The OpenAI Coder Who Spun Tech On its Head
Fixer. Feather ruffler. Coder in chief. OpenAI President Greg Brockman wears a lot of hats at the buzzy startup. But his toughest job may be keeping its bigger rivals at bay after jolting them awake with the breakout success of ChatGPT.
THE TAKEAWAY
• OpenAI’s president has equity in for-profit unit, unlike CEO Altman
• Brockman played a key role in developing GPT-4
• Ex-Stripe CTO has a reputation for rapid execution, pushing for small improvements
Greg Brockman needed a win.
In 2017, Brockman—then chief technology officer of OpenAI, a 50-person nonprofit at the time—was worried about getting left in the dust by bigger rivals like DeepMind, an AI lab owned by Google’s parent company, Alphabet. He also was concerned that OpenAI might fall into the trap that had ensnared other artificial intelligence labs, where researchers fiddled around with projects that were never put into practice.
So Brockman spearheaded a plan to get OpenAI’s researchers and engineers working harmoniously together to create software that could play Dota 2, a complex online battle game. Although both groups had similar duties—researchers focused on training new models, while engineers built software to make those models work—there had been tensions between the two on other projects at OpenAI, a common dynamic inside tech companies, former OpenAI employees said. After months of all-nighters, their software eventually aced the game, and by early 2019 it beat the world’s best human Dota 2 players in thousands of games, generating buzz among gamers and AI researchers.
The success of that effort became the template for creating OpenAI’s biggest sensation yet: ChatGPT, which has captured the tech industry zeitgeist over the past several months. While the chatbot wasn’t based on any radical breakthroughs in AI research, its magic came from a single-minded focus on translating ivory-tower white papers into systems that can perform real-world tasks. And Brockman, more than any other figure at OpenAI, has helped sharpen that focus by pushing technical staff of all stripes to work together on key initiatives, people who know him say.
“If you look at a lot of AI labs in history, people sat around and thought about ideas and [made] really great research breakthroughs, but it’s a difficult thing to take that and actually make it into a product,” said Adam D’Angelo, CEO of question-and-answer site Quora, who sits on OpenAI’s board of directors along with Brockman.
“Greg is the person who’s going to make it actually happen,” D’Angelo said.
While Sam Altman, OpenAI’s CEO, is the company’s public face, Brockman is its secret weapon. He is OpenAI’s president, but his daily responsibilities bear little resemblance to those of most tech leaders with that title. Brockman, 34, has no direct reports, freeing him of the usual drudgeries of management. Instead, he spends about 80% of his time coding, he told The Information in an interview.
“Greg has got a great ability to see both the extreme details of every part of the technical problem, down to the metal at every level, and then also think very strategically about how all the pieces need to come together,” Altman said.
Brockman is an itinerant figure, wandering between different teams to set goals and push them for incremental improvements in software efficiency. He then hops to new projects, leaving them in the hands of managers. “I’m churning up dirt constantly,” Brockman said.
There are downsides to his approach. After he codes a key part of a project on his own, it can take months for other OpenAI engineers to figure out exactly how Brockman solved technical problems, said two people who have worked with him. In early 2022, Brockman insisted on rewriting from scratch a major piece of software the company used to train its models, forcing the rest of OpenAI’s employees to adapt to the new system, one of the people said. Inserting himself into projects and overruling other managers by dictating a change in the team’s priorities could rub colleagues the wrong way, concedes Brockman, who said he has worked on his interpersonal skills to learn to ruffle fewer feathers.
“The failure mode of the way I work could be that people feel thrashed,” he said.
Still, Brockman’s impact explains why others revere him. People who have worked with him describe Brockman as the quintessential “10X engineer”—Silicon Valley slang for a programmer who is 10 times as productive as the average worker. That work has translated into real business for OpenAI, which is now on pace to generate hundreds of millions of dollars per year from selling access to its software, as The Information has reported. Aravind Srinivas, who worked at Google and OpenAI before founding an AI-powered search engine, Perplexity, last year, describes his former colleague’s impact bluntly.
“The hardest-working person at OpenAI is Greg Brockman,” Srinivas said.
Brockman’s combination of coding, product decision-making and team wrangling is a rarity in tech, Srinivas added. “There’s no such person at Google,” he said. (A Google spokesperson declined to comment.)
Brockman’s fear of several years ago—that OpenAI might not be able to compete with bigger rivals like DeepMind—didn’t come to pass. Instead, the surge in interest and sign-ups for ChatGPT caught Google and DeepMind off guard, leaving the much bigger company playing catch-up with development of its own AI chatbot, Bard. And OpenAI, which launched a for-profit arm in 2019, has a powerful ally in one of Google’s biggest rivals. Microsoft, which has struggled to turn its own AI research into revenue-producing products, has poured billions of dollars into the startup to gain access to its technology, valuing the seven-year-old firm at more than $20 billion.
