Gu Quanquan Leaves ByteSeed, Doubao's Paid Model Is a Hidden Agenda

Gu Quanquan, the person in charge of ByteDance's Seed pre - training, posted a short farewell message on X.

However, Gu Quanquan's departure from Seed is not the biggest change in ByteDance's story.

The real change is that Doubao will start charging in June.

The departure of a top - level researcher certainly deserves attention. The greater concern is that ByteDance has entered the second stage, where AI has become the main engine driving ByteDance's growth.

In the past two years, Doubao has helped ByteDance prove one thing: it has the ability to promote AI products to hundreds of millions of users.

However, after the charging begins, Doubao needs to prove another thing: out of these hundreds of millions of users, how many are willing to pay for AI?

From this moment on, the things developed by Seed are not just about papers and scores. They have to be transformed into products, revenues, and user experiences.

This means that ByteDance has to make choices and cannot explore all directions comprehensively.

Gu Quanquan's departure happened right at this turning point.

Gu Quanquan: The One Connecting Biological AI, Foundation Models, and Scaling Abilities

Gu Quanquan completed his undergraduate and master's degrees in the Department of Automation at Tsinghua University. He obtained his Ph.D. in Computer Science from the University of Illinois at Urbana - Champaign in 2014. After graduation, he first worked as a post - doctoral researcher at Princeton University (2014 - 2015), then served as an assistant professor at the University of Virginia in 2015, and joined UCLA in 2018. His research areas cover machine learning, optimization algorithms, and statistical learning theory.

In 2023, Gu Quanquan joined ByteDance's Seed.

That year, ByteDance just completed a major adjustment of its AI organizational structure and established its first large - model team, led by Zhu Wenjia.

At the beginning of 2024, a comprehensive reorganization was completed. Flow and Seed were upgraded to first - level departments on par with Douyin and reported directly to Liang Rubo.

In February 2025, Wu Yonghui joined and took over the basic research of Seed, while Zhu Wenjia shifted to model applications.

Seed needs not only scientists who publish papers but also those who understand model training, can apply their abilities to specific problems, and have led high - intensity projects.

Gu Quanquan exactly fits this profile.

He was involved in two lines at Seed: one was AI4S, and the other was cutting - edge LLM.

AI4S (AI for Science) means using AI to solve scientific problems, such as protein structure prediction, drug molecule design, and material research and development. These tasks, which used to take laboratories several years to complete, can now be solved by AI within a few hours.

AI4S was a relatively popular line at that time. Google's AlphaFold was a milestone product in AI4S.

However, Gu Quanquan's cross - field work is not common in large - company AI organizations. Most scientists either focus on a specific vertical field or on basic model training. Few can achieve results in both directions.

In the AI4S field, he led the development of the SeedFold, SeedProteo, and DPLM series.

SeedFold is a biomolecular structure prediction model developed by ByteDance's Seed. The paper shows that SeedFold outperforms AlphaFold3 in multiple protein - related tasks on FoldBench.

SeedProteo is a de novo all - atom protein design model for protein binder design.

If SeedFold is about "understanding" protein structures, then SeedProteo is about "designing" new protein molecules. The latter has greater commercial potential but also higher technical difficulty.

The DPLM series is a protein language model that attempts to make AI understand and generate protein sequences in a way similar to training large language models.

There has been a lot of exploration in this field in academia, but the characteristic of ByteDance's Seed is to combine it with its own basic model capabilities to form a relatively complete AI4S technology stack.

These achievements carry significant academic weight.

After the publication of the SeedFold paper, it was cited and reproduced by multiple research institutions. The performance of SeedProteo in protein design tasks was also considered one of the strongest models in the industry at that time.

Gu Quanquan's reputation in the AI4S field was largely built up through these projects in the past three years.

However, AI4S is only half of Gu Quanquan's work at Seed.

At the beginning of 2025, he joined the LLM pre - training work, established an LLM optimization and expansion team, and participated in the training of Seed 2.0. This shift seemed sudden at that time, but if one understands the changes in ByteDance's AI strategy, it is actually an inevitable choice.

The popularity of DeepSeek made all large companies realize that pre - training ability is not just a matter of computing power but also of engineering and optimization capabilities.

