Like Minds

Written by David Kushner, Wired

Friday February 1st, 2008

Push Singh and Chris McKinstry had a lot in common: Both Canadian. Both coders. Both obsessed with tapping the Web to create a true artificial intelligence. And both found dead in the same strange way.

On the morning of June 12, 1990, Chris McKinstry went looking for a gun. At 11 am, he walked into Nick’s Sport Shop on a busy street in downtown Toronto and approached the saleswoman behind the counter. “I’ll take a Winchester Defender,” he said, referring to a 12-gauge shotgun in the display. She eyeballed the skinny 23-year-old and told him he’d need a certificate to buy it.

Two and a half hours later, McKinstry returned, claiming to have the required document. The clerk showed him the gun, and he handled the pistol grip admiringly. Then, as she returned it to its place, he grabbed another shotgun from the case, yanked a shell out of his pocket, and jammed it into the chamber.

“He’s got a gun! He’s got a gun!” a woman screamed, as she ran out the front door. The store emptied. He didn’t try to stop anyone.

Soon McKinstry heard sirens. A police truck screeched up, and men in black boots and body armor took up positions around the shop.

The police caught glimpses of him through the store windows with the gun jammed under his chin. They tried to negotiate by phone. They brought in his girlfriend, with whom he’d just had a fight, to plead with him. They brought in a psychiatrist — McKinstry had a history of mental problems and had tried to institutionalize himself the day before. After five hours, McKinstry ripped the telephone from the wall and retreated into the basement, where he spent two hours listening to radio coverage of the standoff. Eventually, a reporter announced that the cops had decided on their next move:

Send in the robot.

McKinstry had stolen the gun because he wanted to end his own life, but now he was intrigued. He’d always been obsessed with robots and artificial intelligence. At 4, he had asked his mother to sew a sleeping bag for his toy robot so it wouldn’t get cold. “Robots have feelings,” he insisted. Despite growing up poor with a single mom, he had taught himself to code. At 12, he wrote a chess-playing program on his RadioShack TRS-80 Model 1.

As McKinstry cowered in the basement, he could hear the robot rumbling overhead, making what he called “Terminator” noises. It must be enormous, he thought, as it knocked over shelves. Then everything went eerily quiet. McKinstry saw a long white plume of smoke arc over the stairs. The robot had fired a tear gas canister, but it ricocheted off something and flew back the way it came. Another tear gas canister fired, and McKinstry watched it trace the same “perfectly incorrect trajectory.” He realized the machine had no idea where he was hiding.

But the cops had had enough. They burst through the front door in gas masks, screaming, “Put the gun down!” McKinstry had been eager to die a few hours before, but now something in him obeyed. The gas burned his eyes and lungs as he climbed from the basement. At the top of the steps, he saw the robot through the haze. It looked like an “armored golf cart” with a tangle of cables and a lone camera eye mounted on top. It wasn’t like the Terminator at all. It was a clunky remote-controlled toy. Dumb.

Three hundred miles away in a suburb of Montreal, Pushpinder Singh was preparing to devote his life to the study of smart machines. The high schooler built a robot that won him the top prize in a province-wide science contest. His creation had a small black frame with wheels, a makeshift circuit board, and a pincer claw. As the prodigy worked its controller, the robot rolled across the floor of his parents’ comfortable home and picked up a small cup. The project landed Singh in the Montreal Gazette.

Push, as everyone called him, had also taught himself to code — first on a VIC-20, then by making computer games for an Amiga and an Apple IIe. His father, Mahender, a topographer and mapmaker who had studied advanced mathematics, encouraged the wüenderkind. Singh was brilliant, ambitious, and strong-willed. In ninth grade, he had created his own sound digitizer and taught it to play a song he was supposed to be practicing for his piano lessons. “I don’t want to learn piano anymore, I want to learn this,” he said.

Singh’s lifelong friend Rajiv Rawat describes an idyllic geek childhood full of Legos, D&D, and Star Trek. One of his favorite films was 2001: A Space Odyssey — Singh was fascinated by the idea of HAL 9000, the artificial intelligence that thought and acted in ways its creators had not predicted.

