19 Japan Wakes the World Up to AI
Sometime in Spring 1982, Ed Feigenbaum and I were having one of our phone chats, me lamenting that a book on computer graphics I’d begun was about to be pulled out from under me by the publisher who proposed it because his underwriter was going bankrupt.
Ed said, you know? The Japanese are doing something really interesting. They’re putting big resources into the next generation of AI. It really might vault them ahead.
Ed sent me a few documents describing what the Japanese hoped to do, and he was right. This was big news. The Japanese Fifth Generation group had decided the time was right to move ahead significantly in AI, design an epoch-making computer to run it, and now had the backing of mighty MITI, the government Ministry of International Trade and Industry. I wrote a proposal for my new agent and went to Utah on a ski vacation.
The Fifth Generation would turn out to be a grand adventure and awaken much of the world to AI. Even the First Culture would finally deign to turn its head and notice, with mixed results.
The moment was good for AI, which was getting more and more public notice. The moment was also good for anything about Japan, because in the early 1980s, a number of books emerged to claim how much smarter the Japanese were in manufacturing and trade than Americans. Ed and I were going to tell the world that their computers might be smarter too.
The Japanese had based their next generation system on Feigenbaum’s recently discovered and empirically proved knowledge principle: intelligent behavior arises from specific and deep knowledge, not just reasoning power, as discussed in Chapter 13.
My new agent, John Brockman, called me at a ski lodge to tell me the proposal had already gone out to sixteen publishers and an auction for the right to publish it would take place on March 31. “Are you skiing?” he asked. “No,” I said, “I’m inside reading Bleak House.” “Good. Keep safe.” I thanked him, went back to my breakfast, and tried to gather my wits.
When I got back to New York, I smiled to see that several of the publishers who planned to participate in the auction had turned down Machines Who Think. Times had changed. Addison-Wesley won the auction and was enthusiastic, urging us on. That spring and early summer I worked feverishly. On my small electric typewriter—I’d buy my first computer with the royalties from The Fifth Generation—I hammered out a first draft. Ed and I collaborated in a way congenial to us both, him with some big ideas (and much first-hand knowledge of Japan, its computer industry, and its education system); me with the questions, awkward and otherwise; the beady eye; the skepticism; a sense of larger context and connections; fascination with the people involved—and willing to write. At the end of July, I made my first visit to Japan, a country I’d love evermore.
In Tokyo I rendezvoused with Ed and his Japanese-born wife, Penny Nii, who had received her masters’ degree in computer science at Stanford and was Ed’s intellectual as well as life partner. She was a knowledge engineer, a vital part of the team to develop any expert system then. A knowledge engineer extracted the knowledge from the head of a human expert and recast it into an executable computer program called an expert system. This task is now automated over large data sets. (Ed claims that Herb Simon was the first knowledge engineer, a man who extracted all the chess expertise from Adriaan de Groot’s book on chess masters and recast it as a chess-playing program.)
The headquarters of the Japanese project, called ICOT (Institute for New Generation Computer Technology), was in a generic high-rise with inspiring views of Mount Fuji in the distance. Japan’s government had dragooned eight of the leading Japanese electronics companies to participate in the project, each contributing researchers. Not all firms participated willingly. The project leader, Kazuhiro Fuchi, wouldn’t allow the reluctant companies to fob off second-rate researchers; he had final say on who’d work at ICOT. He was quite un-Japanese, the strength of will emanating from him like a force field. Although he received Ed, Penny and me in his nicely furnished office with the snow-capped Fujiyama out the window, I knew from one of his researchers, Toshi Kurokawa (he and his wife, Yoko, had earlier translated Machines Who Think into Japanese) that Fuchi usually worked in a crude little cubicle where he could oversee his troops. Despite Fuchi’s formidable will, the eager talent of these young scientists, and the backing of MITI, the whole thing seemed to me terribly fragile.
Over the next few days, Ed, Penny and I visited participating firms. At Hitachi, we heard that they hadn’t wanted to join in this wild scheme; MITI had “made them do it,” whatever that meant. They saw themselves as “followers of IBM,” and only when IBM felt AI was worth doing would they also willingly do it. At the other end of the spectrum was NEC, determined to do everything to make the Fifth Generation succeed.
