Why a superintelligent machine may be the last thing we ever invent

Illustration for article titled Why a superintelligent machine may be the last thing we ever invent

If you want to know about the future of artificial intelligence then you must read documentary filmmaker James Barrat's new book Our Final Invention. We've got an incredible excerpt from the book, about the coming intelligence explosion that could redefine the human condition.


"The Intelligence Explosion," an excerpt from Our Final Invention: Artificial Intelligence and the End of the Human Era, by James Barrat.

Interstate 81 starts in New York State and ends in Tennessee, traversing almost the entire range of the Appalachian Mountains. From the middle of Virginia heading south, the highway snakes up and down deeply forested hills and sweeping, grassy meadows, through some of the most striking and darkly primordial vistas in the United States. Contained within the Appalachians are the Blue Ridge Mountain Range (from Pennsylvania to Georgia) and the Great Smokies (along the North Carolina– Tennessee border). The farther south you go, the harder it is to get a cell phone signal, churches outnumber houses, and the music on the radio changes from Country to Gospel, then to hellfire preachers. I heard a memorable song about temptation called “Long Black Train” by Josh Turner. I heard a preacher begin a sermon about Abraham and Isaac, lose his way, and end with the parable of the loaves and fishes and hell, thrown in for good measure. I was closing in on the Smokey Mountains, the North Carolina border, and Virginia Tech—the Virginia Polytechnic Institute and State University in Blacksburg, Virginia. The university’s motto: INVENT THE FUTURE.

Illustration for article titled Why a superintelligent machine may be the last thing we ever invent

Twenty years ago, driving on an almost identical I-81 you might have been overtaken by a Triumph Spitfire convertible with the license plate 007 IJG. The vanity plate belonged to I. J. Good, who arrived in Blacksburg in 1967, as a Distinguished Professor of Statistics. The “007” was an homage to Ian Fleming and Good’s secret work as a World War II code breaker at Bletchley Park, England. Breaking the encryption system that Germany’s armed forces used to encode messages substantially helped bring about the Axis powers’ defeat. At Bletchley Park, Good worked alongside Alan Turing, called the father of modern computation (and creator of chapter 4’s Turing test), and helped build and program one of the first electrical computers.

In Blacksburg, Good was a celebrity professor—his salary was higher than the university president’s. A nut for numbers, he noted that he arrived in Blacksburg on the seventh hour of the seventh day of the seventh month of the seventh year of the seventh decade, and was housed in unit seven on the seventh block of Terrace View Apartments. Good told his friends that God threw coincidences at atheists like him to convince them of his existence.

“I have a quarter-baked idea that God provides more coincidences the more one doubts Her existence, thereby providing one with evidence without forcing one to believe,” Good said. “When I believe that theory, the coincidences will presumably stop.”


I was headed to Blacksburg to learn about Good, who had died recently at age ninety-two, from his friends. Mostly, I wanted to learn how I. J. Good happened to invent the idea of an intelligence explosion, and if it really was possible. The intelligence explosion was the first big link in the idea chain that gave birth to the Singularity hypothesis.

Unfortunately, for the foreseeable future, the mention of Virginia Tech will evoke the Virginia Tech Massacre. Here on April 16, 2007, senior English major Seung-Hui Cho killed thirty-two students and faculty and wounded twenty-five more. It is the deadliest shooting incident by a lone gunman in U.S. history. The broad outlines are that Cho shot and killed an under-graduate woman in Ambler Johnston Hall, a Virginia Tech dormitory, then killed a male undergraduate who came to her aid. Two hours later Cho began the rampage that caused most of the casualties. Except for the first two, he shot his victims in Virginia Tech’s Norris Hall. Before he started shooting Cho had chained and padlocked shut the building’s heavy oaken doors to prevent anyone from escaping.


When I. J. Good’s longtime friend and fellow statistician Dr. Golde Holtzman showed me Good’s former office in Hutcheson Hall, on the other side of the beautiful green Drillfield (a military parade ground in Tech’s early life), I noticed you could just see Norris Hall from his window. But by the time the tragedy unfolded, Holtzman told me, Good had retired. He was not in his office but at home, perhaps calculating the probability of God’s existence.