To create the character of HAL, the makers of 2001 had consulted with the pioneering AI researcher Marvin Minsky. (In the novel, Arthur C. Clarke predicted that Minsky’s research would lead to the creation of HAL.) Singh devoured Minsky’s 1985 book, The Society of Mind. It presented the high schooler with a compelling metaphor: the notion of mind as essentially a complex community of unintelligent agents. “Each mental agent by itself can only do some simple thing that needs no mind or thought at all,” Minsky wrote. “Yet when we join these agents in societies — in certain very special ways — this leads to true intelligence.” Singh later said that it was Minsky who taught him to think about thinking.

In 1991, Singh went to MIT to study artificial intelligence with his idol and soon attracted notice for his passion and mental stamina. Word was that he had read every single one of the dauntingly complex books on the shelves in Minsky’s office. A casual conversation with the smiling young researcher in the hallway or at a favorite restaurant like Kebab-N-Kurry could turn into an intense hour-long debate. As one fellow student put it, Singh had a way of “taking your idea and showing you what it looks like from about 50 miles up.”

The field of AI research that Singh was joining had a history of bipolar behavior, swinging from wild overoptimism to despair. When 2001 came out in the late ’60s, many believed that a thinking machine like HAL would exist well before the end of the 20th century, and researchers were flush with government grants. Within a few years, it had become apparent that these predictions were absurdly unrealistic, and the funding soon dried up.

In the mid-’90s, researchers could point to some modest successes, at least in narrow applications like optical character recognition. But Minsky refused to abandon the grand Promethean dream of re-creating the human mind. He dismissed Deep Blue, which beat chess grand-master Garry Kasparov in 1997, because it had such a limited mission. “We have collections of dumb specialists in small domains; the true majesty of general intelligence still awaits our attack,” Minsky is quoted as saying in a book called HAL’s Legacy: 2001’s Computer as Dream and Reality. “No one has tried to make a thinking machine and then teach it chess.”

Singh quickly established himself as Minsky’s protégé. In 1996, he wrote a widely read paper titled “Why AI Failed,” which rejected a piecemeal approach to research: “To solve the hard problems in AI — natural language understanding, general vision, completely trustworthy speech and handwriting recognition — we need systems with commonsense knowledge and flexible ways to use it. The trouble is that building such systems amounts to ‘solving AI.’ This notion is difficult to accept, but it seems that we have no choice but to face it head on.”

Singh’s ambitious manifesto prompted an encouraging note from Bill Gates. “I think your observations about the AI field are correct,” he wrote. “As you are writing papers about your progress, I would appreciate being sent copies.”

While Singh was climbing the academic ladder at MIT, McKinstry was trying to put his life back together after spending two and a half months in jail. But the suicidal standoff had given him a new sense of purpose. He liked to think that the police robot had deliberately misfired its tear gas canisters in an effort to save him “Maybe robots do have feelings,” he later mused. By 1992, McKinstry had enrolled at the University of Winnipeg and immersed himself in the study of artificial intelligence. While pursuing a degree in psychology, he began posting on AI newsgroups and became enamored with the writings of the late Alan Turing.

A cryptographer and mathematician, Turing famously proposed the Turing test — the proposition that a machine had achieved intelligence if it could carry on a conversation that was indistinguishable from human conversation. In late 1994, McKinstry coded his own chatbot with the goal of winning the $100,000 Loebner Prize for Artificial Intelligence, which used a variation of the Turing test.

After a few months, however, McKinstry abandoned the bot, insisting that the premise of the test was flawed. He developed an alternative yardstick for AI, which he called the Minimum Intelligent Signal Test. The idea was to limit human-computer dialog to questions that required yes/no answers. (Is Earth round? Is the sky blue?) If a machine could correctly answer as many questions as a human, then that machine was intelligent. “Intelligence didn’t depend on the bandwidth of the communication channel; intelligence could be communicated with one bit!” he later wrote.

