AI Giants Can't Hold On to Star Employees: Departures to Start Up and Multi-Billion-Dollar Valuations
A Huxiu article reviews the phenomenon of more than a dozen star AI technologists leaving leading big tech companies over the past year: Chen Guangyu left Kimi, Yao Shunyu was reported to be preparing an RSI startup, and Lin Junyang left Tongyi Qianwen to found Bulage; the latter two are both valued at about $2 billion according to market reports. The article says at least five core R&D staff left DeepSeek, covering four lines—foundation models, inference, multimodality, and OCR—and that core members have also left Google and Meta. Big tech companies use tactics such as hiding talent, non-compete agreements, desensitization agreements, and stock options to retain people; ByteDance's retention cost is estimated at over 2 billion yuan, yet the primary market can offer multi-billion-dollar valuations to companies that do not even have a product yet.
Why read this
AI score 56
The article traces the paths and valuations of star AI technologists leaving big tech over the past year, and compares the actual effectiveness of retention tactics such as hiding talent, non-compete agreements, and stock options.
Topics
Body · AI translation
**Key Points**
**1. Fame immediately triggers departure to start a business**
Chen Guangyu left Kimi; Yao Shunyu is reported to be preparing an RSI startup, with market reports citing a valuation of about $2 billion; Lin Junyang left Tongyi Qianwen to found "Bulag", with a post-investment valuation of about $2 billion.
**2. Leading tech giants are broadly losing talent**
Over the past year, more than a dozen AI technical stars have left leading tech giants; Google and Meta have both seen core members depart; at least 5 core R&D staff have left DeepSeek, with departures across all four lines: foundation models, inference, multimodality, and OCR.
**3. Retention tools struggle to change employee pricing**
Big companies use talent concealment, non-compete agreements, redaction agreements, options, and legal affairs to retain people; ByteDance's retention cost has been estimated at over 2 billion yuan; but the primary market can offer valuations of billions of dollars to companies with no products, so retention is still at employee prices.
**4. Star employees can counterbalance the organization**
Alibaba had Liu Da Yiheng appear only at official occasions, turning fame into an organizational resource; but after Character.AI's founding team left, the company went bankrupt, and its core founder had previously been brought back by Google at a high price; the Dong Yuhui case also shows that the added value of traffic lies with the individual. 2026-10-27 07:25 AI giants don't believe in star employees AI Story Project quick overview
Geniuses have a pattern: once they become famous, they are not far from leaving.
No sooner had Yao Shunyu been reported to have left to start a business than Chen Guangyu headed for the San Francisco Bay Area. Before that, Lin Junyang said in the early morning, "bye my beloved qwen", unilaterally announcing the end of his relationship with his former employer Tongyi Qianwen.
Becoming famous is like a personal IPO
"Thanks for everything! Especially for believing in and hiring a 16-year-old me back then." On October 15, Chen Guangyu posted on social media, announcing his departure from Kimi.
The comment section quickly turned into a recruiting scene. Anthropic researcher Sholto Douglas publicly called out: "Still planning to go to CMU? You should come to San Francisco and work with the Prime, Base10 crowd!" Chen Guangyu replied that he was indeed heading to the Bay Area.
For a time, the whereabouts of this 17-year-old became a hot topic in the AI circle. Some guessed he had joined Anthropic; after he denied it, an account dedicated to tracking Silicon Valley AI talent flows said he had gone to a startup founded by former Google chief scientist Jeff Dean.
In the AI industry, where technology iterates on a monthly basis, the timetable for professional fame has been compressed. Chen Guangyu is still a high school student. In November 2025, at 16, he was exceptionally hired as an intern at Kimi. Four months later, he became a co-first author of "Attention Residuals", and the related technology was used in the architecture of Kimi K3.
From an ordinary middle school student to being pursued by the market took less than two years.
Fame also means bargaining power. After staying at Kimi for only 11 months, Chen Guangyu headed to his next stop.
If even an intern is like this, how much more so the big names. Almost at the same time as Chen Guangyu, Google DeepMind core researcher Yao Shunyu was reported to be preparing to start a business, in the direction of RSI (recursive self-improvement), with market reports saying his new company is valued at about $2 billion.
Yao Shunyu first became famous in the physics world. During his undergraduate studies in the Physics Department at Tsinghua University, he proposed the "non-Hermitian skin effect" as first author in Physical Review Letters, opening up an entirely new research direction. In October 2024, he shifted from quantum gravity to AI, joined Anthropic's Claude team, and participated in reinforcement learning R&D for Claude 3.7 to 4.5. Eleven and a half months later, he jumped to Google DeepMind.
In May this year, he appeared on a domestic video podcast and became famous on the Chinese internet for his sharp remarks. On the show, the host asked: "Will you stay at Google for a long time?" Yao Shunyu lowered his head and smiled: "I don't think so." At the time, he had been at Google for just half a year.
