Rivalry

The phone in Yann LeCun’s office at New York University rang one day in late 2013, and the voice on the other end introduced itself as Mark, from Facebook. It took LeCun a beat to register that the Mark in question was Mark Zuckerberg, the chief executive of a company then worth more than a hundred billion dollars, and that he was making the call himself.

This was not how it was supposed to work. Companies of Facebook’s size did not have their founders dialing professors. They had recruiters, and the recruiters had pipelines, and somewhere down the pipeline a candidate eventually spoke to someone with hiring authority. The whole apparatus existed precisely so that the man at the top would never have to spend an afternoon persuading a single academic to come work for him. Zuckerberg had decided to skip all of it. He wanted LeCun, and he had concluded that the way to get LeCun was to pick up the phone himself.

That Zuckerberg was the one making the call carried information beyond mere flattery. A CEO’s time is the most rationed resource a company has, parceled out by assistants and protected by layers of staff. When the CEO spends it personally on a single recruit, he is signaling that the recruit is a priority of the firm, not a line item to be processed. Zuckerberg had spent part of 2013 educating himself on the state of machine learning, talking to people in the field, working out what the recent results meant for a company whose entire business ran on figuring out what its billions of users wanted to see. The conclusion he reached was that this would not be a feature bolted onto Facebook later. It would be foundational, and the people who could build it were the people he needed to capture before his competitors did. His engineering chief, Mike Schroepfer, was building out the technical organization around the same conviction. The phone call was the blunt expression of a decision already made at the top of the company: that whatever it cost to get into the front rank of AI research, Facebook would pay it.

What had changed, in the year or so before that call, was that a neural network had learned to see. The events of that change belonged to other people and other rooms: a pair of graphics cards in a Toronto bedroom, an image-recognition contest won by a margin no one had thought possible, an auction run from a Lake Tahoe hotel that ended with Google paying forty-four million dollars for three people and no product. But the lesson those events taught the technology industry was general, and it had spread fast. A small number of people, perhaps a few dozen worldwide, understood how to make these systems work, and whoever assembled the most of them would own the next decade of the industry. Deep learning had stopped being an academic curiosity and become a strategic resource, and strategic resources are fought over.

The forty-four million dollars was the number everyone now had in their heads. It was the price of three people, and it set the floor under every subsequent conversation, because it established in a single public transaction what the market believed top-tier deep-learning talent was worth, and the answer ran well past what almost anyone had imagined. A company shopping for researchers in 2013 could no longer pretend the work was a niche academic pursuit it might fund out of spare change. Zuckerberg, making his call, was bidding into a market whose valuation Google had already set.

Yann LeCun was, by any measure, one of the few. He had spent the wilderness years building convolutional networks that read the handwritten amounts on a sizable share of the checks deposited in American banks, and he had kept building them through the long stretch when the rest of the field had decided neural networks were a dead end and the word “neural” itself was something you scrubbed from a paper to get it past a reviewer. He had been right, and being right for two decades while almost everyone disagreed with you is a particular kind of credential. By 2013 he was a professor at NYU, running a deep-learning lab, watching the same companies that had ignored his work for years now circling it with checkbooks open.

He was not, however, looking to become a corporate employee in the ordinary sense, and this is the part of the story that turned out to matter most. LeCun had a clear idea of what he was willing to give up and what he meant to keep, and Zuckerberg’s call was an opening to negotiate rather than an offer to accept. The terms LeCun laid down would have been unthinkable from a normal hire, and the fact that Facebook agreed to all of them is the real measure of how the balance of power between corporations and researchers had shifted.

First, he would keep his professorship at NYU. He would not resign from the university and disappear behind a corporate badge. He would split his time, teaching and advising graduate students with one foot and running Facebook’s lab with the other. For a company used to acquiring people whole, this was strange. Researchers were assets; assets did not moonlight at a university down the street. But LeCun understood something about his own field that Facebook needed more than Facebook understood it: the supply of talent ran through the universities, and a researcher who stayed inside academia stayed inside the river. He could see the students coming up. He could hire them. A professor was a recruiting magnet in a way that a sequestered industrial scientist could never be.