On July 5, 1996, McKinstry logged on to comp.ai to announce the “Internet Wide Effort to Create an Artificial Consciousness.” He would amass a database of simple factual assertions from people across the Web. “I would store my model of the human mind in binary propositions,” he said in a Slashdot Q&A in 2000. “A giant database of these propositions could be used to train a neural net to mimic a conscious, thinking, feeling human being!”

The idea wasn’t new. Doug Lenat, a former Stanford researcher, had been feeding information into a database called Cyc (pronounced “psych”) since 1984. “We’re now in a position to specify the steps required to bring a HAL-like being into existence,” Lenat wrote in 1997. Step one was to “prime the pump with the millions of everyday terms, concepts, facts, and rules of thumb that comprise human consensus reality — that is, common sense.” But the process of adding data to Cyc was laborious and costly, requiring a special programming language and trained data-entry workers.

Cyc was a decent start, McKinstry thought, but why not just get volunteers to input all that commonsense data in plain English? The statements could then be translated into a machine-readable format at some later date. But McKinstry’s grand vision to harness the collective power of the Internet community to create an artificial intelligence had one serious flaw: The Internet community thought he was nuts.

McKinstry had been posting for years, detailing his research, his theories, and his personal life. He was known in newsgroups primarily for his outlandish rants and tall tales. He claimed to have been a millionaire at age 17. He detailed his police standoff and his experiences dropping acid (“I wandered downtown Toronto thinking and acting as if I was god”).

In December 1996, snarky geeks created a newsgroup in his honor, alt.mckinstry.pencil-dick, taking as its charter “Discussion of Usenet kook McKinstry, aka ‘McChimp.'” Leading the brigade was Jorn Barger, who would later run the site Robot Wisdom (and coin the term weblog). “You write like a teenager, and have shown frequent signs of extreme cluelessness,” Barger emailed McKinstry in May 1995.

McKinstry never shied away from a flame war. “I’m just sick of you spouting your highly uninformed opinion all over the net,” he replied to Barger. He threatened legal action against people who, in an effort to refute his theories, quoted directly from his emails. To those who made fun of his frequent misspellings, he explained that they were caused by dyslexia, not dementia.

But some of McKinstry’s improbable boasts turned out to be true. Many scoffed when he claimed to have moved to Chile to work on the world’s largest telescope, but he soon provided evidence that he was indeed an operator of the Very Large Telescope at the European Southern Observatory. “It’s funny how often I get called a liar,” he once posted. “I will no longer tolerate slander.”

The eccentric researcher made friends among the bohemians and hackers of Santiago. “Chris could make people laugh and wasn’t afraid to make a fool of himself in the process,” recalls his ex-wife. And there was one important person who McKinstry said treated him with respect: Marvin Minsky. McKinstry claimed to have emailed Minsky in the mid-’90s, asking if it were possible “to train a neural network into something resembling human using a database of binary propositions.”

“Yes, it is possible,” Minsky is supposed to have replied, “but the training corpus would have to be enormous.”

That was apparently all the encouragement McKinstry needed. “The moment I finished reading that email,” he later recalled, “I knew I would spend the rest of my life building and validating the most enormous corpus I could.”

On July 6, 2000, McKinstry retooled his pitch for a collaborative AI database. He had a business model this time, one that seemed well suited to the heady days of the dotcom boom. His Generic Artificial Consciousness, or GAC (pronounced “Jack”), would cull true/false propositions from people online. For each submission, participants would be awarded 20 shares in McKinstry’s company, the Mindpixel Digital Mind Modeling Project.

Mindpixel was a term McKinstry invented to describe the individual user-submitted propositions. Pixels, short for “picture elements,” are the tiny, simple components that combine to create a digital image. McKinstry saw mindpixels as mental agents that could be combined to create a society of mind. Gather enough of them — roughly a billion, he estimated — and the mindpixels would combine to create a functioning digital brain.

The criticisms and flames never let up. But McKinstry’s clever stock offer managed to generate mainstream press coverage and hundreds of thousands of mindpixel submissions. He posted regular messages to his “shareholders” and talked up the enormous potential value that the Mindpixel project could have if it achieved its lofty goals. “It’s like inventing teleportation,” he told Wired News in September 2000. “How could you put a value on that?”

