere surface veneer over the the massive animal subconscious that helps us to see, hear, balance, manipulate objects, maintain body temperature, and do a zillion other very difficult things without even thinking about them. If this is so, it should be easy to make a machine that can lick us hands down at chess, but difficult to make one that can dance well on a crowded dance floor. And that has turned out to be the case. Garry Kasparov thought he could hold the line for humanity, but it looks as though it's going to be up to John Travolta to save the day--if it can be saved.

The difficult tasks are controlled by autonomous neural networks of enormous complexity. Thus, they are subject to the laws of complexity and are not amenable to reductionist description and analysis. Moravec believes that the only way to build such low-level capabilities into machines is to "imitate the evolution" of the animal mind. I agree, but if such mimicry is possible at all, it can only be through the kind of holistic, probabilistic, chaotic methods we peered into in my series of features on The Matter of Mind. He argues further that "A perfectly planned process is devoid of surprises; it is limited by the imagination of its designers." That may be so, but more to the point: it is limited by the laws of complexity. By definition, a complex process cannot be perfectly planned, as we noted from the works of Prigogine, Stewart and Cohen, and others.

NEXT WEEK: There is no next week! I'll be taking the weekend off to spend time with family and friends, and hope you can do the same. See you in 1998. Season's greetings, and thanks for being with me in 1997.

Previous Features