Cheaper than printing it out: buy the paperback book.

Out of Control

It wasn't until the mid-1980s that Danny Hillis began building the first massively parallel computer. Just a few years earlier Hillis had been a wunderkind computer science student. His pranks and hacks at MIT were legendary, even on the campus that invented hacking. With his usual clarity, Hillis summed up for writer Steven Levy the obstacle the von Neumann bottleneck had become in computers: "The more knowledge you gave them, the slower computers got. Yet with a person, the more knowledge you give him, the faster he gets. So we were in this paradox that if you tried to make computers smart, they got stupider."

Hillis really wanted to be a biologist, but his knack for understanding complex programs drew him to the artificial intelligence labs of MIT, where he wound up trying to build a thinking computer "that would be proud of me." He attributes to John Holland the seminal design notions for a swarmy, thousand-headed computing beast. Eventually Hillis led a group that invented the first parallel processing computer, the Connection Machine. In 1988 it sold for a cool $1 million apiece, fully loaded. Now that the machines are here, Hillis has taken up computational biology in earnest.

"There are only two ways we know of to make extremely complicated things," says Hillis. "One is by engineering, and the other is evolution. And of the two, evolution will make the more complex." If we can't engineer a computer that will be proud of us, we may have to evolve it.

Hillis's first massively parallel Connection Machine had 64,000 processors working in unison. He couldn't wait to get evolution going. He inoculated his computer with a population of 64,000 very simple software programs. As in Holland's GA or in Ray's Tierra, each individual was a string of symbols that could be altered by mutation. But in Hillis's Connection Machine, each program had an entire computer processor dedicated to running it. The population, therefore, would react extremely quickly and in numbers that were simply not possible for serial computers to handle.

Each bug in his soup was initially a random sequence of instructions, but over tens of thousands of generations they became a program that sorted a long string of numbers into numerical order. Such a sort routine is an integral part of most larger computer programs; over the years many hundreds of man hours have been spent in computer science departments engineering the most efficient sort algorithms. Hillis let thousands of his sorters proliferate in his computer, mutate at random, and occasionally sexually swap genes. Then in the usual evolutionary maneuver, his system tested them and terminated the less fit so that only the shortest (the best) sorting programs would be given a chance to reproduce. Over ten thousand generations of this cycle, his system bred a software program that was nearly as short as the best sorting programs written by human programmers.

Hillis then reran the experiment but with this important difference: He allowed the sorting test itself to mutate while the evolving sorter tried to solve it. The string of symbols in the test varied to become more complicated in order to resist easy sorting. Sorters had to unscramble a moving target, while tests had to resist a moving arrow. In effect Hillis transformed the test list of numbers from a harsh passive environment into an active organism. Like foxes and hares or monarchs and milkweed, sorters and tests got swept up by a textbook case of coevolution.

A biologist at heart, Hillis viewed the mutating sorting test as a parasitic organism trying to disrupt the sorter. He saw his world as an arms race -- parasite attack, host defense, parasite counterattack, host counter -- defense, and so on. Conventional wisdom claimed such locked arms races are a silly waste of time or an unfortunate blind trap to get stuck in. But Hillis discovered that rather than retard the advance of the sorting organisms, the introduction of a parasite sped up the rate of evolution. Parasitic arms races may be ugly, but they turbocharged evolution.

Just as Tom Ray would discover, Danny Hillis also found that evolution can surpass ordinary human skills. Parasites thriving in the Connection Machine prodded sorters to devise a solution more efficient than the ones they found without parasites. After 10,000 cycles of coevolution, Hillis's creatures evolved a sorting program previously unknown to computer scientists. Most humbling, it was only a step short of the all-time shortest algorithm engineered by humans. Blind dumb evolution had designed an ingenious, and quite useful, software program.

A single processor in the Connection Machine is very stupid. It might be as smart as an ant. On its own, a single processor could not come up with an original solution to anything, no matter how many years it spent. Nor would it come up with much if 64,000 processors were strung in a row.

But 64,000 dumb, mindless, ant-brains wired up into a vast interconnected network become a field of evolving populations and, at the same time, look like a mass of neurons in a brain. Out of this network of dumbness emerge brilliant solutions to problems that tax humans. This "order-emerging-out-of-massive-connections" approach to artificial intelligence became known as "connectionism."

Connectionism rekindled earlier intuitions that evolution and learning were deeply related. The connectionists who were reaching for artificial learning latched onto the model of vast webs interconnecting dumb neurons, and then took off with it. They developed a brand of connected concurrent processing -- running in either virtual or hardwired parallel computers -- that performed simultaneous calculations en masse, similar to genetic algorithms but with more sophisticated (smarter) accounting systems. These smartened up networks were called neural networks. So far neural nets have achieved only limited success in generating partial "intelligence," although their pattern-recognition abilities are useful.

But that anything at all emerges from a field of lowly connections is startling. What kind of magic happens inside a web to give it an almost divine power to birth organization from dumb nodes interconnected, or breed software from mindless processors wired to each other? What alchemic transformation occurs when you connect everything to everything? One minute you have a mob of simple individuals, the next, after connection, you have useful, emergent order.

There was a fleeting moment when the connectionists imagined that perhaps all you needed to produce reason and consciousness was a sufficiently large field of interlinked neurons out of which rational intelligence would assemble itself. That dream vanished as soon as they tried it.

But in an odd way, the artificial evolutionists still pursue the dream of connectionism. Only they, in sync with the slow pace of evolution, would be more patient. But it is the slow, very slow, pace of evolution that bothers me. I put my concern to Tom Ray this way: "What worries me about off-the-shelf evolution chips and parallel evolutionary processing machines is that evolution takes an incredible amount of time. Where is this time going to come from? Look at the speed at which nature is working. Consider all the little molecules that have just been snapped together as we talk here. Nature is incredibly speedy and vast and humongously parallel, and here we are going to try to beat it. It seems to me there's simply not enough time to do it.

Ray replied: "Well, I worry about that too. On the other hand, I'm amazed at how fast evolution has occurred in my system with only one virtual processor churning it. Besides, time is relative. In evolution, a generation sets the time scale. For us a generation is thirty years, but for my creatures it is a fraction of a second. And, when I play god I can crank up the global mutation rate. I'm not sure, but I may be able to get more evolution on a computer."

There are other reasons for doing evolution in a computer. For instance, Ray can record the sequence of every creature's genome and keep a complete demographic and genealogic record of every creature's birth and death. It produces an avalanche of data that is impossible to compile in the real world. And though the complexity and cost of extracting the information will surge as the complexity of the artificial worlds surge, it will probably remain easier to do than in the unwired organic world. As Ray told me, "Even if my world gets as complex as the real world, I'm god. I'm omniscient. I can get information on whatever attracts my attention without disturbing it, without walking around crushing plants. That's a crucial difference."