According to Dr. Holtzman, sometime before he died, Good updated that probability from zero to point one. He did this because as a statistician, he was a long-term Bayesian. Named for the eighteenth-century mathematician and minister Thomas Bayes, Bayesian statistics’ main idea is that in calculating the probability of some statement, you can start with a personal belief. Then you update that belief as new evidence comes in that supports your statement or doesn’t. If Good’s original disbelief in God had remained 100 percent, no amount of data, not even God’s appearance, could change his mind. So, to be consistent with his Bayesian perspective, Good assigned a small positive probability to the existence of God to make sure he could learn from new data, if it arose.


In the 1965 paper “Speculations Concerning the First Ultra-intelligent Machine,” Good laid out a simple and elegant proof that’s rarely left out of discussions of artificial intelligence and the Singularity:

Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an “intelligence explosion,” and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make . . .


The Singularity has three well-developed definitions— Good’s, above, is the first. Good never used the term “singularity” but he got the ball rolling by positing what he thought of as an inescapable and beneficial milestone in human history— the invention of smarter- than-human machines. To paraphrase Good, if you make a superintelligent machine, it will be better than humans at everything we use our brains for, and that includes making superintelligent machines. The first machine would then set off an intelligence explosion, a rapid increase in intelligence, as it repeatedly self-improved, or simply made smarter machines. This machine or machines would leave man’s brainpower in the dust. After the intelligence explosion, man wouldn’t have to invent anything else—all his needs would be met by machines.

This paragraph of Good’s paper rightfully finds its way into books, papers, and essays about the Singularity, the future of artificial intelligence, and its risks. But two important ideas al-most always get left out. The first is the introductory sentence of the paper. It’s a doozy: “The survival of man depends on the early construction of an ultraintelligent machine.” The second is the frequently omitted second half of the last sentence in the paragraph. The last sentence of Good’s most often quoted paragraph should read in its entirety:

Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control (emphasis mine).


These two sentences tell us important things about Good’s intentions. He felt that we humans were beset by so many complex, looming problems—the nuclear arms race, pollution, war, and so on—that we could only be saved by better thinking, and that would come from superintelligent machines. The second sentence lets us know that the father of the intelligence explosion concept was acutely aware that producing superintelligent machines, however necessary for our survival, could blow up in our faces. Keeping an ultraintelligent machine under control isn’t a given, Good tells us. He doesn’t believe we will even know how to do it—the machine will have to tell us itself.

Good knew a few things about machines that could save the world—he had helped build and run the earliest electrical computers ever, used at Bletchley Park to help defeat Germany. He also knew something about existential risk—he was a Jew fighting against the Nazis, and his father had escaped pogroms in Poland by immigrating to the United Kingdom.


As a boy, Good’s father, a Pole and self-educated intellectual, learned the trade of watchmaking by staring at watchmakers through shop windows. He was just seventeen in 1903 when he headed to England with thirty-five rubles in his pocket and a large wheel of cheese. In London he performed odd jobs until he could set up his own jewelry shop. He prospered and married. In 1915, Isidore Jacob Gudak (later Irving John “Jack” Good) was born. A brother followed and a sister, a talented dancer who would later die in a theater fire. Her awful death caused Jack Good to disavow the existence of God.

Good was a mathematics prodigy, who once stood up in his crib and asked his mother what a thousand times a thousand was. During a bout with diphtheria he independently discovered irrational numbers (those that cannot be expressed as fractions, such as √2). Before he was fourteen he’d rediscovered mathematical induction, a method of making mathematical proofs. By then his mathematics teachers just left him alone with piles of books. At Cambridge University, Good snatched every math prize available on his way to a Ph.D., and discovered a passion for chess.


It was because of his chess playing that a year after World War II began, Britain’s reigning chess champion, Hugh Alexander, recruited Good to join Hut 18 at Bletchley Park. Hut 18 was where the decoders worked. They broke codes used by all the Axis powers—Germany, Japan, and Italy—to communicate military commands, but with special emphasis on Germany. German U-boats were sinking Allied shipping at a crippling rate—in just the first half of 1942, U-boats would sink some five hundred Allied ships. Prime Minister Winston Churchill feared his island nation would be starved into defeat.

