Author | Beyond the Layout, Author｜Hua Hua
Where do Zhang Yiming's time and energy go?
Recently, Xu Xin, founder of Today Capital, mentioned this in a podcast. On the spot, she advised Gao Jiyang, CEO of Xinghaitutu, to spend less time on daily management and more time focusing on the technical direction.
The example she used was Zhang Yiming.
She said, TikTok is such a huge business, yet Zhang Yiming spends 50% of his time on Seed.
A research team called Seed. Most regular users haven't heard this name.
My first reaction to these words was confusion.
In the first half of this year, Tencent launched WorkBuddy, and Alibaba continuously adjusted its organizational structure. What about ByteDance? Doubao is adjusting, Coze and TRAE are constantly updating, and Seedance is also accelerating.
But in the AI track, that kind of ByteDance product that makes competitors lose sleep the moment it's played hasn't appeared yet.
I. ByteDance Did Not Redefine the Agenda in the First Half of the Year
Let's look at the timeline of a few things.
During the Spring Festival Gala, Doubao monopolized traffic and high exposure, also leaving many competitors behind, which can be considered a victorious battle. But in reality, since the Spring Festival, there haven't been many major version iterations of Doubao. Version 2.0 of the Doubao large model was released in February, and afterwards, Doubao's strategy basically revolved around paving the way for the Pro version, a process also facing considerable public opinion pressure.
Frankly speaking, judged by ByteDance's past standards, this pace can only be described as restrained.
Coze's trajectory is more typical. As an AI development platform, Coze entered the market early. In 2023, when the Agent concept hadn't yet gained popularity, Coze was already in the game. But at the node when Agents truly exploded in the first half of 2026, with OpenClaw becoming an overnight sensation, Alibaba and Tencent scrambling for the lobster frenzy, and the Coding battlefield bustling, Coze appeared unusually silent.
This is surprising.
TRAE by ByteDance goes without saying. Since its release, its reputation in the programming track hasn't been bad; even against the backdrop of fierce competition like Cursor, Claude Code, and OpenAI Codex, TRAE holds its ground. However, it consistently lacks industry-level topics of discussion, confined within the small circle of programmers.
In June, TRAESolo quietly upgraded to TRAE Work, expanding its positioning from a developer tool to AI office for everyone, placing itself in the position of a follower in AI office work.
Feishu's changes, in contrast, only landed recently.
On July 30th, an internal ByteDance email stated: the Feishu product team as a whole was merged into Doubao, and the sales team was assigned to Volcano Engine. Once a first-level BU, it was no longer independent overnight. (Extended reading: Feishu Became Doubao)
On the surface, it's Feishu being downgraded. More fundamentally, ByteDance is acknowledging the limitations of the independent office software form, consolidating computing power, model capabilities, and end-side collaboration under the same resource scheduling framework.
Looking at these events together, ByteDance in the first half of this year doesn't seem like the ByteDance of the past.
What was ByteDance like before? Today's Headlines debuted, directly taking the lead. Douyin launched, experiencing crushing growth in the short video business. TikTok went global, exploding worldwide. Feishu once kept DingTalk and WeChat Work awake at night.
However, in the first half of this year, ByteDance hardly proactively defined an industry agenda once.
Agents, workstations, enterprise productivity—almost every key term was shouted out by others first, with ByteDance following up.
For a company that was once adept at defining tracks, this change itself is an abnormal signal.
II. The Real Bargaining Chip Is Not Doubao
Where Zhang Yiming's 50% of energy is placed is worth in-depth discussion.
Investing half of one's time and energy into the Seed team indicates this is by no means a project occasionally reviewed by management; it seems more like a strategic layout where core energy is poured.
Over the past year or so, ByteDance has continuously recruited AI talent from top-tier labs like Google DeepMind for the Seed layout.
