The split of "Utopia", how the new openai replaces the old openai.
text丨He Qianming Huang Framexin
Data analysis and charting丨Huang Framexin
Editor丨Huang Junjie
Of all the Internet products born so far, only two can achieve 2.5 per week within two years of launch. Hundreds of millions of people use it, one is Douyin and the other is chatgpt. The difference is that chatgpt didn’t have addictive dances when it started, and it basically didn’t buy volume promotions.
As of November this year, it will be two years since the chatgpt product came out, bringing openai US$3.7 billion in revenue, far behind the levels of Google and Facebook in the same period, and higher than when Pinduoduo went public, with a company valuation of US$157 billion. It also broke the historical record.
The not-so-sexy side of the numbers is that OpenAI’s annual revenue of 3.7 billion US dollars relies on nearly 9 billion US dollars in expenditures - spend 2.4 yuan and get back 1 yuan, which is WeWork and SenseTime level operating efficiency.
openai hopes to use radical growth to solve profitability problems. According to the information, OpenAI estimates that by 2026, it will use US$40 billion in expenditures to earn US$26 billion in revenue - equivalent to nearly 1.4 times Baidu's revenue; by 2029, it will use US$100 billion in expenditures to achieve profitability.
From the Roosevelt Room of the White House to the conference rooms of TSMC, Sam Altman describes ambitious goals that require unlimited investment to achieve. He plans to spend US$100 billion to build a data center, install 2 million AI chips, and consume 5 gigawatts of electricity - equivalent to the annual electricity consumption of more than 4 million American households; if he were to manage TSMC, he would have to invest US$7 trillion. Build 36 wafer fabs to produce AI chips. The numbers were so crazy that one TSMC executive mocked his proposal as “podcasting bro”-level: paper-based and unrealistic.
Within openai, radical growth is already under pressure: Anthropic, founded by former employees, is carving up market share with the strong support of Amazon; more and more developers are discovering that problems such as text summary can be solved without the use of the most advanced large models; and Improving model capabilities may also encounter bottlenecks ......
It cannot be concluded that Altman is a visionary. At least the first two years of chatgpt relied on his usual extraordinary ambitions: to control the consumption of infrastructure among several trillion-dollar super giants. In the Internet market, a semi-profit research organization can carve cracks in the giant walls. As for what made openai emerge so suddenly despite the lack of capital at the time, its co-founder and president Greg Brockman once explained: "(Talent) is openai's secret weapon."
However, As the company shifts from non-profit to for-profit, OpenAI solves capital problems but also triggers people problems - organizational turmoil and brain drain that have barely stopped in the past two years. Openai, which has lost its secret weapon, has placed itself in a turbulent and ever-changing market segment, and the key factor in change is precisely people. On the second anniversary of the release of
chatgpt, we study the changes in the world's largest AI startup company from the perspective of people. The flow of these people not only shapes openai, but also leads the global AI industry.
We had hoped to list chatgpt as a co-author, but after two years and three generations of model updates, we still continue to feel the limitations of large models. The list of “contributions” attached to the technical report of the
gpt-4, gpt-4o, and o1 models is the starting point for our research.To make the analysis more complete, we also included authors of papers from the three generations of gpt, gpt-2, and gpt-3 to provide a comparative perspective, a total of 631 people.
We use "name + openai" as keywords to generate linkedin search links, and manually find valid linkedin homepages one by one - large models cannot help us complete this tedious task. Using proxycurl's API, we captured the work and education experiences of contributors to the linkedin homepage, cleaned the data with R language, used a large model to classify functions according to job descriptions, and then manually proofread it.
There are no contributors to the LinkedIn homepage. We searched again with a large model, but the information was incomplete or outdated. Finally, we used an original and reliable method: manually finding media reports and social network pages on search engines, and supplementing the data of about 10 people. We also used the open source python library scholarly to add the Google Scholar data of 214 of them, evaluate their academic influence, and finally built a database of 583 contributors.
gpt iteration, openai changed
In December 2015, openai was established and positioned as a non-profit AI laboratory, relying on the belief of "developing safe AGI that benefits all mankind" and Elon Musk (elon Musk) snatched people from giant Internet companies such as Google, and established a research team of more than 50 technical elites in more than a year: in addition to the initiator Altman, there is also one of the founders of the AI wave, Google Brain, who is responsible for Ilya Sutskever, CTO of Stripe, an emerging unicorn in Silicon Valley, Greg Brockman, and the new generation of AI research genius Andrej Kapaci karpathy) and other researchers. In the early days of
, the team built a robot hand for playing Rubik's Cube with one hand, studied AI for playing Dota, and finally saw the possibility of AI changing human life in the direction of language understanding of the world. In 2019, they bet on big language models.