Do fish have hair? Can blue tits fly? Did Alan Turing theorize that machines could, one day, think? Did Quentin Tarantino direct Terminator 2? Is a neural network capable of learning? Is the Mindpixel project just a scam to make Chris McKinstry famous? — Questions submitted to the Mindpixel database

Meanwhile, in Cambridge, Push Singh was pursuing a similar vision. He had teamed with Stanford researcher David Stork to create a database of commonsense knowledge through open submissions. On the surface it resembled Mindpixel, but instead of yes/no questions it compiled factual statements like “every person is younger than their mother” and “snow is cold and is made of millions of snowflakes.”

In September 2000, two months after McKinstry launched Mindpixel, Singh posted a message on the rec.arts.books newsgroup to announce Open Mind Common Sense. “We have recently started a project here at MIT to try to build a computer with the basic intelligence of a person,” it read. “This repository of knowledge will enable us to create more intelligent and sociable software, build human-like robots, and better understand the structure of our own minds. We invite you all to come visit our project web page, and teach our computer some of the things all us humans know about the world, but that no computer knows!”

But the Web community was dubious. The reason: McKinstry. “How is it any less moronic than Mindpixel?” Barger replied. Another poster agreed: “Should be obvious by now that these [AI] guys … are the most successful con artists of our time.”

“Mindpixel isn’t moronic, it’s courageous,” Singh responded. “I disagree with how McKinstrey is doing it (as a company, giving out shares’ that will never have any value, instead of making it public immediately). [And] the Mindpixel idea of training up a neural network’ with the database is clearly ridiculous. I also believe our interface is better.”

It didn’t take McKinstry long to respond. “First, no e’ in McKinstry,” he fired back. “And second, your statement is misleading. The database is publically [sic] available right now, just not for commercial use.” McKinstry bristled at Singh’s dismissal of his stock plan. “Thems fightin’ words!” And if Singh felt Mindpixel was “clearly ridiculous,” perhaps he’d be willing to bet on whose database would be first to achieve intelligence? As for the dig at his site’s interface, McKinstry conceded that Open Mind’s was better, but blamed it on his lack of resources: “You didn’t have to write it all by your lonesome.”

McKinstry insisted that Mindpixel had one significant advantage over Open Mind: He required his contributor-shareholders to verify the accuracy of each other’s submissions. “The net is a very open place,” he wrote. “How do you keep garbage out without any form of validation mechanism? … All you have to do is try to [imagine] Slashdot without the moderation system to see what’s going to happen to your database.”

McKinstry had been stung by Singh’s criticisms. But the fact that Singh called Mindpixel “courageous” and was pursuing a similar project gave him a sense of validation. His response to Barger sounds triumphant. “How many years have you been fighting this idea of mine here in these news groups?” he crowed. “Now I guess the whole MIT Media Lab is crazy too?”

McKinstry sat at his computer uploading statements to Singh’s Open Mind database. “Don Adams played Maxwell Smart. Trees don’t have neurons. Jesus was a superstar. Marvin Minsky was alive in 2001. Houses don’t eat pork.” One thought led to the next in a revealing free association. “Push is normally a verb,” he typed. “McKinstry is competing with Push.”

Actually, McKinstry was hoping to forge a partnership with Singh. In 2000, he hinted to his shareholders that Mindpixel and Open Mind Common Sense were going to connect their databases. Singh initially did nothing to dispel this impression. “Chris is a good guy,” he told Wired News.

McKinstry’s mind turned often to Singh. They had so much in common: Two young researchers obsessed with simulating common sense. Both Canadian. Both Net-savvy.

Like McKinstry, Singh was convinced that the potential of artificial intelligence was enormous. “I believe that AI will succeed where philosophy failed,” he had written on his MIT homepage. “It will provide us with the ideas we need to understand, once and for all, what emotions are.” According to Bo Morgan, a fellow student at MIT, Singh suggested that giving common sense to computers would solve all the world’s problems.

“Even starvation in Africa?” Morgan asked.

Singh paused. “Yeah, I think so.”

