Dialogue with Fu Sheng: Agent kills the traditional graphical interface

Two years ago , in June 2023 , a debate about the impact of ChatGPT on startups broke out in WeChat Moments. Zhu Xiaohu, managing partner of Jinshajiang Venture Capital, believes that the GPT track is not friendly to entrepreneurs and does not recommend entrepreneurs to choose this direction. Fu Sheng, chairman and CEO of Cheetah Mobile and chairman of Orion Star, bluntly said that Zhu Xiaohu was " ignorant and fearless " and said that half of the startups in Silicon Valley are centered around GPT .

After a year, the debate finally ended with a consensus .

Zhu Xiaohu later explained that the core idea is to stop believing in general big models (because big models) " 3.5 will become a commodity next year , and 4 will also be a commodity three years later ! " For most entrepreneurs, scenarios come first and data is king!

As a veteran with 15 years of Internet entrepreneurship experience, Fu Sheng has personally experienced two technological revolutions: PC (personal computer) Internet and mobile Internet . Under the wave of AI , he is more focused on one thing - empowering robots with model capabilities.

At the beginning of 2024 , the "Six Little Tigers of Large Models" in China were in full swing, and Fu Sheng's team also proposed to develop a 100 billion parameter model on its own. Just three months later, he decisively stopped it. "The competition in large models is too fierce, unless you don't do anything else, just do this . "

In February 2025 , the launch of the DeepSeek-R1 model completely changed the domestic large model competition landscape. The open source and low-cost features have caused the " Big Six Tigers " to adjust their strategies and even abandon the development of basic models. This verifies Fu Sheng's judgment a few years ago that in the future, there will be only a few players in the field of large models, and application companies will become the mainstream.

In his view, basic big models are as common as hydropower, and the real opportunity lies in how to transform them into practical applications. From the hesitation of developing a billion-dollar big model to the decisiveness of embracing open source, Fu Sheng's decision ultimately leads to one core: application is everything.

At the same time, Fu Sheng and Zhu Xiaohu also reached a consensus in the field of humanoid robots: " Bipedal robots are difficult to industrialize in the short term. Laboratory exploration is fine, but implementation still needs to focus on practical scenarios. " Fu Sheng bluntly stated that there is a huge gap between cool demos and stable and replicable products . "If a robot carrying plates has a 99% accuracy rate, it means one is broken every day, which users simply cannot accept."

In May this year , the global AI industry is preparing for the next big wave - Agent. Giants such as Google and Anthropic have made their plans. Microsoft said at the Build conference that just like HTTP changed the Internet, MCP is reshaping the open Agent ecosystem.

In the first half of this year, Fu Sheng has been abroad. In this round of AI competition, he felt the changes in the international reputation of Chinese technology companies and made market predictions. " If you experiment well in China, you should do it overseas. Today, China's ability to make products and applications is very obvious in the world. "

In mid- May , Fu Sheng, who was about to go abroad again , accepted an exclusive interview with "China Entrepreneur" and shared his thoughts on AI technology, entrepreneurial barriers and future trends.

"Chinese Entrepreneurs": Which countries have you visited recently ? What changes in the field of artificial intelligence excite you ?

Fu Sheng: I just came back from the United States. In the first half of this year, I traveled almost all over the Northern Hemisphere, including Europe, the United Kingdom, Japan, the Middle East, and the United States.

The reputation of Chinese technology companies has improved. In the United States, it is very obvious, including in Silicon Valley, that even so-called white funds believe that Chinese teams, especially large-scale model application teams, are very competitive, and Chinese engineers are very competitive . This has basically become a consensus in Silicon Valley.

"China Entrepreneur": Was this consensus established before or after the launch of the DeepSeek-R1 model?

Fu Sheng: It existed before DeepSeek came out. DeepSeek is more like a landmark event, which made more people realize that this Chinese team is so powerful.

Kai-Fu Lee once said that there is a

ChatGPT moment, so soon tens of millions of people in the United States will use AI because it provides a good experience. In the past two years, our model was still a little worse than others. The experience was a bit like water, which produced a lot of steam but did not boil in the end. Users used it for a few times and then felt that it was not so good. But DeepSeek directly made the water boil.

Source: AI generated

"China Entrepreneur": After DeepSeek came out, many basic model companies are adjusting their strategies.

Fu Sheng: In the future, there will be no big model companies. In the future, there will be application companies. Big models are just infrastructure. They will exist , but they will be rare. In the future, the tokens of big models will be like tap water and electricity. They will be provided by the open source community or by some large companies. There will always be someone to fill this strategic ecological niche.