The hiring of Wu Yonghui, former Vice President of Research at Google DeepMind and Google Fellow, in 2025 was just the beginning. Since then, research talents specializing in foundational models, reinforcement learning, and multimodal fields have gathered at the Seed team. The lineup is comparable to a Whampoa Military Academy for China's large model talent.
Tech media outlets like LatePost disclosed that the number of full-time employees at Seed exceeded 200 in 2024, increased to over 300 in 2025, and this number is still growing.
Applications like Doubao, Coze, and Jimeng within the ByteDance ecosystem are almost entirely built upon Seed's model capabilities.
Most ByteDance AI products seen by the outside world today are only the first layer of application of Seed's capabilities. The truly massive investment of manpower, computing power, and budget still points to basic research that ordinary users can hardly perceive.
In other words, Zhang Yiming and ByteDance haven't stopped investing. They are just investing further down the stack.
III. ByteDance Won in the Past with the Foundation, and Is Still Betting on the Foundation Now
This made me rethink ByteDance's success over the past decade. A pattern emerges repeatedly.
When Today's Headlines came out, portals, clients, Sina, and Tencent News were all ahead. What Today's Headlines truly changed wasn't the information itself, but the way information was consumed. The killer feature behind it was the recommendation algorithm.
Douyin is the same. Short videos weren't invented by ByteDance; Kuaishou had already carved out a market. But ByteDance redefined content distribution and short video gameplay.
TikTok's overseas expansion essentially follows the same logic. Jianying's later success also involved redoing automatic subtitles, smart editing, and speech recognition using underlying AI capabilities.
Looking back at ByteDance's path over the past decade, one finds that what it truly bets on has never been a particular product form. What truly remains on the balance sheet is the engineering approach generalizable across products.
What Zhang Yiming truly believes in is redoing the foundational capabilities behind interactions, ultimately letting users flow towards him. This also explains why he places core energy on Seed.
From recommendation algorithms to large models, today's ByteDance still attempts to use generational gaps in the technical foundation to level the first-mover advantage in upper-layer product forms.
IV. In the AI Era, the Foundation is Starting to Lose to Products
But the question is, does this previously tried-and-tested approach still hold in the AI era?
This is more complex than it seems.
In the Internet era, interaction methods didn't change quickly; the form of news apps hasn't changed qualitatively in a decade, and the short video track is also basically stable. Once a gap opened up in the foundation, the advantage at the upper layer could be maintained for a long time.
The AI era is different.
ChatGPT has long not been limited to chatting. Codex brought an Agent-native interaction paradigm. Claude is moving deeper into enterprises. WorkBuddy cuts into the desktop in the form of intelligent agents, letting AI do the work directly.
ByteDance is not without action; its Seedance is a standout, from the 2.0 version at the beginning of the year to the recently released 2.5 version, showing superior speed and effectiveness in technical iteration within vertical tracks.
But the problem is that technical barriers in vertical dimensions can hardly automatically translate into mass product adoption.
This point is perhaps the most fundamental difference between AI and the Internet.
For the first time in AI, a phenomenon appears: the iteration cycle of foundational capabilities is starting to lag behind the reconstruction speed of upper-layer interactions and workflows.
When competitors directly cut into core productivity scenarios through product forms, the lead at the foundation might not have time to transform into product barriers.
The real challenge ByteDance faces today lies right here.
V. Doubao's Success Became ByteDance's Inertia
There's also another problem brought by ByteDance as a first-mover: Doubao was too successful, and succeeded too early.
Success is an advantage, but success changes a company's attention.
In June 2026, Doubao's monthly active users reached 382 million. Based on the Volcano Engine commercialization system calculation, ByteDance's large model segment achieved an Annualized Recurring Revenue (ARR) of $4 billion, exceeding the total ARR of all other large model companies in China combined. Daily token calls for large models exceeded 180 trillion.
For any company, this data would call for champagne. But for ByteDance, it became a subtle burden.
