DeepSeek went to Zhang Yiming’s card table, but the C position must be bean bun

Wen| Student Tang Chen

ByteDance may not accept DeepSeek, but you must face it.

One manifestation of this dissatisfaction is that peers are actively accessing DeepSeek. When they hope to catch this wave of traffic, Byte has shown great restraint and only connects its non-core products to DeepSeek. For example, To C uses Wukong Browser and Instant Dream, etc., while the B side uses Flying Books, Volcano Engine, etc. The two super apps, Douyin and Doubao, are also Byte’s strongest weapons in the era of AI models. They have never let go of DeepSeek and have shown great restraint.

Byte response to DeepSeek: both needs and needs

The outside world can also see from it that Byte’s strategy to deal with DeepSeek’s pressure: smaller businesses ride DeepSeek to compete for some users; core businesses focus on self-developing large models to promote AI ecological collaboration.

From the actual effect, the former has a lot of revenue and has taken over part of DeepSeek’s traffic. Take Feishu as an example. In early February, it announced that relevant product functions fully support DeepSeek-R1. According to Qimai data, its ranking in the free list of Apple’s App Store has soared from outside the 500 before connecting to DeepSeek to Top 100.

The latter’s actions have also accelerated significantly a month after DeepSeek’s popularity. In just two days in the past, Byte has made two major updates: First, Tech Planet reported that Douyin has recently tested the AI ability to access bean bags and opened two super entrances: one is in the Short Video playback interface, which is in the same column as likes, comments, and forwarding functions; the other is in the message list of Douyin App, similar to the current Huawei Xiaoyi homepage.

Tuyuan Tech Planet

On March 6, Byte announced that the reasoning model of the bean bun model was launched, covering question and answer, search, writing and reading scenarios. The input box can be experienced by turning on the deep thinking switch.

One of DeepSeek’s valuable contributions is that while announcing the phased victory of the open source model, it also proves that reasoning is the core capability of future large models. Bean buns kept pace. Previously, kimi, Tencent Yuanbao, and Wenxin Model have all launched their own reasoning models.

Byte used to be good at using the ABtest tool to manufacture apps in bulk, but now it advocates pragmatism and romance, expressing both needs and needs. Byte is obviously not convinced by the emergence of DeepSeek. According to Tiger Smell reports, Byte internally believes that short-term fluctuations cannot represent the most realistic user usage of a product.

In fact, Byte’s response to the big model has been half a beat slower than its peers. Its CEO Liang Rupo criticized this at the annual staff meeting for two consecutive years. In January 2024, he pointed out that the company was slow to organize and less sensitive to opportunities than startups. GPT did not begin to be discussed at the company level until 2023, and the large model startups that did better in the industry were all from 2018 to 2021. Founded;

In January this year, Liang Rupo led the team to reflect on DeepSeek’s popularity, saying that the follow-up speed was not enough and that the smart cap would be pursued this year. Some insiders also revealed that the company had already noticed the reasoning model, but it did not pay enough attention to it at the time.

The key to why Byte was able to become one of the top companies in two years was Zhang Yiming’s belief in working hard to achieve miracles. After setting the goal, he spent money fiercely and invested heavily. According to statistics, in 2024 alone, bean buns spent more than 1 billion yuan on investment. The result was to win the title of AI application with the largest number of domestic users that year. Other bytes are the main AI applications, which basically rank among the top 30 downloads in the seven-wheat annual list, and occupy two places in the top three.

This has also led to the capital expenditure of bytes reaching 80 billion yuan in AI investment in 2024 alone, close to the total of Baidu, Alibaba and Tencent, which is about 100 billion yuan.

“Be restrained like Tencent before DeepSeek came along”

At present, DeepSeek is on Zhang Yiming’s table, but the C position must be bean bun. Also facing DeepSeek, Byte’s attitude is more like Tencent before the emergence of DeepSeek.

You know, for a long time in the past, restraint was almost a unique product design of WeChat. However, in this wave of standard DeepSeek, WeChat and even Tencent are making crazy efforts to rapidly promote their To C applications to connect DeepSeek. Typical representatives are WeChat and AI assistant Yuan Bao.

In May 2024, when Tencent first released Yuanbao, it claimed that it was not ready to push the volume of sales, nor had it plans to increase the overall scale of competing products. But it is facing a situation where strong enemies are surrounded by powerful AI assistants outside, and the DAU of Doubao and Wen Xiaoyan has already exceeded one million.

Tencent did not appear very anxious because of this, and Yuanbao did not take much promotion action after it went online. This has been evaluated by the outside world as Tencent’s deployment of AI was late and not strong enough. If there is an explanation, it is the response of Tang Daosheng, CEO of Tencent Cloud and Smart Industry Business Group (CSIG), in an interview: When we see clear opportunities, we will not hesitate to increase investment, and collaborative operations can be very agile. rdquo;

In other words, Tencent is not reluctant to invest, nor is it not unaware of the significance of this wave of enthusiasm. It is just waiting for an opportunity. After DeepSeek broke the circle, Tencent quickly realized that it was clear that the opportunity had come and began to increase investment without hesitation: spending money, providing resources, buying quantity, and working together to get results.