You can see that OpenAI has recently transformed itself into a non-profit company. Its intention to become a public utility is very obvious. It has appointed a CEO , called the Application CEO . In the future, it will definitely: from large models to AGI , it will become a public utility. It will no longer be an exclusive company, because if it is exclusive, it will compete with others every day. But the applications based on the model can become something that makes money.

In the future, companies will focus on applications regardless of whether they build large models or not.

"Chinese Entrepreneurs": Have you ever thought about making a big model?

Fu Sheng: When our team wanted to train hundreds of billions of parameters, I disagreed. Because we started working on artificial intelligence in 2017 , we spent a lot of effort on the speech-to-text engine at that time, and later found that open source was also very good. I realized at the time that the more technology you have, the more valuable you are, but this is not the case in the business chain. In many cases, the more technology you have , the more valuable you are. The core of technology often becomes a kind of public infrastructure. From the perspective of making money, it doesn't make much money.

Two years ago, we thought that the "Hundred Model War" would eventually vanish. First, there is no need for so many models. Second, I think open source and closed source will definitely rise and fall. There are always a group of people in the world with geek thinking. If you (closed source) get first and second place, then the third place must go open source.

Third, the business model. The big model itself has no business model, so no startup can persist for a long time, except for OpenAI and two other startups. What is the essence of this technology? The essence is manpower, computing power and continuous investment.

"Chinese Entrepreneurs": Isn't DeepSeek also a startup company?

Fu Sheng: I talked to one of the " Six Little Dragons of Hangzhou " . The outside world has overlooked that DeepSeek is not a startup. As its parent company, Huanfang is very rich. First, it has enough profits from its original business. Second, it has enough data and computing power. Third, it is rich enough. It is not like those startups with commercial purposes. The goal of commercialization is stuck on the neck from the first day. The DeepSeek team is really based on a kind of research and interest . The so-called useless use, the useless use may really burst out something.

"Chinese Entrepreneurs": Does this mean that now we should stop thinking about how to become a giant, but think more about how to climb onto the shoulders of giants?

Fu Sheng: The difference in the capability of top-level large models is about 3 to 6 months, not too much. I talked to another domestic company that makes large models, and he said, we are all in the seafood business, and a good model will be outdated in about 3 months, because others will always come up, and the gains and losses will be the same.

Each company is choosing its own path. Now it seems that everyone agrees that

The underlying model, whether it is self-developed or open source, is not the core, the user experience is the core. Make the application itself well and let more users use it, this is the core growth indicator today.

"Chinese Entrepreneurs": The basic model change cycle is 3 to 6 months, which is very fast, and it will also pose different new challenges to application developers.

Fu Sheng: Try not to use the tool flow you have written to design the architecture, but call the big model to help you design the architecture. Because the architecture you want may be a little better than the big model now, but when the big model is updated to a new version, this architecture may not be as good as the one the big model came up with.

Once, I listened to the sharing of the founder of Manus . He said that what he did was completely dependent on the big model. He did not do any traditional workflow process. He just let the big model judge and decide what tools to call . In this way, as the big model becomes more capable, it will grow with it.

"Chinese Entrepreneurs": The entrepreneurial circle is now thinking about a question, that is, how far away from the basic model is safe.

Fu Sheng: We don’t need to think too much about this today. Once a user’s behavior habits are formed, unless a better product comes out, the previous one will be kicked out. In fact, there are many such examples. For example, when ChatGPT first came out, someone immediately developed a product called ChatPDF . The website traffic immediately increased, but OpenAI could upload PDF immediately .

But I want to tell you that at least last year, the traffic of ChatPDF did not decrease, but continued to grow. Once the user's behavior is formed, he will not change his path. Just like if there is a supermarket downstairs and you go there often, even if there is another one next door, unless it is 70% off, if the prices are the same, you will still be used to it.

"China Entrepreneur": You have developed the robot voice interaction system AgentOS , which is connected to DeepSeek 's large model. If they launch a multimodal model in the future, can you just use it directly?

Fu Sheng: From a technical point of view, it is definitely not possible. A big model will definitely appear in the future. Just like us humans, we can handle any noise. It will definitely appear, but it is not important. Recently, I told the team that there are no static barriers to entrepreneurship, only dynamic barriers. It is impossible for something to be done once and for all.

Photography: Deng Pan

I finished this technique today, maybe it's over

In 6 months or a year, new technologies will emerge, but in this round of iterations, users stand with me, I am growing, and users will have new demands. I have to hold on to users more firmly based on the new demands.

We have been doing entrepreneurship for so many years, and finally found that good technology is not thought up by the R&D team, but grows together with users. My (product) functions continue to satisfy users, so if you add this function, you will not be more powerful than my system. My system is always moving forward, so my position is stable.

"Chinese Entrepreneurs": Where are the barriers to this product?

