Why does Zhang Yiming devote 50% of his time to Seed?

Where do Zhang Yiming's time and energy go?

Xu Xin, founder of Capital Today, mentioned this topic in a recent podcast. She persuaded Gao Jiyang, CEO of Xinghaitu, on site to spend less energy on daily management and focus more on the technical direction.

The example she cited was Zhang Yiming.

She said that Douyin is such a massive business, but Zhang Yiming devotes 50% of his time to Seed.

Seed is a research team whose name most ordinary users have never heard of.

My first reaction to these remarks was confusion.

In the first half of this year, Tencent launched WorkBuddy, while Alibaba adjusted its organizational structure continuously. What about ByteDance? Doubao is under adjustment, Coze and TRAE are being updated continuously, and Seedance is also accelerating its iteration.

However, in the AI track, the kind of product that makes peers lose sleep once ByteDance rolls it out has not yet appeared.

I. ByteDance did not define industry topics in the first half of the year

Let's first look at the timeline of several events.

During the Spring Festival Gala, Doubao grabbed massive traffic and high exposure, pulling far ahead of many competitors, which could be called a great victory. But in fact, there have not been many major version iterations of Doubao since the Spring Festival. In February, Doubao Large Model 2.0 was released, and after that, Doubao's strategy basically centered on paving the way for its professional version, and the whole process also faced a lot of public opinion pressure.

Frankly speaking, measured by ByteDance's past standards, this rhythm can only be described as conservative.

The development track of Coze is more typical. As an AI development platform, Coze was laid out very early. In 2023, when the concept of Agent had not yet become popular, Coze had already entered the market. But at the real outbreak node of Agent in the first half of 2026, when OpenClaw became popular overnight, Alibaba and Tencent were competing for the lobster boom, and the Coding battlefield was full of fierce competition, Coze appeared unusually silent.

This point is quite unexpected.

There is no need to mention ByteDance's TRAE. Since its release, it has earned a good reputation in the programming track. Even against the backdrop of fierce competition among Cursor, Claude Code and OpenAI Codex, TRAE can still hold its ground, but it always lacks industry-level topic influence and is limited to the small circle of programmers.

In June, TRAESolo was quietly upgraded to TRAE Work, with its positioning expanded from a developer tool to an all-staff AI office tool, positioning itself as a follower in the AI office track.

The changes of Feishu have only been implemented recently.

On July 30, ByteDance sent an internal email: the entire Feishu product team was merged into Doubao, and the sales team was transferred to Volcano Engine. The former first-level BU was no longer independent overnight.

On the surface, it is a downgrade of Feishu. The more essential reason is that ByteDance is recognizing the limitations of the independent office software form, and converging computing power, model capabilities and end-side collaboration into the same resource scheduling framework.

Connecting these events, we can find that ByteDance in the first half of this year is not the same as it used to be.

What was ByteDance like in the past? When Toutiao came out, it took the lead directly. After Douyin was launched, its short-video business achieved overwhelming growth. TikTok went global and became a hit worldwide. Feishu once made DingTalk and WeCom lose sleep at night.

However, ByteDance hardly took the initiative to define any industry topic in the first half of this year.

Almost every keyword such as Agent, workbench and enterprise productivity was first put forward by others, and ByteDance followed up later.

For a company that has always been good at defining tracks in the past, this change itself is an abnormal signal.

II. The real bargaining chip is not in Doubao

Where Zhang Yiming devotes 50% of his energy is worth in-depth discussion.

Investing half of his time and energy in the Seed team shows that this is by no means a project occasionally checked by the management, but a strategic layout poured with core energy.

Over the past year or more, ByteDance has continued to recruit AI talents from top laboratories such as Google DeepMind for its Seed layout.

Wu Yonghui, former Google DeepMind Vice President of Research and Google Fellow, joined Seed in 2025, which is just the beginning. Since then, scientific research talents with deep expertise in basic models, reinforcement learning and multimodal fields have gathered in the Seed team one after another. Its lineup is comparable to the Whampoa Military Academy of large model talents in China.

