Hubris

“By 2020, they will have caught up. By 2025, they will be better than us. And by 2030, they will dominate the industries of AI.” — Eric Schmidt, on China, Artificial Intelligence and Global Security Summit, November 1, 2017

On the morning of May 23, 2017, in the canal town of Wuzhen, two hours south of Shanghai, the best Go player alive sat down across a board from a machine and lost in a way that was harder to watch than a defeat. Ke Jie was nineteen, ranked first in the world, and he had spent the months since AlphaGo destroyed Lee Sedol in Seoul telling anyone who asked that he would not lose. He believed it. He had studied the Korean games, had seen where Lee Sedol went wrong, had convinced himself that a human who prepared correctly could still beat the thing. He was wrong, and he knew it early. In the first of three games he played a brilliant opening, the kind of aggressive, inventive play that had made him the best, and the machine simply absorbed it and ground forward, and by the middle of the game Ke Jie’s hands were shaking. He resigned all three. After the second loss he put his face in his hands at the board and the room went quiet, and the photograph of that moment traveled around the Chinese internet faster than any official wanted it to.

What almost no one outside China saw was the part that mattered most, which is that most of China did not see it either. The government had restricted live coverage of the match. The summit was held on Chinese soil, organized in cooperation with Chinese authorities, attended by Chinese officials, and yet the live video feeds inside the country went dark, mentions on social platforms were thinned, and the foreign press in the room found that the spectacle they had come to cover was being quietly managed out of view of the people it was about. A nation that had watched the Lee Sedol match the year before in enormous numbers was not permitted to watch its own champion fall. The censorship was the tell. It said that the people running China had understood something about the machine on the board that the people who built it had not yet fully reckoned with. They had decided that a Western company humbling the pride of Chinese intellect, on live television, inside China, was not entertainment. It was a national-security event.

To understand why a board game triggered that reflex, you have to go back fourteen months, to the match this book has already described, the one in Seoul in March 2016 where AlphaGo beat Lee Sedol four games to one and played, in the second game, a move so strange that the experts called it inhuman. That match was watched by more than two hundred million people worldwide, tens of millions of them in China. It is difficult to convey to a Western reader what Go means there. It is not a hobby. It is a two-thousand-year-old emblem of a particular kind of intelligence, patient and deep and civilizational, the game scholars played, the game that stood for the mind itself. And a piece of software built by a London lab owned by an American search company had just taken it apart on a stage, in front of an audience the size of a continent. In Silicon Valley the match was a triumph of engineering. In Beijing it was a starting gun.

The two reactions reveal how differently the same event can be read depending on what you already fear. The Western coverage of AlphaGo fixated on the philosophy: the alien move in game two, the question of whether the machine was creative, the spectacle of human intuition meeting something it could not parse. The story was about intelligence in the abstract. The Chinese coverage, and more importantly the Chinese government’s reading, fixated on something narrower and more practical. A foreign company had demonstrated a decisive lead in a technology that everyone in the room understood was not really about Go at all. Go was the demonstration. The thing on display was a method, a way of building systems that learned superhuman skill from data and computation, and that method was going to be applied to economics, to weapons, to surveillance, to everything. The Americans watched the match and saw a milestone in the history of ideas. The Chinese watched it and saw a foreign power planting a flag on terrain they intended to own.

The man who would later put the clearest words to this was not a government official but a venture capitalist named Kai-Fu Lee, and his biography is its own argument about which way the talent had begun to flow. Lee was born in Taiwan, raised partly in Tennessee, educated at Columbia and Carnegie Mellon, where in the 1980s he built one of the first speaker-independent speech-recognition systems for his doctorate. He had been a star at Apple, then at Silicon Graphics, then he had run Microsoft Research’s new lab in Beijing in the late 1990s, then he had been the founding president of Google China. He was, in other words, exactly the kind of person the American technology industry had spent decades producing and exporting: a brilliant Chinese-born engineer who made his name in American institutions. And by the mid-2010s he had gone home, founded a venture firm in Beijing called Sinovation Ventures, and begun pouring money into Chinese AI startups. The flow of talent that had carried people like him to California for thirty years had reversed, and he was both a symptom of the reversal and one of its most effective promoters.

Lee’s framing, which he would lay out at length in his 2018 book AI Superpowers, was that AlphaGo had been China’s Sputnik moment. The comparison was deliberate and exact. When the Soviet Union put a satellite into orbit in 1957, the shock was not really about the satellite. It was about what the satellite implied: that a rival had quietly built the capacity to do something the United States had assumed only it could do, and that the gap was real and the response had to be national. Sputnik produced NASA, the National Defense Education Act, a generation of American children pushed into math and science by a frightened government. Lee argued that AlphaGo did the same thing in reverse, in China, and that the Lee Sedol match and then the Ke Jie match together lit a fire under a country that had been quietly preparing the kindling for years. The difference was speed. The American response to Sputnik took years to organize. China’s response to AlphaGo took months.