But Singh’s ambitions were modest and grounded compared with McKinstry’s. The man behind Mindpixel was certain that his database would become a thinking machine in the near future. The father of a son from his brief marriage in the ’90s, he sometimes referred to GAC as his second child. He believed that he would be recognized as one of the great scientific minds in history. “He thought he deserved a Nobel Prize,” says a friend who blogs under the handle Alphabet Soup. “He compared himself to Einstein and Turing. He said GAC would make him immortal.”

McKinstry meant that part about immortality literally. “The only difference between you and me is the same as the difference between any two MP3s — bits,” he wrote in an Amazon.com review of How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics. (He gave the book three stars.) McKinstry often told friends that he intended to upload his consciousness into a machine: He would never die.

Do teenagers think they know everything? Is MIT the best tech school in the world? Did HAL 9000 … go nuts and try to kill everyone? Does me got bad grammar? Does Wired magazine mostly write about different types of wire? Is death inevitable? — Questions submitted to the Mindpixel database

McKinstry’s hopes for a partnership with the MIT project were soon dashed. “McKinstry was fundamentally different than us,” Singh’s collaborator, Stork, recalled. “We thought people wouldn’t participate in the project if they were making some guy in Chile rich.”

McKinstry didn’t let it go. On July 16, 2002, he tried to reconnect with Singh, emailing him a link to a paper on language models. It suggested a way that statements submitted to Open Mind and Mindpixel could be understood by machines. “This is what I’ve been babbling inarticulately about all these years. It just needs to be trained on a corpus of validated propositions,” he wrote. The paper’s author was Canadian. “Another coincidence,” McKinstry noted.

Four days later, Singh sent an unenthusiastic reply. “Current statistical approaches are still too weak to learn complex things,” he wrote. “We need some really new ideas in machine learning that go beyond what people are doing today. It helps to have the large datasets like mindpixel or openmind, but we’re still missing the right learning component.”

Open Mind, which would eventually garner more than 700,000 submissions in five-plus years, was now part of a Commonsense Computing division at the MIT Media Lab. Singh was pursuing another research project for his PhD. He was also coauthoring papers with Minsky and presenting his ideas at conferences and symposia around the world.

Privately, McKinstry began speaking of his resentment of Open Mind. Singh’s project, he felt, had gotten all the attention simply because it was affiliated with MIT. He complained that Singh had copied his statistical model for collecting data and claimed that he had contacted a dean at MIT asking that Singh’s work be taken down. (There is no evidence to support this allegation.)

Mindpixel would eventually receive roughly 1.5 million submissions, but McKinstry’s lack of business skills had become apparent. He had lined up no commercial partners or applications and apparently had no intention of honoring any of the promises he’d made to his “shareholders.” All he had was an enormous collection of questions ranging from “Does Britney Spears know a lot about semiconductor physics?” to “Is McKinstry a media whore with no real credentials or expertise?”

McKinstry, who said he was diagnosed as bipolar, went into decline. A fight with his latest girlfriend led to a few nights in a Chilean mental hospital. His mood was briefly buoyed when an article he’d written, entitled “Mind as Space,” was chosen to run in a 2003 anthology that would feature contributions from many of the luminaries in the AI field. But as the publication of the book was repeatedly postponed, he grew more frustrated and despairing. He started wondering about his old rival again.

On January 12, 2006, McKinstry hit Singh’s personal blog. “It has been hard to give this blog any attention while finishing my dissertation,” Singh had written some six months earlier. “I am now Dr. Singh!” Singh also wrote about “some new ideas [Minsky] has been developing about how minds grow. The basic idea is called interior grounding,’ and it is about how minds might develop certain simple ideas before they begin building articulate connections to the outside world.”

New ideas? McKinstry commented on Singh’s blog that it sounded similar to a 1993 paper in the journal Cognition, and he provided a link to the PDF. On his own blog, he wrote, “The idea reminded me strongly of some neural network experiments that I replicated in 1997.” Singh never replied.

“So what exacty does a web suicide note look like?” McKinstry wrote on January 20, 2006, a week after he posted to Singh’s blog. “Exctly like this.”

