Since DeepSeek's meteoric rise, large tech companies that were once methodically advancing AI have accelerated their efforts. Amidst the apparent hustle and bustle, there is more anxiety than joy.

"Why can't big companies create another DeepSeek?" Such questions have lingered over the heads of large tech firms. After all, they possess ample financial and human resources, as well as business scenarios to support their endeavors, yet a small team has stolen the spotlight.

At this juncture, the only viable option is to hop on the bandwagon first and address the details later. However, before they could catch their breath, the AI agent Manus stirred up waves online, with praise and skepticism intertwined.

The current situation is that while AI large model applications are popular among consumers, there is no clear path to commercialization. What is certain is that AI cannot tread the old path; the crux of commercialization lies in creating entirely new interactions and scenarios.

Through the fog, we discern that the AI racetrack requires not only cutting-edge innovation teams but also the abundant resources and business scenarios of large companies. Only through robust collaborations can the most competitive landscape be forged.

Anxiety to "Hurry and Get on Board"

The landscape of the AI large model industry is undergoing subtle shifts.

After DeepSeek's popularity, other large companies have tried every means to hop on this bandwagon and compete for AI large model users. Especially after DeepSeek introduced off-peak discounted pricing, the pressure on big companies increased sharply.

Observing the actions of big companies, they can be divided into two categories: the proactive and the restrained.

Tencent is representative of the proactive camp. As the last big company to enter the AI market, Tencent's Yuanbao initially seemed like a strategically defensive "AI pawn." It has always been self-developing with a laid-back attitude and has not received much attention from the outside world.

However, under the influence of DeepSeek, the globally popular "catfish," Tencent's Yuanbao seized the opportunity to soar and emerged from the crowd of AI large model apps as a dark horse that arrived late but surpassed others.

In last year's AI native app rankings, Yuanbao's monthly active user scale ranked only eleventh, far behind Doubao, Kimi, and ERNIE Bot, which occupied the top three positions.

After announcing the integration of DeepSeek on February 13 this year, Yuanbao began to make significant strides. It not only underwent multiple updates but also launched its self-developed deep thinking model Hunyuan T1 and fast thinking model Hunyuan Turbo S. Simultaneously, it initiated its most proficient strategy of ad blitzes, with ads recommending the download of Yuanbao appearing in various Tencent applications. Douyin, Bilibili, and Zhihu also made substantial investments in ads.

The results were immediate. Yuanbao's downloads surpassed those of Doubao and DeepSeek successively, briefly ranking first in the free app download rankings in Apple's China App Store in early March before being overtaken again.

ByteDance is representative of the restrained camp.

ByteDance has been slow to react in the AI racetrack, with CEO Liang Rubo reflecting several times on the company's insufficient sensitivity to opportunities. Currently, it continues to adopt the strategy of "miracles happen with great effort" for AI large models, entering the field late but investing heavily. Reports say that its investment in large models is "unlimited," with the Doubao team increasing salaries to poach talent from other big companies, currently boasting a team size of several thousand people.

The "money power" helped Doubao shine last year, surpassing Kimi to become the AI large model app with the highest daily active users. However, just one month later, DeepSeek surpassed it to become the industry leader.

From its attitude towards DeepSeek, it can be seen that ByteDance prefers to stabilize its foundation in AI large models. Compared to its competitors, it exudes a more "reluctant" demeanor.

Currently, it only integrates non-core products such as Wukong Browser and Jimeng with DeepSeek, as well as B-end products like Feishu and Volcano Engine. The two super apps, Douyin and Doubao, do not intend to rely heavily on DeepSeek and are quite "reserved," closely aligned with ByteDance's goal this year of "pursuing the upper limit of intelligence."

Even though their philosophies and paths are not entirely the same, under the pressure of DeepSeek's "open source + low-cost API," big companies have unanimously accelerated their advancements. Enhancing their large model capabilities is particularly crucial. For example, Tencent has launched its self-developed Hunyuan T1 deep thinking model on Yuanbao, building autonomous model capabilities around A-search scenarios.

Regardless, entrants must keep moving forward at all times. Facing new industry dynamics, big companies are also adjusting their advancement paths in a timely manner. Previously, OpenAI was a firm supporter of closed-source large models, while Alibaba adhered to the open-source route. With the popularity of DeepSeek's open-source large model, open source has gradually gained the upper hand, and Baidu has also begun to pursue the open-source route for large models.

Big companies' access to DeepSeek is not to rely on it from now on but to reap the benefits and seize the initiative in AI to C. This reveals an anxiety of "although we haven't figured it out yet, we must get on board first."

The Commercial Model is Not Yet Connected

Even though they are popular, DeepSeek and big companies have not yet figured out how to monetize C-end users.

DeepSeek has leveraged its own efforts to stimulate the demand for reasoning. The model can demonstrate the thinking process, which not only enhances users' trust in the AI system but also provides users with opportunities to deeply understand and adjust the model's reasoning details.

This capability gives AI large models a higher efficiency advantage in information retrieval. Therefore, some believe that AI large models can try to achieve commercial breakthroughs through search functions and view this as "the first ray of light for the commercial landing of large models."

Currently, in addition to Baidu sticking to its old business, internet giants such as 360 and iFLYTEK, as well as emerging forces like Dark Side of the Moon and Zhipu Qingyan, are competing for the AI search market.

However, even though traditional search has proven a huge space for commercialization, whether AI large models can follow the same path remains a question.