On February 13, Tencent Yuanbao connected to the full-blooded version of DeepSeek-R1, supporting both Tencent Mixed and DeepSeek models, and switched to a dual-engine drive of self-developed Mixed and open source DeepSeek;

On February 15, Tencent borrowed WeChat, the king of traffic on China’s social media, and connected DeepSeek-R1 for grayscale testing;

On March 1, Tencent Yuanbao computer version was officially launched. At present, Yuanbao has covered multiple ports such as mobile, computer/Mac, and web to meet the AI needs of different scenarios.

What is more noteworthy is that Tencent’s largest traffic portal, WeChat, in addition to searching for DeepSeek, actually provides services from Yuanbao. It also provides download promotion to Yuanbao in the life service Jiugong Grid. It is also marked with red dots, new reminders, and full of blood DeepSeek.“”“” People couldn’t help but exclaim. Tencent went all out, but Zhang Xiaolong couldn’t restrain himself.“”

In other words, in less than 20 days, Tencent Yuanbao quickly surpassed DeepSeek through investment in ecological resources, high-intensity and high-frequency technical iterations, and high-cost investment in buying volume, reaching the top of the download rankings, and the bean buns retreated to third place. According to Caijing World Weekly, as of February 27, Tencent Yuanbao spent a total of 281 million yuan on launch in 27 days, most of which occurred after February 13.

Tencent’s abnormal anxiety has also been interpreted as that when it comes to AI business, Tencent adheres to investment logic, and the core issue of concern is when to wait and see, when to enter, and when to exit.

In Tang Chen’s view, this is essentially Tencent’s speculative inertia. Although the ripening Yuanbao has temporarily made up for the shortcomings of Tencent’s large model in C-side applications, how to maintain it in the future and play a greater role in Tencent’s social ecosystem, especially the WeChat content ecosystem, will be Ma Huateng’s top priority in 2025.

How long can the self-research of bean buns last?

This is also the problem that Byte needs to solve. Liang Rupo mentioned at the annual meeting that compared with pursuing a specific product, making intelligence itself the most important goal can stimulate more attempts and not ignore key technology nodes.

This means that Byte will not bet on a third-party hit, but will focus more on the intelligence level of the entire product system and conduct more radical integration attempts.

The company’s strategic guidance is implemented at the business level, as Wu Di, head of the Volcano Engine Intelligent Algorithm, said. He believes that customer needs will not be fully met by a model company, and the technical route of bean buns is quite different from DeepSeek; Moreover, evaluating a model is like evaluating a person. It is difficult to be perfect. Evaluating a model has dozens of different dimensions. No model can do the best in all aspects. OpenAI, Claude, and Google are all strong in some aspects and weak in some aspects.

Tuyuan March 7 App Store free latest rankings

This can also be regarded as the starting point for Byte to adhere to its self-developed model in its core products. It can also be seen from the way bean buns cut into Douyin that bean buns have excellent performance in terms of copywriting generation and polishing, information extraction, especially vision (visual understanding) capabilities, which can provide Douyin creators with excellent creative tools. It can also meet the immediate needs of users to search for answers. Byte hopes to realize product jump and interoperability within the Douyin ecosystem without causing user experience fragmentation.

But the question that Byte still needs to answer is, how long can the self-research of bean buns last? How much ecological power can the self-research of bean buns exert in the ByteApp factory? Another pressure brought by DeepSeek is that the higher the user acceptance of DeepSeek and the deeper the connection of large manufacturers, the more obvious the outside world will have about the self-developed + platform model and self-developed capabilities.

Moreover, as the models of big factories are iterating, DeepSeek’s progress will not be small. According to DeepSeek’s model update speed, perhaps in the first half of 2025, Liang Wenfeng is expected to lead the team to launch a new generation of models.

There are already negative cases in the industry. Before Musk released Grok-3, he publicly commented on DeepSeek-R1. He stepped on Liang Wenfeng and said that xAI will launch a better AI model and Grok-3 will be the smartest AI on earth. However, the model did not appear smart enough in actual testing, nor did it set off a DeepSeek-like download craze around the world.

This rollover also reminds big manufacturers that Musk’s embarrassment may recur. If model performance cannot surpass DeepSeek too much, it will be difficult to attract users to jump to their own products.

I also noticed that Wu Di also said that AI is like a marathon. If it goes faster or slower now, it will not make much difference in a long historical cycle.

His words are also understandable. Byte’s current restraint may also reverse, and like Baidu, he wants to open up, change his technical route, and run faster. There is no difference, that is, pragmatism takes precedence, and bean buns may also be demoted and kicked out of the C position. Whether such a situation will occur depends on how determined and invested Zhang Yiming is in self-research.

References:

Hu sniffed,”Byte refuses to accept DeepSeek, Douyin has other plans”

Alphabet list,”Ma Huateng left the side of his bed to Liang Wenfeng”

Hedgehog Commune,”Tencent Yuanbao relied on DeepSeek to reach the top, but Doubao couldn’t follow suit”

Student Tang Chen,”Gray testing is connected to DeepSeek, and WeChat AI search is about to come out?”