He was sitting in a café near his home in Santiago, pounding the keys on his Mac laptop. He posted the message on his blog and a slightly different version on a forum at Joel on Software, a popular geek hangout.

McKinstry’s rant was florid and melodramatic. “This Luis Vuitton, Parada, Mont Blanc commercial universe is not for me,” he wrote. He talked about his history of suicidal feelings and botched attempts, and he insisted that this time things would be different. “I am certain I will not survive the afternoon,” he wrote. “I have already taken enough drugs that my alreadt weakened liver will shut down very soon and I am off to find a place to hide and die.”

The online forum members were understandably skeptical. McChimp was flinging bananas again, they figured. “Have a nice trip! Let us know if there’s anything beyond the 7th dimension!” read the first comment. “Typical of his forum,” McKinstry replied. “I am having more trouble than usual typing due to the drugs. I have to go die not. bye.” Then, “It is too late. I will leave this cafe soon and curl up somewhere.” A few minutes later: “I am leaving now. People are strating to notice I canot type and I am about to vomit. Take to go. Last post.” Later still: “I am leave now. Permanently.”

“I don’t buy this for a minute,” replied a familiar detractor named Mark Warner. It was enough to pull McKinstry back into the fray for one last flame war. “Warner, you were alway an ass,” he replied. “I have to go vomit now and take more pills.” His final post continued the theme: “I am feeling really impaired. And yes, time will tell what happens to me. I really have to get out of here. I cannot type. and want to vomit. Time to go hide.”

Three days later, on January 23, after calls from panicked friends, the police checked McKinstry’s apartment and found his body. He had unhooked the gas line from his stove and connected it to a bag sealed around his head. He was dead at age 38.

McKinstry’s few friends say he occasionally spoke of suicide, but no one knew why he had gone through with it this time. Carlos Gaona, a younger hacker who had become his protégé, raced over to the apartment and convinced McKinstry’s girlfriend to give him his laptop, his journal, the dog-eared books. And, of course, the Web was full of his thoughts, rants, dreams, and nightmares. He never got to upload his consciousness into a thinking machine, but in a sense he had been uploading himself his entire adult life. Before he died, he had replaced the home page of chrismckinstry.com with the words “Catch you later.”

One blogger wondered, “If not for his belief in the permanence of the internet, that his suicidal proclamation would remain on the World Wide Web for posterity — would Chris McKinstry be alive today?”

Others wondered how this would affect the idea of collaborative AI databases. On January 28, Bob Mottram, who had once been offered the unpaid position of chief software developer at Mindpixel, wrote in a post memorializing McKinstry: “For the present, the last man standing in this game is Push Singh.”

After completing his dissertation, Singh was offered a job as professor at the MIT Media Lab. He would be teaching alongside his mentor, Minsky, who credited him with helping to develop many of the ideas in his new book, The Emotion Machine: Commonsense Thinking, Artificial Intelligence, and the Future of the Human Mind. He would have the resources to pursue his dream of “solving AI.” Before assuming his position, though, he decided to take time off, as he told a friend, “to think.”

Everything in Singh’s life seemed to be going well. He was enjoying a relationship with a girlfriend who worked at the lab. The IEEE Intelligent Systems Advisory Board, a consortium of top AI figures around the world, had selected him as one of the top 10 researchers representing the future of the field.

But privately, Singh was suffering. He had severely injured his back while moving furniture, and though he did his best to stay engaged on campus, colleagues noticed that he was distracted. He told a friend, Eyal Amir, that there were times when he was incapable of doing anything because of the excruciating pain. Some thought it was clinical depression. Colleague Dustin Smith asked, “How much of your attention is on the pain at a given moment?”

Singh replied, “More than half.”

In The Emotion Machine, Minsky suggests that chronic pain is a kind of “programming bug.” He writes that “the cascades that we call Suffering’ must have evolved from earlier schemes that helped us to limit our injuries — by providing the goal of escaping from pain. Evolution never had any sense of how a species might evolve next — so it did not anticipate how pain might disrupt our future high-level abilities. We came to evolve a design that protects our bodies but ruins our minds.”