Multiple tech media including LatePost disclosed that the number of full-time employees of Seed exceeded 200 in 2024, and rose to more than 300 in 2025, and this number is still growing.

Doubao, Coze, Jimeng and other ByteDance application scenarios are almost all built on the model capabilities of Seed.

Most of the ByteDance AI products the outside world sees today are only the first layer of application of Seed's capabilities. The largest amount of manpower, computing power and budget are still invested in basic research that ordinary users can hardly perceive.

In other words, Zhang Yiming and ByteDance have not stopped investing, but the direction of investment is getting more and more underlying.

III. ByteDance won in the underlying layer in the past, and is still betting on the underlying layer now

This makes me rethink ByteDance's success over the past decade, and a rule appears repeatedly.

When Toutiao came out, portals, news clients, Sina and Tencent News were all ahead of it. What Toutiao really changed was not the information itself, but the way people read information, and its killer feature behind it was the recommendation algorithm.

The same is true for Douyin. ByteDance did not invent short video, and Kuaishou had already occupied the market. But ByteDance redefined content distribution and the gameplay of short video.

The essence of TikTok's global expansion still follows the same logic. Jianying, which rose later, also rewrote automatic subtitles, intelligent editing and speech recognition with underlying AI capabilities.

Looking back at ByteDance's path over the past decade, you will find that what it has really bet on has never been a certain product form. What really stays on the balance sheet is the engineering methodology that is universal across products.

What Zhang Yiming truly believes in is to redo the underlying capabilities behind interactions, so that users will eventually flow to his products. This also explains why he devotes his core energy to Seed.

From recommendation algorithms to large models, today's ByteDance still tries to use the generational gap of the technology base to offset the first-mover advantage of upper-layer product forms.

IV. In the AI era, the underlying layer can hardly outpace product iteration

But the problem is, does this methodology that has worked repeatedly in the past still hold true in the AI era?

This is far more complicated than it seems.

In the Internet era, the interaction mode did not change fast. The form of information apps has not had qualitative changes for ten years, and the short-video track has also been basically stable. Once the underlying layer opens a gap, the upper-layer advantage can be maintained for a long time.

The AI era is different.

ChatGPT is no longer limited to chatting. Codex has brought the native interaction paradigm of Agent. Claude is moving deeper into the enterprise scenario. WorkBuddy cuts into the desktop in the form of intelligent agents, enabling AI to work directly.

ByteDance has not taken no action. Its Seedance is outstanding, and from version 2.0 at the beginning of the year to the newly released version 2.5, its technical iteration speed and effect in the vertical track are both top-tier.

But the problem is that the technical barriers in the vertical dimension can hardly be automatically transformed into popular product penetration.

This is perhaps the most essential difference between AI and the Internet.

For the first time, AI has shown a phenomenon that the iteration cycle of basic capabilities begins to lag behind the reconstruction speed of upper-layer interactions and workflows.

When competitors directly cut into core productivity scenarios through product forms, the underlying leading advantage may not be able to be transformed into product barriers in time.

This is the real challenge ByteDance is facing today.

V. The success of Doubao has become ByteDance's inertia

There is another problem brought by ByteDance as a first-mover: Doubao is too successful, and it succeeded too early.

Success was originally an advantage, but it will change a company's attention.

In June 2026, Doubao's monthly active users reached 382 million. Calculated based on Volcano Engine's commercial system, ByteDance's large model sector has an annualized recurring revenue (ARR) of 4 billion US dollars, exceeding the sum of ARR of all other large model enterprises in China. The daily average Token call volume of large models exceeded 180 trillion.

For any company, this set of data is worth celebrating with champagne. But for ByteDance, it has become a subtle burden.

Doubao has verified for ByteDance that AI assistants can have hundreds of millions of users. So the organizational resources continue to tilt here, focusing on the professional version, reasoning and Agent mode. The product iteration route is all centered on how to make the AI assistant more usable.