Fu Sheng: If you really want to talk about barriers to entry for many so-called applications today, it’s hard to find any. For example, everyone uses TikTok to make videos.

Everyone can come up with an algorithm to make a recommendation based on a 15- second video today, but it doesn’t work because Tik Tok already has so many users and has formed an ecosystem. At this point, even if you have a super powerful technology, you will not be able to compete with the original product form.

Because in the end, we present it in the form of a product, but in the end it is a system, and the continuous growth of this system is the most important thing. Today's entrepreneurs really don't have to worry too much about which model will have a new function one day. The new function still needs to be spread, which takes time, and users need time to change their habits. You still have a long time to make the product experience better.

"Chinese Entrepreneurs": Agents are very popular recently. Do you think this will be a good thing?

Fu Sheng: I feel the word Agent is a bit overheated now . It is essentially a technical principle and a technical framework . The biggest difference between this framework and previous ones is that it is centered on large models, and many of its capabilities are handled by large models, rather than being centered on our processes.

Agents decompose tasks, judge and call tools by themselves. The improvement of model capabilities this time has changed the whole idea of ​​R&D or application. It will be closer to users. In the past, we often said, what is your technical barrier? I will first establish a technical barrier, develop a very strong technology, and then find users. But today it may be the other way around. What is your insight into users? If you have a good insight into users, you can quickly form a set of services for them to use.

I often think that Agent is getting closer and closer to the early days of the Internet, which is about rapid implementation and small steps. The earliest PCs were given to a group of programmers, and no matter how difficult the interface was, it could be used. The same is true for mobile Internet, which allows users to quickly use previously complex functions.

" Chinese Entrepreneurs ": This is a very important change in thinking.

Fu Sheng: The whole idea of ​​making applications today will change. It is a bit like the industrial software of the past. For example, when I was working on antivirus, there were already Kingsoft, Rising, and Kaspersky when I joined. At that time, they released a version a year, and they used a lot of very complex technical codes to complete the construction of the security engine, and then faced security issues.

When I was working on security software, 360 , there were few people and the Internet was booming, so I could quickly get the needs of users and solve their problems with some simple methods. The growth and iteration were very fast. The idea of ​​using big models to develop applications today may be a little different from the past when working on the Internet, and it will be faster.

"Chinese Entrepreneurs": Has this change in development paradigm been reflected in the entrepreneurial circle?

Fu Sheng: We invested in a company that built an AI digital live broadcast room. The entire team has only a dozen people, and it has become profitable this year, with revenue growing very fast. They may not know how to write code, but they understand users and how to meet user needs through AI .

Today, people often ask me, "Do you think Agent is a good direction?" I say, "I can't answer that. Of course I think it is, but it doesn't mean you win just because you become an Agent . The essence of the end is to start from user needs."

"Chinese Entrepreneurs": Mobile Internet has killed tool manuals, but will Agents also destroy the tools themselves?

Fu Sheng : Agent kills the traditional interfaces of tools, such as a button .

The emergence of large models has allowed machines to revolve around humans for the first time. Previously, humans revolved around machines. Today, much of our work is aimed at making machines understand human language and know human intentions through their specifications.

"Chinese Entrepreneurs": Given that humanoid robots are so popular this year, what is your basic judgment?

Fu Sheng: It’s just like not training a large model with hundreds of billions of parameters back then, or even more so. The large model with hundreds of billions of parameters itself is still a paradigm shift, and it will definitely become something really useful, at least it’s worth a try.

Humanoid robots should essentially be laboratory products. They are good for some technical exploration, but they will not form an industry in a long time.

"Chinese Entrepreneurs": But you will make robot hands in the future.

Fu Sheng: Yes, let me add that some companies that make humanoid robots also believe that bipedalism is not necessary. Everyone has a fantasy that one day, bipedalism will be stable and cheap, and it will be like a human being, and it will be able to enter thousands of households, and then the war will be over, right? I think this is a fantasy. It is more in line with our expectations for a certain technology, that is, it is closer to humans, including artificial intelligence.

Source: AI generated

"China Entrepreneur": This year you launched the robot voice interaction system AgentOS. Do you think this is a core change in empowering robot carriers with large models?

Fu Sheng: I think the main thing is to create a smart brain and a good interactive mode, including the ability to move. This is where we can see robots being put into use today, such as in exhibition halls, museums, restaurants, hotels, etc.

At present, the basic conditions for making non-factory robots are mature. It depends on who can come up with a good product, who has better costs, better performance, and more stable quality. Some other robots can be derived, such as some for cleaning and some for mowing. This is the first stage. I think this battle will take about 3 to 5 years.

"Chinese Entrepreneurs": For example, what about adding an arm to the sweeping robot?

