Zhang Yiming and Ma Huateng's AI talent war: The competition has expanded from recruiting PhDs all the way down to high school students, but can these talents be retained?

On October 9, 2025, in Caohejing, Xuhui District, Shanghai, an opening ceremony brought Zhang Yiming, who had not appeared in public for a long time, back into the public eye.

Wearing a black T-shirt, he stood on the podium of the Zhichun Innovation Center in Xuhui, Shanghai, using a machine learning term called overfitting to explain why he was looking for AI talents among 16-year-olds.

The so-called overfitting means that the model performs perfectly on the training data, but its performance drops sharply when facing new data that it has never seen before.

In other words, it does not learn the rules behind the data, but only memorizes the noise in the training set by rote. To put it more bluntly, it remembers the answers but does not understand the questions. Zhang Yiming's solution is the 14 first batch of reserve researchers sitting under the stage, the youngest is 16 years old and the oldest is 18 years old.

The eagerness of large tech companies for AI talents is simply a real-life version of the "claw machine game for talents".

This summer vacation, Tencent Youth Science Training Camp took a more radical move. It opened product practice opportunities for school students aged 13 to 18 around the world, with directions directly aligned with core businesses such as fintech, WeChat Mini Program AI+Education, and WeChat Search AI, and middle school students are led by director-level mentors.

It is not a visit, not a lecture, but directly getting hands-on to make products.

Almost at the same time, Geely Holding launched the Cross-era Leap Talent Training Program, which does not take the college entrance examination scores as a reference and has no academic threshold. High school graduates can directly enter the four cutting-edge fields of new energy, artificial intelligence, low-altitude flight and low-orbit satellites, with paid training and an annual salary of up to 300,000 yuan.

On the other side of the ocean, Silicon Valley AI giant Palantir opened direct recruitment positions for high school graduates, with 500 people competing for 22 spots, and the annual salary after regularization is 170,000 US dollars (about 1.2 million yuan).

Therefore, a seemingly strange but not inconsistent fact is that even if the model leaders of large tech companies are already occupied by post-90s generations, they still hope to recruit younger AI geniuses, and for this reason, they do not hesitate to move the battlefield of talent recruitment forward at an unprecedented speed.

Moving forward from campus recruitment to the doctoral stage, from doctoral stage to undergraduate stage, and then from undergraduate stage to high school stage or even earlier. This is not an individual action of a single enterprise, but a collective shift of the talent strategy of an entire generation of technology companies.

1

From "Picking Ripe Fruits" to "Growing Crops"

In the past two decades, the basic logic of large tech companies' talent strategy was "picking ripe fruits": waiting for universities to cultivate talents to maturity, then poaching them with high salaries.

The Genius Youth Program launched by Huawei in 2019 is the ultimate version of this logic, with no restrictions on academic background or schools, only focusing on whether the candidates can challenge world-class topics, with a maximum annual salary of 2.01 million yuan. Zhi Huijun, who later founded Agibot, was among the selected talents.

But this logic is somewhat outdated in the AI era.

According to the traditional 5 to 6-year training cycle for doctoral students, the AI direction a doctoral student chooses at admission may have iterated 3 to 4 rounds by the time he graduates. When he publishes a paper in a top journal, the industrial circle is already discussing a brand new paradigm.

Jiang Jie, Vice President of Tencent, observed a key change: the generation gap between the solutions made by young people in schools and the industrial circle is getting smaller and smaller, their understanding of large models is very close to that of the industrial circle, and their knowledge system is completely connected with the industry.

This means that the university training system is changing from being half a step ahead of the industry to being one step behind the industry. When the training cycle itself exceeds the technology iteration cycle, "waiting for maturity before picking" has become a luxury.

Therefore, large tech companies began to "grow crops" on their own.

Tencent's layout is the most systematic. The Spark Program launched in 2019 selects talents from high school students, was upgraded to the Spark Challenge Camp in 2025, and the Youth Science Training Camp in 2026 further extended its reach to 13-year-olds, forming a complete talent supply chain covering from high school students to doctoral students.

As of 2026, more than 800,000 people have joined the Youth Science Training Camp platform, which has cooperated with more than 2,000 schools, and the participants include 114 students who have won gold medals in national subject competitions.

ByteDance took another path. Zhang Yiming and Yu Yong founded the Zhichun Innovation Center, recruiting 30 full-time reserve researchers aged 16 to 18 every year, with a 5-year training cycle.

