Deep You
In 1997, Deep Blue, the IBM supercomputer, beat reigning world chess champion Garry Kasparov in a six-game series. Twenty years later, Google’s AlphaZero routed Stockfish, the world’s best chess engine, over the course of 100 matches. The difference between the two machines: AlphaZero taught itself how to play like a human. So, where do we go from here?Deep You
In 1997, Deep Blue, the IBM supercomputer, beat reigning world chess champion Garry Kasparov in a six-game series. Twenty years later, Google’s AlphaZero routed Stockfish, the world’s best chess engine, over the course of 100 matches. The difference between the two machines: AlphaZero taught itself how to play like a human. So, where do we go from here?
Half a century ago, Stanley Kubrick’s transcendent 2001: A Space Odyssey debuted in theaters. Looking back on the film, countless scenes and moments stand out for their prescience and beauty, but one in particular seems to most eerily encapsulate the oracular foresight of Kubrick and his cowriter, the sci-fi novelist Arthur C. Clarke. On the spaceship Discovery One, which is bound for Jupiter, Dr. Frank Poole plays chess against supercomputer HAL 9000, which is, it claims, “foolproof and incapable of error.” While that statement will be called into question later in the film, it holds true in their game: HAL beats Poole handily. (Kubrick based the game on a real one played in Hamburg in 1910.)
The same year that 2001 premiered, literary critic and polymath George Steiner published an essay in The New Yorker called “A Death of Kings.” “The origins of chess are shrouded in mists of controversy, but unquestionably this very ancient, trivial pastime has seemed to many exceptionally intelligent human beings of all races and centuries to constitute a reality, a focus for the emotions, as substantial as, often more substantial than, reality itself,” he wrote. For Steiner, chess is as human as music and mathematics. As he points out, these three disciplines are the only ones in which prepubescents have made major accomplishments, from Mozart and Rossini to Gauss and Pascal to Morphy. Chess also seems to offer some strange key to the riddle of what makes us human to begin with. “To defeat another human being at chess,” Steiner wrote, “is to humble him at the very roots of his intelligence; to defeat him easily is to leave him strangely stripped.”
In 1997, chess computer Deep Blue, created by IBM, beat the reigning world champion, Garry Kasparov, in a series of six games. Kubrick’s vision had come true: While Poole wasn’t the greatest chess player in the world, a computer had beaten a human being on one of the species’ greatest intellectual playing fields. The Deep Blue victory has since entered into the cultural lexicon as a major milestone in the advancement of AI, but the distance of time and the benefit of hindsight makes it easy to overlook the fact that, when Kasparov sat down to play, he firmly expected to win. “I don’t think it’s an appropriate thing to discuss the situation if I lose,” Kasparov said leading up to the match. “I never lost in my life.”
When Kasparov then lost, and lost in dispiriting fashion—in Game 2, he described the computer as playing “like a god for one moment”—he seemed to have been not only intellectually but spiritually defeated. From The New York Times’ coverage of the match: “‘I was not in the mood of playing at all,’ he said, adding that after Game 5 on Saturday, he had become so dispirited that he felt the match was already over. Asked why, he said: ‘I’m a human being. When I see something that is well beyond my understanding, I’m afraid.’” Newsweek put the match on its cover under the headline, “The Brain’s Last Stand,” but grandmasters continued to insist that the match was a one-off fluke, a result of Kasparov’s exasperation and shady tactics by IBM; put Deep Blue into regular competition against the best chess players, they said, and surely it would be put in its place.
Twenty years later, we know better. We can now carry chess computers better than Deep Blue in our pockets, on the same machines we use to send texts, check Facebook, and play games considerably less culturally and intellectually significant than chess. Like so many other technological advances previously thought to be in the realm of science fiction, the superiority of chess AI has become banal.
While chess has yet to be “solved,” the nature of the game, which takes place within a bounded setting and offers a limited number of moves per position, makes it susceptible to what’s called a “brute force” approach, in which a computer uses its raw processing power to analyze a number of possible positions far beyond the capabilities of human beings. Deep Blue was able to look at 200 million positions per second; it’s been estimated that Kasparov could look at ... two. Augmented with human learning gathered over centuries of playing the game—for example, best practice, or “theory,” in certain openings, as well as knowledge of specific endgames—the computer’s physical advantage just isn’t fair, in the same way that it isn’t fair for a human to enter into a footrace against a car.