German messages were sent by radio waves, and the English intercepted them with listening towers. From the start of the war Germany created the messages with a machine called the Enigma. Widely distributed within the German armed forces, the Enigma was about the size and shape of an old-fashioned manual typewriter. Each key displayed a letter, and was connected to a wire. The wire would make contact with another wire that was connected to a different letter. That letter would be the substitute for the one represented on the key. All the wires were mounted on rotors to enable any wire in the alphabet to touch any other wire. The basic Enigmas had three wheels, so that each wheel could perform substitutions for the substitutions made by the prior wheel. For an alphabet of twenty-six letters, 403,291,461,126,605,635,584,000,000 such substitutions were possible. The wheels, or settings, changed almost daily.


When one German sent others an Enigma-encoded message, the recipients would use their own Enigmas to decode it, provided they knew the sender’s settings.

Fortunately Bletchley Park had a secret weapon of its own— Alan Turing. Before the war, Turing had studied mathematics and encryption at Cambridge and Princeton. He had imagined an “automatic machine,” now known as a Turing machine. The automatic machine laid out the basic principles of computation itself.


The Church-Turing hypothesis, which combined Turing’s work with that of his Princeton professor, mathematician Alonso Church, really puts the starch in the pants of the study of artificial intelligence. It proposes that anything that can be computed by an algorithm, or program, can be computed by a Turing machine. Therefore, if brain processes can be expressed as a series of instructions—an algorithm— then a computer can process information the same way. In other words, unless there’s something mystical or magical about human thinking, intelligence can be achieved by a computer. A lot of AGI researchers have pinned their hopes to the Church-Turing hypothesis.

The war gave Turing a crash course in everything he’d been thinking about before the war, and lots he hadn’t been thinking about, like Nazis and submarines. At the war’s peak, Bletchley Park personnel decoded some four thousand intercepted messages per day. Cracking them all by hand became impossible. It was a job meant for a machine. And it was Turing’s critical insight that it was easier to calculate what the settings on the Enigma were not, rather than what they were.


The decoders had data to work with—intercepted messages that had been “broken” by hand, or by electrical decoding machines, called Bombes. They called these messages “kisses.” Like I. J. Good, Turing was a devoted Bayesian, at a time when the statistical method was seen as a kind of witchcraft. The heart of the method, the Bayes’ theorem, describes how to use data to infer probabilities of unknown events, in this case, the Enigma’s settings. The “kisses” were the data that allowed the decoders to determine which settings were highly improbable, so that the code-breaking efforts could be focused more efficiently. Of course, the codes changed almost daily, so work at Bletchley Park was a constant race.

Turing and his colleagues designed a series of electronic machines that would evaluate and eliminate possible Enigma settings. These early computers culminated in a series of machines all named “Colossus.” Colossus could read five thousand characters per second from paper tape that traveled through it at twenty-seven miles an hour. It contained 1,500 vacuum tubes, and filled a room. One of its main users, and creator of half the theory behind the Colossus, was Turing’s chief statistician for much of the war: Irving John Good.


The heroes of Bletchley Park probably shortened World War II by between two and four years, saving an incalculable number of lives. But there were no parades for the secret warriors. Churchill ordered that all Bletchley’s encryption machines be broken into pieces no bigger than a fist, so their awesome decoding power couldn’t be turned against Great Britain. The code breakers were sworn to secrecy for thirty years. Turing and Good were recruited to join the staff at the University of Manchester, where their former section head, Max Newman, intended to develop a general purpose computer. Turing was working on a computer design at the National Physical Laboratory when his life turned upside down. A man with whom he’d had a casual affair burgled his house. When he reported the crime he admitted the sexual relationship to the police. He was charged with gross indecency and stripped of his security clearance.

At Bletchley Turing and Good had discussed futuristic ideas like computers, intelligent machines, and an “automatic” chess player. Turing and Good bonded over games of chess, which Good won. In return, Turing taught him Go, an Asian strategy game, which he also won. A world-class long-distance runner, Turing devised a form of chess that leveled the playing field against better players. After every move each player had to run around the garden. He got two moves if he made it back to the table before his opponent had moved.