ByteDance's Top Seed program is also iterating: the annual salary of campus recruits rose from about 1.5 million yuan in 2024 to 3 million to 5 million yuan in 2025, and the annual salary for core positions reached 6 million yuan in 2026.

Baidu's AIDU program represents another idea, which does not compete for talents in the early stage, but deepens the cultivation. After the 2026 upgrade, it adopts a dual-mentor mode, with experts from the technical committee and business experts providing guidance respectively, the CEO and business group leaders personally participate in interviews, the annual salary starts at one million yuan, and there is no upper limit for compensation.

The above three paths, among others, actually point to the same fact: large tech companies are no longer satisfied with being downstream buyers of talents, but want to become upstream producers of talents, or at least complete a deep binding before talents are discovered by others.

2

Talent Pricing for Gifted People and Defensive Hoarding

Why are large tech companies willing to invest a 5-year training cycle for a 16-year-old child? The answer lies in the economics of AI R&D.

Wang Hao, founder of an AI startup, did the math: Alibaba's large model Qwen is backed by hundreds of thousands of computing power cards, representing an investment of nearly 10 billion US dollars. If a doctoral student can help optimize chip usage efficiency, an annual salary of several million yuan is extremely cost-effective in comparison.

Yang Ling, assistant professor at the International Machine Learning Research Center of Peking University, further explained that a single large model experiment may use hundreds or even thousands of GPUs, and the cost of a single failure can be as high as millions to tens of millions of yuan. If top researchers can reduce invalid training, enterprises are naturally willing to pay high salaries.

This logic determines that the pricing method of AI talents is fundamentally different from that of traditional technical positions. The value of traditional technical positions grows linearly, and the more experience you have, the higher the output. But the value of top AI talents is non-linear: a person who can optimize training efficiency can directly save tens of millions of yuan in computing power costs.

At the end of 2025, the first thing Yao Shunyu did after joining Tencent was to break the wall. In the past, the SFT data of Hunyuan was not deduplicated, and the number of repeated redundant data could reach tens of millions of entries.

He led a data team of more than 20 people to focus on data review, and in a few months, the redundant data was compressed from tens of millions of entries to more than 10,000 entries, cutting it by two orders of magnitude.

The result exceeded expectations: the training cost of Hy3 Preview was reduced by 47%, the task resolution rate of the official version of Agent soared from 72% to 90%, the hallucination rate was cut from 12.5% to 5.4%, and the input price was as low as 1 yuan per million Tokens.

With his own efforts and smaller parameters, Tencent has been made one of the domestic large model players with the strongest Agent capabilities.

This non-linear pricing is directly reflected in the salary. It is not uncommon for AI doctors from top laboratories of Peking University and Tsinghua University to join large tech companies for model R&D with an annual salary of several million yuan in their first year. Palantir in Silicon Valley is even more radical: the salary level of its direct recruitment for high school graduates is already the treatment of top AI doctors in China.

But beyond the pricing logic, there is a more ruthless dimension: defensive hoarding.

An observation from an industry insider is very straightforward: this advanced talent scouting is ostensibly respect for talents, but essentially a kind of defensive hoarding. Whoever locks in the potential talents first will be less passive in the next round of competition.

When all leading enterprises begin to move talent scouting forward, the focus of competition changes from "who offers better treatment" to "who takes action earlier". The salary of AI doctors from Peking University and Tsinghua University has risen from hundreds of thousands to millions of yuan in a few years, and the next thing to be pushed up is the signing price of high school students with Olympic gold medals.

A more extreme signal comes from ByteDance. According to reports from LatePost, Seed established a special department around 2026 to build a full-coverage talent pool, exhaust the information of outstanding students and fresh graduates in China, and accurately grasp the student resumes of key universities, laboratories and tutors.

When an enterprise even needs to accurately grasp the resumes of outstanding high school students, "talent scouting" is no longer a recruitment activity, but an intelligence work.

There is a thought-provoking difference between China and the United States: the younger talent trend in Silicon Valley mostly occurs at the level of startups and research institutions, while the younger talent trend in Chinese large tech companies directly occurs in the core management layer.

Yao Shunyu took charge of Tencent's Basic Model Department at the age of 27, and Luo Fuli reached level 22 of Xiaomi at the age of 31. The reason for this difference is that China's AI industry did not really start until around 2018, and the stock of existing talents itself is small. Large tech companies do not have the time window to "wait for veteran employees to transform". Silicon Valley has veterans with decades of experience such as Geoffrey Hinton and Yann LeCun sitting in, but China's AI industry cannot afford to wait.

