Source: Intelligent Emergence
Interview: Zhou Xinyu, Deng Yongyi
Author: Zhou Xinyu
Editors: Su Jianxun, Yang Xuan
It's been ten months since ChatGPT sparked the AI craze in China.
Every investor and headhunter familiar with the domestic AI landscape already has a mental map of talent. If you had the chance to sit down with them for a chat, the conversation might go something like this:
"Where did Zhang Qianchuan from Toutiao go?"
"He was poached by MiniMax."
Zhang Qianchuan, former head of user products at Toutiao, announced the establishment of an AI application company shortly after leaving in early 2023. However, not long after his departure, he joined the big model startup company MiniMax. According to 36Kr, MiniMax offered a considerable reward.
"Did Seven start her own business after leaving?"
"No, she joined DeepSpeech Technology as the head of product."
Seven (Wang Jingjin), former head of product at Douyin, recently joined DeepSpeech Technology, a AI startup valued at over 1.5 billion US dollars. 36Kr learned that before officially joining, Seven was invited to serve as a consultant for six months.
"Are there any other notable talents to watch?"
"Yes, Louis. Compared to the previous two, he is the one truly valued by the industry."
Louis, Yang Luyu, former co-founder of Musical.ly and a key figure in TikTok's overseas market. After leaving ByteDance in mid-2022, his whereabouts became a mystery. In June of this year, Louis entered the AIGC gaming track with a company called "Farlight Digital Technology."
An investor commented to 36Kr, "As long as Louis starts a business, I want to talk to him. If he ventures into AI, even more so."
AI entrepreneurship is booming, but in technology startups, product development and trial and error in implementation, amidst many uncertainties, talent has become the bargaining chip for pricing and transactions, and is also the most important factor in the current battle of big models.
The talent war quickly ignited the race track. At the beginning of the year, Wang Huiwen made a high-profile entry, followed closely by Wang Xiaochuan. More than ten entrepreneurs who posted AI "hero posts" soon followed.
According to several AI experts from top universities, entrepreneurs represented by the "Two Wangs" have almost talked to all AI talents in the market. In an interview with 36Kr in April of this year, Wang Xiaochuan, who heard that there were more than 20 big model companies in Beijing, asked, "Can you put our (Baichuan Intelligence) recruitment email in the article?"
In the field of general technology, AI is a mature hot spot, and talents are not concentrated in laboratories like room-temperature superconductors, but are evenly distributed among universities, institutions, and tech giants. Experienced investors and headhunters have long been familiar with the titles, past experiences, and latest trends of AI talents over the past ten months.
For example, to recruit technical executives, one cannot avoid a trip to the birthplace of ChatGPT: Silicon Valley.
Before heading to Silicon Valley, headhunter Mia packed two cans of finely packaged West Lake Longjing tea and four bottles of Laoganma in her suitcase. Upon arrival, she went to a bar in Mountain View, offering Chinese specialties, along with domestic company offers and annual salaries of nearly ten million RMB, to invite several algorithm scientists to return.
In Mountain View, where Google, Microsoft, and other tech giants are concentrated, it is not difficult to encounter venture capital bigwigs looking for AI talents, in addition to headhunters.
Some have seen Baidu's Vice President Jing Kun (the creator of Baidu's "Xiaodu" series of smart products) on the street; at Stanford University's lawn and Sand Hill Road, known as the "Wall Street of the West Coast," they have encountered Huang Yungang, a partner at Source Code Capital, and Dawson Dayson, managing partner of ZhenFund.
In early May, Wang Xiaochuan appeared at an AI technology party in Silicon Valley. Facing dozens of Chinese engineers, Wang Xiaochuan shared his entrepreneurial ideal of "AI reshaping search," attempting to build closer relationships with talents. In an interview with 36Kr, he mentioned, "There are some good leaders in domestic large companies and universities, and some more star-studded talents are indeed in the United States."
Helping domestic founders connect with Silicon Valley AI talents has also become a business opportunity. A headhunter in Silicon Valley recalled that in March alone, there were almost nightly parties organized by Chinese intermediaries in Mountain View, with a registration fee of over ten thousand RMB and a threshold: successful entrepreneurship or investment in good projects.
While companies are spending heavily to recruit talent, they also need to prevent employee poaching.