Four weeks after Chris McKinstry committed suicide, the police were dispatched to an apartment at 1010 Massachusetts Avenue near MIT. Inside, they found the 33-year-old Singh. He had connected a hose from a tank of helium gas to a bag taped around his head. He was dead.

Mahender Singh still has the robot that his son created in high school. “He thought that computers should think as you and I think,” he says. “He thought it would change the world. I was so proud of him, and now I don’t know what to do without him. His mother cries every day.”

“If anyone was the future of the Media Lab, it was Push,” wrote the director of the lab, Frank Moss, in a mass email on March 4, 2007. A memorial wiki page was set up, and friends and colleagues posted dozens of testimonials as well as pictures of the young researcher. “His loss is indescribable,” Minsky wrote. “We could communicate so much and so quickly in so very few words, as though we were parts of a single mind.”

Singh’s childhood friend Rawat, with whom he had watched 2001 as kids in the ’80s, posted too. “This might sound corny,” he wrote, “but I felt at the funeral that they should play Amazing Grace’ [as in] Spock’s death scene in Star Trek II, where Kirk eulogized him as being the most human’ being he had ever met in his travels.” It would have been appropriate to Push, he said, “who was at once intellectually curious and logical (or as he put it, sensible) and deeply human.”

Privately, Rawat cites a different movie. “Sometimes I think this totally ridiculous thought,” he says, “that he was bumped off like the end of Terminator 2.” He refers to the fate of the character Dr. Miles Dyson, who creates a neural network processor that eventually achieves sentience and turns against mankind. When a cyborg from the future warns of what’s to come, an attempt is made to kill Dyson before he can complete his work. Ultimately, the scientist nobly sacrifices himself while destroying his research to prevent the machines from taking over the world. “That’s a fantasy [Push] would have gotten a kick out of,” Rawat says.

Amid the grieving, there were whispers about the striking parallels between Singh’s and McKinstry’s lives and deaths. Some wondered whether there could have been a suicide pact or, at the very least, copycat behavior. Tim Chklovski, a collaborator with Singh on Open Mind, suggests that perhaps McKinstry’s suicide had inspired Singh. “It’s possible that he gave Push some bad ideas,” he says. (The rumors are likely to begin again: The fact that Singh committed suicide in nearly the same way McKinstry did has not been reported or widely known until this writing.)

Details have not been forthcoming from MIT. After initial reports in the media of an “apparent suicide” by Singh, a shroud of secrecy descended. Minsky and others in the department declined to be interviewed for this article. The school has long been skittish about the topic of suicide. MIT has attracted headlines for its high suicide rate in the past, and the family of a 19-year-old student who set herself on fire sued the school in 2002. A week after Singh’s suicide, a columnist in the student paper urged school officials “to take a more public and active role in acknowledging and addressing the problem of mental health at the Institute.” Singh’s bio page and personal blog remain online, but shortly after Wired began making inquiries, MIT took down the tribute wiki.

Many say the greatest tragedy is that neither young man lived long enough to see his work bear fruit. Recently, the Honda Research Institute in Mountain View, California, began using Open Mind data to imbue its robots with common sense. “There is a nice resurgence of interest in commonsense knowledge,” Amir says. “It’s sad that Push didn’t live to see it.”

After McKinstry’s long struggle for academic legitimacy and recognition, his “Mind as Space” article will finally appear in the book Parsing the Turing Test, whose publication was delayed from mid-2003 to this February. “McKinstry himself was a troubled soul who had mixed luck professionally,” the book’s coeditor, Robert Epstein, says. “But this particular concept is as good as many others.”

In his acknowledgments, McKinstry credits Marvin Minsky for his “encouragement of my heretical ideas”; his colleagues at the European Southern Observatory’s Paranal facility, “who tolerated my near insanity as I wrote this article”; and “of course the nearly fifty thousand people that have worked so hard to build the Mindpixel Corpus.”

McKinstry and Singh were both cremated. Singh’s sister scattered his ashes in the Atlantic, not far from MIT. McKinstry’s remains are said to be under his son’s bed in the UK. Meanwhile, someone is posting to newsgroups under McKinstry’s name. “I have always been and will always be,” one message read. “I am forever.”