However, Silicon Valley's "talent scouting" efforts are even greater, which means that if Chinese large tech companies' "talent scouting" is benchmarked against Silicon Valley, there is still much room for improvement.

At present, when high school students in China enter large tech companies, except for ByteDance's Top Seed which offers a daily salary of 2000 yuan, most projects are still at the level of practice and training, and have not entered the formal employment pricing system. This gap itself is a direction worthy of further exploration.

3

After Getting the Talents, Then What?

There is a cruel paradox in this competition: large tech companies are paying higher and higher prices to lock in younger and younger people, but the locked-in targets themselves are losing at an increasingly fast speed.

A frequently cited fact is that nearly 70 technical talents left ByteDance's Seed team in the past year, nearly 30 of whom joined Tencent to be responsible for AI Infra and data infrastructure work.

Ge Hao, a Top Seed intern at ByteDance, turned into a core member of the RL Infra training of Alibaba's Qwen team. Earlier, Li Yukun, the first employee of DeepSeek, left ByteDance's search team, and Xu Mingyu, a former Seed Edge member, also joined DeepSeek.

These figures illustrate a fact covered by the "talent scouting" narrative: the talent war between large tech companies is not a one-way hoarding, but a zero-sum cycle.

The talented young person you recruited from high school today may be poached by another company five years later with a higher salary, better computing power, and more free research directions. Locking a 16-year-old child requires a 5 to 10-year training cycle, but losing him only requires a better offer.

What pushes this absurdity to the extreme is the "garden leave" phenomenon in Silicon Valley.

According to reports, Google once paid high salaries to support some AI researchers, letting them do nothing for a year, with the only requirement that they do not join competitors.

Comments from the legal and economic circles summed up this behavior sharply: these scientists are "hoarded like Pokemon cards". A company pays nine-figure salaries, not to get people to work, but to prevent competitors from getting them to work.

When the strategic value of talents exceeds their use value, the competition for talents among large tech companies degenerates from competing for productivity to competing for exclusivity.

A 2026 working paper by Harvard Business School and Georgetown University formally modeled this: dominant enterprises sometimes hire and retain cutting-edge researchers, not for the purpose of making them produce output, but to prevent competitors from getting them.

Every researcher who is put on the bench is a person that challengers cannot hire. The incumbent protects its own profits, while the society loses the discoveries that these researchers could have made elsewhere.

Applying this logic back to the "talent scouting" of Chinese large tech companies, the nature of the problem changes.

Tencent's Youth Science Training Camp opens product practice opportunities for 13-year-old middle school students, ByteDance's Top Seed offers a daily salary of 2000 yuan, and Geely offers an annual salary of 300,000 yuan to high school graduates. Each of these actions looks like cultivating talents when viewed alone, but when viewed together, they are more like taking positions in advance.

A 16-year-old child is locked into the training track of a certain large tech company, and even if he wants to go elsewhere five years later, the sunk cost and path dependence have already formed. What the enterprise buys is not his output at the age of 16, but the discount of his future options.

What this child loses is not just the space for trial and error in college, but also the ability to repeatedly negotiate prices in the talent market. And what the large tech company cultivates over five years may be a person with high loyalty but whose market bargaining power has been overdrawn in advance, or a person who is ready to change jobs at any time and takes the large tech company as a springboard.

Yang Ling's judgment has a new meaning here. He said that the supply of AI talents will gradually increase in the next 5 years, and the demand for talents cannot expand endlessly.

But what he did not say is that when the supply increases, the "talented young people" locked in at sky-high prices today may become costs that large tech companies are eager to get rid of tomorrow. The reverse of the garden leave logic is: the people you hoard at a high price today may not even need to be put on the bench tomorrow.

4

Epilogue

The real stake in this competition is not which large tech company can grab the most geniuses, but a more fundamental question: can creativity be owned?

The underlying logic of large tech companies' "talent scouting" is to treat talents as an asset that can be locked, hoarded, and discounted.

From picking fruits to growing crops, from recruitment to intelligence work, from employment to "garden leave", every step turns human creativity into a line of numbers on the balance sheet.

But creativity has a weird property: it is only stimulated when there is freedom of choice, and shrinks when it is locked. The more you want to own it, the less it belongs to you.

This means that "talent scouting" faces a structural paradox: an enterprise spends five years cultivating a 16-year-old child, aiming to make him make breakthrough work, but the premise of breakthrough work is precisely that he