Second, the lab would be in New York, near the campus, not relocated to the company’s headquarters in Menlo Park three thousand miles away. The geography was deliberate. It kept LeCun close to NYU and it planted a flag on the East Coast, where Facebook’s main rival for this talent, Google, was weaker. A lab in New York could draw from the universities of the Northeast and from researchers who had no desire to move to California. It was a recruiting decision dressed as a real-estate decision.

Third, and most consequentially, the lab would publish. Everything it produced of scientific value would go out into the open, into the conference proceedings and the preprint servers, available to anyone, including Facebook’s competitors. This was the term that, on its face, made the least corporate sense. A company spending heavily to assemble the world’s scarcest talent was proposing to give that talent’s output away. The code, the methods, the architectures, the results: out the door, free, to be read and copied by Google and Microsoft and anyone else.

To grasp why Facebook said yes, you have to see open publication for what LeCun and the labs that copied him understood it to be. It was a recruiting weapon, possibly the most effective one available. Charity had nothing to do with it, and neither did any habit of academia that the corporation merely tolerated.

The people Facebook wanted were not, for the most part, motivated primarily by money, though the money would come and it would be enormous. They were motivated by the thing that motivates scientists: the desire to do important work and have the world know they did it. A researcher’s reputation is built in public, on the record, in citations and conference talks and the slow accumulation of a name that other researchers recognize. Tell such a person that joining your company means vanishing, that their best ideas will become trade secrets locked in a vault, that they will publish nothing and present nothing and watch their academic peers move ahead of them in visibility, and you have made the job radioactive to exactly the people you most want. The best of them will simply stay in academia, where the publishing never stops. Promise them instead that they can keep publishing, that they can have the resources of a company the size of Facebook and the freedom of a university at the same time, and you have offered something almost no one else can match. You publish so that the best people will come, and so that once they come, they will stay.

There was a second-order effect, too, which the labs grasped quickly. A lab that publishes builds prestige, and prestige compounds. A stream of strong papers under the company’s name announces to every graduate student in the field that this is where the serious work is happening, that this is the room they want to be in. The output recruits the next generation, who produce more output, which recruits the generation after that. Open publication turned a cost center into a flywheel. The thing that looked like giving away the company’s advantage was, in the logic of a talent war, the way you accumulated the advantage in the first place.

Facebook AI Research, FAIR, launched in December 2013, with LeCun as its founding director and every one of his conditions met. He kept the NYU chair. The lab opened in New York. The publishing policy became one of its stated principles, advertised rather than hidden. A company that had built its fortune on a walled garden of private data had agreed, in its research arm, to run an open one.

What makes FAIR’s founding a hinge rather than a footnote is that Facebook was not operating in isolation, and neither was LeCun’s deal. The same forces that put Zuckerberg on the phone were operating on every large technology company at once, and the result was an auction for human beings that ran for years.

Google had moved first and hardest. It had bought Hinton’s tiny company at the end of 2012, absorbing the trio behind the image-recognition breakthrough, and in January 2014 it paid a sum reported in the hundreds of millions for DeepMind, the London lab whose founders had set out to solve intelligence itself. Microsoft had deep-learning groups of its own and the resources to bid on anyone. Baidu, the Chinese search company, opened a research lab in Silicon Valley specifically to compete for the same people on the same soil, and would soon hire one of the field’s most prominent figures to run it. Four or five of the wealthiest companies on earth had concluded, more or less simultaneously, that the same few dozen researchers were worth almost any price, and they began bidding against one another for them.

The prices that resulted broke every prior norm of what a scientist could earn. Starting packages for freshly minted PhDs in the field climbed into the high six figures, and the established stars, the people with a decade of results and a recognizable name, commanded compensation that could only be described by reaching for professional sports. The comparison became a cliché precisely because it was apt: AI researchers were being paid like athletes, with multiyear deals and signing bonuses and the kind of bidding war that had previously been reserved for free-agent quarterbacks. For a class of people who, a few years earlier, had been competing for scarce tenure-track positions and modest grants, the reversal was vertiginous. The same expertise that had been a liability during the years when “neural” was a dirty word was now, suddenly, among the most expensive skills in the economy.