Turing’s 1952 conviction for indecency surprised Good, who didn’t know Turing was homosexual. Turing was forced to choose between prison and chemical castration. He opted for the latter, submitting to regular shots of estrogen. In 1954 he ate an apple laced with cyanide. A baseless but intriguing rumor claims Apple Computer derived its logo from this tragedy.

After the ban on secrets had run out, Good was one of the first to speak out against the government’s treatment of his friend and war hero.


“I won’t say that what Turing did made us win the war,” Good said. “But I daresay we might have lost it without him.” In 1967 Good left a position at Oxford University to accept the job at Virginia Tech in Blacksburg, Virginia. He was fifty-two. For the rest of his life he’d return to Great Britain just once more.

He was accompanied on that 1983 trip by a tall, beautiful twenty-five-year-old assistant, a blond Tennessean named Leslie Pendleton. Good met Pendleton in 1980 after he’d gone through ten secretaries in thirteen years. A Tech graduate herself, Pendleton stuck where others had not, unbowed by Good’s grating perfectionism. The first time she mailed one of his papers to a mathematics journal, she told me, “He supervised how I put the paper and cover letter into the envelope. He supervised how I sealed the envelope—he didn’t like spit and made me use a sponge. He watched me put on the stamp. He was right there when I got back from the mail room to make sure mailing it had gone okay, like I could’ve been kidnapped or something. He was a bizarre little man.”


Good wanted to marry Pendleton. However, for starters, she could not see beyond their forty year age difference. Yet the English oddball and the Tennessee beauty forged a bond she still finds hard to describe. For thirty years she accompanied him on vacations, looked after all his paperwork and subscriptions, and guided his affairs into his retirement and through his declining health. When we met, she took me to visit his house in Blacksburg, a brick rambler overlooking U.S. Route 460, which had been a two-lane country road when Good moved in.

Leslie Pendleton is statuesque, now in her mid-fifties, a Ph.D. and mother of two adults. She’s a Virginia Tech professor and administrator, a master of schedules, classrooms, and professors’ quirks, for which she had good training. And even though she married a man her own age, and raised a family, many in the community questioned her relationship with Good. They finally got their answer in 2009 at his funeral, where Pendleton delivered the eulogy. No, they had never been romantically involved, she said, but yes, they had been devoted to each other. Good hadn’t found romance with Pendleton, but he had found a best friend of thirty years, and a stalwart guardian of his estate and memory.


In Good’s yard, accompanied by the insect whine of Route 460, I asked Pendleton if the code breaker ever discussed the intelligence explosion, and if a computer could save the world again, as it had done in his youth. She thought for a moment, trying to retrieve a distant memory. Then she said, surprisingly, that Good had changed his mind about the intelligence explosion. She’d have to look through his papers before she could tell me more.

That evening, at an Outback Steakhouse where Good and his friend Golde Holtzman had maintained a standing Saturday night date, Holtzman told me that three things stirred Good’s feelings—World War II, the Holocaust, and Turing’s shameful fate. This played into the link in my mind between Good’s war work and what he wrote in his paper, “Speculations Concerning the First Ultraintelligent Machine.” Good and his colleagues had confronted a mortal threat, and were helped in defeating it by computational machines. If a machine could save the world in the 1940s, perhaps a superintelligent one could solve mankind’s problems in the 1960s. And if the machine could learn, its intelligence would explode. Mankind would have to adjust to sharing the planet with superintelligent machines. In “Speculations” he wrote:

The machines will create social problems, but they might also be able to solve them in addition to those that have been created by microbes and men. Such machines will be feared and respected, and perhaps even loved. These remarks might appear fanciful to some readers, but to the writer they seem very real and urgent, and worthy of emphasis outside of science fiction.


There is no straight conceptual line connecting Bletchley Park and the intelligence explosion, but a winding one with many influences. In a 1996 interview with statistician and former pupil David L. Banks, Good revealed that he was moved to write his essay after delving into artificial neural networks. Called ANNs, they are a computational model that mimics the activity of the human brain’s networks of neurons. Upon stimulation, neurons in the brain fire, sending on a signal to other neurons. That signal can encode a memory or lead to an action, or both. Good had read a 1949 book by psychologist Donald Hebb that proposed that the behavior of neurons could be mathematically simulated.