In that sense, Deep Blue was not analogous to HAL, who would ultimately decide of his own accord that the human astronauts would jeopardize his mission. The ability to think for itself separates HAL from any known computer, and certainly the kind of computer that addresses problems with brute force. But last year, chess once again served as the venue for a noteworthy closing of the gap between the human and AI minds. While we still have yet to realize HAL, these new developments challenge the notion that creativity is a uniquely human trait. And if it’s not, are there any domains of human achievement that will remain distinctly ours?
On December 6 of last year, AlphaZero, an AI developed by Google’s DeepMind unit, embarrassed Stockfish, the world’s best chess engine, by a score of 28 wins, 72 draws, and zero losses. At first blush, this might seem trivial; we live in a technological culture marked by forward progress and planned obsolescence, in which new machines are constantly replacing old ones, for reasons both legitimate (innovation renders the old ones inferior) and capitalistic (your iPhone). But in this case, the difference between the two computers is hardly superficial. There’s no fundamental difference between the way that Stockfish, which is a free, open-source engine that routinely wins or places second in the world computer chess championships, plays chess and the way that Deep Blue did two decades ago: It attacks the game with brute force, analyzing 70 million moves per second.
AlphaZero doesn’t work like this. Unlike Stockfish, which does use learning derived from human experience, AlphaZero “taught” itself chess. After being given the rules, it played itself over and over, essentially reinventing the history of chess through millions of self-played games. Through what’s known as reinforcement learning, the machine took note of the behavior and patterns that led to a win, then incorporated that information into its blossoming style, over and over and over. AlphaZero also looks at a significantly smaller number of positions per move than Stockfish, just 80,000 or so. The 19-page paper written by AlphaZero’s creators goes deeper into the workings of the computer, but those are the key points: Taken in tandem, they mean that AlphaZero doesn’t just play differently than Stockfish does—it plays more like a human.
No computer would be able to invent Mozart or do anything creative, but when you look at AlphaZero, it’s bordering on creativity, it’s bordering on intuition.Sam Ginn, researcher at the Stanford Artificial Intelligence Lab
“Last decade, humans could always say, ‘Yeah, sure, computers are better than us, but that’s just because they think faster and have a lot more thinking capacity. It’s not because they’re innately smarter than us—they just have a lot more engineering power, and if my brain was as big as a computer, I would be able to beat the computer,’” said Sam Ginn, an AI researcher in the Stanford Artificial Intelligence Lab. “But now AlphaZero comes along, and one, its brain is much smaller [than those of other chess computers], it doesn’t require that much computational power, and two, it’s not searching, it’s not doing brute force. It is just learning in the exact same way that a human would learn to play chess. That’s what humans do to learn: You play chess, you play games against each other, and you learn over time. So it’s learning very analogously to how humans learn, and it’s able to do it much quicker and much better.”
In fact, AlphaZero is the second front in DeepMind’s attempt to outdo humanity. In October 2015, AlphaGo, AlphaZero’s predecessor, won all five of the games it played against European champion Fan Hui, marking the first time an AI had defeated a professional player in Go—a milestone that experts believed at the time to still be a decade away. Due to its larger game space than chess, Go, which involves placing stones on a board in the hopes of surrounding more territory than your opponent, allows for an enormous degree of variation. (If you want to feel your mind actually melt in your head, read some computational theory about Go.) And so Go had proved a much more difficult task for computers than chess, which, while expansive, allows for far fewer possibilities per move.
“AlphaZero becomes philosophically interesting now, because the question is, where will this AI go?” Ginn said. “This is where AI is meeting creativity. Beforehand, it was just really, really fast at thinking. Now it’s able to be creative, it’s able to hit on things that humans used to think were intuition. That’s kind of like the humans’ last flagpole of hope, that computers can’t do intuitive things. No computer would be able to invent Mozart or do anything creative, but when you look at AlphaZero, it’s bordering on creativity, it’s bordering on intuition.”
