And so the AI pioneers of the twenty-first century will most likely rely on a kind of artificial natural selection to create their thinking machines. A bank of massivelyparallel supercomputers designed to simulate the multiple connections of our neuronal system will be connected to some kind of external world, and programmed to experiment with more or less random responses to the data coming in from that world. The humans will establish criteria that reward behavior that shows signs of intelligence and punish less promising behavior. Innovative new strategies generated by a computer will be preserved and shared with all the other machines; less successful strategies will be eliminated. Over time, the virtual gene pool will start to accumulate intelligent strategies for dealing with--and predicting events in--the outside world. A machine connected to some kind of light detector might begin to register the twenty-four-hour cycle of light and dark; that "insight" would then be passed on to the machine's virtual descendants, one of whom might then develop an awareness of seasonal cycles, by noting changes in the length of days. That strategy would be passed down to future generations, where it could be enhanced or expanded by new innovations. Presumably, somewhere at the end of this assisted evolution, a digital Copernicus awaits us.

For every successful assessment of the world, of course, there are a million failures, just as there are in the real-world fossil record. The difference here is that the pace of evolutionary time has been sped up dramatically; the supercomputers can churn through thousands of generations of artificial minds in a matter of minutes, if required. Already, this type ofevolutionary approach has "bred" software programs that are faster and more stable than programs written using conventional techniques. For a single, focused task like number sorting or fingerprint recognition, these evolutionary techniques can be extremely effective. But for a more open-ended, multidimensional goal like general intelligence, the problem with the evolutionary approach--if problem is the right word for it--is that the human overseers don't have much control over the direction in which the software will evolve. You can push it in certain directions, particularly in the early stages of evolution, but before long it is bound to develop a mind of its own. Already, natural selection has seen fit to evolve countless forms of intelligence on the planet. If evolution now turns its invisible hand to growing machine intelligence, it seems unlikely that something resembling human smarts will emerge, no matter how much we rig the system.

So perhaps the sci-fi narratives have it the wrong way around: The sentient machines won't act like little boys or obedient butlers after all, but they'll be even more unnerving for their radical difference. (How multiculturalism will deal with the "new flesh" is a fascinating question.) Our creations won't be disappointed children, à la Frankenstein and AI, but rather some other species altogether--as familiar as a dolphin's mind, or whatever kind of distributed intelligence helps a termite colony build a nest. But if natural selection is to play such a crucial role in the creation of genuine AI, it's fair to ask about the environmental pressures that will shape the development of these new creatures. In what virtual petri dish will they do their learning? Where do you take a supercomputer when you want to teach it more about the world? One place to start might be that epic collection of human wisdom, hardcore porn and shameless self-promotion--the World Wide Web. A machine trying to bootstrap into intelligence might well begin by surveying those endless connected pages and soaking it all in. The web itself might not become self-aware, but perhaps a machine will cross that threshold by immersing itself in the datasphere for enough evolutionary cycles to start making sense of it all.

Although that machine intelligence would emerge out of studying our man-made information networks, there would be differences of kind, not just magnitude. Perhaps most important, it would be a thoroughly textual intelligence, having evolved in a universe of words. While language is certainly a central--and arguably defining--characteristic of human intelligence, it is a relatively new innovation. Our minds are also shaped by the ancient emotional centers of the limbic system and the advanced visual processing that we share with our primate cousins. (It's no accident that words for knowledge overlap so frequently with words for sight.) A machine trained to think for itself by reading data from web servers would lack the ballast and value judgments of our emotions, and the spatial logic that we borrow effortlessly from the shape recognizers of the visual system; the machine's smarts would be largely lexical, the intelligence you get when you live in a world of words (though perhaps a rudimentary visual intelligence could be evolved by studying the web's endless supply of Britney Spears pix). You can glimpse this lexical intelligence already in software applications like Kurzweil's Cybernetic Poet--a program that uses a kind of neural net technology called Markov models to detect textual patterns in poetry. Based on the patterns it associates with each poet, the program can compose its own poetry, in the style of John Ashbery or Robert Frost. (It can even compose authorial "blends.") The program seems to do best with elliptical and associative forms like haiku--it's not likely to churn out a cybernetic Odyssey anytime soon--but it suggests that word immersion can lead to remarkable advances. Yet if that immersion someday includes all the data stored on the web, not just the work of a few handpicked poets, the eventual results will vary wildly from the intelligence of the human race.

There's a striking premonition of this idea near the end of 2001, as Dave Bowman shuts off the homicidal HAL 9000. As HAL winds down to nothingness--with his haunted, I'm-afraid-Dave refrain growing slower with each removed chip set--we learn that the state-of-the-art AI unit was "born" in 1992 in Urbana, Illinois. As it turns out, the modern web as we know it was born in Urbana in 1992, with the creation of the Mosaic browser. It's a fitting correlation: True artificial intelligence of the sort that HAL was supposed to embody may well evolve out of the web's vast archives. But that intelligence will not harbor the human failings--or strengths--that have historically been projected onto our fictional robot offspring. They will not love their mothers, nor dream of electric sheep.

• « Previous
• 1
• 2
• 3

• Get The Nation at home (and online!) for 75 cents a week!
• If you like this article, consider making a donation to The Nation.