Ed and Penny were demigods in Japan, and I was considered a kind of retainer in their wake. No wonder. Expert systems, with their real-world applications, were sweeping AI in Japan as well as the United States (and elsewhere), so a visit from Feigenbaum and Nii was a descent to earth of beings who normally dwelled in celestial regions.
Ed and Penny repaid this rapturous devotion by offering infinite patience with their disciples (they had many in every company we visited). After each talk, each demo, they asked careful questions, gave guidance, and abundant encouragement. The general ideas of expert systems weren’t too difficult to grasp; you could attack problems at many levels of expertise and at the end have something to show for it. Thus the most gifted people in AI did the innovative work but left opportunities for the less brilliant to make useful contributions.
However, this method only worked when the deities came around and did regular evangelism, patting, encouraging, and applauding. “Yeah,” Ed said wearily late one night. “Sometimes I feel like the slab in 2001: A Space Odyssey. I come down from time to time to see how everybody’s doing. I see they’re not there yet, but I tell them to keep at it and go away until the next time.”
On one of our many long car trips from an outlying firm back to Tokyo, Ed told me that it had taken much missionary work to make people see the value of actual knowledge to make thinking machines successful. “I took my cue from Herb Simon,” he said. “Anytime Herb had an opportunity, he’d write a paper, or give a talk, popularize his ideas, and put them into language that a given audience could understand, tell them why it was significant to them. It was—it is—exhausting work, but the only way to have an impact.”
The Fifth Generation was fun to write. Ed and Penny knew much about Japan, and I was learning. Within eighteen months, I brought the finished manuscript to my agent John Brockman at a local falafel stand—John was never big on fancy literary lunches—carrying both copies in a couple of Zabar’s shopping bags.
The editorial back-and-forth was more daunting than usual, because our assigned editor was determined to snuff from the manuscript every possible sign of life. V. S. Pritchett had long ago told me that I’d been lucky to publish my first book in England. English editors welcomed the writer’s idiosyncratic voice and tried only to make sure the prose was reasonably clear. American editors were “ridiculously meddlesome.”
The whole process was somewhere between melodrama and opera buffa. The publishers had bought our names and ideas but wanted to write their own book. The editor was ungifted and oppressive. (“Is editorial heaven,” I asked this man sweetly, “a place where manuscripts appear without the inconvenience of authors?”) The revisions he insisted on were so awful that from a Caribbean holiday, Ed emailed me: “I’ve dragged this thing around like a dead dog: I can’t read it, and I can’t get rid of it.” Matters got so bad that Ed and I finally told the editor-in-chief that we were withdrawing the book and would of course return the handsome advance.
This got editorial attention in a useful way. The oppressive editor was fired (“I was leaving anyway”) and I attempted restoration. Writing with original verve is one thing, but I soon realized why resurrection is properly considered a miracle. Even the final product appeared with a blunder so big on the printed cover that we insisted the publisher recall the book and fix it. Eventually, in authorized and unauthorized editions, in many languages around the world, the book would sell very well.
In the long run, the Fifth Generation didn’t quite turn out as the Japanese hoped. Some argue it was before its time (its multiple levels of programming anticipated deep learning). Others say that the evolution of off-the-shelf components made obsolete the special-purpose machines the Japanese proposed to build or that the programming language chosen was too cumbersome.
In the only English-language evaluation of the Fifth Generation, Ed Feigenbaum and Howard Shrobe (1993) of MIT laid out the project’s detailed technical achievements and failures. To wit, the project did little to advance the state of knowledge-based systems, AI as such. Its natural language goals and other human interface goals were dropped, and its hopes of useful applications did not materialize. The Fifth Generation project’s research and development of parallel reasoning machines, as opposed to linear machines, was almost unique and very useful to parts of AI that require heavy signal processing (vision, speech, robotics). However, lack of parallelism wasn’t the biggest challenge for AI; the lack of ways to deal with large-scale knowledge bases was a much bigger problem. (This was to change, but not for another decade or so.)
But ICOT’s achievements showed that Japan could innovate in computer architecture; at peak performance, its specialized machines reached the original goals the project set. Above all, the project created an attractive aura for AI, knowledge-based systems, and innovative computer architecture. “Some of the best young researchers have entered these fields because of the existence of ICOT,” Japanese scientists reported. Japanese roboticists are now world leaders.
