Genius Makers: The Mavericks Who Brought A.I. to Google, Facebook, and the World

by Cade Metz

When he joined Google in 2015, Levine already knew Ilya Sutskever, another Russian émigré, and Sutskever introduced him to Alex Krizhevsky, who started working with the new robotics group. If he ever ran into a problem, Levine would ask Krizhevsky for help, and Krizhevsky’s advice was always the same: Collect more data. “If you have the data and it’s the right kind of data,” Krizhevsky would say, “then just get more of it.” So Levine and his team built what they called the Arm Farm.

After setting up their lab above a tiny chocolate factory in the San Francisco Mission District, researchers like Zaremba spent weeks walking around this old, rapidly gentrifying Hispanic neighborhood, debating what lofty goal they should chase. They eventually settled on two: a machine that could beat the world’s best players at a three-dimensional video game called Dota, and a five-fingered robotic hand that could solve the Rubik’s Cube.

Both projects were also conspicuous stunts, a way for OpenAI to promote itself as it sought to attract the money and the talent needed to push its research forward. The techniques under development at labs like OpenAI were expensive—both in equipment and in personnel—which meant that eye-catching demonstrations were their lifeblood.

Musk himself poached a top researcher from the lab, lifting a computer vision expert named Andrej Karpathy out of OpenAI and installing him as the head of artificial intelligence at Tesla so he could lead the company’s push into self-driving cars.

As the company developed this technology, further funding was required, so Abbeel decided he would ask the biggest names in AI. Yann LeCun visited the lab in Berkeley and agreed to invest after pouring a few dozen empty plastic bottles into a bin and watching the arm pick them without a hitch. Yoshua Bengio declined to invest. Though he had taken only part-time gigs with the big tech companies, he said that he had more money than he could ever spend and that he preferred to focus on his own research. But Geoff Hinton invested. He believed in Abbeel. “He is good,” Hinton says. “And that’s very surprising. After all, he’s Belgian.”

But by the ’90s, when Cyc showed little sign of real progress, the idea of rebuilding human intelligence was not something the leading researchers talked about, at least not in public, and that remained true for the next two decades. In 2008, Shane Legg said as much in his PhD thesis. “Among researchers the topic is almost taboo:4 it belongs in science fiction. The most intelligent computer in the world, they assure the public, is perhaps as smart as an ant, and that’s on a good day. True machine intelligence, if it is ever developed, lies in the distant future,” he wrote. “Perhaps over the next few years these ideas will become more mainstream, however for now they are at the fringe. Most researchers remain very sceptical about the idea of truly intelligent machines within their lifetime.” Over the next few years, such ideas did become more mainstream, and this was largely thanks to Shane Legg, who went on to build DeepMind alongside Demis Hassabis, and who, with Hassabis, convinced three significant figures (Peter Thiel, Elon Musk, and Larry Page) that the research was worth investing in. After Google acquired DeepMind, Legg continued to argue in private that superintelligence was nigh, but he rarely spoke about it in public, in part because people like Musk were intent on stoking concerns that intelligent machines could destroy the world. But in spite of his reticence, his ideas continued to spread.

Although researchers like Sutskever were initially reticent about voicing their views, Elon Musk did not hold back. Nor did the lab’s other chairman: Sam Altman.

Sam Altman was a Silicon Valley archetype: In 2005, he founded a social networking company as a twenty-year-old college sophomore.15 The company was called Loopt, and it eventually raised $30 million in venture capital, including one of the first investments made by Y Combinator and its founder Paul Graham. Seven years later, Loopt’s social networking service was shut down after being sold at a loss for its investors. But this was a successful exit for Altman, a trim, compact man with sharp green eyes and a particular talent for raising money. Graham soon announced that he was stepping down as president of Y Combinator, and he named Altman as his replacement, an appointment that surprised many across the family of Y Combinator companies. This made Altman an advisor to an endless stream of start-ups. In exchange for advice and capital, Y Combinator received a stake in each company, and Altman personally invested in some companies, too, becoming very wealthy, very quickly. He felt that any monkey could run YC, but he also felt that in running it, he developed an usually sharp talent for evaluating people, not to mention the skill and opportunity needed to raise large amounts of capital. During his rapid ascent, he was motivated first by money, then by the power over the people and the companies in his orbit, and then by the satisfaction that came from building companies that had a real impact on the larger world.

Altman knew that what he believed wouldn’t always come to pass. But he also knew that most people underestimated what time and rapid expansion could bring to seemingly small ideas. In the Valley, this was called “scale.” When Altman decided an idea would scale, he was not afraid to bet big on its progress. He might be wrong time and again, but when he was right, he wanted to be breathtakingly right. For him, this attitude was encapsulated by an oft-repeated quote from Machiavelli: “Make mistakes of ambition and not mistakes of sloth.”

In April of 2018, he and his researchers released a new charter for the lab,19 describing a very different mission from the one he laid out as it was founded. Altman had originally said OpenAI would openly share all its research. That’s why it was called OpenAI. But after seeing the turmoil created by the rise of generative models and face recognition and the threat of autonomous weapons, he now said that as time went on, it would hold back some technologies as it gauged their effect on the world at large.

the new OpenAI charter said—explicitly and matter-of-factly—that the lab was building AGI. Altman and Sutskever had seen both the limitations and the dangers of the current technologies, but their goal was a machine that could do anything the human brain could do. “OpenAI’s mission is to ensure that artificial general intelligence (AGI)20—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity.

Then OpenAI built a system that mastered Dota 2,23 a game that plays like a more complex version of capture the flag, requiring collaboration between entire teams of autonomous agents. That spring, a team of five autonomous agents beat a team of the world’s best human players. The belief was that success in the virtual arena would eventually lead to automated systems that could master the real world. This was what OpenAI did with its robotic hand, training a virtual re-creation of the hand to solve a virtual Rubik’s Cube before moving this know-how into the real world. If they could build a large enough simulation of what humans encounter in their daily lives, these labs believed, they could build AGI.

The hope was that researchers could change the equation with new kinds of computer chips—chips that could drive this research to levels beyond both Nvidia’s GPUs and Google’s TPUs. Dozens of companies, including Google, Nvidia, and Intel, as well as a long line of start-ups, were building new chips just for training neural networks, so that systems built by labs like DeepMind and Open AI could learn far more in far less time. “I look at what is coming in terms of new computing resources, and I plot this in relation to current results, and the curve keeps going up,” Altman says.

Two labs now said they were building AGI. And two of the world’s largest companies said they would provide the money and the hardware they would need along the way, at least for a while. Altman believed he and OpenAI would need another $25 billion to $50 billion to reach their goal.

ON March 27, 2019, the Association for Computing Machinery, the world’s largest society of computer scientists, announced that Hinton, LeCun, and Bengio had won the Turing Award. First introduced in 1966, the Turing Award was often called “the Nobel Prize of computing.” It was named for Alan Turing, one of the key figures in the creation of the computer, and it now came with $1 million in prize money. After reviving neural network research in the mid-2000s and pushing it into the heart of the tech industry, where it remade everything from image recognition to machine translation to robotics, the three veteran researchers split the prize three ways, with LeCun and Bengio giving the extra cent to Hinton.