Source: 36 krypton
writing ｜ Ksssssss
edit ｜ Li Motian
Source | Yuanchuan Technology Review (ID: kechuangych ）
Cover Source ｜ Pixabay
At the press conference in the morning, OpenAI showed a new generation of large model GPT-4o, which made a lot of colleagues feel on pins and needles. However, Altman, who has always liked the whole big news, did not take part in the battle in person this time, but left the spotlight to Chief Technology Officer Mira Murati.
Like Ilia, the chief scientist, Mira is also a typical technological elite:
She studied mechanical engineering at Dartmouth College and worked as a senior engineer at Zodiac Aerospace, the French airline. Later, Mira joined Tesla and led the development of Model X motor system. In this process, Mira began to contact automatic driving, and her interest in artificial intelligence grew day by day [1] 。
Finally, she was recruited by Altman.
The competition of artificial intelligence always follows the rule that talent comes first, data comes second and computing power comes third. The top management of OpenAI are all technical elites like Mira. Such a huge talent arsenal is the secret that OpenAI is far ahead.
However, in the "arsenal" of OpenAI, there are not only traditional technology elites, but also more "unconventional talents".
For example, Prafulla Dhariwal, the multimodal leader of GPT-4o, actually has only a bachelor's degree.
Among the authors of Sora's paper, there is also a researcher who only has a high school diploma. He began to learn to write code at the age of 17. He has just turned 21 this year and still has acne marks on his face [2] 。
Even Christopher Olah, the former technical director of OpenAI, hasn't studied seriously for a few days. He wrote "University of Real Life Experience" in the column of educational experience on his LinkedIn homepage.
Obviously, OpenAI has its own understanding of what is "scientific research talent".
Talent arsenal
After turning through a large number of resumes, we found that OpenAI's talent recruitment has two significant characteristics, namely, "two do not look":
First, education is not considered. In large research institutions, the doctoral degree is usually the stepping stone to job hunting.
Yang Likun once mentioned that Meta divides research posts into two categories: basically only people with doctorate degrees have the opportunity to work as research scientists, and the rest are research engineers, which forces them to drop to a higher level [4] 。
But OpenAI is not so particular.
For example, Aditya Ramesh, the younger brother of India who created the liberal arts graph model DALL · E, has only a bachelor's degree from New York University. In fact, my Indian brother has plans to continue his studies, but OpenAI has been urging him to join the company as soon as possible.
The "senior high school student researcher" Will DePue mentioned earlier is more extreme.
When he was in high school, his mind was not on studying and he started a data analysis company. After the company was acquired, he went to the civil radical organizations in the United States and volunteered for seven months. Therefore, when Will DePue applied for OpenAI, he could not get a decent resume.
In February last year, Spectrum Research counted the educational background distribution of the ChatGPT team, and found that the number of undergraduates, masters and doctors was equal, accounting for 33%, 30% and 37% respectively [5] 。
Second, it does not depend on qualifications. OpenAI dares to let newcomers take the lead.
When Aditya Ramesh was studying DALL · E, she actually became a full member only a short time ago. Now, this young Indian with only 6 or 7 years of work experience has successively worked in DALL · E 2 DALL·E 3、GPT-4， And Sora's paper, leaving his own name.
This once made Indian media extremely excited, "DALL · E was originally of Indian origin" [7] 。
In the Sora team, this phenomenon is more obvious. The researcher who led the project is Bill Peebles, who graduated as a doctor in 2023 and is a fresh student.
Of course, there are many traditional super elites in OpenAI.
For example, Tim Brooks, another leader of Sora, is young but has a lot of experience. His teacher Alyosha Efros is a leading figure in the field of computer vision. He has also been engaged in AI research in Google, Nvidia and other large factories. However, the management led by Altman has more than one record of experience in Silicon Valley.
OpenAI usually allows a few super elites, with young and talented technical talents, to engage in scientific research together.
In a sense, the life of OpenAI is really in line with the secular romantic imagination of geek genius.
However, the reality is not an inspirational film after all, and the word "genius" will not be engraved on the face of applicants; What on earth does OpenAI, which has chosen "two no see", rely on to recruit people?