On July 20, 2017, less than two months after Ke Jie wept at the board in Wuzhen, the State Council of the People’s Republic of China issued a document called the New Generation Artificial Intelligence Development Plan. It was not a white paper or a think-tank proposal. It was a directive from the top of the Chinese government, and it set three deadlines with the kind of bluntness that Western technology strategy almost never permits itself. By 2020, the plan said, China would draw level with the world’s most advanced AI technology. By 2025, it would achieve major breakthroughs and lead the world in some areas. By 2030, China would become the world’s primary center of artificial-intelligence innovation, the place where the frontier was set. The plan named the technology as a strategic priority on the order of a national project, directed provincial governments and state banks and universities to align behind it, and let it be understood that the money would follow, in subsidies and government-guidance funds and the kind of patient state capital that no Silicon Valley venture firm could match.

A Western reader’s instinct is to discount this as the boilerplate of a planned economy, the sort of grand five-year proclamation that gets issued and forgotten. That instinct was wrong, and the people who held it paid for it. China had been building the substrate for years. It had more internet users than any country on earth, which meant more data, and in machine learning data was the fuel. It had a mobile-payment and super-app ecosystem, built around WeChat and Alipay, that generated behavioral data of a richness and granularity that no Western system produced, because Western consumers had not collapsed their entire economic lives into a single app the way Chinese consumers had. It had a regulatory environment, particularly around personal data and surveillance, that placed almost no friction in the way of collecting and using that data. And it had, by this point, a deep bench of engineers, many of them trained in the United States and now coming home, drawn by the money, the mission, and a government that had declared their field the future of the nation.

The data argument deserves a moment, because it was the heart of the case that China would do more than catch up, that it would pull ahead, and it was genuinely persuasive. A neural network of the kind this book has been describing learns by example. Show it ten thousand labeled images and it learns a little; show it ten million and it learns a great deal. For most of the field’s history the bottleneck had been the algorithm, the question of how to get the network to learn at all, and that was a question that American and British and Canadian labs were best at answering. But the AlexNet result and everything after it had shifted the bottleneck. Increasingly the limiting factor was not cleverness but quantity, the sheer volume of real-world examples a system could train on, and on that axis China had a structural advantage that no amount of Western ingenuity could erase. Its population was larger. Its citizens lived more of their lives on instrumented platforms. Its laws did not stop companies or the state from gathering and pooling what those platforms recorded. If the future of AI was going to be decided by who had the most data, the contest was already over, and the West had lost it without noticing.

That argument turned out to be too simple, as the later chapters of this book will show, because raw data quantity mattered less than its proponents claimed and algorithmic breakthroughs kept mattering more than they predicted. But in 2017 it was the consensus, and it was frightening, and it was the engine under everything Beijing was doing.

The companies arrived to match the plan. Baidu, the Chinese search giant, had positioned itself as the country’s AI champion, and its most visible signal had come three years earlier, when it hired Andrew Ng as chief scientist, a move this book has already described. Read in 2014 as a curiosity, it looked different against the plan that followed. The symbolism was impossible to miss and was meant to be. A Chinese company had reached into the heart of Silicon Valley and hired one of its most famous brains, to do its research on American soil, for the explicit purpose of competing with the American companies down the road.

It did not last, and the way it ended was as instructive as the way it began. In March 2017, Ng announced he was leaving Baidu. He was gracious about it, and the company was gracious back, and the official story was simply that he wanted to pursue new things, which he did, founding a series of AI ventures and education efforts. But by then Baidu had already brought in the figure who would actually run its AI ambition, and that hire was the more telling one. In January 2017, Baidu announced that Qi Lu, a senior executive who had spent years running search and other businesses at Microsoft, was joining as chief operating officer to drive the company’s pivot to artificial intelligence. The pattern that Andrew Ng had embodied, the American-trained engineer recruited home to build China’s frontier, had become a strategy with an org chart. The brains the United States had trained were being hired, one by one, to compete against the country that trained them.

Qi Lu’s path was the same path as Ng’s, run in reverse and at a higher altitude. He had grown up poor in rural China, won his way into Fudan University in Shanghai, come to the United States for a doctorate at Carnegie Mellon, and risen through Yahoo and then Microsoft to become one of the most senior Chinese-born executives in American technology, the man who had run Bing against Google. Going home to Baidu was a promotion, not a retreat: a bet that the most interesting work in his field was about to happen in the country he had left, and that bet was being made by exactly the people best positioned to know. Every such move was, in isolation, a personal decision about where to do the most exciting work. In aggregate they were a transfer of human capital that no policy in Washington had authorized and few in Washington had noticed, and the people making the moves were among the most capable engineers the American system had ever produced.