Doubao proved one thing for ByteDance: an AI assistant can have hundreds of millions of users. So organizational resources continued to tilt here: Pro version, reasoning, Agent mode. Product iteration routes all revolved around making the AI assistant more useful.
The problem is, the industry has entered the next stage. The discussion is no longer about whether AI can chat; competition is shifting from conversation to Agent desktops.
This is not the same logic as Doubao.
Doubao is not an isolated case; all successful products encounter this issue. The inertia of success is sometimes harder to break than failure.
In Feishu's new customers in Q2, over 90% simultaneously purchased AI products. This proves the combination of Doubao and Feishu is commercially viable.
But commercial success and strategic correctness are sometimes two different things.
When a team gets used to making incremental improvements around a product with hundreds of millions of MAU like Doubao, they often overlook those disruptive new entry points.
VI. Tencent and Alibaba Build Systems, ByteDance Still Builds Capabilities
Looking at Tencent and Alibaba's moves together makes it clearer.
Tencent had no model advantage this year; its Hunyuan large model still lags behind Doubao and Qwen in capability. But Tencent did one thing right: seeing WorkBuddy take off, it immediately merged QClaw, then integrated Tencent Docs, Meeting, IMA, and QQ Mail. The core of this move isn't making a single product; it's using 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 personally leading. In April, it set up a group technology committee. In June, DingTalk changed leaders, with 29-year-old Chen Yusen taking over as CEO. Simultaneously, it integrated three Agents—QoderWork, Wukong, and MuleRun—into "Qwen Office," all in one go.
Tencent and Alibaba are doing the same thing: turning products into systems. Their biggest commonality this year is reorganizing existing products, integrating AI into the entire production system.
What about ByteDance? The models are strong, and the products exist. But integration at the system level has only just begun.
After Feishu merged into Doubao and Volcano Engine, it was ByteDance's first time placing office, AI, and cloud under the same framework. This is more than just a step behind Tencent and Alibaba's integration.
The question is, how far is Seed, the foundational model ByteDance is betting on to support products, from truly changing mass productivity?
VII. What is Zhang Yiming Waiting For?
According to ByteDance's previous path to victory, this is somewhat abnormal. What exactly is Zhang Yiming waiting for? Or more directly, what is he betting on?
Is he betting on models? Not entirely. Models are the means. What he is truly betting on is that models will ultimately become the common foundation for all products, just like the recommendation algorithm.
If he bets right, all Agents, whether Doubao, Coze, or any future form, will eventually grow on Seed. By then, the outcome of the entry-point battle becomes unimportant. Because no matter what the usage path changes to, there's only one foundation.
If he bets wrong, the entry-point war is already over, and Agent usage habits solidify. No matter how strong the model is, it can only retreat to the background. The role of a supplier is clearly not one ByteDance wants to play.
This question is sharp enough for any founder. For Zhang Yiming, it's especially so.
In Zhang Yiming's methodology, "delayed gratification" is about extending the timeline to exchange for certain excess returns. But in this AI marathon with extremely high variables, excessive delay might also mean losing the entry ticket to the upper-layer ecosystem.
The real question is, will Zhang Yiming accept temporary product lag and still spend 50% of his time on Seed?
This is not an impulse.
He has been talking about these four words, "delayed gratification," since the first day of entrepreneurship. It's just that this time, he turned these four words from a personal principle back into a company's strategic choice.
Words from [Beyond the Layout]:
Competition among internet companies is like a 100-meter sprint. Today, AI is slowly turning this race into a marathon.
Products can change generations in months; Agents can change interactions in half a year.
But what truly determines a company is perhaps increasingly not the product, but those things users will never see.
Recommendation algorithms, large models are like this. Likely, future new infrastructure will be the same.
Looking at ByteDance's strategy today, it's easy to think it has slowed down.
But the real question might not be slowness. It's whether this generation of AI is a product war or an infrastructure war.
This answer might only be known many years later.