The philosophy of OpenAI
In fact, OpenAI has always been known for its strict recruitment. In 2017, someone shared the interview experience of OpenAI on Reddit, an overseas forum:
After passing the preliminary screening, he has gone through four rounds of interviews, including one speech, two research interviews, and one programming interview, which is comparable to five passes and six generals. Moreover, the direction of the two research interviews was different. One was to investigate the technical knowledge reserve, and the other was to focus on philosophical discussion, requiring candidates to share their thoughts on the evolution of AI technology [8] 。
On the US recruitment website Glassdoor, nearly half of the people gave negative comments on their job hunting experience.
Because the interview process of OpenAI is extremely long, and the interviewer always likes to make some strange moves. At the beginning of last year, Diane Yoon, vice president of human resources at OpenAI, publicly explained the reasons for doing so:
OpenAI focuses more on "problem-solving ability".
She mentioned that although OpenAI is a research institution, its behavior style is not outdated. OpenAI believes that the purpose of research is to solve practical problems and encourage researchers to try the simplest way, rather than blindly pursuing academic innovation, because the former is usually more effective.
Many achievements of OpenAI are the continuation of this culture.
For example, Sora, which shocked the world, improved and carried forward the technical route proposed by Google. Last year, Sora's basic paper "Scalable diffusion models with transformers" was even rejected by the top AI academic conference CVPR because of "lack of innovation".
The same is true for the GPT-4o just released. OpenAI didn't make any academic innovation, but with the help of its strong engineering ability, it turned the AI that can talk freely in science fiction movies into reality.
Source: Screenshot of social software However, this ability to solve problems will not be directly reflected in the resume. Therefore, OpenAI has designed many routines. Diane Yoon, for example, said that she often asked candidates to provide, and had done "influential work", in order to observe whether candidates have the awareness of solving problems and promoting innovation [9] 。
In fact, this practice is not uncommon. Many technology companies have very unique "recruiting posture".
As recorded in the Biography of Jobs, early Apple never recruited "honest people". Jobs often asked some strange questions to test whether candidates had a sense of humor and rebellious spirit. Sometimes, Jobs even teased the other party in the interview, asking him "where he is" and "how many times he has taken drugs" [10] 。
Because the more "lunatics" are recruited, the more "wild" innovation will be made.
Despite the endless emergence of quirky routines, OpenAI does not prevent it from wiping out the best technical talents. For example, the Indian brother Aditya Ramesh mentioned earlier, who was once a student of Yang Likun, did some research in Meta. However, he finally issued a "good man card" to his tutor.
Source: Screenshot of social software For those who want to engage in scientific research, OpenAI is a natural organization.
The charm of grand narrative
OpenAI just has an Oppenheimer idealism.
Oppenheimer not only personally accelerated the birth of nuclear weapons, but also strongly opposed the misuse of nuclear weapons.
Altman's original intention of founding OpenAI is also to worry about AI being abused by technology giants and harming human beings. Therefore, they set up a non-profit laboratory that is not controlled by large technology companies as a check and balance.
Altman has always believed that it is not difficult to promote such a seemingly crazy and unrealistic idea, "because people will think it is too cool and take the initiative to join in to help." The actual trend is also as he expected.
OpenAI was founded at the end of 2015, but it started quite late. At that time, Google, Facebook and other technology giants had already divided up AI talents. However, with its romantic corporate philosophy, OpenAI has successfully recruited top academic talents such as Ilia.
At that time, Google offered Ilia an annual salary of $2 million. After much thought, he finally felt that "saving humanity" was more important.
In 2018, OpenAI released a Charter of the Company, which further defined its mission, namely, "to ensure that general artificial intelligence benefits all mankind". After the release of GPT-4o, Altman did not forget to redraw the big cake in the announcement.
Oppenheimer's idealism is like a banner, gathering countless enthusiastic technical talents.
For most ordinary people, human welfare is a topic that is far away, and people care more about when they can raise wages and pay off the mortgage. However, the current group of technical talents really believe that they are engaged in a career related to the future of mankind.
Under the iceberg
However, the romantic company concept is just the gorgeous coat of OpenAI; The ideal seed can thrive because it is planted on solid soil.