From gpt-1 to gpt-2 and gpt-3, the model parameters have expanded from 100 million to 175 billion, the cost investment has increased, and the number of contributors has increased from 4 to 30. In order to raise resources, the company has to go to Africa. Profit transformation.
Of the more than 30 researchers who laid the foundation for chatgpt, 2/3 of them have left openai. As of October, 80 of the 583 contributors no longer work at openai, including more than 20 founding team members and mid-level and senior management.
"The most compelling artificial intelligence systems, in addition to requiring algorithmic innovation, also consume the most computing resources," openai Zeng once wrote in an article, "In the next few years, we will invest billions of dollars in purchasing cloud computing resources to attract and retain talented people."
Paradoxically, every time the company successfully raises more resources , will trigger a round of departures of contributors. Former OpenAI chief security researcher Dario Amodei has spoken about the core of this contradiction, they are worried about "commercial interests affecting security priorities." The first batch of
contributors left in early 2020. At that time, Altman introduced a US$1 billion investment from Microsoft to promote OpenAI to form a for-profit entity. Amoudi and a group of colleagues left their jobs to start the AI company Anthropic. That time, openai lost 8 of the 30 gpt-3 paper authors.
Altman believes in the concept that "economic growth can make democracy more universal." The pursuit of growth is the instinct of this former yc entrepreneurial mentor in running a company. Chatgpt went online on November 29, 2022, openai raised tens of billions of dollars from Microsoft, and held a press conference in November 2023 that looked like early Apple. Altman seemed to become the star of a new generation of technology companies.
However, less than a month later, the contradiction between the pursuit of growth and commercial scale and the AGI belief of benefiting mankind completely broke out within the company, and shocked the world in a super dramatic way - the then chief scientist Suzkewei Joint Board of Directors Altmann's expulsion was suddenly announced.
"He has repeatedly provided false information about security measures. The board of directors is basically unable to understand the operating effectiveness of security measures and cannot judge which aspects may need improvement." Helen Toner, a director of OpenAI who participated in the action, said .
In just five days, this revolution that shocked the world came to an end, and Altman won. He returned to the company and a majority of the board resigned. Six months later, Sutskvi left openai, once again triggering the departure of a group of contributors who were concerned about security.
Altman is still determined to move forward with the next step: turning OpenAI into a for-profit company.
The new openai comes from a big company.
New and old alternate, and the way openai attracts talents has also changed: using generous salaries to compete with google, meta, etc. for talents. At the beginning of 2023, OpenAI has offered an annual salary of US$800,000 to fresh doctoral graduates. The annual salary of senior researchers can be several times higher, comparable to Google and Meta, which have tens of billions of dollars in annual profits.
Money has become the driving force for controlling talents, and it is also a time bomb that cannot retain people. As soon as the company stops growing, compensation becomes a burden and options lose their appeal. The cost of choosing the business route became clear. Now, the pursuit of growth is not only Altman’s pursuit, but also the foundation for maintaining openai’s operations.
The newcomers not only reshaped the company's team, but also reshaped the chatgpt product.
Only one person, Suzkovi, participated in the development of gpt-1 to gpt-o1. His departure leaves only Alec Radford with similar experience. Radford has participated in the research and development of gpt-1 to gpt-4o. He joined openai in 2016 and was one of the first researchers to apply the transformer architecture proposed by Google to large-scale language training. He also established the current large language model. One of the people in training mode.
"Only Alec Radford's resignation can we be certain that openai will collapse." Emad Mostaque, founder of the large model company stability ai, once said on social media.
The rapidly accelerating flow of talents in the past two years has pushed openai to quickly adapt to the stage that all startups must go through - a company that rises because important people create products will definitely no longer have the role of individuals in determining the life or death of the product during development. New models of
openai are still being steadily introduced to the market. There are a total of 390 core contributors of the gpt-4o and o1 generation models. More than 70% joined the company after the release of chatgpt, and they shaped the new openai. In fact, high salaries have indeed solved the pain problem of the loss of veterans. The three months after the struggle in November 2023 will be the peak period for contributors to the two generations of gpt-4o and o1 models.
openai’s recruitment strategy has not completely continued the proposal given by Musk when it was first established: focus on recruiting students who have not yet graduated but are very smart. “It is best to let them join before they achieve breakthrough results.” The contributors who joined chatgpt after its release were only more than 10 recent graduates.