But things can change quickly in tech. Lumbering tech giants have reinvented themselves before and could do so once again, to OpenAI’s detriment. Brockman and his colleagues at the startup could also fail to deliver on their ultimate vision of developing artificial general intelligence—AGI— which refers to systems that display human-level intelligence. Such a system would be far more advanced than ChatGPT, with the ability to do everything from groundbreaking research to starting companies.
More ominously, some AI researchers are frightened that AGIs could eventually break free from their human masters and wreak havoc. But Brockman is more optimistic about humans’ ability to build AI that won’t be a threat. “We would be naive to think that the way we think about problems today is going to be the same way we think about them five years from now,” he said.
College Dropout
Brockman grew up in Thompson, N.D., a town of about 1,100 people near the Minnesota border, where he distinguished himself in science and mathematics. As a student at Red River High School, he earned a silver medal for his performance at the 2006 International Chemistry Olympiad. After maxing out the math curriculum at his high school by 10th grade, he began taking most of his classes at the University of North Dakota in Grand Forks.
“My principal required me to take three classes a semester at the high school, and then I could take anything I wanted at the university,” Brockman said.
Brockman enrolled in 2008 at Harvard University, where he planned to major in math and computer science (he had learned to code while spending a gap year abroad in Russia and China). But when he felt he couldn’t learn anything more about software from his peers at the school, he dropped out of Harvard. He transferred to the Massachusetts Institute of Technology, where students “knew way more about software than I did,” Brockman said.
He didn’t stay long there, either. After setting his sights on launching his own software company, Brockman met entrepreneurs Patrick and John Collison, who were working on a fledgling payment startup, which they later named Stripe. He decided to talk to them about a job, even though he had his doubts about their business.
“I was pretty sure it wasn’t possible to succeed, and that it would be necessary for us all to wait until we were older to have any real impact,” Brockman said. “But I was curious to find out which of us was wrong, so I took the meeting.”
Brockman was the one who was wrong. Dropping out of MIT on a Thursday in 2010, he was on a plane to join the Collisons’ company in San Francisco by the next Monday. Brockman spent five years at their company, later named Stripe, eventually becoming its CTO during a period in which it saw explosive growth and became one of the most highly valued startups in the tech industry.
Becoming a Fixer
The genesis for OpenAI was a dinner in late 2015 at the Rosewood Hotel, a popular gathering place for tech investors and entrepreneurs in Menlo Park, Calif. Brockman attended, along with a dozen or so other tech luminaries, including Tesla CEO Elon Musk and then-Google research scientists Ilya Sutskever and Dario Amodei. Altman, who had met Brockman through Patrick Collison, had brought the group together to discuss assembling some of the world’s top AI researchers to create a new nonprofit lab to develop AGI, Brockman said.
At the time, Altman pitched the project as a critical counterweight to the AI brainpower technology giants like Alphabet were hoarding. Putting advanced AI in the hands of powerful corporate interests could lead to a technology that benefited only a select few instead of humanity at large, Altman believed.
Brockman, who had left Stripe recently, was all in on the project from the start. “As we were leaving, Sam and I looked at each other and said, ‘Let’s do it. Let’s actually start this lab,’” Brockman said.
OpenAI’s earliest backers—including Brockman, Altman, Musk, LinkedIn co-founder Reid Hoffman, Y Combinator founding partner Jessica Livingston and PayPal co-founder Peter Thiel—collectively pledged $1 billion to fund the new nonprofit. Brockman, who became OpenAI’s CTO, took the lead on recruiting researchers to staff the new organization, inviting a group of candidates to join him in Napa Valley over a weekend to pitch them on the project, said one former OpenAI employee.
In OpenAI’s early days, his responsibilities were fluid—and sometimes mundane. Brockman spent his first day on the job scrounging around for a whiteboard to write on during a brainstorming session with fellow OpenAI co-founders Sutskever and John Schulman, a computer scientist who joined from University of California, Berkeley.
“That was what I thought my job would be—that I would be just the fixer, the guy who will say, ‘Whatever it is that you need to be better at what you’re doing, I will go do that,’” Brockman said. “It took a couple of weeks for me to find the first engineering problem, and I became a fixer for that.”
In OpenAI’s first years, Brockman and Sutskever made most of the day-to-day decisions, with Brockman managing the startup’s software engineers and Sutskever overseeing its researchers, two people who were there at the time said. Meanwhile, Brockman was on the phone constantly with Altman, who was running Y Combinator at the time, Brockman said. Altman and Musk were co-chairs of the company’s board.
One puzzle Brockman had to solve was getting its researchers and engineers playing from the same score. Some OpenAI engineers minimized the significance of researchers’ contributions, while its researchers saw engineers as technicians rather than scientists, a reflection of the experience many of them had working in larger corporate labs, former employees said.
But the Dota 2 project spearheaded by Brockman helped break the ice. To push the effort along, Brockman spent hours on the phone with Dota 2’s developers to figure out how to make the software for playing the game work, a technical challenge given that the game and OpenAI’s algorithms were written in different programming languages, three people involved said.
