Author | LI Shuiqing
Edit | Heart
Why was the world's popular app ChatGPT born in OpenAI?
Recently, an in-depth article by foreign media The Information tried to explain why. According to multiple interviews, a five-year ChatGPT product development landing story surfaced, and the first promoter behind it also appeared.
Back in 2017, when OpenAI was a 50-person nonprofit, Greg Brockman, then chief technology officer of OpenAI, took the lead in a protracted battle for fear of being left behind by rivals such as Google's DeepMind.
The ChatGPT project team trial was in the Dota 2 game, and a very new way of collaboration was born, and Brockman "disrupted" various departments as the "wandering president", writing code 80% of the time. OpenAI CEO Sam Altman and other board members called him the number one advocate for the productization of ChatGPT, "a rarity in tech" and "no such person at Google." Still, he created confusion in OpenAI's code, bringing new trouble.
Greg Brockman (left) and Sam Altman (right)
If Altman is OpenAI's hallmark, Brockman is OpenAI's secret weapon.
Brockman, a standard Silicon Valley legend who came out of a small town and dropped out of Harvard and MIT to start a business, was involved in the OpenAI startup plan at the Rosewood Hotel dinner in 2015 and became OpenAI's most powerful "problem solver." Among the many projects, he has excelled at "optimizing the speed and cost of OpenAI systems" to the extreme, including the ChatGPT project.
The progress of the ChatGPT project has not been smooth sailing. In 2017, OpenAI faced great pressure due to its non-profit nature, and Elon Musk, one of the founders, withdrew and turned to AI development based on the commercial company Tesla. Brockman started the war on DeepMind in this context, and like the problems faced by all AI startups, problems such as doubling the cost of AI development and "ideal or bread" are in front of Brockman. Vice President Dario Amodei and others led the departure of the team, which also made the work more difficult, and anxiety and worry spread within OpenAI.
Sam Altman (left) and Elon Musk (right)
But Brockman's fears didn't come true, and instead, ChatGPT's surge in sign-ups in 2023 caught Google and DeepMind off guard to catch up with Bard, their own AI chatbot.
Microsoft has struggled to turn its AI research into revenue-generating products and has poured billions of dollars into OpenAI to acquire its technology, putting OpenAI at a 7-year valuation of more than $20 billion.
Looking ahead, general-purpose artificial intelligence (AGI) systems will be much more advanced than ChatGPT, capable of doing everything from groundbreaking research to starting companies. Brockman is optimistic about the prospects of this technology and will continue to step up research and development with the team. "It would be naïve to think that the way we think about things today will remain the same five years from now," he said. ”
First, the birth of ChatGPT: traced back to the Dota 2 game project, the president wandered "spoiler"
The success of a product is not the credit of one person, but there is often a competent person who is mainly driven by it. Behind the success of ChatGPT's product, it is known that nearly a hundred of the world's top scientists and engineers have contributed, but little is heard about how the people behind it are in control.
1. AI game software project, providing a model for ChatGPT's research and development project
In 2017, when OpenAI was a 50-person nonprofit, Brockman, then chief technology officer of OpenAI, feared being left behind by rivals such as Google's DeepMind, so he spearheaded a plan to have OpenAI researchers and engineers collaborate to develop OpenAI Five, a software that could play the complex online battle game Dota 2.
Former OpenAI employees say the two teams have similar responsibilities — researchers focus on training new models while engineers develop software to make those models work—but on other OpenAI projects, the relationship between the two teams has been strained, a prevalent state of affairs within tech companies. After months of overnight work, their software finally found success in the game, and by early 2019, it had beaten the world's most advanced human Dota 2 players in thousands of games.
The success of this project quickly caused a stir in the gaming and technical circles and became a model for the ChatGPT project.
2. President of "Wandering": the first promoter of ChatGPT productization
People close to Brockman say that Brockman is the person who has contributed the most to the cooperation of various technicians in the ChatGPT project.
The magic of ChatGPT lies not in major breakthroughs in technological development, but in the application of AI technology in ivory towers to real-world tasks. Adam D'Angelo, CEO of Quora, a question-and-answer site that sits on OpenAI's board with Brockman, said that in the past, most of the AI research was in the lab, and it was difficult to turn the technology into a product, and Brockman was the one who made the AI successfully productized.
While OpenAI's CEO, Sam Altman, is the company's face, Brockman, 34, is the company's secret weapon. He is the president of OpenAI, but has no direct reports, which frees him from the usual administrative drudgery; Instead, he said in an interview with The Information, he spends 80 percent of his time programming.