The problem is that the industry has entered the next stage. The discussion is no longer about whether AI can chat, and the competition is shifting from dialogue to Agent desktop.

This is not the same logic as Doubao.

Doubao is not an isolated case. All successful products will encounter this problem. The inertia of success is sometimes harder to break than failure.

Among Feishu's new customers in the second quarter, more than 90% purchased AI products synchronously. This proves that the combination of Doubao and Feishu is commercially viable.

But commercial success and strategic correctness are sometimes two different things.

When the team is accustomed to making progressive improvements around Doubao with hundreds of millions of MAUs, it often ignores those disruptive new entrances.

VI. Tencent and Alibaba are building systems, while ByteDance is still building capabilities

It will be clearer if we look at Tencent and Alibaba's actions together.

Tencent has no model advantage this year, and Hunyuan Large Model still lags behind Doubao and Qwen in capabilities. But Tencent did one right thing: when WorkBuddy proved its value, it immediately merged QClaw into it, and then connected Tencent Docs, Tencent Meeting, IMA and QQ Mail. The core of this series of actions is not to make a single product, but to use AI as a thread to connect all existing productivity products.

Alibaba's path is more radical. In March, it established the ATH Business Group, with Wu Yongming taking charge personally. In April, it set up the Group Technical Committee. In June, DingTalk changed its CEO, with 1992-born Chen Yusen taking over. At the same time, it integrated three Agents including QoderWork, Wukong and MuleRun into "Qwen Office" in one go.

Tencent and Alibaba are doing the same thing: turning products into systems. Their biggest common point this year is to reorganize existing products and integrate AI into the entire production system.

What about ByteDance? It has competitive models and products, but the system-level integration has just started.

After Feishu was merged into Doubao and Volcano Engine, ByteDance put office, AI and cloud into the same framework for the first time. It is far behind Tencent and Alibaba in terms of integration progress.

The question is, how far is Seed, the base model that ByteDance bet on, from truly changing the productivity of the public with the products it supports?

VII. What is Zhang Yiming waiting for?

According to ByteDance's past winning paths, this is somewhat abnormal. What on earth is Zhang Yiming waiting for? Or more directly, what is he betting on?

Is he betting on the model? Not exactly. The model is just a means. What he is really betting on is that the model will eventually become the common underlying base that all products depend on, just like the recommendation algorithm.

If he wins the bet, all Agents, Doubao, Coze, and any product form in the future will eventually be built on Seed. At that time, the result of the entrance war will become unimportant. No matter how the usage path changes, there is only one underlying base.

If he loses the bet, the entrance war will be over, and the usage habits of Agent will be solidified. No matter how strong the model is, it can only retreat to the background. The role of a supplier is obviously not what ByteDance wants to play.

This question is sharp enough for any founder, especially for Zhang Yiming.

In Zhang Yiming's methodology, delayed satisfaction means stretching the time axis in exchange for excess returns with certainty. But in the AI long-distance race with extremely high variables, excessive delay may mean losing the ticket to the upper-layer ecosystem.

The real question is whether Zhang Yiming accepts the temporary lag of products and still devotes 50% of his time to Seed.

This is not a whim.

He has been talking about "delayed satisfaction" since the first day of starting his business. Only this time, he turned this personal principle into a corporate strategic choice again.

Words beyond the layout:

The competition among Internet companies is like a 100-meter sprint. Today, AI is slowly turning this game into a long-distance race.

Products can be iterated to a new generation in a few months, and Agent can change an interaction mode in half a year.

But what really determines a company may be less and less about products, but those things that users can never see.

So are the recommendation algorithms and large models. The same may be true for new infrastructure in the future.

Looking at ByteDance's strategy today, it is easy to feel that it is slowing down.

But the real problem may not be the slowdown. It is to figure out whether this generation of AI is a product war or an infrastructure war.

The answer may not be known until many years later.

This article is from the WeChat official account "Beyond the Layout", author: Huahua, published with authorization from 36Kr.