A 29-year-old outstanding female leader from ByteDance has gone viral.
The story of the prodigy girl now belongs to ByteDance.
Zeng Yan is a researcher in ByteDance's Seed team who has barely appeared in public. Born in 1997, she currently serves as the head of pre-training for Seedance, leading the Seedance 2.0 pre-training pipeline. These days, rumors are spreading that Zeng Yan has been promoted to a core manager of Seedance, taking over an increasing number of responsibilities from Zhou Chang, the former head of Seed's visual generation and multimodal division.
Scrolling through social platforms, posts discussing Zeng Yan are going viral: 29 years old, joined ByteDance via campus recruitment in 2021, became the core head of Seedance in just five years, approved the 2000-billion-parameter model, got two consecutive promotions within one year... The full picture of her career is gradually becoming clear.
This is ByteDance's internal talent cultivation success story — Zeng Yan is not a high-paid external technical talent poached by ByteDance, but a campus recruit who grew up step by step within the system. The 5 years she spent at the company also witnessed a leading internet giant redistributing decision-making power in the AI era.
Xi'an Jiaotong University, Fastest Promotion Record for Campus Recruits
One of the impressions Zeng Yan left to the public is her extremely fast promotion pace.
According to public information, she completed her undergraduate study at Xi'an Jiaotong University, then pursued a master's degree in computer-related major at the University of Montreal in Canada. In September 2021, the 24-year-old joined ByteDance AI Lab as a campus recruit, starting her career as an algorithm engineer.
At that time, ChatGPT had not been released yet, and ByteDance's AI was mainly used for recommendation and content understanding. Only two months after joining the company, Zeng Yan published the paper Multi-Grained Vision Language Pre-Training: Aligning Texts with Visual Concepts on arXiv as the first author. This research applied multi-grained alignment to enable the model to establish image-text correspondence between the overall scene and local visual concepts simultaneously, and proposed the well-known X-VLM model later.
After that, Zeng Yan continued to focus on vision-language pre-training, participated in projects such as CCLM and Lynx, and shifted from image-text understanding to more complex multimodal fields.
In 2023, ByteDance established the large model team Seed, and Zeng Yan transferred to the new team.
In November of that year, Zeng Yan published the paper Make Pixels Dance: High-Dynamic Video Generation as the first author — the method uses "first and last frame + text" for joint guidance, enabling the model to maintain consistency of characters and scenes during complex motion. She locked in the multimodal and video generation track at a very early stage, which was a relatively unpopular field at that time but later became ByteDance's core strategic track.
PixelDance, corresponding to this paper, kicked off the two-track parallel competition for ByteDance's internal video generation route: Zeng Yan led the team to advance PixelDance, while another team explored the Seaweed route in parallel. The two routes represented different technical trade-offs. After both routes were verified successfully in parallel, the teams completed convergence and merger at the end of 2024, shifting from the original 3D-UNet architecture route to DiT, and the Seedance pipeline was thus finalized.
For Zeng Yan, this was a critical career turning point: she grew from a solid executor in multimodal pre-training and video generation projects to the route leader who decides the development direction of the entire video generation pipeline.
In 2024, Zeng Yan's position was adjusted from algorithm engineer to algorithm researcher, making her one of the youngest researchers in the Seed team.
Seedance 1.0 was released to the public in June 2025. At the team dinner at the end of that year, Zeng Yan, who was already the head of video models, proposed to directly train a 200B-parameter model, which caused considerable disagreements within the team. However, Wu Yonghui, the general head of Seed, and Zhou Chang chose to support her judgment after discussion. This seemingly radical choice at that time later proved to be one of the key factors that enabled Seedance 2.0 to open up a generation gap with competitors.
In February 2026, Seedance 2.0 was released, and quickly became viral for its multi-shot narrative, higher picture fidelity, and the ability to produce cinematic-level finished videos within 60 seconds.
What followed was Zeng Yan's almost "exceptional" promotion pace — according to comprehensive public information, between 2025 and 2026, Zeng Yan completed the rank leap from 3-2 to 4-2. In ByteDance's technical rank system, 4-2 means entering the core expert level, where one can independently take the lead in defining technical routes, schedule large-scale computing power resources, and hold decision-making power over the R&D of models with parameters at the 10-billion and 100-billion level.
As a campus recruit who joined the company in 2021, such a rapid rise is not common in ByteDance.
Zhang Yiming Pours Heavy Investment in AI Talents
Zhang Yiming, who rarely appears in the public view nowadays, every piece of news about him is mostly related to AI strategy.
According to Bloomberg, this month Zhang Yiming put forward a long-term judgment at an all-hands meeting for ByteDance's AI team: ByteDance has the foundation to compete with OpenAI and Anthropic, but to catch up with the cutting-edge base model capabilities, "the earliest possible time will be 2029 or 2030", and the goal of the Seed large model is to benchmark against the first-tier international players.
On the choice of technical routes, he clearly opposed relying on the distillation shortcut to quickly climb the leaderboard, and advocated native base model training, accepting short-term gaps.
It can be seen that in the past two years, ByteDance's strategic focus has fully shifted to AI, with bets laid on computing power infrastructure, base models, cutting-edge hard technologies and application layer layout.
Market rumors say ByteDance's capital expenditure in 2025 exceeded 1.5 trillion yuan, of which about 900 billion yuan was used for AI computing power procurement; in 2026, this figure was raised to more than 2 trillion yuan, of which about 850 billion yuan points to AI chips. In the same round of competition, Alibaba completed a share placement of about 80 billion Hong Kong dollars early to increase investment in AI infrastructure, and Tencent's capital expenditure in the second quarter of this year rose to 52.784 billion yuan.