A while ago, a blogger claimed that when using Tencent's Yuanbao, which is integrated with DeepSeek, there were prominent link ad content in the generated answers, such as "It is recommended to compare prices and make appointments through formal platforms like 58 Home." Subsequently, relevant personnel from Tencent clarified that the links were actually part of the original website.

Behind the controversy and the quelling of doubts lies the dilemma of C-end commercialization for AI large models. Vendors dare not act rashly, considering user acceptance.

Compared with traditional search, the results directly given by AI search are highly unique, equivalent to a strong recommendation to users. This raises concerns that if clients with strong marketing capabilities do not have high-quality products and services, the results of AI search may lack fairness and practicality.

Moreover, the information provided by AI cannot be completely accurate. Even strong models like DeepSeek cannot avoid the "hallucination" problem. After users repeatedly emphasize the details of their questions, AI often solemnly spouts nonsense, even attaching irrelevant reference materials. If ads are added on top of this, the likelihood of users encountering pitfalls increases.

Even if ads are placed in small windows beside the content, whether they can attract users to click is another question.

From this perspective, AI search and ads seem to be contradictory.

At this stage, for C-end users, with competitors offering free services, some products usually adopt a method of limiting the number of uses, and few apps dare to charge boldly.

As for the subscription model, it lacks a survival soil in China. This can be inferred by referring to the survival status of long-video platforms. Yang Zhilin, the founder of Dark Side of the Moon, also believes: "Charging based on the number of users cannot create greater commercial value as the product develops. Subscriptions will not be the ultimate business model."

Therefore, the commercialization of AI technology cannot follow the old path; the core lies in creating entirely new interactions and scenarios.

Is the AGI Landscape Emerging?

Some believe that 2025 will be the first year of AI agent commercialization, but the technology explosion is far from over, and it is uncertain who will win the AGI racetrack.

Recently, Manus has received much attention, and the industry has once again focused on AI Agents, as if a second "DeepSeek moment" has arrived.

△Source: Manus AI official X account

Compared to traditional AI assistants like DeepSeek, Doubao, and Kimi, Manus is no longer limited to information retrieval and language interaction. Instead, it directly delivers complete results to users by autonomously executing complex tasks. This closed-loop capability fills a certain market gap to some extent.

Objectively speaking, Manus is not a groundbreaking technological innovation; it essentially optimizes the process experience and opens up the imagination for intelligent agent products. This will also have some impact on other vendors, such as promoting the development of multimodal interaction (text, charts, code, etc.), accelerating the shift from conversational interactive products to task execution products, and driving the iteration of Agent technology.

After the interaction changes, the scenarios will also change, and huge commercial opportunities will follow. Among them, internet big companies remain a force that cannot be ignored.

On March 11, Ji Yichao, the founder of the company behind Manus, said: "Manus uses different fine-tuned models based on Alibaba's Qianwen large model (Qwen)."

This means that innovation is not out of thin air; the stunning performance of Manus is also built on the shoulders of giants. Facing the breakthrough point of intelligent agents, a super application of large models, many "giant" companies are making layouts.

"We believe that in 2025, we will see the first batch of AI agents join the workforce," predicted Sam Altman, co-founder of OpenAI. On March 12, OpenAI released a series of new tools and APIs to help developers more easily create AI Agents that can automatically complete tasks.

Focusing on China, Alibaba, Baidu, and other big companies are also betting on the future of AGI.

"The primary goal of Alibaba Group's AI strategy is to achieve AGI, which is defined as AI being able to complete more than 80% of human capabilities. So if AGI becomes a reality, the artificial intelligence industry may become the largest industry globally," said Alibaba CEO Wu Yongming recently during an earnings conference call.

"Intelligent agents are the most mainstream form of AI applications and are about to reach a tipping point," said Baidu founder Li Yanhong at the end of last year. Currently, Baidu's ERNIE Bot intelligent agent platform has attracted over 800,000 developers.

AI large models are the cornerstone of intelligent agents, and large model vendors should have the advantage of being in a prime position.

Xiao Hong, the founder of Manus, expressed: "I think the original manufacturers are worth doing it, but they don't seem to be doing it well enough." Whether Manus is doing well enough, whether it is just a temporary buzz or has gained a first-mover advantage, it is currently still entangled in a web of praise and doubt.

After testing and experiencing Manus, industry insiders found deficiencies in its stability, environmental adaptability, and task closure. Another industry blogger revealed that the Agent product of a big company, after internal testing, was stronger than Manus.

It can be seen that AI entrepreneurs should not rush to show off their achievements but should focus more on building technological barriers and optimizing product experience; otherwise, the gap between ideal and reality may backfire on them.

Fortunately, DeepSeek's popularity has pointed the way for the industry, and Manus has confirmed that there is a definite and widespread demand for user-friendly intelligent agent products in the market. With a host of internet giants and startups actively engaging in functional and interactive innovation, once the commercial closed loop is established, AI will usher in explosive development.

It remains to be seen who will be the first to cross the finish line on the path of innovation.

Conclusion

It is difficult to say who will be the next to become popular, but this voyage towards AGI requires not only DeepSeek and Manus but also more innovative forces to work together.

The real winners will be those who can continuously convert technological advantages into unique commercial value and build a complete industrial closed loop.

DeepSeek's popularity has indeed brought pressure to big companies like Alibaba and Baidu, but in the long run, the benefits outweigh the drawbacks. Commercialization still relies on breakthroughs in underlying technologies and ecosystem coordination capabilities.

The journey towards the final outcome of AGI may not be limited to just one DeepSeek or Manus, nor to just Alibaba and Baidu.