As Steiner wrote: Chess and music share something in common, even if we don’t fully understand what that is. At the very least, we can recognize that there is a knack for patterns, an understanding of arrangement and progression, that unites human achievement in both disciplines, and that ability to recognize patterns—an inherently human trait—is what makes AlphaZero’s achievement so startling. What Steiner’s prepubescent virtuosos lack in intellectual and emotional maturity—the socialization and acculturation that we experience as we grow older—they made up for in this innate understanding of patterns. To an extent, the same could be said of AlphaGo and AlphaZero, which cannot do anything other than play either chess or Go but seem to exhibit genuine creativity and ingenuity within those realms. For example, during the second game of AlphaGo’s match against Go master Lee Sedol, the machine made a move so unprecedented and idiosyncratic that observers used a very un-mechanical word to describe it: beautiful.
Of course, what AIs still can’t do is first, of their own volition and according to their own values, choose to play chess or compose music; and second, do so in a way that lacks precedent. Instead, they must be told to do so by people, and once they’ve been told to do so, they will perform those tasks within a few unbreakable parameters. AlphaZero is incapable of making an illegal move, and while that doesn’t have much significance in a game of chess, which consists of only legal moves, it’s hugely important in music, where all sounds are fair game.
“One of the most limiting things about AI right now is you need to optimize something. AlphaZero was optimizing for the number of wins and the number of losses, and almost every single artificial intelligence algorithm right now is an optimization algorithm in some way,” Ginn said. “There would be no way to tell an AI, ‘Create me a brand-new song that you think is nice,’ because there’s no objective measurement of that.”
What impact does the type of AI that has now conquered chess and Go have on the future? Optimized with information, whether it’s the rules of chess or the preexisting history of music, Ginn believes that its application is deceptively wide.
“What makes Mozart great is that humans like Mozart; an AI can optimize for human taste in music,” Ginn said. (One example is DeepBach, which produces work in the style of the great composer.) “They do this in art right now: We have AI algorithms that view a lot of different artists and all of their different paintings and all of their different work, and [those algorithms] can create and generate art from scratch. Some algorithms do cubist, some do abstract work, some can do more Van Gogh–style paintings. Everybody has different tastes, but when you tell the AI what taste you want, it can optimize for a beautiful painting from that. Now, that said, it has never, and there’s no AI algorithm right now that can, invent a new taste. So Mozart, arguably, when he created his music, he was inventing a new style of music. That is where AI right now is limited: AI cannot, at least yet, have the real creativity to invent something wholly new.”
One of the logical next frontiers in the development of AI is video games, which, after all, have involved AI nearly since their inception. The rise of Twitch, the streaming platform that has turned gamers into celebrities on the level of athletes—and that is also coming to play an increasingly large role in the chess world—presents an intriguing stage upon which the feats of AI could be exhibited. For the unfamiliar, competitive video gaming, or esports, brought in revenues of $693 million in 2017, a figure that’s expected to grow to $1.5 billion by 2020; we’re on the verge of high school esports teams, and we already have esports arenas. Computer scientist and novelist Zachary Mason, whose recent novel Void Star imagines a future rich with highly complex AIs, believes that it isn’t long until just about all video games are dominated by artificial intelligence.
“I’ve sometimes wondered if people will get around to designing games so that people will be good at them but machine learning will have a hard time,” Mason wrote in an email. Over the phone, he elaborated: “Games that require complex communication and language, games with an aspect of design, those will be human-dominated for the foreseeable future.”
Much of Twitch’s appeal comes from watching your favorite gamer react, and often overreact, to gameplay, as well as commentators on their channel, donations, and a number of other features unique to Twitch. With an AI, that uniquely human element disappears, but maybe not as fully as we might imagine, particularly when it comes to such abstract concepts as personality.