To retain talent, a startup raised the salaries of nearly ten core employees by 30%, and almost all engineers within the company use pseudonyms.
Large companies are also on guard. The night Baidu announced the establishment of the large language model "Wenxin Yiyuan," the participating members were called in by HR for a meeting and had to re-sign non-compete agreements.
The consensus quickly formed: the company with the most talent reserves is closest to being "China's OpenAI."
Million-dollar salaries, hard to recruit Silicon Valley talent
Given the scarcity of AI talent today, Mia believes that million-dollar salaries are not an exaggeration.
In the first quarter of 2023, over 170,000 Chinese AI companies emerged like mushrooms. A popular saying in the industry describes the entrepreneurial fervor in the AI race: "There are over 20 companies in Beijing alone claiming to be able to develop big models."
However, this is in stark contrast to the 300,000 AI talent gap reported by the Chinese Ministry of Industry and Information Technology. To be more stringent, how many people in China have complete experience in developing large models? Hai Lang, a senior talent consultant in the AI industry, told 36Kr, "Not more than 100 people."
If you want to recruit a team leader capable of leading a hundred people in engineering, there are only a few CTOs and chief scientists left in the big factories in China. "Companies looking to recruit technical executives can only look to Silicon Valley," explained Hai Lang.
Unfortunately, whether it's tea, Laoganma, or a million-dollar salary, it's difficult to impress Silicon Valley talent.
Before Mia could even mention the salary conditions, a researcher bluntly refused: "My wife and children are here, and it's difficult to solve the education problem for my children after returning to China."
In a little over a year, Joshua, an algorithm engineer at Microsoft Bing, will be able to obtain a U.S. green card as planned. In his life plan, after accumulating two more years of work experience, he will move from Washington to Silicon Valley to start a business.
He is unwilling to leave Silicon Valley—even though since February of this year, he has received almost 99+ unread messages and notification red dots on job search apps almost every day. In addition to salary, Joshua also sees companies offering hidden benefits: no OKRs for the first half of the year.
After being bombarded with job information for nearly a month, Joshua closed the pop-ups of several job search apps.
In the talent war, companies with mature landing scenarios have a greater advantage. In Mia's view, landing scenarios are the resumes of companies and are more attractive to talents.
For companies whose landing solutions are not yet mature, they either need a founder like Wang Huiwen and Wang Xiaochuan, who can bring in money and talent, or they rely on "pie in the sky" to attract talent.
However, the gap between the domestic research environment and the reality of Silicon Valley makes the ideal "pie" drawn by big model entrepreneurs less convincing.
During her time in Silicon Valley, Mia visited Intel's headquarters. There, each algorithm engineer had access to four to five hundred GPUs. "But in China, four to five hundred GPUs are often the upper limit of computing power that a project team can apply for," Mia said.
Similarly, the generosity of Google's AI lab towards talent once surprised Lan Zhenzhong. There, he could freely access TPUs equivalent to several thousand A100s. Even in relatively obscure research groups, the department could afford the most expensive wine in high-end restaurants for department gatherings.
After founding the AI company "West Lake Heart" in China, Lan Zhenzhong experienced the scarcity and high cost of computing resources for the first time: "The company and school's research funds can only cover a small part, and the rest has to be rented from public cloud services." To raise funds to buy computing power, he had to have five or six phone calls with investment institutions and clients every day.
Silicon Valley entrepreneurs returning to China are anxious about resources, while entrepreneurs going to Silicon Valley repeatedly hit walls. A well-known entrepreneur once admitted in public that his recent trip to Silicon Valley was not to recruit talent, but to exchange technical experience.
However, someone who had exchanged ideas with the entrepreneur in Silicon Valley told 36Kr, "Don't believe him. Because it's very difficult to recruit people, and most people who go there can only establish contacts first."
Meta's open source, headhunters worked in vain for half a year
At the beginning of the year, AI technical experts were still the most sought-after targets in the industry. Large language models like ChatGPT were still considered "imports." For most companies, to become "China's OpenAI," they had to aggressively recruit technical talent.
However, an unexpected event occurred shortly after.
The one who stirred up the situation was Meta (formerly Facebook), which had long been betting on AI big models. On March 8, 2023, the large language model Llama, known as the "strongest open source model," was leaked, and anyone could download and use it; in July, Llama's developer Meta took the initiative to become a "disruptor" and open sourced the more powerful Llama 2, which almost all companies could directly use for free.