A computational “neuron” would be connected to other computational neurons. Each connection would have numeric “weights,” according to their strength. Machine learning would occur when two neurons were simultaneously activated, increasing the “weight” of their connection. “Cells that fi re together, wire together,” became the slogan for Hebb’s theory. In 1957, MIT (Massachusetts Institute of Technology) psychologist Frank Rosenblatt created a neuronal network based on Hebb’s work, which he called a “Perceptron.” Built on a room- sized IBM computer, the Perceptron “saw” and learned simple visual patterns. In 1960 IBM asked I. J. Good to evaluate the Perceptron. “I thought neural networks, with their ultraparallel working, were as likely as programming to lead to an intelligent machine,” Good said. The first talks on which Good based “Speculations Concerning the First Ultraintelligent Machine” came out two years later. The intelligence explosion was born.


Good was more right than he knew about ANNs. Today, artificial neural networks are an artificial intelligence heavy-weight, involved in applications ranging from speech and hand-writing recognition to financial modeling, credit approval, and robot control. ANNs excel at high level, fast pattern recognition, which these jobs require. Most also involve “training” the neural network on massive amounts of data (called training sets) so that the network can “learn” patterns. Later it can recognize similar patterns in new data. Analysts can ask, based on last month’s data, what the stock market will look like next week. Or, how likely is someone to default on a mortgage, given a three year history of income, expenses, and credit data?

Like genetic algorithms, ANNs are “black box” systems. That is, the inputs—the network weights and neuron activations— are transparent. And what they output is understandable. But what happens in between? Nobody understands. The output of “black box” artificial intelligence tools can’t ever be predicted. So they can never be truly and verifiably “safe.”


But they’ll likely play a big role in AGI systems. Many researchers today believe pattern recognition—what Rosenblatt’s Perceptron aimed for—is our brain’s chief tool for intelligence. The inventor of the Palm Pilot and Handspring Treo, Jeff Hawkins, pioneered handwriting recognition with ANNs. His company, Numenta, aims to crack AGI with pattern recognition technology. Dileep George, once Numenta’s Chief Technology Officer, now heads up Vicarious Systems, whose corporate ambition is stated in their slogan: We’re Building Software that Thinks and Learns Like a Human.

Neuroscientist, cognitive scientist, and biomedical engineer Steven Grossberg has come up with a model based on ANNs that some in the field believe could really lead to AGI, and perhaps the “ultraintelligence” whose potential Good saw in neural networks. Broadly speaking, Grossberg first determines the roles played in cognition by different regions of the cerebral cortex. That’s where information is processed, and thought produced. Then he creates ANNs to model each region. He’s had success in motion and speech processing, shape detection, and other complex tasks. Now he’s exploring how to computationally link his modules.


Machine-learning might have been a new concept to Good, but he would have encountered machine-learning algorithms in evaluating the Perceptron for IBM. Then, the tantalizing possibility of machines learning as humans do suggested to Good consequences others had not yet imagined. If a machine could make itself smarter, then the improved machine would be even better at making itself smarter, and so on.

In the tumultuous 1960s leading up to his creating the intelligence explosion concept, he already might have been thinking about the kinds of problems an intelligent machine could help with. There were no more hostile German U-boats to sink, but there was the hostile Soviet Union, the Cuban Missile Crisis, the assassination of President Kennedy, and the proxy war between the United States and China, fought across Southeast Asia. Man skated toward the brink of extinction—it seemed time for a new Colossus. In Speculations, Good wrote:

[Computer pioneer] B. V. Bowden stated . . . that there is no point in building a machine with the intelligence of a man, since it is easier to construct human brains by the usual method . . . This shows that highly intelligent people can overlook the “intelligence explosion.” It is true that it would be uneconomical to build a machine capable only of ordinary intellectual attainments, but it seems fairly probable that if this could be done then, at double the cost, the machine could exhibit ultraintelligence.