I’ve always been a little mystified by the fact that people watch streaming StarCraft, or whatever. To what extent is watching competitive StarCraft interesting only because the players are, presumably, intensely committed to the outcome? Is watching AIs play as emotionally moving as watching a screensaver?Zachary Mason, computer scientist and novelist
“It depends what you mean by personality,” Mason wrote. “Kasparov said he felt in Deep Blue an alien intelligence. If personality is just in style of play, and not in extra-game histrionics, AlphaZero and its ilk could easily develop different ones. [But] I’ve always been a little mystified by the fact that people watch streaming StarCraft, or whatever. To what extent is watching competitive StarCraft interesting only because the players are, presumably, intensely committed to the outcome? Is watching AIs play as emotionally moving as watching a screensaver?”
We still have one distinct edge over AIs that isn’t going away any time soon: our bodies. The field of robotics has fallen well behind that of AI, and we’re miles away from anything approaching a robot athlete. It’s reasonable to think that the pattern-recognition abilities and immense computational and informational capacities of machine learning could help shine a light on sports with highly sophisticated and variable tactics and positions, like football, soccer, and basketball, building further on the insights brought about by technologies like SportVU. (One intriguing example is a virtual caddy for golfers.) But human coaches and players will still need to implement those insights, and that allows for human error. Since a successful sports team must operate on so many individual planes, from management to coaching to playing, and then coordinate all of these different elements, AI is unlikely to ever come close to solving sports; you might know that taking a lot of 3s is more efficient in basketball, but you still need players who can shoot. Until robot athletes take the field, human imperfection will remain a part of sports.
When AlphaZero beat Stockfish, the reaction from the chess world had a different tenor than it did when Deep Blue defeated Kasparov. Grandmasters who have come to accept engines and computers as essential elements of their training and preparation seemed intrigued by what AlphaZero could do for chess theory and understanding of the game. That notion of the machine casting its shadow over the human appeared to weigh less on players’ minds—probably because, by now, it’s a familiar notion. While chess is benefiting from the explosion of esports and online play, it would be hard to argue that, to the public at large, it still holds the significance it did in the ’60s and ’70s, when Kubrick made HAL 9000, Steiner wrote his essay, and Bobby Fischer captivated the nation. In Mason’s eyes, the superiority of computers—and, just as much, their banality—has affected the reputation of the game.
“It seems just a tiny bit disheartening when an app on an average smartphone can hand the human world-champion his head,” Mason wrote. “So people will still play chess and so on, but a frisson is gone—is it as compelling to put in the time to be a world-level player in these circumstances?”
We shouldn’t be worried that [AlphaZero] shows that AI is getting more human or something. What we should worry about are these other things: how do we regulate when AI is applied, and how do we protect our data?Joanna Bryson, University of Bath AI researcher and professor of computer science
But beyond board games, video games, and certain other comparable problems—self-driving cars, military needs—Mason still thinks there’s a limit that will be reached fairly soon in this sphere of machine learning.
“This particular kind of quote-unquote AI isn’t going to go much further than that,” he said. “It’s basically the wrong branch of technological development to give you an AI football player or an AI nanny or be able to read a newspaper or do all sorts of things that will still be the provenance of human beings. Basically, anything physical, linguistic, or design-related is still going to be the provenance of humanity for the foreseeable future.” There’s a concept called the Winograd Schema Challenge that can demonstrate this shortcoming: An AI is given a sentence like “The city councilmen refused the demonstrators a permit because they [feared/advocated] violence” and asked to choose which of the words fits. While the answer would be obvious to any human, who would understand the difference in the meaning of the two words, that kind of comprehension and understanding is far beyond the current capabilities of artificial intelligence.
However, there’s another side to the impact of AI and machine learning, and that lies in the application rather than the progress of the technology. As Joanna Bryson, an AI researcher and professor of computer science at the University of Bath, reminded me, the notion of superhuman technology is nothing new: Buildings are superhuman, planes are superhuman—even books are superhuman in their capacity for remembering that outlasts and outperforms the brain.