Altman said: "Greg has a great ability to see the ultimate detail of each part of a technical problem, but also the focus of each level, and then think very strategically about how to put all the pieces together. ”
Brockman is a "mobile employee" who moves between different teams, sets goals, and pushes them to improve software development efficiency. He then moves on to new projects and hands them over to his manager. Brockman said: "I've been 'spoiling'. ”
▲ Greg Brockman
3. It is a top 10 engineer, but it also causes OpenAI code confusion
Brockman's approach inevitably brings some drawbacks. Two people who have worked with Brockman say it could take months for other OpenAI engineers to figure out how Brockman solved technical problems after he wrote the key parts of the project himself. In early 2022, Brockman insisted on rewriting one of the company's main software used to train models from scratch, forcing other OpenAI employees to adapt to the new system, one of the people said.
Brockman admits that meddling in projects himself, vetoing other managers by ordering changes to the team's priorities, can annoy colleagues. He says he's already improving his interpersonal skills to learn to make colleagues less angry. "The failure of my way of working can be a blow."
Still, Brockman's influence explains why others revere him.
People who worked with him described Brockman as the quintessential "10X engineer," a Silicon Valley slang term for programmers who were 10 times more prolific than the average employee.
Aravind Srinivas, who has worked at Google and OpenAI, founded the artificial intelligence search engine Perplexity last year. He bluntly describes the former colleague's influence: "The hardest working person at OpenAI was Greg Brockman, who combined coding, product decisions, and team arguments in a rare way in the tech world. Google has no such people! He said.
Second, the early entrepreneurial era of OpenAI: dropped out of MIT, planned to fight Google's technology monopoly, and became a problem solver
Where is Brockman from? Why is it so capable? It starts with his upbringing.
1. Dropped out of MIT to start a business and studied programming in China
Brockman grew up in a small town of about 1,100 people near the Minnesota border. A student at Red River High School, he won a silver medal at the 2006 International Chemistry Olympiad, reached the limit of his math curriculum in 10th grade, and began studying at the University of North Dakota in Grand Forks.
"My principal asked me to take three classes per semester in high school, and then I could take any class I wanted in college," Brockman said. ”
In 2008, Brockman entered Harvard University, where he planned to major in mathematics and computer science, learning to program during his Gap years in Russia and China. But when he felt he couldn't learn more about software from his classmates at school, he dropped out of Harvard. He transferred to MIT, arguing that students there knew more about software.
Brockman's goal was to start his own software company, and he met entrepreneurs Patrick and John Collison, entrepreneurs at the payments company Stripe, to work with him.
On a Thursday in 2010, Brockman dropped out of MIT and boarded a plane the following Monday to join the Collisons in San Francisco. Brockman spent five years at Stripe, eventually becoming the company's chief technology officer. During this time, Stripe experienced explosive growth to become one of the most valuable startups in the tech industry.
▲ Greg Brockman
2. 2015 Rosewood Hotel Dinner, meet the OpenAI initiators
The origins of OpenAI can be traced back to a dinner at the Rosewood Hotel in late 2015, where Brockman was joined by more than a dozen other tech celebrities, including Tesla CEO Elon Musk and then-Google research scientists Ilya Sutskever and Dario Amodei. Brockman said Altman knew him through Patrick Collison.
At the time, Altman positioned the project as a key balancing force hoarded by AI tech giants like Alphabet. He argues that putting advanced AI in the hands of powerful corporate interests could result in a technology benefiting only a few, rather than all of humanity. Brockman had just left Stripe and was excited about the project.
The earliest proponents of OpenAI include Brockman, Altman, Musk, LinkedIn co-founder Reed Hoffman, Y Combinator founding partner Jessica Livingston, and PayPal co-founder Peter Thiel, who together pledged $1 billion in funding for the new nonprofit.
In the early days of OpenAI, everyone's responsibilities were fixed, sometimes prosaic and ordinary. On Brockman's first day on the job, he was in a brainstorming session with OpenAI co-founders Satskefer and John Schulman looking for a whiteboard to write on.
3. Become a "problem solver" and optimize the speed and cost of AI systems to the limit
"I'm just a problem solver, spending weeks finding the first engineering problem and then being the solver of that problem," Brockman said. ”
During OpenAI's first few years of existence, Brockman and Satzkelfer made most of the day-to-day decisions, Brockman managed the startup's software engineers, and Satzkefer supervised researchers, according to two people involved. Brockman said he has been on the phone with Altman at work.
One of the challenges Brockman had to solve was how to get researchers and engineers to work from the same perspective. Insiders say some OpenAI engineers underestimate the importance of researchers' contributions, and researchers see engineers as technicians rather than scientists.
But Brockman's Dota 2 project helped break the ice. To push the effort, Brockman spent hours on the phone with Dota 2's developers to figure out how to get the software to play the game work, a technical challenge because the algorithms for the game and OpenAI are written in different programming languages, three people familiar with the matter said.
Former employees say the advances will also help Brockman become a key player in the development of OpenAI's generative AI models. Work on these models also began in 2017 after Google researchers published a research paper titled "Attention Is All You Need." This paper shows how a deep learning model called "Transformer" can help people build larger models more efficiently.
A person familiar with the matter said that while GPT doesn't primarily rely on OpenAI's underlying research breakthroughs, it has helped the startup develop a product that won't collapse because millions of people use it. Dedicated to optimizing the speed and cost of OpenAI systems, Brockman says he's "obsessed with every bit of gain you can get."
Third, the AI arms race begins: the dispute between "ideals and bread" is calmed, and ChatGPT becomes famous in the battle
The birth of ChatGPT was not a smooth ride, but it was accompanied by fierce internal divisions and constant organizational turmoil at the beginning.
1. R&D costs have doubled, and non-profit OpenAI is under pressure
In late 2017, Musk held a meeting in OpenAI's office to express concerns about OpenAI.
▲ Elon Musk
According to one attendee, OpenAI was slow at the time and could lose out to DeepMind. Musk worries that DeepMind's technology is accelerating to the brink of unreachability.
Musk felt that Tesla was a for-profit company with the ability to raise more money and therefore better develop AGI.
Some OpenAI employees believe that Musk's doing so will fuel an AI arms race that prioritizes speed over safety. Everyone wants to be the first to arrive at AGI.
After the dispute, Musk left OpenAI's board in 2018.
Despite some disagreements, Altman and Brockman took Tesla's founder's comments to heart, especially as the cost of training AI software continues to rise.
According to previous reports, OpenAI spent about $34 million a year on Google cloud services in 2018, and directly doubled the spending the following year.
According to a person with direct knowledge of the work, at one point, the Dota 2 project team alone spent more than $2 million per month on computing power, accounting for more than 90% of its monthly computing budget at the time.
But to develop AGI, OpenAI requires more money, up to billions of dollars.
2. The ChatGPT project caused controversy, and the vice president took the lead in leaving
Brockman recalled that he was met with intense skepticism from investors who were reluctant to put millions of dollars into research nonprofits because they had no hope of reaping financial benefits. "We just realized that this was not possible and we needed to adjust the entity."
As a result, OpenAI created a for-profit subsidiary in March 2019 that has the potential to generate returns for outside backers, but capped at 100 times the initial investment.
Altman does not own an equity stake in the for-profit unit. Brockman said he did acquire the equity, but did not say how much he owned, and that his stake may have limited his influence on OpenAI in one way: He is a board member of OpenAI, and the company's bylaws stipulate that only board members without financial shares can vote in cases where the interests of investors in the nonprofit and investors in for-profit subsidiaries may diverge.
Within OpenAI, some employees are uneasy about creating for-profit departments. Some believe OpenAI is abandoning its ideal of developing AI systems that are not influenced by enterprises. Eventually most agreed with the change, arguing that company leaders were trying to strike a balance so that OpenAI remained responsibly pursuing its goal of creating AGI.
As OpenAI continues to refine GPT, some of its employees have lost their illusions about Brockman's approach to releasing new products. For example, ChatGPT has come under fire from some critics for spreading misinformation and behaving bizarrely.
At the end of 2020, a group of employees led by Dario Amodei, vice president of research at OpenAI, left to form a competitor called Anthropic.
According to Brockman, the debate with Amodei and other Anthropic co-founders is over the ideological difference in how long to wait before releasing AI software. Brockman prefers to open up products and improve them through customer feedback.
3, ChatGPT suddenly succeeded, Brockman became OpenAI
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, it doesn't work for that, it doesn't work for this – but you're kind of missing the fact that it's useful for everything, and actually a lot of people will find surprising utility in it." He said.
The tech industry is scrambling to develop similar chatbots.
DeepMind and Google Brain, Alphabet's two AI lab rivals, began working together for the first time to try to catch up with OpenAI. In March, Anthropic released ChatGPT competitor Claude.
Meanwhile, Musk recruited DeepMind researcher Igor Babuschkin earlier this year to launch a new "anti-woke" AI lab within Twitter, another company owned by Musk.
Brockman also continues to push OpenAI's products forward. In March, he demonstrated the latest AI model, GPT-4, to OpenAI, showing how this advanced model can generate code for websites based only on rough sketches of their designs.
"Without his involvement, I don't think the quality of the whole project would have reached this level." Altman said.
Conclusion: Through the fog, OpenAI has become an AI entrepreneurship model
With the rise of ChatGPT to fame, many people began to pay attention to the secret of OpenAI's success behind it. From The Information's storytelling, we can see that this startup is actually facing the same problems as many AI peers, including the run-in of technology and product teams, the contradiction between ideals and bread, etc., and the process is also very difficult.
OpenAI's success has its natural and inevitable reasons, such as superior and determined leaders, as well as organizational and capital factors, which also provide a reference for domestic AI entrepreneurship.
Source: The Information