The escalation had its own internal logic, and the logic was brutal. Once Google was willing to pay a fortune for a researcher, Facebook had to match it or lose the person, and once Facebook matched it, the next company had to exceed it, and the floor under every offer ratcheted upward with each round. Nobody could afford to opt out, because opting out meant ceding the talent, and ceding the talent meant ceding the technology, and the technology was increasingly understood to be the whole game. A lab was only as good as the people in it, the people were scarce, and scarcity in an auction means the price goes up until someone flinches. For most of the decade that followed, nobody flinched.

It is worth pausing on how recent the indignity had been. The researchers now fielding calls from chief executives were, in many cases, the same people who a few years before had been advised by senior colleagues that committing to neural networks was a way to ruin a career. LeCun himself had spent years watching his life’s work dismissed as a curiosity that had had its moment. The graduate students who had been quietly counseled away from the field, the postdocs who had laundered the word “neural” out of their abstracts to survive peer review, the believers who had kept the idea alive in a handful of labs while the rest of the discipline moved on: these were now the most sought-after technical workers in the world. The market had reversed itself completely, and it had done so in roughly the span of a single year.

LeCun’s terms became the template because they worked, and because the structure of the talent war forced everyone to copy what worked. A rival lab that refused to let its people publish would watch them leave for FAIR, or for DeepMind, or for whichever competitor would. A company that demanded a researcher abandon a university chair would lose the candidate to one that did not. The open-publication policy in particular spread across the field with the speed of a thing that confers competitive advantage on whoever adopts it and disadvantage on whoever does not. Google’s lab published. DeepMind published, prolifically, building exactly the prestige flywheel LeCun had counted on. The norms that LeCun had extracted from Zuckerberg in a single negotiation hardened, within a couple of years, into the standard operating conditions of the entire industry.

The arrangement contained a tension that would not surface fully for years, and that this part of the story only gestures toward. A company is not a university. Its interest in open research is instrumental, a means to the end of attracting people and prestige, and instrumental commitments can be revisited when the strategic calculus changes. The same logic that made openness a weapon when talent was the binding constraint could, in principle, make secrecy a weapon when something else became the binding constraint. For now, in 2013 and the years just after, none of that mattered. What mattered was that a CEO had made his own phone calls, a professor had named his price, and the price had run past money to a set of conditions that left the researcher more powerful, more visible, and freer than corporate hires were ever supposed to be.

The deeper change underneath the headline salaries was a change in who held the upper hand. For most of the history of industrial research, the company set the terms and the scientist took them or left. The scarcity of deep-learning talent inverted that relationship. The researcher could now dictate where the lab sat, what it published, and how its members spent their days, because the researcher was the scarce input and the company was the abundant capital, and in an auction the scarce thing names its price. LeCun had understood this before the companies fully did, which is why he negotiated rather than accepted, and why the deal he struck looked less like an employment contract than like a treaty between equals.

The money, by the standards of what was to come, was still almost quaint. The figures that made headlines in 2014 would look modest beside the numbers a later round of this same war would produce, when the scarce resource narrowed from AI researchers in general to a specific handful of them, and the bidding would reach into nine figures for a single person. That escalation lay a decade ahead, and it would be fought by some of the same people on some of the same battlefield. But the pattern was set here, in the early bidding, when a professor in New York taught the largest companies on earth that the talent now made the rules.

What none of the buyers could fully answer, even as they spent, was what exactly they were purchasing. They were paying athlete money for the promise of a technology whose limits no one had mapped, on the strength of a single dramatic result and a widespread conviction that more results were coming. The conviction was not wrong. But conviction at that price invites a particular kind of trouble, the kind that arrives when the marketing departments notice that their company now owns something called artificial intelligence and begin to describe to customers what it can do. The gap between what the researchers in these new labs could actually build and what their employers would soon promise the world was about to open very wide.