The open sourcing of Llama quickly lowered the threshold for training large models. AI practitioners found that there was no need to spend so much money to recruit technical talent to train models from scratch. In theory, as long as a company has enough high-quality data, it can fine-tune Llama at a lower cost to train a model with good performance.
Soon, many companies' large models emerged like mushrooms, and some even claimed to be "self-developed" based on fine-tuning Llama. A widely circulated joke in the industry is: "If Llama hadn't been 'miserably open sourced,' there wouldn't be so many 'self-developed' models in China."
Subsequently, the value of technical talents in the field began to decline.
Several companies that had previously wanted to recruit from Silicon Valley changed their requirements to hiring engineers domestically, with annual salaries capped at 400,000 RMB. Mia sang a song titled "I'm Not Going to Work Tomorrow" at a KTV: "I feel like I wasted my time in Silicon Valley in the first half of the year."
In the first half of the year, due to the immaturity of the technology, the commercialization of large models did not go smoothly. The open source of Llama solved the bottleneck of large model technology and quickly shifted the progress bar of AI enterprise development from model refinement to application landing.
Correspondingly, the enthusiasm of companies to recruit technical talent in the first half of the year shifted to product managers in the second half. In Netflix's latest job posting, the annual salary for AI product managers was raised to 900,000 USD, exceeding the 650,000 USD for AI technical directors.
However, market supply and demand are not the only measure of top talent, and the heat for icon-level technical experts remains undiminished.
For companies, technical experts are not just employees who code, but also a symbol full of connotations: technical talent signifies the technical ceiling, as well as a facade for continuously attracting investors, clients, and talent.
Kunlun Wanwei's CHO, Yang Shu, has always believed that talent is the most valuable asset for an AI company. The company not only needs solid developers, but also an iconic figure. Just like Steve Jobs was to smartphones, "The difference between an icon and a developer is that an icon has market appeal and can attract even more talented people to come over."
In 2020, Kunlun Wanwei entered the AIGC and AGI fields, and the overall size of the related teams is now close to a thousand people. However, by 2023, in order to compete for talent, Yang Shu and her HR colleagues communicate with nearly a hundred candidates every week. Recently, Kunlun Wanwei also brought in an "icon" - top AI scientist Yan Shuicheng, to serve as the co-CEO of Tiangong Intelligence and the director of the Kunlun Wanwei 2050 Global Research Institute, attracting AI talent from around the world.
"Before Dr. Yang Hongxia came, I felt that there weren't many people in the market who thought that ByteDance could do big models," commented a HR representative from a major company.
As the former head of the M6 project at Alibaba DAMO Academy, Yang Hongxia's move to ByteDance's AI camp earlier this year - this news also led many to believe that ByteDance now has the potential to compete with "old players" Baidu and Alibaba in the AI field.
As for whether it's worth spending millions to recruit a technical expert for AI, the HR representative replied to 36Kr, "Before achieving technological innovation, it's important to ensure that the company's image keeps up with the trends of the times."
Everyone is wary of the bubble
After June, the temperature gradually cooled. In the first half of the year, money had flowed to the early players in the big model field.
According to incomplete statistics, there were probably more than 20 big model companies that raised money in the first half of the year, but after June, the number dropped by more than half.
An AI investor from a dual-currency fund terminated the investment process for four or five big model companies. She told 36Kr that they are currently only looking at AI applications.
Unfortunately, to this day, the AI track still does not have a "killer-level" application - the market and investors are both waiting to see if the highly valued big model technology can continue to provide substantial returns.
"CV (computer vision) has been hot for a year or two, but the big model cooling down is extremely fast," said Hai Lang. "This year, it seems like everyone has taken a lot of money, but in reality, they are under a lot of pressure."
Cooling down along with the race track is the enthusiasm of companies for AI talent.
The early enthusiasm for recruiting people was more about the excitement and FOMO (Fear of Missing Out) of companies facing new technologies. "No one really cares whether recruiting so many people is useful, let's just hype up the atmosphere first," said Mia, who received demands at the beginning of the year, most of which did not specify the number of talents to be recruited. "One is that it is indeed difficult to recruit people, and the other is that companies don't know how many people to recruit."
