Dreamland

“To advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return.” — OpenAI, introductory announcement, December 11, 2015

In the late summer of 2015, Ilya Sutskever was one of the most valuable researchers alive, and he was unhappy about being recruited away from a job he loved. He had helped build AlexNet in a Toronto bedroom. He had gone to Google with the rest of Geoffrey Hinton’s group when the company won the Lake Tahoe auction. At Google Brain he sat near the center of the most capable machine-learning operation on earth, with all the compute he could ask for and colleagues who were the best in the world. The people now trying to pull him out of it were a thirty-year-old startup investor who had never run a research lab and a serial entrepreneur whose previous ventures were rockets and electric cars. Their pitch was that he should help them build a counterweight to the company that already employed him.

The dinner is the part everyone later remembered. It happened at the Rosewood, the hotel on Sand Hill Road in Menlo Park where venture capitalists do their deals, in the summer of 2015. Sam Altman, then the president of the startup accelerator Y Combinator, had organized it. He had invited a small group of people who cared about the long-term trajectory of artificial intelligence, and the conversation kept circling the same anxiety: the most powerful AI in the world was being built inside a handful of profit-seeking companies, and almost no one outside those companies had any say over how it would be used. Elon Musk was there. He had spent the previous year warning, loudly and in public, that advanced AI might be the greatest existential threat humanity faced. The two of them had been talking for months about whether the answer to a dangerous technology controlled by a few corporations was to build a different kind of organization to develop it in the open, for everyone.

Altman was an unlikely person to be assembling an AI lab, which was part of his appeal. He was not a researcher and made no pretense of being one. He had dropped out of Stanford to start a location-sharing company called Loopt, sold it for a modest sum, and then risen, by way of unusual energy and an unusual instinct for whom to bet on, to run Y Combinator, the accelerator that had seeded Airbnb, Dropbox, and Stripe. His talent was for organizing ambitious people around an idea and finding the money to keep them fed. He had spent the previous months convinced that artificial general intelligence was both achievable and dangerous, and that the right response was to stop wringing his hands and build the institution he wished existed. He was, in temperament, an operator who had decided to operate on the largest available problem.

Sutskever listened. He was sympathetic to the worry. But he had Google’s resources, Google’s salary, and, when it came, Google’s counteroffer, which was reportedly close to twice what the new outfit could pay. For weeks he wavered. Altman kept after him. The recruitment mattered out of proportion to a single hire because of what Sutskever represented. He was one of the three names on the AlexNet paper, a co-author of the result that had triggered the whole gold rush, and he carried the imprimatur of Hinton’s lineage. Landing him would tell every other researcher in the field that the new lab was a serious scientific enterprise and not a vanity project funded by famous men. Losing him to Google’s counteroffer would have told them the opposite. The founders understood this, which is why the pursuit went on for weeks and why Altman treated it as the decisive test of whether OpenAI could exist at all.

What finally moved Sutskever, by his own later account, was the mission. The new place had less money, not more, and although he would get a grand title, that was not it either. What pulled him was the mission, the specific people who had signed up to pursue it, and the sense that this was a chance to shape the most important technology of the century from inside an institution built for that purpose rather than as a side project of an advertising business. In November, Sutskever told Google he was leaving. Google tried again to keep him. He left anyway. The departure was a quiet repeat of the lesson the Lake Tahoe auction had taught three years earlier: the best people in this field could not simply be bought and kept, because the thing they wanted most was not always for sale.

On December 11, 2015, the organization he had joined introduced itself to the world. It was called OpenAI, and it was a nonprofit.

The announcement read like a manifesto written by people slightly embarrassed to be writing a manifesto. OpenAI’s goal, it said, was to advance digital intelligence in the way most likely to benefit humanity as a whole, unconstrained by any need to generate a financial return. Because the research would be free of financial obligation, the lab could chase a positive human impact for its own sake. It would collaborate openly with others, publish its work, and share its patents. The structure was the message. Every other serious AI lab in the world was either a university department starved of compute or a corporate division that answered, in the end, to shareholders. OpenAI was trying to be a third thing: an institution with the resources of a company, the openness of a university, and the time horizon of neither.

The roster was the proof that the thing was real. Altman and Musk were co-chairs. Greg Brockman, who had just left the payments company Stripe, where he had been chief technology officer, took the same title at OpenAI and ran the place day to day. Brockman had built much of Stripe’s engineering organization from a handful of people into one of the most admired teams in the Valley, and he had walked away from it because he had decided that AI was the only thing worth working on and that he wanted to be present at the start of something rather than maintaining something already built. He was the kind of engineer who would later be described as having personally interviewed nearly every early hire. Sutskever was research director, the scientific center of gravity, the catch that told other researchers the lab was serious. Beneath them was a founding research team assembled with the same logic Facebook and Google had used in their hiring wars, except pointed the other way. Wojciech Zaremba had turned down offers from both Google and Facebook to come. John Schulman was a young Berkeley reinforcement-learning researcher. There were others: Trevor Blackwell, Vicki Cheung, Andrej Karpathy, Durk Kingma, Pamela Vagata. Some of them took real pay cuts. Brockman had spent part of the autumn flying candidates out to a retreat at a winery north of San Francisco, asking each of them, in effect, to bet a career on an organization that did not yet exist and could not match their market price. Most of those he asked said yes, which told him something about how the field’s best people weighed money against the chance to work on the thing that mattered most to them.