The Westerners who saw it most clearly tended to be the ones who had spent the most time inside both worlds, and the one whose warning carried the furthest was Eric Schmidt. Schmidt had run Google as its chief executive through the years when it acquired DeepMind and built Google Brain, and by 2017 he was executive chairman of Alphabet and chaired a Pentagon advisory board on innovation. He was, in short, a man positioned to see both the technology and the geopolitics at once, and what he saw frightened him. On November 1, 2017, at a summit in Washington on artificial intelligence and national security, he laid out a timeline as blunt as Beijing’s own. China’s 2030 plan, he told the room, should be taken literally. “By 2020, they will have caught up. By 2025, they will be better than us. And by 2030, they will dominate the industries of AI.” He offered it as a forecast read off the document and the data, not a possibility to be hedged, and he warned his own country that it was sleepwalking, that the United States still thought of AI as an industry it owned when it had become a contest it might lose.

Schmidt’s warning sat in the strange double position that defined the whole American response to China’s rise in those years. It was, on one reading, a genuine alarm from a man who understood the stakes. On another reading it was the oldest move in Washington, an industry leader telling the government that a foreign threat demanded more money and fewer constraints for his own sector, that the way to beat China was to unleash American companies and stop worrying so much about the things the critics in the earlier chapters of this book had begun to worry about. Both readings were true at once, and that was the problem. The China threat was real, and it was also useful, and once it became useful it became impossible to tell how much of the alarm was about China and how much was about the budget. The hubris of the American technology industry in this period lay deeper than dismissal. By the end of 2017 it had stopped dismissing China. The harder thing to shake was an assumption: that whatever China did, the answer was simply for America to do more of what it was already doing, faster, with fewer questions asked.

There was a deeper failure of imagination underneath the strategic one, and it had been building since the field’s earliest days. The American technology industry had spent a generation believing that it would always be the source of the world’s important software, that talent and capital and the frontier all naturally pooled in a forty-mile stretch of California, and that the rest of the world’s role was to adopt what the Valley invented. That belief had been roughly true for a long time, which made it harder to see when it stopped being true. The engineers in the labs had grown up inside an assumption of permanent leadership so total that they did not experience it as an assumption at all. It was simply the shape of the world. And so when a rival announced, in writing, that it intended to take the lead by a specific date and then began assembling the data, the money, and the people to do it, the instinctive American response skipped past any question of whether leadership was permanent. The challenge looked like an aberration to be corrected, a race to be won by running harder on the same track, rather than evidence that the track itself had more than one runner and always had.

What the China plan exposed was that the field of artificial intelligence had quietly become a domain of national power, and almost no one building it had decided that on purpose. The researchers in the labs in California and London and Montreal had spent their careers thinking of themselves as scientists, publishing openly, moving freely between countries and companies, attending the same conferences, sharing the same code. That open, borderless, faintly utopian culture had been the field’s defining trait through every chapter of this story, from the connectionist underground to the auction in the Lake Tahoe hotel room. And now two governments were beginning to look at that culture and see a strategic resource where the scientists had seen a commons, the way an earlier generation of governments had looked at uranium. The Chinese plan made the strategic reframing explicit. It would take Washington a few more years, and a few more shocks, to make its own.

The Chinese champions that grew up under the plan made the stakes concrete in a way the document’s abstractions could not. Companies most Americans had never heard of were becoming, in narrow domains, the best in the world. SenseTime and Megvii, whose Face++ platform could identify a face in a crowd, were building computer-vision systems of remarkable accuracy and selling them, among other places, to the security apparatus of the Chinese state. iFlytek dominated Chinese-language speech recognition. The Valley had long dismissed Chinese technology as copies of American products. These were frontier systems, in some cases ahead of anything in the West, built on the data advantage the country’s scale and surveillance had handed them. Kai-Fu Lee called China’s move from copycat to innovator a fact rather than a boast, and in computer vision and speech, by 2018, the description held. And the uses to which those systems were being put, the cameras and the checkpoints and the scoring of citizens, were a preview of a question the rest of the book would have to confront: that the same technology that read a chest X-ray or translated a sentence could also watch a population, and that the people deciding which it would do were not the engineers who built it.

The deepest irony of the AlphaGo moment was that the company that built the machine wanted nothing to do with the race it had started. DeepMind’s people talked about solving intelligence and curing disease; the geopolitics embarrassed them. They had played the game in China as a gesture of respect for a culture that revered it, and the gesture had been received as a provocation. Ke Jie, for his part, recovered. He went back to playing humans, and beating them, and he made a kind of peace with the machine that had broken him, saying later that studying its games had made him a better player, that it had shown the whole world of Go moves that no human had ever imagined. The technology that humiliated him also taught him. That doubleness, the same system as both threat and teacher, ran through everything the field touched in these years, and it was about to surface in places far more fraught than a Go board, in the photographs people uploaded and the faces the cameras learned to read. The machines had started to see. The next reckoning was over what, and whom, they saw.