OpenAI’s engineers also struggled to apply their machine-learning algorithms to a game with a seemingly limitless number of variables, from how players moved to features on the game’s map. They eventually figured out how to do it through careful engineering and expensive computing power.
Those advances would also help Brockman as he became a key figure in the development of OpenAI’s generative pre-trained transformer models, former employees said. Work on those models, which generate humanlike answers in response to text-based prompts, also began in 2017, after Google researchers published a research paper titled “Attention Is All You Need.” The paper showed how a deep-learning model called a transformer could help people build larger models more efficiently.
While GPT didn’t rely on any research breakthroughs at OpenAI, it helped the startup build a product that would not break down with millions of people using it, one person said. Bent on optimizing the speed and cost of OpenAI’s systems, Brockman would “obsess about every few gains that can be had,” said Perplexity’s Srinivas, who worked on GPT at OpenAI.
“It’s an engineering innovation in the sense that they got to an unprecedented scale,” said Yoshua Bengio, a professor of computer science at Université de Montréal, who has published some of the foremost research on a branch of machine learning that enabled transformers and GPT.
AI Arms Race
In late 2017, Musk held a meeting at OpenAI’s office to share a concern: The startup was moving too slowly and was at risk of losing to DeepMind, whose technology he feared was on the brink of accelerating beyond reach, according to one person who attended the meeting.
Musk suggested that Tesla, because it was a for-profit company, would be better positioned to develop AGI because of its ability to raise more capital. Some OpenAI employees responded that by doing so, Musk would be feeding an AI arms race in which companies prioritized speed over safety as each sought to reach AGI first. Musk left OpenAI’s board in 2018.
Musk didn’t respond to a request for comment. Altman declined to comment on the meeting.
Despite those disagreements, Altman and Brockman took the Tesla founder’s comments to heart, especially as costs of training AI software were rising. OpenAI was spending roughly $34 million per year on Google Cloud services in 2018, a figure that doubled in 2019, The Information previously reported. At one point, the team working on the Dota 2 project alone was spending over $2 million a month on computing power—more than 90% of the company’s monthly computing budget at the time, said a person with direct knowledge of the effort.
But to develop AGI, OpenAI would need much more capital—billions of dollars. Brockman recalled encountering heavy skepticism from investors reluctant to put millions into a research nonprofit with no hope of financial upside for them. “We just realized it wasn’t possible and that we would need to pivot entities,” he said.
The result: OpenAI in March 2019 created a for-profit subsidiary with the potential for a payday for outside backers, though returns would be capped at 100 times their initial investment.
Altman took no equity in the for-profit arm, as The Information first reported. For his part, Brockman said he did take equity (he didn’t respond to an email asking how much). Brockman’s stake may limit his influence over OpenAI in one respect: He is a board director at OpenAI, and the company’s bylaws specify that only board members without financial stakes can vote in situations where the interests of the nonprofit’s and the for-profit subsidiary’s investors may diverge.
Inside OpenAI, some employees were uneasy about the creation of the for-profit arm. Some felt OpenAI was abandoning its ideals to develop AI systems that would be free of corporate influence. In the end, most went along with the change because they felt the company’s leaders were attempting to strike a balance that would allow OpenAI to still responsibly pursue its goal of creating AGI.
As OpenAI has continued to refine GPT, some of its employees have become disenchanted with Brockman’s approach to releasing new products. ChatGPT, for example, has been blasted by some critics for spewing misinformation and bizarre behavior.
In late 2020, a group of employees led by OpenAI’s vice president of research, Dario Amodei, left to launch a rival called Anthropic. According to Brockman, the dispute with Amodei and other Anthropic co-founders hinged on ideological differences about how long to wait before releasing AI software. Brockman preferred to ship products and use customer feedback to improve them.
“This picture [that] you can solve these problems in a vacuum…it’s a fallacy,” Brockman said.
Amodei didn’t respond to a request for comment.
When ChatGPT launched in November, its success caught Brockman off guard. “For us, we look at all the flaws—it doesn’t work for this, doesn’t quite work for that, doesn’t work for this—but then you kind of miss the fact that because it’s so useful for everything, that actually lots of people are going to find surprising utility in it,” he said.
The response touched off a scramble across the tech industry to come up with similar chatbots. Alphabet's two rival AI labs, DeepMind and Google Brain, began cooperating for the first time in order to try to catch up to OpenAI. In March, Anthropic released Claude, a ChatGPT competitor. In the meantime, Musk earlier this year recruited DeepMind researcher Igor Babuschkin to launch a new “anti-woke” AI lab within another company Musk owns, Twitter.
Meanwhile, Brockman continues to push OpenAI’s products forward. In March, he conducted a demonstration of GPT-4, the company’s latest AI model, showing how it could generate code for a website based only on a rough sketch of the site’s design. The demonstration exceeded the capabilities of any model publicly released by another tech company.
“I don’t think the whole project would have shipped near this level of quality without his involvement,” Altman said.