The number attached to all of this was a billion dollars. The announcement said its backers had committed one billion dollars to OpenAI, and the figure did exactly the work the founders wanted it to do. It told the world this was not a hobby, and that the lab could compete with anyone for talent and machines. The list of backers was its own statement: Altman, Brockman, and Musk personally; the LinkedIn co-founder Reid Hoffman; Jessica Livingston of Y Combinator; the investor Peter Thiel; Amazon Web Services; and the Indian technology firm Infosys. It is worth being precise about what the billion dollars actually was, because the precision matters to the rest of the story. It was a pledge, not a bank balance. It was the sum the backers had said they were prepared to give over a span of years, not money that had been wired into an account. The amount OpenAI actually spent in its early years was a small fraction of it, and the gap between the announced figure and the delivered one would become one of the sources of the strain that eventually split the founding partnership. But in December 2015 the billion was a flag planted in the ground, and the flag said: a nonprofit can play at the frontier.

The reasoning behind the openness was not naive, though it was easy to read as naive. Musk and Altman had absorbed the argument that ran through the previous several years of worry about AI, the argument that if a sufficiently powerful system were ever built by a single actor pursuing its own narrow interest, the rest of the world would have no recourse. They thought slowing or stopping powerful AI was impossible. The conclusion they drew was that it should instead be distributed. If the technology was going to be built regardless, better to build it in the open, with the results shared, so that no one company and no one government could corner it. The threat was concentration. The antidote was diffusion. Whether that logic held up was a question the field would argue about for the next decade, and OpenAI itself would later quietly abandon much of the openness in its name. But in 2015 the position had a clean internal coherence, and it gave a generation of safety-minded researchers an institution to belong to.

There was a specific company on the founders’ minds when they talked about concentration, and they were not subtle about it. The thing they feared most was a single firm reaching transformative AI first and keeping it private, and the firm best positioned to do that was Google, which by 2015 held DeepMind in London and the absorbed Toronto group and the deepest compute on earth. Musk in particular had grown alarmed about DeepMind specifically and about the prospect of one company controlling a technology he believed could end the human story. Building a credible rival, even a nonprofit one, was a way to deny Google a monopoly on the future. There was something almost paradoxical in the structure of the worry: the safest outcome, as Musk and Altman saw it, was more AI rather than less, spread more widely, so that no single actor could run away with the prize. The fear that had filled the previous year of public warnings, the talk of summoning a demon, did not resolve into caution. It resolved into a billion-dollar bet that the way to be safe was to build faster, in the open, with the right people, before anyone less careful built it in secret. The doom argument had become an institution.

It also gave the press an irresistible story, because of who was bankrolling it. Musk’s involvement guaranteed coverage, and Musk supplied the quotable lines. He had spent 2014 telling audiences that with artificial intelligence humanity was summoning the demon, and now he was helping fund a billion-dollar effort to build the very thing he feared, on the theory that the only safe way to handle a demon was to make sure everyone had one. The contradiction was obvious, and the founders did not pretend otherwise. The bet was that the danger came from a single uncontrolled superintelligence in private hands, and that the way to avoid that future was to ensure many capable systems existed in many hands, openly. It was, in its way, the same instinct that had governed nuclear strategy in the previous century, repackaged for software.

One line in the early OpenAI thinking captured the premise more cleanly than the mission statement did. The lab’s founders did not believe the giants had some secret sauce that could not be reproduced. They believed the opposite, that the talent was the thing, that a small group of the right people with enough machines could match a corporation’s best research division. It was not, the sentiment went, that the people at Google drink different water. The brains were portable. They had just been shown to be portable, in fact, by the very migration OpenAI represented. Sutskever had walked out of Google Brain and the lights had not gone out at Google Brain, and now the lights had come on at OpenAI. If intelligence was a matter of people and compute rather than one firm’s proprietary magic, then the org chart of the field was not fixed. It could be redrawn by anyone willing to write the checks and make the calls.