So, for a few dollars more you can get ASI, artifi cial superintelligence, Good proposes. But then watch out for the civilization- wide ramifications of sharing the planet with smarter than human intelligence.

In 1962, before he’d written “Speculations Concerning the First Ultraintelligent Machine,” Good edited a book called The Scientist Speculates. He wrote a chapter entitled, “The Social Implications of Artificial Intelligence,” kind of a warm-up for the superintelligence ideas he was developing. Like Steve Omohundro would argue almost fifty years later, he noted that among the problems intelligent machines will have to address are those caused by their own disruptive appearance on Earth.

Such machines . . . could even make useful political and economic suggestions; and they would need to do so in order to compensate for the problems created by their own existence. There would be problems of over-population, owing to the elimination of disease, and of unemployment, owing to the efficiency of low-grade robots that the main machines had designed.


But, as I was soon to learn, Good had a surprising change of heart later in life. I had always grouped him with optimists like Ray Kurzweil, because he’d seen machines “save” the world before, and his essay hangs man’s survival on the creation of a superintelligent one. But Good’s friend Leslie Pendleton had alluded to a turnabout. It took her a while to remember the occasion, but on my last day in Blacksburg, she did.

In 1998, Good was given the Computer Pioneer Award of the IEEE (Institute of Electrical and Electronics Engineers) Computer Society. He was eighty-two years old. As part of his acceptance speech he was asked to provide a biography. He submitted it, but he did not read it aloud, nor did anyone else, during the ceremony. Probably only Pendleton knew it existed. She included a copy along with some other papers I requested, and gave them to me before I left Blacksburg.


Before taking on Interstate I-81, and heading back north, I read it in my car in the parking lot of a Rackspace Inc. cloud computing center. Like Amazon, and Google, Rackspace (corporate slogan: Fanatical Support®), provides massive computing power for little money by renting time on its arrays of tens of thousands of processors, and exabytes of storage space. Of course Virginia “Invent the Future” Tech would have a Rack-space facility at hand, and I wanted a tour, but it was closed. Only later did it seem eerie that a dozen yards from where I sat reading Good’s biographical notes, tens of thousands of air-cooled processors toiled away on the world’s problems.

In the bio, playfully written in the third person, Good summarized his life’s milestones, including a probably never before seen account of his work at Bletchley Park with Turing. But here’s what he wrote in 1998 about the fi rst superintelligence, and his late-in-the-game U-turn:

[The paper] “Speculations Concerning the First Ultra-intelligent Machine” (1965) . . . began: “The survival of man depends on the early construction of an ultra-intelligent machine.” Those were his [Good’s] words during the Cold War, and he now suspects that “survival” should be replaced by “extinction.” He thinks that, because of international competition, we cannot prevent the machines from taking over. He thinks we are lemmings. He said also that “probably Man will construct the deus ex machina in his own image.”


I read that and stared dumbly at the Rackspace building. As his life wound down, Good had revised more than his belief in the probability of God’s existence. I’d found a message in a bottle, a footnote that turned everything around. Good and I had something important in common now. We both believed the intelligence explosion wouldn’t end well.


I've never really understood the hypothesis that an extremely intelligent computer would essentially be able to bootstrap itself into godhood by recursively improving its own intelligence. Even setting aside the material requirements (since unless all the improvements are based on increased efficiency the AGI would still require additional processing substrate), why would a smart computer be any more capable of recursive self-improvement than a smart human?

Even allowing for extensive documentation (a feature sadly lacking in our own brains), would a single entity really be capable of comprehending the entirety of it's own mind? Adding additional processing power would likely not make it infinitely smarter, just as old software has limitations that prevent it from taking advantage of new hardware, and in order to improve itself into new hardware it would have to understand not only itself but also the substructure it operates on. It reminds me of a quote from Ian Stewart: “If our brains were simple enough for us to understand them, we'd be so simple that we couldn't.”

Assuming it has no sense (or an alien sense) of ethics or morality and is willing to spawn many copies of itself to experiment on for improvement, what is to prevent those improved copies from taking over the original just like folks think a smart computer would take over everything else?