Among the senior researchers of OpenAI, there is a Chinese named Li Jing.
He graduated from Peking University with an undergraduate degree, and then chose to go to MIT for further study and obtained a doctor's degree. After graduation, he got the opportunity to engage in post doctoral research in Meta, and directly followed Yang Likun, the winner of Turing Prize. After nearly three years of Meta research, Li Jing turned to OpenAI.
In hindsight, almost every choice of Li Jing is the best solution for his career.
The combination of these "optimal solutions" has built a huge "talent hematopoietic machine" in the United States. From academia to industry, the United States provides a complete one-stop service:
At the front line of teaching in Berkeley, MIT and other colleges and universities, there are a large number of well-known leading figures, including the Daniel Alyosha Efros mentioned above, and He Kaiming, the developer of ResNets.
The professors in these top universities are close to the industry, which is enough for "work package distribution". Yang Likun of New York University is the most typical. Most of his students have worked in Meta.
The top academic conferences in the field of artificial intelligence are basically organized by the United States. For example, CVPR (International Conference on Computer Vision and Pattern Recognition) and ICCV (International Conference on Computer Vision) in the field of computer vision are hosted by the Institute of Electrical and Electronic Engineers (IEEE) headquartered in New York.
When a large number of young people aspire to the computer industry and prepare to make a great achievement, they are often surprised to find that the United States has already laid one resplendent supply station after another on the track of scientific research.
When they are tired of walking alone in the wilderness, they will inevitably turn around and drive to another track.
Therefore, we do not have to criticize those researchers who go to the United States.
With a strong industrial foundation, the United States has gathered the most computer talents in the world.
MacroPolo, a think tank, once made statistics. They defined the researchers who received papers from NeuroIPS as "top AI researchers". They found that by 2022, 57% of the top AI researchers were working in the United States; In contrast, China, ranked second, accounted for only 12%.
Of course, we have made rapid progress - in 2019, China was in the "other" column.
However, if classified by nationality, only 28% of the top researchers of American nationality [12] 。 The continuous influx of Chinese, Indians and Europeans constitutes half of the US AI.
Therefore, it is hard to say how much enlightenment the idealistic story of OpenAI has for the pursuers outside the United States.
Epilogue
In 2020, Christopher Olah, then technical director of OpenAI, shared a blog. In the circle of computer scientists, Ola is a legend. He has only a high school diploma, and has successfully launched a career relying on self-study and guidance from the big man.
In this blog entitled "Do I need to go to college?", Ola shared his way of self-learning artificial intelligence [13] : Actively attend lectures, participate in academic conferences, visit laboratories, etc.
In the United States, these academic resources are open to all. In the process, Ola was lucky to know a quantum physicist and completed his first paper under his guidance.
Later, he was recognized by Joshua Bengio, the "leader of deep learning", and once wanted to enroll him in the university.
During this period, Ola also received a financial subsidy of US $100000, so she didn't have to be distracted by practical problems. This money comes from Thiel Fellowship, which was founded by Peter Thiel and is dedicated to funding young people who drop out of school and are eager to engage in the technology industry.
There is no doubt that Ola's success is due to personal efforts, but also benefits from a more inclusive talent environment.
Such an environment is the real source of innovation for a company and an economy.
reference material:
[1] Where We Go From Here with OpenAI's Mira Murati，a16z
[2] Will Depue: 20 Year Old OpenAI Researcher Shares The Secret To Building Anything， THE DOCK with Omar Waseem
[3] Linkedin
[4] Yann LeCun，X
[5] ChatGPT team background research report, wisdom spectrum research
[6] Two years after DALL-E debut, its inventor is “surprised” by impact，VentureBeat
[7] Aditya Ramesh: The Inventor Of AI Text-To-Visual Tool Dall-E Has Indian Origins，HomeGrown
[8] What is the job interview process like at OpenAI，Reddit
[9] So you want to work at OpenAI? Here’s what it takes，FastCompany
[10] Biography of Steve Jobs, Walter Isaacson
[11] Deep Learning Revolution, Kate Metz
[12] The Global AI Talent Tracker 2.0，MacroPolo
[13] Do I Need to Go to University?