That, more than any single technical result, is what the founding of OpenAI did. It redrew the org chart and then froze it into a shape that would hold. For the previous three years, since the Lake Tahoe auction, the structure of serious AI had been settling. Google had bought Hinton’s group and then, in early 2014, bought DeepMind in London for a sum reported in the hundreds of millions, giving it two of the strongest research operations in existence under one corporate roof. Facebook had stood up FAIR at the end of 2013 with Yann LeCun directing it on terms that let him keep his university chair and publish everything. Microsoft and Baidu had their own labs and their own war chests. By the time OpenAI announced itself, the American side of the field had hardened into a small set of well-funded poles: Google with DeepMind, Facebook with FAIR, the corporate research arms at Microsoft and elsewhere, and now a nonprofit designed as a counterweight to all of them. That configuration, four or five serious players each with its own theory of what it was for, would prove durable. Companies would be founded and acquired and renamed in the years that followed, but the basic geometry, a handful of richly resourced labs racing one another while arguing about whether the race itself was wise, was set by the end of 2015.

There was a fifth pole, and it was not in California. While the American labs were bidding one another up for researchers, the most consequential single hire of 2014 had gone east. On May 16, 2014, Andrew Ng, one of the few people whose name carried weight on both sides of the Pacific, announced that he was leaving his Stanford and Silicon Valley footing to become chief scientist at Baidu, the Chinese search company often described in the Western press as the Google of China. Baidu gave him a budget, a mandate, and, pointedly, a research presence in Silicon Valley itself, staffed in part with engineers the American companies would have liked to keep. A Chinese firm had planted a flag in the heart of American technology and hired American talent to work on Chinese problems.

Ng’s move was reported in the United States mostly as a curiosity, a star professor taking an interesting job abroad. Read against what came later, it was something larger. It was the moment the talent war stopped being an American family argument. The flow of researchers, which until then had moved among Google and Facebook and Microsoft and the universities, now had a destination on the other side of the world. And it sat uneasily next to the founding gesture of OpenAI, because the same instinct ran through both. The opportunity at Baidu, as Ng described it, was access to a scale of users and data that was hard to match anywhere else; for a researcher who believed data and compute, more than theory, would decide the next decade, the logic of going where the data was had a certain force. That was OpenAI’s bet too, pointed the other way.

What neither side liked to dwell on was that openness was symmetric. A paper published in Mountain View could be read in Beijing as easily as in Menlo Park, and a model whose weights were released could be downloaded anywhere there was an internet connection. The same logic that made OpenAI a counterweight to Google inside the United States made the entire American field legible to anyone else who was paying attention. Ng’s move was an early sign that the board on which the game would be played was not national but global, and that the openness the Western labs prized as a safeguard might also be the channel through which their lead leaked away. What that meant for the balance of power between the two countries was a question the next decade would force, and it would be argued most sharply by a venture investor named Kai-Fu Lee, who had watched the Ng hire and concluded that the West’s assumption of a permanent lead was already wrong.

None of this was obvious at the time, and it would be dishonest to pretend the founders saw it whole. In December 2015, OpenAI was a few dozen people in an office in San Francisco’s Mission District, with a billion-dollar promise and almost nothing to show for it yet. It had no product. It had published no landmark result. It had a research director who had bet his career on it, a chief technology officer who had quit a great job to run it, two famous co-chairs whose partnership would not survive three years, and a mission statement that asserted a great deal and proved none of it. By the most ordinary measure, it was a startup that had not started.

In its first months it behaved less like a company than like a graduate seminar with a famous benefactor. There was no agreed-upon research agenda, because no one was certain what the fastest path to general intelligence even looked like. Some of the team thought the answer was reinforcement learning, agents that learned by trial and error in simulated worlds. Others thought it was unsupervised learning, models that taught themselves the structure of the world from raw data without being told what to look for. The early work spread across video games and robotic hands and toy environments, much of it published openly, none of it yet pointing at a clear destination. What held the place together was a conviction rather than a plan, shared by everyone who had taken the pay cut to be there, that whatever the path turned out to be, they wanted to be the ones walking it.

But the shape was now set. There was Google, holding both DeepMind and the descendants of Hinton’s group, the deepest bench and the most compute. There was Facebook, with LeCun’s open lab. There was Microsoft, and there were the others. There was a nonprofit founded on the conviction that the safe response to a dangerous technology was to build it faster, in public, before anyone less careful could. And across the Pacific there was a credible challenger, with one of the field’s best-known scientists, a war chest, and a thesis that the future belonged to whoever could deploy at the largest scale. The players were seated. The money was committed, or at least pledged. The question of who owned intelligence had been answered, provisionally, in the only way it could be: by a small number of institutions that had each decided, for reasons of profit or fear or ambition, that they would be the ones to find out what intelligence could do.

What it would actually do, once it was loosed on the world rather than benchmarked in a lab, was a separate matter, and a darker one. That same year, in a lab in Montreal, a graduate student had already worked out how to make a neural network generate images of things that did not exist, faces of people who had never been born. The board was set. The game was about to get strange.