Fu Sheng: I don’t think it has arrived yet. The second stage is to add arms to the machine body. Now adding arms is just a gimmick. Why? Because today’s arms and large model control, including its stability, have not reached the standard of real implementation .

When we talk about implementation, we mean that the robot should do the work, not just shoot a video. Users have very high requirements for the success rate. For example, in the matter of picking up a cup, a 90% accuracy rate is useless. If the accuracy rate is 99% , then the robot has to pick up 100 plates a day, and if it drops a plate, you can't use it.

Therefore, the second stage is to add arms, which involves issues such as load and accuracy. This will take at least 3 to 5 years. In the third stage, structures like two feet or something similar may appear. In most scenarios, as long as you complete the second stage, you don’t need two feet, because adding two feet will increase the power cost by at least twice, so the third step is nowhere in sight.

"Chinese Entrepreneurs": So you always emphasize not to be cool, stability is very important.

Fu Sheng: A technology makes a demo , then a finished product, and then a finished product that can be mass-produced. Finally , the finished product can run stably in different scenarios and environments . There are three huge gaps here. In each gap , some industries may take several years, ten years, or even twenty years.

"Chinese Entrepreneurs": Now I see that your products are increasingly moving from being cool to being everyday .

Fu Sheng: To be a company that continues to grow , you can’t be a company that only has a lot of PR articles . The more you work in this industry, the more you realize that no matter it is called a robot, a big model, a mobile Internet, or even a food factory, the underlying rules are the same, which is whether you can provide customers with a good service and whether this service can satisfy customers.

So I later found that this high-tech trap is actually quite scary, because if you make something cool, you will hear applause every day, and you will be confused, am I really getting better and better on the road to progress? The fact may not be like this .

"Chinese Entrepreneurs": Starting from 2025 , there are fewer quarrels in the field of AI , and you and Zhu Xiaohu no longer quarrel. Is it because there is more and more consensus, or is there other reasons?

Fu Sheng: AI 's progress this year has been more under the surface, unlike before, when much of it was at the public level. For example, in addition to my quarrel with Zhu Xiaohu, I also had a quarrel with Zhou Hongyi. When Sora came out, Zhou Hongyi said that it would take ten years to become one year, but I said that was nonsense and impossible.

I think AI has made a lot of progress this year. As for reasoning, the importance of O1 to O3 is underestimated. In the GPQA (Graduate Level Reasoning Test), a human doctoral student can only get 75 points with the help of Google search engine and 30 points with his own research , while O3 can reach more than 80 points . If this is called logical ability surpassing humans, I think it has reached a watershed , but not many people know this.

"Chinese Entrepreneurs": What changes have taken place in your internal organization?

Fu Sheng: We are promoting programming for all employees recently, breaking down the positions of programmers. We call them AI special forces, which means that in theory, you can do any job assigned to you. It is not like before, when front-end developers only wrote front-end, Android developers wrote Android, and Apple developers wrote Apple.

Now the entire interface of our product website was written by the business manager, two newly recruited young people. They were not programmers at all before, but they spent a week to create it using AI to generate code. In the past, it would probably take two or three months to do it.

Source: Respondents

"Chinese Entrepreneurs": To become a company where everyone is AI- focused, what corresponding adjustments should be made to the organizational structure?

Fu Sheng: Not every company needs all employees to be AI professionals , but at least AI should become something similar to an IT department. As a technology company, we are also exploring. At least we have done a few things. First, we have made AI capabilities an assessment item for all employees, and we will start with mid-level managers.

Second, in terms of organizational structure , we recently established a new R&D department in Xi'an . This department recruits all young people and trains them all in AI . They are called AI special forces, and many new projects are assigned to them.

"China Entrepreneur": Last year you launched the policy of "every employee writing agents", and this year you launched the policy of "every employee programming".

Fu Sheng: Why don't we promote programming for all employees in 2023 ? Because at that time, the big model is still in the programming stage . In fact, if you don't have any basic knowledge, it is still difficult to make it run. Now the model is more powerful. In the future, I would like to make adjustments, which is to rewrite the job descriptions and redivide all positions, but I definitely don't have that much energy now.

"Chinese Entrepreneurs": Provide some advice to entrepreneurs.

Fu Sheng: Go to Silicon Valley more often to see their ecosystem. Silicon Valley has regained its prosperous ecosystem. There are all kinds of fantastic ideas for AI , which are what our entrepreneurs lack today. Our entrepreneurs are still relatively homogeneous.

Software export can combine the good ideas of Silicon Valley with our execution capabilities.

Second, you can often visit GitHub , where there are often a lot of new things. At this time, you should also look more at AI , because its scope is very broad, so don’t be too limited.

Third, go overseas more often if you have the chance. Hardware exports are entirely an opportunity for Chinese companies, and we must respect our partners.