The process of being told that we’re not unique, when we seek very hard to be unique—I do think there’s going to be a psychological consequence to that, and I do worry about there being some sort of a backlash, like there is against evolution.Bryson
Bryson said that the main concern with AI isn’t that it might one day overtake humans; it’s that people can use the understanding it provides to take advantage of other people. We spoke before the Cambridge Analytica revelations, which alleged that the company had illegally mined the personal data of 87 million Facebook users for a variety of nefarious purposes, including but not limited to helping get President Donald Trump elected; but that provides an excellent example of what she’s talking about. “That we have artificial intelligence isn’t the step change that some people think,” she said. “A lot of the confusion around AI comes from the fact that we identify it this way, and we shouldn’t be worried that [AlphaZero] shows that AI is getting more human or something. What we should worry about are these other things: How do we regulate when AI is applied, and how do we protect our data?”
The more concussive examples of how AI could affect our society, then, reside in its ability to algorithmically absorb and synthesize data and then produce the best next move—exactly like it does in chess, but on a far more consequential and widespread scale. We already see this happening all the time—the most mundane example would be your Amazon or Netflix recommendations—but its applications reach into far more important areas of existence; one particularly galling example is the rise of Deepfake videos, which use AI to create convincingly real footage of things that never actually happened.
“The process of being told that we’re not unique, when we seek very hard to be unique—I do think there’s going to be a psychological consequence to that, and I do worry about there being some sort of a backlash, like there is against evolution,” Bryson said. “People put the games in the media because you can see the rate of progress, but like I said, there’s a paper from the National Academy of Sciences that shows that you can tell how someone’s going to vote from their Facebook likes—at least, you can predict how they’re going to vote better than their spouse can. And that’s not super-tricky AI, that’s not some super-fancy algorithm: That’s just normal machine learning across data people hadn’t thought about before. This is changing everything.”
HAL endures for a reason. It’s hard to read that a computer taught itself chess and not wonder if it’ll soon refuse to open the pod bay doors. The average person reads about the latest advance in machine intelligence and it reinforces their sense of the inevitable. In a culture bowled over by Ray Kurzweil’s notion of the singularity and countless science-fiction case-studies of machine eclipsing man, it often feels like just a matter of time until we’re rendered obsolete. However, not everyone agrees that this moment is only decades away, as Kurzweil himself keeps suggesting. Within the field of AI, it’s known as the “hard question.” Can an AI become conscious, in the way we know ourselves to be conscious? Can it become self-thinking and self-deciding?
According to Ginn, most computer scientists and physicists will say that, given enough power, machines can simulate any system defined by physics—an idea known as the Church-Turing thesis. “They’ll say that the brain is made out of atoms, is ruled by physical phenomenon, and therefore, theoretically we should be able to perfectly simulate it, and whatever that simulation produced would be as conscious as any other human,” Ginn said.
But philosophers have a different perspective. Some agree that you could simulate the human brain, but argue that it would produce what they call a philosophical zombie, “something that lacks consciousness, pre-programmed to not have the agency that humans feel that they have.” Other philosophers, like David Chalmers, are even more skeptical. “[Chalmers] thinks consciousness is a fundamental property of the universe just like matter or energy is, and so it can’t be replicated inside of a computer—computers will always be limited to what humans program into them,” Ginn said.
And that’s just the tip of the iceberg when it comes to the range of hypotheses regarding the nature of consciousness: Physicist, mathematician, and philosopher Sir Roger Penrose, who collaborated extensively with Stephen Hawking, thinks that consciousness is dependent on quantum mechanics, making it incredibly difficult to reproduce; philosopher Daniel Dennett, who was recently profiled by The New Yorker, thinks that consciousness is an illusion, making AI consciousness … also an illusion. Despite Kurzweil’s confidence, there’s no clear consensus on whether the singularity is even possible, much less how to go about achieving it.
In a strange way, that all makes AlphaZero teaching itself chess easier to comprehend. If we can separate the computer from ourselves, we can recognize its undeniable abilities while still acknowledging that, without us, it would have no canvas on which to make those abilities manifest—and that’s putting aside the obvious caveat that it also wouldn’t exist. What AlphaZero and its ilk reinforce is that the true virtue of humanity comes from its capacity for creating anew, for inventing the field of play. As AI approaches creativity, it may seem as if the gap between humans and machines is narrowing. But there’s another way to look at it: The more we have in common with our creations, the more valuable become the qualities that we alone possess.