AI is undoubtedly the largest battlefield at present. At the beginning of this year, CEO Liang Rubo set the annual key phrase as "climb the peak bravely", judging that AI is an industrial-level opportunity comparable to PC, Web and mobile Internet, and required ByteDance to concentrate resources to tackle the base model capabilities of large models, "There are many important opportunities in the AI era, we need to pursue the most important ones and climb the highest peak."
A single tree cannot make a forest, and "climbing the peak bravely" requires a prerequisite — talents, whose importance in the AI era is unprecedented.
It is remembered that at the all-hands meeting earlier this year, Liang Rubo made it very clear: increase talent density, increase incentives, and make salary and incentives "ahead of the top level" in every market around the world. The supporting policies were implemented via global internal email on December 19, 2025: bonus investment increased by 35%, salary adjustment investment increased by 1.5 times, and the upper and lower limits of the total compensation for all ranks were raised simultaneously.
One of the reasons Wu Yonghui once talked about for joining ByteDance is that he wanted to build a first-class AI research team. His consistent idea is to "reuse young people" — more programs for young people have been launched one after another, for example, the Top Seed research intern program offers a maximum salary of 2000 yuan per day, and the eligibility is extended to undergraduate and even high school students; in the previously launched campus recruitment program of Seed, the team plans to recruit 100 graduates of the class of 2027.
In this context, Zeng Yan's promotion has a clear footnote.
When talking about AI talents, Zhang Yiming once used an analogy: people trained in the traditional mode "have professional knowledge and proficient skills, but cannot handle innovative tasks", which is like overfitting in machine learning; what he wants is people who are "active in thinking, passionate and resilient" — the core is independent thinking and emphasis on practice, focusing on long-term and overall situation, learning in exploration, keeping a normal mind, being relaxed and confident, and embracing uncertainty.
Post-95s Generation Taking Power Collectively Is Rewriting China's AI Development Direction
Pull the camera away from ByteDance alone, and you will see a rearranging seat table — over the past year, large tech companies including Tencent, Alibaba, Xiaomi, Baidu and others have almost in the same window handed over the core model R&D work to young people around 30 years old.
Yao Shunyu is a name that cannot be avoided when talking about Tencent. Born in Hefei, Anhui in 1998, he studied at Hefei No.1 High School, and was admitted to Tsinghua University's Yao Class with a score of 704 in 2015. During his PhD study at Princeton University, he proposed the Tree of Thoughts framework and the ReAct method — the latter is now the mainstream technical route for global language agent development; he joined OpenAI in 2024 and participated in the R&D of core agent products such as Operator and Deep Research.
Until December 2025, Yao Shunyu joined Tencent as the chief AI scientist in the CEO/President's Office. In July 2026, Tencent merged the Hunyuan large language model department and the multimodal model department into the base model department, which is under his overall management. At a research symposium where leaders of Guangdong Province visited Tencent's headquarters in October this year, the 28-year-old Yao Shunyu sat next to Ma Huateng.
A little older, Liu Dayiheng from Alibaba was born in 1993 in Zigong, Sichuan. He completed his bachelor, master and doctoral studies at Sichuan University, was selected into the first tier of Huawei's "Genius Youth" program in 2020, and joined Alibaba Damo Academy the next year, taking charge of pre-training since the early stage of Qwen. He has participated in the entire series from Qwen1 to Qwen3.5, participated in the open source of more than 300 models in total, and won the Best Paper of NeurIPS 2025 as the corresponding author.
At the Yunqi Conference this September, Liu Dayiheng, as the head of the ATH business group's Token Foundry Qwen LLM project, took the stage to share after Wu Yongming.
Xiaomi set a new ceiling for job ranks. Luo Fuli, born in Yibin, Sichuan in 1995, got her bachelor's degree from Beijing Normal University and master's degree in computational linguistics from Peking University. She used to work on VECO at Alibaba Damo Academy, then joined High-Flyer and DeepSeek (core developer of DeepSeek-V2), and officially announced to join Xiaomi to lead MiMo in November 2025. On September 22, 2026, she was promoted to Xiaomi's rank 22 — the highest level in the current rank system, only about 10 months after she joined the company.
Nowadays, there are rumors that Xiaomi's large model business is facing discussions on whether to be spun off, and Luo Fuli's next step is highly anticipated.
The list goes on: Sun Tianxiang, born in 1997, took over as the head of Baidu's base model R&D department this summer; Chen Yusen, born in 1992 at Alibaba, took over as CEO of DingTalk in June, becoming the youngest division CEO in Alibaba's history; post-2000s researchers at Zhipu AI have already participated in technical discussions on GLM pre-training and infrastructure...
These names confirm the same fact — the decision-making power of AI campaigns in large tech companies is being handed over to a group of young people. In the final analysis, the competition of AI is never just a competition of models and computing power, but a competition of talent density. Large tech companies not only need to "snatch talents", but also need to "grant power to them".
It is recalled that Luo Fuli once gave advice to young people:
"Make the most of cutting-edge AI tools, and cultivate your own unique judgment and scientific research aesthetics in the process of continuous trial and error. In the era of rapid technological change, unique cognition, judgment and aesthetics are the core, most irreplaceable competitiveness for young people."
This article is from the WeChat Official Account Pedaily (ID: pedaily2012), written by Yang Jiyun, authorized to release by 36Kr.