Innovation requires a certain degree of ambiguity and disorder, while competing effectively demands order and discipline.
Upon taking over ByteDance's foundational model research department, Seed, in early 2025, Wu Yonghui encountered a situation where a large, well-funded research team that had spent two years developing a top-tier domestic model was quickly surpassed by a much smaller team with fewer resources. The department head acknowledged the misstep, and the company's CEO noted in a company-wide meeting that better results were possible. According to multiple Seed insiders, Wu was expected to elevate the model's capabilities to be the best in China and competitive with leading international model companies.
Colleagues describe Wu as "steady." After assuming leadership, he conducted intensive one-on-one meetings with over 100 core researchers and subsequently promoted several researchers focused on model architecture.
Throughout his first year, Wu concentrated on two primary objectives: enhancing the capabilities of the foundational model and improving research efficiency to ensure timely delivery, and fostering a research-oriented culture aimed at conducting first-class research and building a premier AI research team. A person close to Wu described him as "both pragmatic and romantic," noting that he seems to have grown within the ByteDance system rather than being a product of Google.
Wu Yonghui joined Google after earning his PhD in 2008. He spent his first seven years there as a software engineer working on core search ranking systems before transitioning to Google Brain for AI application research, where he contributed to transforming machine translation with deep learning and integrating it into search ranking algorithms. By 2023, he had become a Vice President of Research at Google DeepMind, involved in the development of the Gemini model and its early-stage competitive efforts. Another individual familiar with Wu stated, "He has a deep understanding of technology and can quickly identify which directions are likely to yield results." Over the past year, Seed's text-to-image and text-to-video models have ranked among the global leaders in certain benchmark tests, and the Doubao mobile assistant model has become an industry focus.
In the highly competitive and critical area of foundational models, Seed has iterated through four versions (including the upcoming Doubao 2.0), showing significant improvement over previous models and continuously narrowing the gap with leading overseas models. However, a Seed insider noted that there is still ground to cover to address "technical debt" accumulated over past years. While pursuing the long-term goal of building a "first-class research team," Wu must also constantly balance this with achieving short-term objectives.
In January 2025, Seed established the "Seed Edge" virtual team with a three-year evaluation mechanism to encourage key researchers to tackle more fundamental, long-term AGI challenges. Wu Yonghui was involved in forming this team. He later assembled a Focus team by reassigning researchers to break down departmental silos and existing divisions of labor, tasked with overcoming core challenges and developing improvements for the next model version. The remaining teams working on the current generation of the foundational model were organized into a Base group, encompassing engineering, data, and evaluation. That March, following the suspension of the then LLM team leader Qiao Mu after a reported incident, teams under LLM such as Pre-train, Post-train, and Horizon began reporting directly to Wu Yonghui.
Two Seed sources explained that Wu's plan allows for the concurrent development of three model generations, with personnel and projects able to rotate, thereby optimizing internal resources. "Achievements from Edge can be directly implemented; long-term projects identified by Focus can transition to Edge, and their results can be used to enhance the current model generation." To improve efficiency, Wu also promoted greater internal transparency of data and code libraries, while maintaining external confidentiality, stating internally that this was to "address some internally raised issues."
Within ByteDance, Seed operates separately from revenue-generating departments and reports directly to group management, making it easier to avoid internal politics. However, within Seed itself, multiple teams often work on similar research directions, and communication between teams has historically been challenging. Two former Seed researchers said the department previously resembled a collection of research groups rather than a unified department, where obtaining access to documents from another group required approval from the document owner and their manager—a process so difficult that even department leaders sometimes could not secure full access. While improved information sharing boosted communication efficiency, it also led to complications. In the second half of 2025, Seed experienced at least two incidents of information leakage by interns, after which the policy of open internal document authorization was revised.
Outside the core foundational language model research direction, Seed largely maintained its existing organizational structure. The previous head of Seed, Zhu Wenjia, reported to Wu Yonghui after a period of joint leadership for several months, taking responsibility for large model applications. The multimodal interaction and world model team is led by Zhou Chang, who joined ByteDance from Alibaba in 2024. The team's representative achievement is the Doubao mobile assistant model. Over the past year, following the leave of visual multimodal generation head Yang Jianchao and the departure of visual foundational model research head Feng Jiashi, Zhou Chang's management responsibilities expanded to include new models like the text-to-image Seedream and text-to-video Seedance. The Infrastructur team is led by Xiang Liang. ByteDance's AI Lab, with its remaining three research focuses—AI for Science, Robotics, and Responsible AI—was integrated into Seed, with its head, Li Hang, reporting to Wu Yonghui.
Over the past year, Seed's overall headcount has remained around 1,500 people, with expansion slowing compared to the previous two years. The division has almost ceased external hiring for mid-to-senior level technical managers, instead placing greater emphasis on recruiting recent graduates and promoting young talent. For instance, a PhD graduate from Tsinghua University in 2024 now reports to both Zhou Chang and Wu Yonghui.
The upcoming Doubao 2.0 model is the most significant output during Wu Yonghui's first year leading Seed. It is a multimodal model similar to Gemini, with 1 trillion parameters, making it the largest model trained by Seed since its establishment. After Wu's arrival, many Seed researchers noted a significant increase in meeting frequency. Teams that previously met every three months now convene monthly, with the core team meeting bi-weekly. A ByteDance employee mentioned frequently seeing Wu in the cafeteria, tray in hand, sitting down with researchers to discuss progress over a meal. After research team presentations, Wu typically asks a few brief questions—usually three. He does not provide direct solutions but guides the team to contemplate more fundamental issues, according to one Seed researcher.