And yet. After the deadly Tohoku tsunami, when the Fukushima Daiichi atomic power plant melted down in 2011, Japanese robots should have been on the spot. Unfortunately, government agencies that guided and funded research had believed a disaster on the scale of Three Mile Island or Chernobyl could never happen in Japan; thus government decision-makers saw no point in financing research into robots that could withstand high levels of radiation. It was a tragic miscalculation with ghastly consequences.
Without counting the human cost (as if you could), Fukushima Daiichi is a giant demolition project that will require an estimated 40 years to complete and cost $15 billion. Three years after the meltdown, demolition work began. Robots that could climb over debris were deployed but had to be controlled by cables that got easily tangled because radioactivity interferes with wireless transmission. Planned next were robots that could cut through obstacles and pick up debris, followed by janitor robots that used high-pressure water jets and dry ice to clean wall and floor surfaces (Strickland, 2014).
The spent fuel rods must be removed, the radioactive water contained, and finally, the three damaged nuclear cores removed. That job alone might take twenty years. For the most part, humans control the robots I’ve described, and improved autonomous robots might be more successful—but nothing can speed the half-life of the radioactive cores. In 2017, a small (shoe-box sized) radiation-hardened robot was built and deployed to find the melted down fuel cores. Like an aerial drone, Manbo used tiny propellers to navigate through the radioactive water used to cool down the reactors and finally found and videoed the three reactors whose cores had melted down during the disaster (Fackler, 2017).
Alarmed by the failure of conventional robots at the time of the disaster, the U.S. Department of Defense’s Advanced Research Projects Agency (DARPA) conducted a grand competition in robotics between 2013 and 2015, aimed at producing radiation-proof autonomous robots to go into catastrophic sites like Fukushima, open doors, move debris out of their path, turn off valves, climb ladders, connect a firehose to a hydrant, and perform other difficult tasks in a human environment. The competition kept roboticists up late all over the world, subsisting on Cokes, pizzas, and instant ramen, working to win. Roboticists at SCHAFT, a firm owned by Google but originally Japanese, handily won the midterm round of challenges in December 2013, but in June 2015, a team from South Korea’s Advanced Institute of Science and Technology won the grand competition (and its prize of $2 million) with a humanoid robot, DRC-HUBO. In 44 minutes and 28 seconds, the robot completed all eight of the competition’s tasks flawlessly (Guizzo & Ackerman, 2015).
As we know, catastrophe can happen anywhere.
One reason, though hardly the main one, the Japanese were keen to ramp up AI in the early 1980s was a stark demographic fact: the Japanese population was growing older rapidly, and this large cohort of elders must be cared for. Perhaps because of this, certainly because I thought our book was getting tech-heavy and needed some levity, I introduced the Geriatric Robot.
The great thing about the Geriatric Robot, I wrote, is that it doesn’t just clean you up, feed you, or wheel you out into the fresh air. The great thing is that it listens. Tell me again, it says, about how wonderful/awful your kids are. Tell me again, it says, about that great coup of ’73. It listens patiently and sincerely—again and again. It isn’t hanging around to inherit your money or because it can’t get any other job. You are its job. It doesn’t get distracted or bored. It doesn’t judge you. It’s an attentive caregiver who will be there long after your biological family has lost its serenity or your hired help is fed up. “We humans can’t help it,” I added. “It’s part of our charm” (Feigenbaum & McCorduck, 1983).
In the past few years, I’d often been invited to give talks to college students and needed to illustrate AI with something that would be vivid to people that age, and better, make them laugh. My dear friend, the novelist Hortense Calisher, who was in her seventies, thought the Geriatric Robot was hilarious and ought to find a wider audience. If Hortense, at her age, didn’t find it offensive, then I imagined other people wouldn’t either. In the book, I flagged it with all sorts of rhetorical signals that I was just kidding.
Ed took a look at it and said, maybe not. I insisted. The editor excised it. But that editor had tried to throttle every sign of life the manuscript showed, so I put it back. Fun or not, it seemed appropriate, given the Japanese plans to meet responsibilities of eldercare with AI.
The Geriatric Robot was a small part of the book—nothing compared to other, more significant challenges raised by the Japanese, which soon brought an invitation to Feigenbaum to testify before a Congressional hearing—but for the First Culture, it was proof positive that AI people, me included, occupied in the Great Chain of Being the level of insensible brutes.
That’s just above the plants.