After careful consideration, companies gradually realized that the internet strategy of recruiting people in large numbers and expanding aggressively does not apply to big models.
Wang Huiwen once told 36Kr that he felt that having too many people working on big models would have a negative impact, and the smallest team should have around 30 people. In July of this year, Elon Musk made a high-profile announcement about entering the big model field, and his new company, xAI, has only 12 members.
The counterexample of the people strategy is Meta - even though it has two star big model teams, Llama and OPT, due to the uneven distribution of computing resources, over half of the Llama authors have chosen to resign.
In the scarce resources of the big model field, reducing the size of the team can not only "increase the amount of disposable resources per capita," as Musk said, but also improve management efficiency. Wang Xiaochuan mentioned in a media interview that after managing 3000 people at Sogou, he found that now at Baichuan, with only 100 to 300 people, it is very easy to improve efficiency.
Moreover, research on big models is a field that urgently requires talent and insight, and the people strategy has little effect.
"A smart brain is better than thousands of soldiers and horses," said Gao Yizhao, CEO of Zhi Zi Engine, telling 36Kr that their team's self-developed multimodal meta-multiplication ChatImg 2.0, with core algorithm development, required less than 5 people.
As the supply and demand of big model talent tends to balance, companies' urgent desire for AI talent has quickly returned to calm.
"Companies basically only need to recruit a powerful CTO, or a single-digit technical leader," said Mia, who recently received a sharp decrease in recruitment demands. The flow of top talent has basically become a foregone conclusion in the "hot battle" of the first half of the year, and Mia found that it is not difficult to recruit engineers from domestic and foreign major companies or computer science departments at universities with an annual salary of three to four hundred thousand.
The cautious expansion of companies for AI talent is also rooted in the talent expansion bubble that occurred in the field of computer vision (CV).
In 2018, the CV trend brought about the rise of the four AI dragons. At that time, the valuation of Megvii, which soared to 6 billion USD, raised over 2 billion USD within a year.
Where did most of the financing go? The answer is recruiting people. At that time, even recent graduates with a background in CV could earn an annual salary of 600,000 RMB.
However, many companies soon realized that CV did not have very high technological barriers. The business of AI companies To B and To G was quickly eroded by upstream cloud providers. The difficulties of the four AI dragons in recent years are there for all to see, with the earliest listed company, SenseTime, losing 2 RMB for every 1 RMB earned in 2022.
Under pressure to generate revenue, many top talents who went to research institutes at major companies have returned to universities. Hai Lang found that Chinese-American scientists who had previously helped major companies poach talent are now starting to publish papers as professors.
Almost all companies do not want to see another talent bubble.
Before the release of Llama, because big models were a high-barrier new technology, the performance goals set for talents were not specific, such as "surpassing GPT-3.5 by the end of the year, and surpassing GPT-4 in the next 1-2 years." But Llama quickly brought the progress bar to the stage of application development, and the performance of talents quickly shifted towards commercialization.
Making money became the primary indicator. Mia received many demands from companies in the second half of the year, shifting from recruiting people to helping talents with project management.
A researcher from Silicon Valley complained to Mia, "Didn't the company promise to give enough space for research? Why do they now want to focus on revenue?" Mia quickly responded, "The company believes in your ability."
The end of 2023 is a critical point for many companies and investors to test the value of talent.
"Both investors and founders need to see what results the people they spent money to recruit can achieve by the end of the year, and then decide whether to continue entering the market," said Hai Lang. From the ease of giving titles, he can feel the caution of companies this year: in 2018, it was not difficult to negotiate a position of T10 or P10 for a technical expert. But this year, even a P9 position requires a boost to reach.
Getting up at six or seven in the morning, swimming for an hour, then working in the lab until 9 pm, and then going home to be with the family - this was a day for Lan Zhenzhong when he was studying for his Ph.D. at Carnegie Mellon University, "Never working overtime, and never bringing work home."
But in this wave of enthusiasm, Lan Zhenzhong broke his habit. This experienced technologist and new entrepreneur has recently taken the initiative to meet with many investors, learning management and strategy from scratch: "Before the final round, I am working hard to avoid being eliminated by the market."
(Due to the request of the interviewees, the names Hai Lang, Mia, and Joshua are pseudonyms)
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