Multiple Seed sources indicated that training the Doubao 2.0 model faced challenges at the infrastructure level. They analyzed that during the intense catch-up phase of the previous two years, foundational capability building had been relatively neglected. Consequently, scaling up the parameters for Doubao 2.0 led to instability, at one point stalling progress. A Seed researcher analogized, "It's like building a house; if the foundation is unstable, adding more bricks on top is prone to cause problems." Weng Jiayi, head of RL Infra at OpenAI, remarked in a podcast that every model team's infrastructure has bugs, and the essence of competition among model companies is the speed at which they fix these bugs. This speed determines how many ideas can be tested within a given timeframe, while generating ideas is solvable by increasing talent density. OpenAI began重构 its three-year-old infrastructure system last year to address accumulating "technical debt."
Over the past year, the large model teams at Alibaba and Tencent have also increased their focus on infrastructure. The Qwen model team began forming an internal Infra team mid-last year, work previously reliant mainly on Alibaba Cloud's AI platform PAI. Tencent announced the establishment of an AI Infra department and a data computing platform department at the end of last year, directly overseen by its large model head, Yao Shunyu. For the Seed team, overhauling its infrastructure system presents greater difficulty. Seed's Infra team, comprising hundreds of personnel, supports the research and experimentation of dozens of models within Seed and is considered by senior management to be top-tier domestically. A Seed insider stated that a comprehensive重整 would require significant investment in manpower and resources, along with substantial "trust costs," necessitating a "fix the wheel while the car is moving" approach. Following the issues encountered during Doubao 2.0's training, multiple teams collaborated for three months, primarily addressing problems through model architecture and training data adjustments, to ensure the model's launch before the Chinese New Year.
Demis Hassabis, CEO of Google DeepMind, once stated, "We are committed to creating a unique cultural blend: the focus and agility of a startup, combined with the blue-sky thinking of academia." Multiple Seed sources said this is also Wu Yonghui's goal—he wants to establish Seed as a first-class research brand. ByteDance has provided an environment conducive to this. Over the past two years, ByteDance has gathered a pool of research talent at Seed, often hiring promising individuals even without an immediately suitable position. ByteDance has implemented more relaxed assessment mechanisms for Seed, largely eliminating OKRs for most. In mid-2025, ByteDance issued "Doubao virtual shares" as an incentive, independent of company stock options, and raised compensation multiple times. A relaxed work atmosphere is also fostered at Seed, where interns can communicate directly with top leadership. A top Seed intern shared on social media that during a nearly two-month project, they mostly slept at the company, often waking up excited with new ideas at 2 a.m., coding for two hours before going back to sleep, feeling "like discovering new phenomena every moment."
In a department-wide meeting last March, Wu Yonghui stated that Seed had done much excellent work that was not well known externally and encouraged researchers to publish their findings in the form of papers and blog posts. "He also suggested everyone 'decorate' their personal homepages," said a former Seed member. According to preliminary statistics, the number of papers published by the team in the three months after Wu joined exceeded the total for the entire year of 2024. The ability to showcase research成果 externally further motivated Seed researchers. "It filtered out a group of very diligent, focused, or say, research-obsessed young people," commented a core researcher who left Seed last year. Most of them follow a simple routine between work and home, eating only in the company cafeteria for convenience, finding their primary enjoyment in the "feedback" from their work. It is understood that a practice exists at Seed where several researchers can decide to spontaneously investigate a direction. If their direct superior disagrees or refuses resources, they can escalate the proposal. If approved from above, it can be pushed forward top-down. "They were 'top dogs' in their previous fields, coming to Seed with ego, each wanting to achieve something great," the same source estimated, suggesting that perhaps only 20% of resources are allocated to short-term, immediately useful projects, adding that "waste might be the right approach; ByteDance isn't short on resources."
However, Seed is not a research organization isolated from competition. It needs to provide ByteDance with the "ammunition" to compete against companies like Tencent and Alibaba. Team resources inevitably tilt towards those demonstrating short-term results. Multiple Seed sources said that starting in the second half of last year, researchers working on isolated, single-point studies felt their work, if perceived as not aligning with Seed's mainstream direction, received little attention. From the third quarter of last year, ByteDance management imposed new requirements on Seed's paper publications, including "high quality" and "content unrelated to core technologies currently under iteration." Subsequently, the monthly number of published papers decreased. "Yonghui wants to create a research atmosphere, but it's hard to avoid deadline-driven priorities," said one Seed researcher. Their team originally planned to dedicate more resources and time to developing a proprietary algorithm to solve a specific technical problem. However, to meet a faster launch schedule, shortly after project initiation, they were instructed to change approach and build upon an open-source project instead, "even if it meant a performance compromise."
In January of this year, ByteDance CEO Liang Rubo stated in a company-wide meeting that the keyword for the year is "Scaling New Heights," with the short-term peak being the "Doubao/Dola assistant application." The focus is on ensuring "AI model capabilities are industry-leading and integrating existing businesses effectively through the assistant." In the company-wide meeting a year prior, Liang had said the AI research team should "explore the upper limits of intelligence." Innovation requires appropriate灰度 and chaos, while应对 competition demands order and discipline. Balancing these two objectives remains the ongoing management challenge for Wu Yonghui at Seed.
Comments