ByteDance's Seed needed someone who could systemize pre - training.

Gu Quanquan understands statistical learning theory, optimization algorithms, and has experience in large - scale training, which makes him the most suitable candidate for this position.

The goal of the LLM optimization and expansion team he established was to build a "highly scalable pre - training stack" to enable the stable training and iteration of Seed 2.0 and subsequent frontier - scale models.

According to his LinkedIn post, this goal was basically achieved. He mentioned that "I led the team to build a highly scalable pre - training technology stack, which successfully supported the training of Seed 2.0 and subsequent cutting - edge large models."

This means that ByteDance's Seed has no longer relied on external technologies or single - point breakthroughs in pre - training ability but has formed an engineering system that can be continuously iterated.

In the three years at ByteDance's Seed, Gu Quanquan connected ByteDance's scientific problems, basic model capabilities, and large - scale training capabilities.

Gu Quanquan's value lies in his ability to achieve results in both the AI4S and LLM directions and systemize these abilities.

However, this is also the problem.

The things Gu Quanquan did have strong "long - term value" but unclear "short - term product value."

After Doubao starts charging, the value ranking of ByteDance's Seed will inevitably become more realistic.

AI4S Organizational Adjustment

A few days before Gu Quanquan posted his departure message, there was news that the AI4S team under ByteDance's Seed was undergoing an organizational adjustment.

People close to ByteDance said that "spin - off is not considered," and the AI4S team will be led by Yang Zhenyuan. Subsequently, there were news that core members related to AI4S, such as Xiao Wenzhi and Gu Quanquan, had left or were preparing to leave to start their own businesses.

"Not considering spin - off" means that ByteDance does not plan to spin off AI4S as an independent business or make it an independently operating entity.

AI4S is still part of Seed and must follow Seed's overall strategic priorities. Yang Zhenyuan's takeover indicates that ByteDance still invests in this direction, but the investment method and goals may have changed.

AI4S is a very special field. Its value is difficult to measure with our current product indicators.

Developing a protein structure prediction model that outperforms AlphaFold3 is of course a major academic breakthrough. However, there is still a long way to go to transform this breakthrough into commercial revenue.

Another point is that the achievements of AI4S are likely to "follow the people."

For products like Doubao, the capabilities are more embedded in the company's system. For example, the model platform, training cluster, inference architecture, and product entrance are all in ByteDance's hands. The departure of people will have an impact, but the system remains.

However, AI4S is different. It often depends on the cross - understanding of biology, chemistry, protein structure, drug discovery, and model methods by several core researchers. Many achievements are not just an app function but a set of research routes, model assumptions, data processing methods, experimental judgments, and industrial resources.

A large part of these things exists in people's minds, resumes, and networks.

So, after Gu Quanquan's departure, the impact on Seed 2.0 may be small.

But it's different for AI4S.

The departure of core members like Xiao Wenzhi and Gu Quanquan has a great impact on the AI4S team.

They take away not only technical capabilities but also reputation, connections, and judgment on future directions in this field.

After Yang Zhenyuan takes over, it is still unknown whether the AI4S team can maintain its presence in academia and the industry.

The more crucial question is how much resources and patience ByteDance is still willing to give to AI4S.

AI4S is a direction that requires long - term investment and has unclear short - term returns.

Its value may appear in three or five years, or it may never appear. For a commercial company, this kind of uncertainty is difficult to bear in the long run.

Especially when Doubao starts charging, the situation of AI4S will become even more awkward.

ByteDance is willing to invest in AI, and it is among the top in domestic large companies.

In May, there was news that ByteDance raised its capital expenditure plan for 2026 by at least 25%, increasing the AI - related capital expenditure from the previously discussed approximately 160 billion yuan to more than 200 billion yuan.

However, patience is limited. Coupled with ByteDance's internal horse - racing mechanism, the team must achieve results within a certain period.

When the organizational priorities change, the directions with strong long - term value but unclear short - term contributions may have their priorities lowered.

The organizational adjustment of the AI4S team is essentially part of this process.

It doesn't mean that ByteDance is giving up AI4S, but rather that ByteDance wants AI4S to serve product and commercialization goals more clearly.

For scientists, this is a very realistic choice.

Staying means accepting the new logic of the organization and adjusting the research direction closer to product requirements. Leaving means moving to an environment that respects "long - term scientific value" or starting a business on their own to turn their technological ideals into a new organization.

Xiao Wenzhi and Gu Quanquan chose the latter. It's not because they don't recognize ByteDance, but because they have more confidence in their judgment in the AI4S direction. They believe that this direction has long - term value and that they have the ability to realize this value.

Doubao's Charging: The Real Turning Point

The root cause of all this is that Doubao is going to start charging.

Actually, almost all AI products are charging. OpenAI's ChatGPT Plus, Anthropic's Claude Pro, Google's Gemini Advanced, and domestic products like Kimi, Zhipu, and MiniMax all have paid versions.

I think there is nothing wrong with Doubao charging.

In the past, Seed could tell a grand story about its technological landscape.

LLM is the foundation, multi - modality is the extension, video generation is the breakthrough point, voice is the experience enhancement, AI4S is the long - term layout, and Agent is the future direction.

Each direction has its own value, and each team is doing its own thing. Seed's role is to integrate these capabilities to form a complete AI technology system.

This story was valid in 2023 and 2024.

At that time, the entire AI industry was exploring directions, and all large companies were casting a wide net, trying to occupy positions in every possible technological direction. ByteDance's Seed had a wide and deep technological landscape, which was itself a competitive advantage.

But now, the situation has changed.

In Doubao's own words, "I will tell you in the most direct, honest, straightforward, heart - piercing, hardcore, decisive, no - nonsense, pain - hitting, unsparing, incisive, and straightforward way. Charging means that the product has to be responsible for the user experience."

At this time, each team within Seed has to answer this question: Can the function you develop be included in Doubao's membership benefits?

Obviously, the value of AI4S is long - term, strategic, and targeted at specific industry customers. It may become ByteDance's core competitiveness in the fields of biomedicine, materials science, and chemical synthesis in three or five years.

But currently, it's difficult for Doubao to prove its product value.

This is not a problem specific to ByteDance but a problem that all commercial companies face in the AI commercialization stage.

The same is true for Google. DeepMind used to be able to carry out long - term projects like AlphaGo and AlphaFold. But later, Google found that DeepMind's research direction should also be closer to products. So Google merged DeepMind and Google Brain into Google DeepMind and made Gemini the core of Google's AI strategy.

After that, projects that cannot contribute value to Gemini, Google Search, and Google Cloud will find it difficult to get enough resource support.

ByteDance's Seed is now on this path.

Whoever can reduce the inference cost can increase Doubao's gross profit margin; whoever can improve the user experience can increase user retention and make users pay.

Because these capabilities can be quantified, evaluated, and directly correspond to revenues and profits, which also conforms to ByteDance's internal value system.

Meanwhile, there is a problem that we need to view dialectically. AI development requires long - term investment, tolerance for failure, and enough freedom for scientists to explore unknown directions, which is beyond reproach.

However, AI commercialization requires short - term returns, clear value proof, and each ability to correspond to scenarios where users are willing to pay.

In the early stage of the AI industry, all large companies were casting a wide net, trying all possible directions, and could tolerate some teams conducting long - term research without considering short - term returns.

But when AI enters the commercialization stage and products start charging, the company has to be responsible to shareholders, investors, and users.

This is the turning point that ByteDance's Seed is facing now.

Gu Quanquan's departure is not a bad thing for both parties. It's more like a mutual achievement.

For Gu Quanquan, he has proved himself during his three - year tenure at ByteDance's Seed. These experiences and reputation have enriched his resume. He can return to academia to continue in - depth research on AI4S or start a business, just like Lin Junyang, turning his technological ideals into a new story.

For ByteDance, the transformation of Seed's organizational nature is irreversible. The subsequent iteration of Seed 2.0 has not been affected, Doubao is still being iterated, and Volcengine is still launching new capabilities.

A healthy AI ecosystem naturally requires different types of organizations.

Gu Quanquan has found a more suitable position for himself, and ByteDance's Seed has found a clearer direction. This is a two - way choice and a two - way achievement.

This article is from the WeChat official account "Zimu AI", written by Miao Zheng and published by 36Kr with authorization.