The Bay Area is famed for nurturing speculative investments like flying cars, floating cities, and the notion that a ride hailing service can turn a profit. A new utopian investment opportunity arrived Monday: Shovel dollars into a San Francisco artificial intelligence lab cofounded by Elon Musk and you’ll receive a share of the profits when (or if) it figures out how to create machines smarter than humans.
That pitch comes from OpenAI, an independent AI research lab cofounded as a nonprofit in 2015 by Musk and Sam Altman, the president of startup incubator YCombinator. Its stated mission was to safely create software as capable as people, which it terms artificial general intelligence or AGI, and share the benefits with the world. The founders argued society shouldn’t have to hope that profit-seeking tech giants would do that. The project’s initial backers, which also included early Facebook investor Peter Thiel, pledged $1 billion.
Monday, OpenAI’s leaders said that a paltry $1 billion wouldn’t be enough to compete with the well-resourced AI labs at companies such as Google and Facebook after all. They announced the new investment vehicle, a company called OpenAI LP, as a way to raise extra money for the computing power and people needed to steer the destiny of AI. Musk left the board of OpenAI last February, and is not formally involved in OpenAI LP.
“In order to build AGI you need to have billions of dollars of investment” in computing resources, says Ilya Sutskever, initially OpenAI’s research director, but now chief scientist of the new for-profit, OpenAI LP. “We created this structure in order to raise this capital while staying aligned with our mission.”
AI software is often marketed as the ultimate money-saving technology, in the eyes of some posing a threat to human workers and even society. Yet developing advanced new AI systems can be extremely expensive because of the powerful computer processors used in cutting-edge machine learning projects, and the high demand for the relatively small—overwhelmingly male—band of people qualified to work on them.
Financial filings from Alphabet’s London AI unit DeepMind help illustrate the challenge. The company would burn through a billion in only a few years if it continued to operate as it did in 2017, when it spent £334 million ($442 million). OpenAI’s most recently available IRS filing, for 2016, when it was significantly smaller, show that the organization spent $11.2 million, with more than $1.9 million of that going to Sutskever, who had left a promising career in Google’s AI research team to join OpenAI.
Sutskever’s compensation won’t be a matter of public record any more. He and most of OpenAI’s employees now work for the new for-profit company, OpenAI LP. Altman is CEO. The original nonprofit entity remains as a kind of overseer to keep the new creation on mission.
OpenAI LP was established in such a way that it has a fiduciary responsibility to uphold the nonprofit’s charter, which pledges to “use any influence we obtain over AGI’s deployment to ensure it is used for the benefit of all.” The nonprofit’s board has some representation from OpenAI LP, but legal protections have been added to bar it including more than a minority with financial stakes in the partnership.
Greg Brockman, who recently transitioned from CTO at the original OpenAI to the same role at OpenAI LP, says those arrangements are designed to attract AGI-curious dollars that could not legally or practically be given to OpenAI in its original incarnation, such as venture funds with a fiduciary responsibility to seek profits for their investors. Venture capital firm Khosla Ventures and LinkedIn cofounder Reid Hoffman’s charitable foundation have committed funds to the new OpenAI LP. OpenAI declined to reveal how much they had committed. Khosla did not respond to a request for comment.
OpenAI LP employees will also be granted stakes in the company, a move intended to make it easier for top researchers to leave tech companies and their employee stock grants behind. “Everyone agrees that if we build AGI it will create orders of magnitude new value,” Brockman says. “It’s important that people can say ‘I’m going to Open AI but in the long term it’s not going to be disadvantageous to us as a family.’” Holders of stakes in OpenAI LP will be limited to a return of 100 times what they put in, with any excess going back to the nonprofit OpenAI to sustain its mission.
AI bots capable of doing any human task, but better, do seem like a solid way to make money—once they exist. Brockman theorizes a network of low-cost AI bot doctors. For now, OpenAI’s investors and employees must shoulder considerable uncertainty. Although AI technology has improved greatly, today’s systems can take on only narrowly defined problems, and can be easily tricked.
When asked how long they think it will be until AGI arrives, Brockman and Sutskever demur. “Who knows how long it’s going to take?” Sutskever says. “We’re committed to the mission no matter how long,” Brockman says, adding that people are motivated to join OpenAI for its mission, not just money.
He also says OpenAI is open to commercializing some of its technology before AGI arrives if the technology takes a long time to work out. “If the timeline is longer,” says Brockman, “we’ll do what we have to.” OpenAI LP employees may be given a chance to sell some of their stake in the company, similar to schemes at some startups.
Oren Etzioni, CEO of the Allen Institute for AI, a nonprofit established and funded by Microsoft cofounder Paul Allen, knows first-hand that it’s tough to hire top AI researchers when they’re constantly stalked by corporate recruiters. He’s less sure that using the profit motivation to act more like a tech company is the only answer. “I disagree with the notion that a nonprofit can’t compete,” he says, gesturing to influential research papers published by his own lab and OpenAI in areas such as language processing.
Making big leaps beyond state of the art techniques often requires creative thinking about novel ideas, not just putting more iron behind old ones, Etzioni says, and leaner, independent organizations can be the best at that. “If bigger and better funded was always better then IBM would still be number one,” he says.
IBM is not number one in AI right now. That crown probably belongs to Alphabet, which despite helping inspire OpenAI’s restructuring may also stand to gain from the move.
Brockman says one priority for funds freed by OpenAI’s restructuring will be buying hardware to build up its in-house computing infrastructure. He also expects to spend a lot more on cloud computing. When OpenAI put up a team of bots against professional players of the videogame Dota 2 last summer, it trained the underlying code on 128,000 conventional computer processors, and 256 powerful graphics processors for weeks at a time—rented from Google.