Chapter 86 15.4 The Arms Race in Computing

It wasn't until the mid-1980s that Danny Hillis began building the first massively parallel computing computer.In fact, as early as a few years ago, Hillis was already a prodigy majoring in computer science.His pranks and hacking deeds are legendary even in MIT, a school known as the originator of hackers.With his customary clarity, Hillis summed up the bottleneck of von Neumann computers to the writer Steven Levy: "The more knowledge you feed into a computer, the slower it runs. Say, the more knowledge you give him, the sharper his mind becomes. So we're in a kind of paradox where the smarter you make a computer, the dumber it becomes."

What Hillis really wanted to be was a biologist, and his knack for understanding complex programs drew him to MIT's artificial intelligence lab.There, he finally decided to try and design a thinking computer that "would be my pride."He credits John Holland with the seminal idea of ​​designing a lawless, three-headed and six-armed computer monster.In the end, the group led by Hillis invented the first parallel processing computer, the "Connection Machine". In 1988, each "connector" could sell for a high price of one million U.S. dollars, making a lot of money.Armed with the machine, Hillis began serious research in computational biology.

"We know that there are only two ways to make something extremely complex," Hillis said. "One is through engineering and the other is through evolution. Of the two, evolution can make things that are more complex." ’” If design doesn’t produce the computers we’re proud of, then we’ll have to rely on evolution. Hillis's first massively parallel "link machine" enabled 64,000 processors to run simultaneously.He couldn't wait to start the evolution, so he injected the computer with 64,000 very simple programs.Like Holland's genetic algorithm or Ray's "Earth," each individual is a string of symbols that can mutate.In Hillis' Connected Machine, though, each program has a dedicated processor to process it.As a result, populations can respond extremely quickly, in numbers that are simply not possible with serial computers.

In his culture medium, the initial "little guys" were random instruction sequences, but after tens of thousands of generations of evolution, they became programs (sequencers) that could sort a long series of numbers.Most larger software will include such a sort routine.Over the years, countless human efforts have been spent in the field of computer science designing the most efficient sorting algorithm.Hillis multiplied thousands of sequencers in a computer, mutating randomly and occasionally swapping sexual genes.Then, as is usual with evolutionary strategies, his system tests these programs, terminating those that are inefficient, and only the shortest (best) sorters get a chance to replicate.Cycled through tens of thousands of generations, his system bred a program that was nearly as short as the best sorting programs written by human programmers.

Hillis then restarted the experiment, but this time with an important difference: while testing the evolved sequencer, the test program (the tester) itself was allowed to mutate.The strings used to test can be made more complex to resist those simple collating bodies.The sequencer had to aim at a moving target, while the tester had to dodge a deflecting arrow.In effect, Hillis transformed the test list of numbers from a rigid, passive environment into a proactive organism.Like foxes and hares, monarch butterflies and milkweeds, sequencers and testers form a classic coevolutionary relationship.

Hillis, still a biologist at heart, sees the mutating test body as a parasite trying to interfere with the sequencing program, and sees his world as an arms race—parasite attack, host defense, parasitism The worm counterattacks, the host defends and counterattacks, and so on.Conventional wisdom has it that such a stalemate arms race is a foolish waste of time, or doomed to get bogged down.However, Hillis found that the introduction of the parasite did not hinder the development of the sequencer, on the contrary, it accelerated the rate of evolution.Parasitic arms races may be ugly, but they dramatically speed up the rate of evolution.

Like Tom Ray, Danny Hillis also discovered that evolution can transcend ordinary human capabilities.The parasites developed in the "connectors" stimulated the sequencers to devise more efficient solutions.After 10,000 cycles of co-evolution, Hillis' little monsters evolved an orderform that computer scientists had never seen before.Most ironically, it happens to be one step short of the shortest human-designed algorithm.Seemingly dull evolution has devised an ingenious and very effective software program. The single processor in the Connector is stupid, about as intelligent as an ant.No single processor alone can come up with an ingenious solution to any problem, no matter how many years it takes.Even stringing together 64,000 processors isn't much better.

And when 64,000 stupid, dumb ant brains form a vast interconnected network, they form an evolutionary population, which looks like a large collection of neurons in the brain.Those difficult problems that make human beings exhausted often find wonderful solutions here.This artificial intelligence method of "order emerges from massive connections" has been dubbed "connectionism". Early intuitions that evolution and learning are closely related have been reawakened by connectionism.Connectionists exploring artificial learning have made their mark by linking dull neurons into gigantic networks.They developed a connection-based approach to parallel processing—running on a virtual or hardware-implemented parallel computer—similar to a genetic algorithm in that it can perform a large number of calculations simultaneously, but with a more sophisticated (smarter) evaluation mechanism.These greatly "enlightened" networks are called neural networks.So far, neural networks have had limited success in producing "intelligence", although their pattern recognition abilities are very useful.

However, the idea that everything comes from a lower connection is astonishing.What magical change has taken place within the network that has given it almost godlike power to breed organization out of interconnected dumb nodes, or programs out of interconnected dumb processors?What is the magic touch that happens when you tie it all together?One minute you have a rabble of simple individuals; the next minute, connected, you have emergent, useful order. For a moment, connectionists speculated that perhaps all that was needed to create sanity and consciousness was a network of interconnected neurons large enough that rational intelligence could assemble itself.As soon as they tried, their dream was shattered.

But artificial evolutionists are still chasing the connectionist dream.It's just that, with the slow pace of evolution, they will be more patient.And this slow, very slow pace of evolution really disturbs me.I expressed my concerns to Tom Ray in this way: "Off-the-shelf evolutionary chips and parallel evolutionary processors make me a little anxious because evolution takes an incredible amount of time. Where does this time come from? Look at nature's Running speed, think about how many tiny molecules have been stuck together during our conversation. The parallel speed and scale of nature is incredible, and we are going to try to surpass it. In my opinion , there simply wasn’t enough time to do it.”

Ray replied, "Oh, I have the same anxiety. But on the other hand, I'm amazed how fast the evolution can go in my system even with just one virtual processor. Again, time is relative. The time scale of evolution is determined by the time span of a generation in evolution. For humans, a generation is thirty years, but for my little ones, a generation is a fraction of a second. And , when I play God, I can speed up the overall mutation rate. I'm not sure, but maybe I can get more evolution on the computer." There are other reasons for doing evolution on computers.For example, Lei can record the genome sequence of each "little thing" and preserve complete demographics and population pedigree.It generates massive amounts of data that simply cannot be collected in the real world.Although as the complexity of the man-made world explodes, so does the complexity and cost of extracting information, it may still be easier to do than in an uncontrollable organic world.As Ray told me, "Even though my world has become as complex as the real world, I am God and I know everything. I can get any information I'm interested in without disturbing it or walking around Trample the plants. That's a fundamental difference."