Big companies' attitude toward geniuses is contradictory. Geniuses' papers and open-source projects are the most intuitive showcase of a company's strength. Major companies at home and abroad initially welcomed technical talent stepping into the spotlight. Unlike technical personnel of the past, almost every AI practitioner of this generation has their own social media accounts and is accustomed to sharing project progress and technical judgments on podcasts and X.
Lin Junyang is a typical example. While serving as technical lead of Tongyi Qianwen, he always stood at the forefront and was Qwen's spokesperson in the global open-source community. When a new model was released, he personally explained technical details on X; when developers asked questions, he stepped in to answer.
Precisely because of his reputation in the industry, his departure once put his former employer in a passive position. On March 4, Lin Junyang posted on X, "bye my beloved qwen", unilaterally announcing the end of his relationship with his former employer, without even leaving the company time for handover and public relations. On the night of the announcement, Qwen core members Yu Bowen and Kaixin Li successively announced their departures.
Geniuses who become famous are like a large-value asset: after shining in papers and podcasts, their value keeps rising, and then they leave. Becoming famous is like a personal IPO.
Google is the hardest-hit area. In February, AlphaGo core architect David Silver left to start a business; in June, Noam Shazeer, who had been brought back for $2.7 billion, moved to OpenAI, and Nobel Chemistry laureate John Jumper joined Anthropic; in August, chief scientist Jeff Dean, who had served Google for 27 years, founded Discovery Loop with several veterans.
Meta also failed to retain people. In August, Yu Jiahui, who was reportedly recruited for $100 million, announced his departure to start a business.
Domestically, the biggest spender is Tencent. At the end of last year, it brought in OpenAI researcher Yao Shunyu as chief AI scientist, and in February this year, it poached senior research scientist Pang Tianyu from Sea AI Lab to lead Hunyuan's multimodal reinforcement learning.
Recently, the news that former ByteDance intern Tian Keyu's new company completed nearly $30 million in financing sparked heated discussion in the AI circle.
In 2024, during his internship at ByteDance, Tian Keyu became embroiled in the "poisoning the large model" controversy and was sued for 8 million yuan; the case is still under trial. But he is also a NeurIPS Best Paper winner, and investment institutions are still willing to bet on him.
The big companies' encirclement and interception
English text is being prepared.
Liang Wenfeng has never publicly discussed this departure, but rumors in the circle say he has been brooding over it. After that, DeepSeek changed its stance of not raising funds and not discussing an IPO, and began looking for money in the market.
Once a name becomes valuable, the display window turns into a shelf. Reports say that after the DeepSeek-V4 technical report was published, headhunters found the signature list overnight and traced people accordingly.
In this round of the talent war, DeepSeek has lost considerable blood. According to public reports, from the second half of 2025 to the first half of this year, at least five core R&D members confirmed their departures: Wang Bingxuan, a core author of the first-generation large language model, went to Tencent; Guo Daya, a core researcher of R1, joined ByteDance Seed; Ruan Chong, a core multimodal researcher, joined Yuanrong Qixing; plus Luo Fuli—across the four main technical lines of foundation, reasoning, multimodality, and OCR, people have left.
The most thoroughly hidden is DeepSeek. This company holds almost no press conferences, and its researchers are nearly invisible on social media; for a long time, all outside imagination about it came only from the models themselves. Founder Liang Wenfeng has rarely appeared in public since July 2024. It was not until the publication of the nearly 300-name signature list in the V4 technical report that the outside world saw its lineup clearly for the first time.
When Wu Yonghui took over ByteDance Seed's basic research in March 2025, he still encouraged researchers to publish papers and write blogs, and even suggested everyone "decorate their personal homepages." But as the talent war intensified, the wind shifted, and companies began to protect core researchers' information as much as possible and reduce their communication with the outside world.
There were also reverse operations. After Lin Junyang left, Alibaba changed its "way of using" technical stars. Liu Daheng, who succeeded as head of the Qwen LLM project, came from the first tier of Huawei's "Genius Youth Program" and is a core member of the Qwen founding team and one of the corresponding authors of the NeurIPS 2025 Best Paper.
The day after the appointment was announced, he was pushed onto the main forum of the Yunqi Conference, ranked second only to Joseph Tsai and Wu Yongming among guests. Unlike his predecessor, who commanded attention on X, Liu Daheng's name has so far appeared only on stages built by the company, each time bearing an official imprint.
This operation is somewhat like separating fame from the private property of geniuses and turning it into a resource distributed by the organization. How famous one becomes, and when, is up to the company.
What cannot be hidden can still be locked down. By 2025, the foundation model teams of major companies all had to sign non-compete agreements, barring them from joining competitors for 6 to 24 months after leaving. Many companies added a "desensitization agreement," requiring people who resign to stay three more months before being released.
As for options, they are like a golden shackle and one of the means of locking people in.