The new contributors mainly come from large companies, and 54% have worked in the seven largest technology companies in the United States.
In its early days, openai relied on its faith to attract a group of employees from large companies to join.But between 2017 and 2020, it was almost difficult for openai to maintain operations, and not many employees from large factories were willing to join. These years have also been a low period for OpenAI to recruit people.
By 2023, the probability of OpenAI becoming the next big technology company looks higher, and its attraction to employees of big companies has increased significantly. More than 40% of contributors with working experience in large factories joined this year, far exceeding previous levels.
has deep accumulation in the field of AI, but two companies that are slow to move in the field of large models - Google and Meta have sent the largest number of contributors to OpenAI, with 119 and 103 people respectively. These two companies alone account for nearly 4% of the total. become.
They also contributed the technical backbone to openai. Among the three generations of gpt-4, gpt-4o, and o1 models, there are a total of 118 leads, 45% of whom have worked in Google and Meta.
followed by yc companies, stripe ranked third, dropbox ranked fifth, plus airbnb, instacart, twitch, cruise and other companies, a total of 59 people were transported. These companies share values similar to Altman's: growth above all else.
Altman has said on social media that he is interested in a college alternative: "Find the smartest, most driven 18-year-olds in the world and give them 10 With more than 10 years of salary and resources, let them do whatever projects they want, paired with smart peers - in exchange for a few percentage points of their future income. "
In openai, his preferences are mapped to the educational background of employees. . Among the contributors to openai, we found the schools where 543 of them graduated from. Most of them came from top universities. Graduates from Stanford University, California Berkeley, and MIT accounted for 30%. Graduates from Chinese universities also have a strong presence. There are more than 20 graduates from two universities in northern Qing Dynasty.
High-level purge, Altman’s team
In 2017, openai, which had just been established two years ago, discussed whether to transform into a for-profit organization. Musk and Altman competed for the position. CEO, Sutskwei and Brockman, who were in charge of day-to-day operations, resisted Musk but were also suspicious of Altman's motives.
In an email, they threw the question directly to Altman: "We cannot fully trust your judgment...We do not understand why the title of CEO is so important to you...agi Is it really your main goal? What does it have to do with your political ambitions?”
Musk was angry and withdrew, but Altman won trust with his statement that he was “still passionate about non-profit institutions”. , and expanded its influence in openai.
openai Later developments proved that Sutskevi’s intuition was not wrong. What Altman calls "nonprofit enthusiasm" lasted only two years before beginning to fade.
Since becoming the CEO of openai as he wished, Altman has been free to transform openai according to his own ideas.
In November 2020, half a year after the release of the gpt-3 model paper, Altman appointed Peter Welinder, OpenAI’s first vice president of product, to be responsible for making the gpt-3 model into an API product for external sales. Soon after, Altman hired Anna Makanju, a former senior policy adviser to the Vice President of the United States, from Facebook to be responsible for policy and public relations. Later, Macanjiu used his experience in foreign affairs to arrange a trip for Altmann to travel around the world and meet with dignitaries from various countries.
This round of high-level changes is relatively mild. The larger chain reaction is that Dario Amodei, vice president of research and responsible for security research, left to found Anthropic, but the core team is still stable.After the release of
chatgpt, openai received a US$10 billion investment from Microsoft. Altman once again elevated its position in product growth and commercialization. Within half a year, three vice presidents were hired to take charge of related matters:
Srinivas Narayan narayanan), joined in April 2023, responsible for the engineering work of chatgpt and API construction of various generations of models. He previously worked on similar tasks at Facebook.
Deng Xiuping (peter x. deng), joined in May 2023, is responsible for the products of chatgpt and chatgpt enterprise version. He previously worked as a product manager at Facebook for more than 10 years, and later became the product leader at the database company Airtable.
krithika muthukumar, joined in August 2023, is responsible for openai product gtm work. Previously, she worked in marketing for more than 10 years in yc companies such as Dropbox and Stripe.
Peter Verinder, the original vice president of product, was transferred to be responsible for new product development. The number of openai’s product executives has caught up with its technical executives. The reason why
has not been exceeded is not because of restraint, but because of external constraints. The research team headed by chief scientist Sutskwei has a religious fanaticism for "safe AGI". A report in "The Atlantic Monthly" wrote: Sutskvi will lead employees in offline gatherings to shout "feel the agi! feel the agi!". He is also the organizer of OpenAI's "Super Alignment Project" and pursues Safety.
Sutskvi’s subsequent expulsion failed, giving Altman the opportunity to expand his influence and take full control of openai. In 2024, openai's senior management entered a period of large-scale purges:
