Byte AI's re-startup: independent organization, full-chain saturation attack

When it comes to AI, ByteDance is still the same ByteDance: once it sees a direction with potential, it will double, saturate, and launch an all-out attack.

A recent example is that before and after the release of the intelligent agent application Manus, ByteDance had at least five teams developing different intelligent agent products, some of which were internal tools. Manus is an intelligent agent application that was just launched in internal testing by the startup Monica on March 6.

Among all the big tech companies in China, ByteDance was the last to start large language models. Before OpenAI ChatGPT went online at the end of 2022, Baidu, Huawei, Alibaba, and Tencent (in chronological order) had all released large language models, but ByteDance had not.

However, after deciding to rely on its own team to develop AI in mid-2023, ByteDance has accelerated its investment in the entire chain and all applications.

In the era of mobile Internet, ByteDance, which was still a startup, could not invest in infrastructure from the beginning, and the chain of mobile Internet was also shorter than that of AI industry. This time, ByteDance has invested in the whole chain of computing chip layer, cloud computing layer (Volcano Engine), model R&D layer, and application layer.

The AI ​​products launched by ByteDance in the past two years have also involved most mainstream application directions such as chatbots, AI search, AI browsers, agent platforms, AI companion social networking, and AI education. Among them, the main product Doubao will become the AI ​​application with the most daily active users in China by the end of 2024.

But at this moment, the interim results of ByteDance, which has accumulated a lot of resources, have been overshadowed by the sudden emergence of DeepSeek. DeepSeek has produced amazing models with a small local technical team of less than 200 people, which is the opposite of ByteDance's routine of gathering a large number of smart people and accustomed to producing results with great efforts.

An even bigger threat is that Tencent, another Chinese Internet giant with a strong core business, abundant resources and ample room to seize AI opportunities, has finally reached its AI turning point.

ByteDance has never monopolized good ideas, but it has always tried to monopolize the commercialization of good ideas. Most of the time in the past, it has succeeded. Whether it is information recommendation, short videos, free novels or short dramas, ByteDance is not the first company to do it, but it can often catch up with better product experience and higher commercialization efficiency. Behind this is rich resources, a large team of talents and product capabilities accumulated over many years.

Will ByteDance continue to dominate as technology versions change? This is a big suspense in China's AI market.

Last November, we wrote in an article: “ByteDance is not the only company in China that has strong product capabilities and traffic resources. WeChat has not yet made a move.” Now Tencent, which owns WeChat, has finally made a move in an unexpected way: fully integrating DeepSeek.

This has had a more substantial impact on ByteDance. On March 19, Tencent President Martin Lau said at an earnings conference that Yuanbao's daily active users increased 20 times from February to March, ranking third among Chinese AI applications. The top two he did not mention were DeepSeek and ByteDance Doubao.

With only one-tenth of the time of ByteDance and a much smaller advertising budget, Tencent's user base has reached about one-fifth of Doubao's.

The competitive situation ByteDance faces is also changing in terms of model capabilities that cannot be quickly improved by surprise attacks.

After nearly two years of intensive updates, the number of ByteDance models, the speed of iteration, and performance have all improved significantly. At the ByteDance all-staff meeting in February this year, Zhu Wenjia, one of the leaders of the ByteDance model development team Seed, said: "At present, our model is definitely in the first echelon and can be compared with the best models at home and abroad."

But because of the many directions to explore, ByteDance missed the opportunity to launch China's ChatGPT moment.

Also at the February all-staff meeting, ByteDance CEO Liang Rubo said: After the progress in AI, the team was actually "a little relaxed" internally, and did not feel that they had to be the first to reproduce o1 (the inference model released by OpenAI in September last year). At that time, they believed that "a month earlier or later doesn't matter much."

Later, DeepSeek-R1, which was the first to reproduce o1, set off an Internet craze, popularized AI among the entire Chinese population, and brought unexpected surprises to Tencent.

Between ByteDance and Tencent, which both have super apps, the competition is also about the synergistic efficiency between the model's comprehensive experience and product growth.

From the blog post when ByteDance released Doubao 1.5 Pro, we can see that ByteDance believes that large-scale user feedback is still important, and has built an efficient, data-driven optimization mechanism around the chain from user feedback to model experience. This is a strength that ByteDance has accumulated in the mobile Internet battlefield.

The surge in the number of users with Yuanbao now also provides Tencent with similar opportunities to practice. Tencent may also take this product surprise as an opportunity to speed up the pace of self-development of models.

In the past two years, Tencent has only updated its main model twice. In 2025, in March alone, Hunyuan updated its 3D generation model Hunyuan 3D-2.0 and its inference model Hunyuan T1.

At the end of last year, an investor of a large model startup company once lamented that their invested capital had a perfect entrepreneurial story, "if it weren't for ByteDance." And the story in the past two months is that ByteDance has shown the agility of an elephant to dance, if it weren't for Tencent's swift leap with DeepSeek.

This is a shock from different species and different dimensions.

DeepSeek is open source and neutral. It has never received investment from any large company and is not in any "camp". More importantly, it has no intention of becoming an entry-level application. It did not quickly expand its capacity after becoming popular to try to catch the overwhelming traffic. This is why Tencent can connect to DeepSeek without hesitation. Due to its open source nature, DeepSeek did not even earn any income from Tencent's access.

If we say that the impact of the Internet on some traditional businesses back then was “destroy you, but it has nothing to do with you”, DeepSeek, which is now open source and not pursuing application transformation for the time being, has brought a new experience to all companies investing in AI: impact you, but it has nothing to do with me.

Before Tencent's access to DeepSeek brought about unexpected changes, ByteDance found its own approach to investing in the AI ​​big model after exploration in the first half of 2023: establishing an independent new organization and returning to ByteDance's value of "always being entrepreneurial."

Tencent and Alibaba have invested in many large-scale model startups, while ByteDance has not invested in any large-scale model companies. Before June 2023, ByteDance considered investing in Step Star and MiniMax, but two factors ultimately made ByteDance executives determined not to invest: one was that TikTok was facing pressure in the United States at the time; the other was that ByteDance executives believed that in the face of the great opportunity of AI underlying technology transformation, ByteDance should do it itself and can do it well.

After making it clear that it would do what it wanted, ByteDance accelerated the establishment of new AI organizations, Flow and Seed, which were independent of the original organizational structure. The former developed AI products, while the latter developed large models, changing the previous situation where ByteDance's AI research and development was scattered across multiple departments.

By the end of 2023, Flow and Seed will have become an organization on par with ByteDance's major businesses such as Douyin, TikTok, and Volcano Engine. It does not report to any existing department, but is a founder project.

Both Tencent and Alibaba have placed large-model R&D in existing organizations, with the Hunyuan team in TEG (Technology Engineering Group) and the Tongyi R&D team in Alibaba Cloud.

The independent organization and founder participation allow ByteDance AI to break away from ByteDance's original expansion rhythm, salary structure and assessment mechanism.

At the end of 2022, ByteDance, which had expanded rapidly in the previous decade, slowed down its recruitment. For example, the head of TikTok's e-commerce business was told at the time that the teams would no longer expand on a large scale in the new year, and the principle was "one person leaves and one comes in." After 2023, ByteDance expanded AI aggressively: According to our analysis of more than 2,000 AI research talents from ByteDance, Tencent and Alibaba (including those who left) from public channels, by the end of 2024, more than 40% of ByteDance's AI researchers were newcomers who joined in the past two years, and the proportion of new additions was higher than that of Tencent and Alibaba.

Starting from the fourth quarter of 2023, ByteDance founder Zhang Yiming also returned to the early days of entrepreneurship and began to frequently meet with AI researchers. An investor told us that in the winter of 2023, when he visited the authors of AI papers as usual, he was surprised to find that many people had just had a one-on-one chat with Zhang Yiming, including doctoral students who had not yet graduated.

ByteDance’s management team has recruited people including Jiang Lu, the former head of Google’s video generation project Video Poet, Zhou Chang, the former head of Alibaba’s Tongyi, and Huang Wenhao, the former head of pre-training at Zero One Everything.

As far as we know, the Byte AI goal conveyed to these R&D personnel is to explore the boundaries of intelligence and reach the highest AGI, pursuing global leadership.

As the core of ByteDance's model research and development, Seed has also gradually integrated some of ByteDance's existing AI teams.

ByteDance established the AI ​​Lab, an AI research and development organization, in 2016, which previously served ByteDance's applications. Before 2023, the NLP (natural language processing) group under AI Lab will have about 100 people, of which only a small team of 10 people will be researching large language models, while the others will mainly focus on translation and optimization of Douyin Xiaoan (a security assistant built into Douyin that has the function of preventing online exposure and online fraud).

After 2023, the AI ​​Lab NLP group shifted its focus to large language models, and in the second half of the same year, the group was transferred to Seed; in 2024, the AI ​​Lab PixleDance group that developed video generation models was also transferred to Seed. Li Hang, the head of AI Lab who previously reported to Yang Zhenyuan, vice president of Byte Technology, began to report to Zhu Wenjia, the head of Seed, in 2023.

ByteDance has also restarted its acquisition radar, which is one of ByteDance's recruitment methods. Since 2023, ByteDance has acquired at least the smart headset brand Oladance and a storage and computing hardware company. It has also negotiated acquisitions with Monica, which made the Agent product Manus, and an AI vision company.

In terms of salary, rank and assessment methods, ByteDance provides its AI R&D teams with conditions that exceed the existing system.

In the Top Seed recruitment program for recent PhD graduates, ByteDance will give outstanding candidates a 3-1 job level (ByteDance has 5 major levels, 1-5, each level is divided into two smaller levels, the larger the number, the higher the job level), with a salary of no less than one million yuan.

In January this year, ByteDance launched the Seed Edge cutting-edge research program aimed at exploring the boundaries of AI technology. The program set up a more relaxed assessment mechanism: ByteDance originally assessed performance every six months, while Seed Edge will make a final assessment after the project has made breakthrough progress.

Even in product projects that are usually determined by "ROI (return on investment)", ByteDance now gives AI a break. In the past, when ByteDance established a new project, the business analysis team would demonstrate the expected revenue and future market space. ByteDance management would require the human resource investment cost of each product to be calculated. If a new project did not meet expectations after a period of operation, the business layer would issue an adjustment order to shut down or replace the product manager.

When evaluating AI projects, ROI is still important, but the cycle is longer. A person from ByteDance's AI product team told us that ByteDance will now use the single-user value of AI products after a certain period as an evaluation coefficient to calculate future revenue. The evaluation cycle for different products is different, and the longest can be up to 5 years; but ByteDance has no mandatory requirements for this evaluation method.

This has inspired some members to be highly motivated. A TopSeed intern shared on Zhihu that during the nearly two months of developing a project, he almost slept in the company, often waking up at 2 a.m. with new ideas, writing code for two hours, and going back to bed at 5 a.m. When he was most engaged, he and his mentor would continue to discuss until late at night. "No one wanted to leave the company. It felt like they were constantly discovering new phenomena every moment."

In a sense, the ByteDance AI department composed of Flow and Seed is more like an independent new company, and the founder of ByteDance has once again invested most of his energy like an entrepreneur.

It is understood that starting from the second half of 2024, ByteDance’s senior management will come to Beijing once a month to convene ByteDance’s core AI decision-making level and current key project team members to review progress and discuss directions.

Zhang Yiming also reads papers and learns technology on his own. ByteDance has a dedicated research team to assist him in understanding cutting-edge technology. One of his mentors is Feng Jiashi, former professor at the National University of Singapore and researcher at ByteDance.

When a company has 120,000 employees and wants to seize a new opportunity that may be very different in nature, it may be a better choice to establish an independent new organization. The new organization can better fight against the inertia of the past.

After the Spring Festival, the AI ​​market environment changed dramatically. A Byte AI person described their state at the end of the year and the beginning of the next year: just as they were about to applaud, their hands stopped in mid-air.

ByteDance's current response is to continue to attack from all sides. They have set multiple goals, brought together ByteDance's original key executives, and introduced heavyweight talents.

According to the February all-staff meeting, the three major goals of ByteDance AI in 2025 are: exploring the upper limit of intelligence, exploring new forms of UI interaction, and strengthening scale effects. Undertaking these goals is the new ByteDance AI organization consisting of three major divisions, Flow Seed and Stone, with a total size of nearly 2,500 people; ByteDance's middle office department is also supporting AI business.

Seed, which is responsible for model development, is currently jointly managed by Zhu Wenjia and Wu Yonghui, who officially joined ByteDance in February this year. This division is responsible for "exploring the upper limit of intelligence", especially the large model development led by Wu Yonghui.

Wu Yonghui is the latest executive to be directly parachuted into the CEO-1 level after CFO Gao Zhun joined in 2022. Compared with Zhu Wenjia, who joined in 2015 and served as the head of product technology at TikTok and also in charge of ByteDance AI Search, Wu Yonghui's R&D experience is more focused on AI.

After receiving his Ph.D. from the University of California, Riverside in 2008, he joined Google and led projects such as Neural Machine Translation (GNMT) and RankBrainn over the past 17 years, with his papers cited more than 40,000 times. In 2023, Wu Yonghui became a "Google Fellow", the highest honor Google gives to experts who have made outstanding contributions in the field of technology.

Under Wu Yonghui, Seed's model research and development mainly focuses on four directions: large language model, visual model, speech model, multimodal & world model. Li Hang, head of Byte AI Lab, also switched from reporting to Zhu Wenjia to reporting to Wu Yonghui.

According to AI Technology Review, Wu Yonghui recently adjusted the Seed structure, setting up several virtual teams that are not in the formal organizational structure. A flatter reporting system is taking shape, and some people who originally reported to Wu Yonghui-2 now report directly to Wu Yonghui. For example, according to our understanding, Gu Quanquan, the head of Byte AI for Science, who previously reported to Li Hang, now also reports to Wu Yonghui on a dotted line.

An investor who has contact with many ByteDance technical personnel observed that in the core R&D process, ByteDance is replacing the personnel previously transferred from the search and promotion department with "AI-native" technical talents like Wu Yonghui.

The current main technical leader of Seed Edge, which was launched in January this year, is also a newcomer to ByteDance who joined in 2024. This research plan is the part of Seed that focuses more on the frontier exploration of the upper limit of intelligence, targeting those studies that may not see obvious benefits in the short term but may lead the long-term development of AI, such as the next generation of model design, the next generation of learning paradigms, and the next Scaling direction.

After Wu Yonghui joined, Zhu Wenjia shifted his focus to research and development of model application capabilities that are more relevant to experience, such as question and answer, creation, problem solving, code, and the ability to use AI tools.

These capabilities are common modules required by different AI applications. They are equivalent to a "capability middle platform" that connects models and applications. ByteDance has done this in the past: in 2013, the second year of its establishment, it began to set up algorithm, engineering and other middle platforms, encapsulating common technologies and functions for front-line product teams to use. This was the basis for ByteDance to quickly build multiple products at opportunity nodes such as short videos.

Flow, which develops AI products, and Stone, which provides back-end R&D support for products, are mainly related to the other two of the three major goals: exploring new UI interaction methods and enhancing economies of scale.

New interaction methods include new hardware terminals and promoting cooperation with mobile phone manufacturers. For example, in October last year, ByteDance launched the smart headset Ola Friend based on the acquired Oladance team, and ByteDance is also developing AI glasses.

The focus of strengthening the scale effect is Doubao, the main AI product of ByteDance. Doubao was first developed by Zhu Jun, the general manager of Flow, in March 2023, and was codenamed Grace at the time. As the earliest new AI product launched by ByteDance, the progress of Doubao has always been the first item in Zhu Jun's OKR.

In addition to Doubao, ByteDance's AI exploration covers almost all directions of current AI products: there are intelligent character interaction product Maobox, image generation product Xinghui, AI mathematics education overseas product Gauthmath, and AI education product Doubao Aixue (formerly called Hippo Aixue) for primary, middle and high school students.

Products directly related to developers, such as the Agent platform Coze and ByteDance's AI coding product Trae, which was launched in January this year, are now under Stone, a product development team that was separated from Flow in the second half of last year. The person in charge of Stone is Hong Dingkun, the person in charge of ByteDance's engineering technology.

Jimeng, a content sharing platform based on pictures and videos, is another important AI product export of ByteDance. It does not belong to Flow, but is managed by Zhang Nan, who is in charge of Jianying. It is understood that Zhang Nan actually spent more energy on Jimeng than Jianying.

In these new AI applications, ByteDance continued its past habits: employing experienced serial entrepreneurs and product managers who have won battles at ByteDance, bringing together many of ByteDance's "heroes".

For example, Zhu Jun, the general manager of Flow, was one of the founders of the short video product Musical.ly, and later served as the product manager of TikTok. Zhang Nan, the head of Jimeng, was the founder of the picture community product Tuba. In 2016, he launched products such as Douyin and Huoshan from 0 to 1, and later served as the CEO of Douyin. Qi Junyuan, who is responsible for the desktop version of Doubao, once founded the task management tool Teambition. Kayden, who is responsible for the cooperation between AI headphones and glasses and mobile phone manufacturers, is well-known as the founder of 36Kr. In fact, he also has experience in starting a business in hardware such as tablets. There are also people with entrepreneurial experience in Zhu Jun's -2.

ByteDance has placed more strategic goals on Doubao, its main product: strengthening the scale effect of AI products and creating an optimization closed loop from large-scale user feedback to improving the actual usage experience of the model.

Doubao has evolved from an initial chatbot to a comprehensive application that combines chat, search, image generation, and AI browsers. ByteDance has aggressively targeted Doubao's scale: in the third quarter of last year, Doubao's DAU growth target was 30%, and in the fourth quarter, this number became 150%; we understand that this year Doubao's DAU target is to exceed 50 million.

In terms of other product exploration directions, ByteDance has shown that models belong to models and products belong to products. Many ByteDance AI products do not exclude the use of models developed by other companies in pursuit of better experience and growth.

Jimeng, Coze, Trae, and Jianying have all been connected to DeepSeek. Shortly after Coze went online in early 2024, it provided developers with the GPT-4 model in the overseas version. Trae has also recently used the Claude 3.7 model, which has the strongest programming capabilities. The original model requires payment, while Trae provides everyone with Calude 3.7 for free. This is a powerful means of attracting new users, which once caused a large number of users to queue up.

With a strong talent lineup, the Byte Group has also mobilized resources that startups cannot match for the AI ​​sector. It can make Douyin work for me: Since April last year, other AI products were not allowed to place advertisements in Douyin and other Byte systems, but later they were allowed to do so.

It has abundant computing power reserves. By 2023, ByteDance's reserve of GPUs will exceed 100,000; in the first half of 2023 alone, ByteDance ordered more than $1 billion worth of GPUs from Nvidia. Foreign media reported that ByteDance's AI computing power procurement budget in 2025 is at least 90 billion yuan.

We also learned that ByteDance now gives priority to high-end GPUs for large model teams, while in relatively stable but business-critical recommendation systems, it uses more domestic AI chips or GPUs.

There were also some frictions and process delays within ByteDance due to too many teams, resources and directions.

“There are too many capable people here, and everyone has their own ideas. It takes some time to reach a consensus on the ideas, so product evolution will be slow,” said a person familiar with ByteDance’s AI products.

A person on the ByteDance model side said that because they have to connect with different product teams, they often have to repeat the same sentence to several different teams in one day.

The recent adjustments to the top management and reporting lines have also made some employees feel uneasy and unable to concentrate. The easier results have been explored almost to the point where further progress is needed. As a result, when there are a lot of total resources, some R&D personnel feel that there are not enough cards, and some teams need to wait in line, "each relying on their own ability."

Whether it is the rapid changes in external competition or the inevitable organizational tension when large companies start new businesses, it is unlikely to affect ByteDance's attitude towards AI. ByteDance's approach to AI is not much different from that of other new businesses: it is still a saturated investment in everything from talent to infrastructure, exhausting all business directions and not missing any possibilities.

Some people doubt the sustainability of this. ByteDance has also invested heavily in AR, education and games, but has come up empty-handed. But there is no doubt that the changes brought about by AI are incomparable to AR or games. In the long run, the value it creates will surpass that of mobile Internet.

When an organization that has grown into China's largest Internet company over the past 13 years encounters a new game with a high enough ceiling, they cannot miss this opportunity.