If that still cannot stop them, there is legal pursuit. In 2024, after Zhou Chang, former technical lead of Tongyi Qianwen, resigned, Alibaba filed a non-compete arbitration, which ultimately ended with ByteDance paying the compensation on his behalf.
There are countless reasons for a breakup, but most come down to one word: money.
There is a saying in the industry that a reliable Infra talent can increase computing utilization by as much as 50%. For a company like OpenAI, which spends about $7 billion a year on computing, that is ten-billion-dollar-level elasticity. The pricing of the top tier has long surpassed that of NBA stars.
But total compensation is a combined equation, including cash, options, stock, and signing bonuses. Therefore, whether a company goes public has become a variable in the talent war.
ByteDance is repeatedly raided, and one frequently mentioned reason is that it is not listed. Options have no public-market pricing, and cashing out has no timeline. To fix this bug, starting in September last year, ByteDance began paying Seed employees monthly option allowances of up to 135,000 yuan; in October, it launched "Doubao shares." Some media estimates put its retention cost at more than 2 billion yuan.
DeepSeek, which claims not to lack money, also sat down at the capital market negotiating table this year. Calling it financing, it may well also be a way to use the public market to renegotiate a price for the geniuses.
This, too, is a way of retaining people.
In the face of the bubble, there is nothing left to retain
The problem with this retention approach is that no matter how deeply people are hidden, how tightly they are locked, or how high the pay is, it is still an employee price.
The job-level promotion and equity incentives that big companies have used for years are an old pricing system measured in years and premised on organizational loyalty. But in this round of talent revaluation, the external market has long switched algorithms.
In January this year, Tian Yuandong, former director of research scientists at Meta FAIR, officially announced his startup and joined Recursive as a co-founder, in the same direction as Yao Shunyu: RSI. This company has not yet released a public product, but already holds $650 million in financing at a valuation of $4.65 billion.
In June, Lin Junyang's new company "Borage" completed a $220 million financing round at a post-money valuation of about $2 billion. Its focus is world models and embodied intelligence. No product, no revenue—only the name Lin Junyang.
Yao Shunyu had not even submitted his resignation letter, yet a $2 billion valuation had already arrived in advance. As for Tian Keyu, with an 8 million yuan lawsuit hanging over him, nearly $30 million in financing still came in.
One paper, one résumé, one technical judgment—and you can raise money. Five years ago, that was a curiosity; now, it is standard practice.
Back to Lin Junyang. Structural adjustments during an expansion phase, changes in reporting lines, and a shrinking scope of authority are common plot points in the departure narratives of many star researchers. Lin Junyang's resignation was triggered by exactly this.
Lin Junyang's predecessor, Zhou Chang, moved to ByteDance in 2024. According to multiple reports, the underlying reason was that he could not coordinate enough computing power. Translated, it still comes down to autonomy and the room to call the shots. And the primary market can provide all of these in one go.
That is only the visible account. Even more tempting is that the cost of startup failure has also been smoothed away.
Character.AI is the most complete sample of this cycle. It was founded by Noam Shazeer, one of the authors of the Transformer paper, and at its peak had more than 20 million monthly active users. In 2024, Google paid $2.7 billion to license its technology and brought Shazeer back to DeepMind, appointing him as a co-lead of Gemini.
In October this year, Character.AI, having lost its founding team, filed for bankruptcy reorganization. The buyer, Disney, offered only $20 million—less than a fraction of what Google paid back then.
The company sank; the people did not. Long before this company reached its endgame, Shazeer had already left Google and joined OpenAI.
If it works out, you get the upside of a startup; if it does not, a top-tier big company is willing to spend billions of dollars to bring you back. Anyone would make that choice.
Once the math reaches this point, the big companies no longer have many chips in hand.
Among the previous generation of entrepreneurs, Yu Minhong understands most deeply what it is like for star employees to turn around and constrain the organization. Over New Oriental's thirty years, the famous teachers who left to start businesses ranged from Luo Yonghao all the way to Li Xiaolai—enough to start another New Oriental.
In the livestreaming era, this old drama reached its most expensive episode. To retain Dong Yuhui, Yu Minhong once offered three titles—East Buy's first senior partner, cultural assistant to the group chairman, and vice president of the cultural tourism group—plus an independent livestream room. Seven months later, Dong Yuhui still left.
After Hui Tong Xing became independent, in its first year it sold more than 10 billion yuan of goods on Douyin, and third-party estimates for 2025 exceed 20 billion yuan. One investment institution, using a 20x price-to-earnings ratio, estimated the company's valuation at nearly 6 billion yuan.
"In the traffic era, the added value of traffic lies with the individual, not the company," Yu Minhong, who paid 358 million yuan in tuition, once said on an earnings call.
When the primary market's quote for one person exceeds a listed company's full-year profit, retention loses its meaning.
This article is from the WeChat public account: AI Story Project, author: Tian Mi, editor: Zhang Xia. AI Industry Signal Channel: Frontier Technology. This content is authorized for publication by the author. The views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP [ ]