Image source: Generated by Unbounded AI

Unlike AI companies such as DeepSeek that build underlying capabilities from big models, Manus AI is a startup that has only done AI applications since day 1.

Founder and CEO Xiao Hong (English name: Red), was born in 1992. After graduating from Huazhong University of Science and Technology in 2015, his first entrepreneurship was in the direction of mobile Internet to B; he started his second AI entrepreneurship in June 2022, and has been around for almost 3 years.

In the venture capital industry, many people recognize him as a "founder with product feel", rather than a founder with a technical background like Liang Wenfeng and Yang Zhilin. Up to now, Manus' parent company Butterfly Effect has completed two rounds of financing, with a total scale of more than 10 million US dollars. The first round of investors is Zhen Fund (Zheng is an angel investor who started Xiao Hong's previous business. They invested all the funds earned from Xiao Hong's previous project into his project); the second round of investors include Sequoia China, Tencent, Zhen Fund and Wang Huiwen.

In 2023, under the matchmaking of investors, the company introduced core members such as Ji Yichao (formerly founder and CEO of Peak Labs, now chief scientist of Manus AI), Zhang Tao (formerly product manager of Light Ai, now head of Manus AI product manager) to join.

Since last year, I have had many conversations with Xiao Hong at different times and held a "relay interview". The model capabilities of large models are changing rapidly, and entrepreneurs in it must constantly adjust their postures softly according to the external environment. I hope to record the continuous thinking process of an AI application entrepreneur in the midst of technological change and when everything is in a state of unstable state - and the charm of this process is that it is changing and will continue to change.

The 2025 we are opening may be the first year of the outbreak of AI applications and the first year of the outbreak of Agents. Manus fired the first shot of the domestic Agent. This interview is from the cutting-edge voices of the front-line "AI application explosion" and "Agent explosion".

The interviews posted by the podcast included two times:

● The first time was in October 2024, Monica.im, the first product of Xiao Hong's first entrepreneurship and the second entrepreneurship;

● The second time was in February 2025, when DeepSeek rewrites the underlying ecosystem of China's AI application. This time around their second product, also a new product (not released at the time) Manus.im.

The text version focuses on sorting out the content of the second interview: Behind the birth of this Agent product, Manus, the founder's complete thinking chain is displayed.

In the past two years, Xiao Hong has thoughts and summed ups on big models, products and entrepreneurship, such as:

1) "Andy Beal's Law of the New Era": Models2) Open Source Entrepreneurship Idea: What is your next capability? First do this part of the application well and wait for the model capability to become stronger; 3) OpenAI is very sorry, DeepSeek Thinking and Display is the first time that the entire human being has seen it, because OpenAI o1 is charged; 4) What product should be made in the original factory of big models? What should be done by the application company? 1. Vertical fields and specific fields may not be done by the original manufacturer; 2. Dirty and tiring work may not be done by the original manufacturer; 3. Some original manufacturers may do it in the future, and there is a window period;

5) Founder's thinking model should "think in a game way" rather than "think in a logical reasoning way";

6) When you realize that you are innovating and leading, you should be more radical and super radical;

7) Today's Chinese entrepreneurs should be more radically globalized.

In the interview, you can also see that the real psychological state of an AI application founder is amid technological change, in the AI ​​jungle surrounded by giants, and the local foundation is in an extremely unstable state.

"It should be understood with the craziest fantasy," Xiao Hong said, "It is better not to short it." Xiao Hong said that his partner Ji Yichao met Huang Renxun at this year's Nvidia client thank-you meeting. Ji asked Lao Huang a question: What will surprise you in the next few years?

Huang Renxun’s answer is: Basically nothing.

(If you are also interested in his entrepreneurial story before becoming Manus, welcome to listen to the podcast "Zhang Xiaojun Jùn | Business Interviews", which has many frank expressions.)

"Andy Beal's Law of the New Era"

Zhang Xiaojun: The just-passed Spring Festival is your busiest Spring Festival?

Xiao Hong: Very busy. Of course, everyone has already had a holiday, but they have also done a lot of communication and many colleagues are working.

Zhang Xiaojun: What big moves are you holding on during the Spring Festival?

Xiao Hong: The Agent product I mentioned to you before is rushing to iterate. Unexpectedly, this Spring Festival, DeepSeek has another hot phenomenon-influenced from all kinds of meanings. Discussion, a lot of work was done.

Zhang Xiaojun: Can you explain the Agent product you are about to release?

Xiao Hong: I want to share a recent observation.

We see that at the beginning, the most popular AI application before ChatGPT was Jasper, then ChatGPT, then Monica (which is a browser plug-in) and beanbao. Cursor becomes popular again next - you will find AIUsed for rapid changes, new AI applications will appear every year.

I am trying to summarize the rules and make predictions myself. Although many times rely on a few data points and the description rules are not necessarily accurate, humans are always used to making such predictions.

I myself feel that if we look at these applications as data points, there are some rules.

For example, Jasper may not have been used before. His product form is that you want to write a marketing article, and you want to fill in the blanks - who is the listener and what the topic is - and after filling in, you can output it for you.

ChatGPT is a dialogue-style, from fill-in-the-blanks to dialogue, which is more intuitive.

From ChatGPT to Monica, which comes with its own context and context, bean buns and quarks are all in this category - not only a Chatbot (chatbot), but also a context for Chatbot. The articles on the screen that the user sees, or in a certain application or email address, can help him reply to emails under the user's authorization. It is equivalent to not just a Chatbot, but a Chatbot with context.

Further back to Cursor. After it became popular, two groups of people were using it: one was engineer, the other was close to product managers, and not even necessarily product managers. For example, he is operating a public account and analyzing public account data through Cursor - this is obviously not a requirement for engineers.

He may also use him as a Chatbot—the left is the code content, and the right is the co-pilot (co-pilot) area, and he will not look at the left. When you write the code, you only tell Chatbot if you have any problems; he will not say there is a problem here, so he will manually change the code.

To some extent, it is also used as Chatbot, and the difference between this Chatbot and the previous few Chatbots is that it can not only talk and chat with context, but also helps you solve this problem by writing code.

When I saw Cursor myself - of course, many teams should understand it in the direction of programming, I think this is certainly necessary, and engineers are also a large group - but I understand it myself as the needs of ordinary users.

What is this rule I see?

First of all, it is becoming more and more in line with the habits of ordinary people. From forms to ordinary conversations is more in line with the habits of ordinary people, and bringing context can make it more convenient for everyone.

For example, you don't have to copy and paste the article like you do with ChatGPT, it itself brings your context. Earlier, some people used ChatGPT to write code, copy and paste the code into a Python script, and then run it. If there is a bug, they will report the error to ChatGPT. ChatGPT writes the code and then merge it into the code manually.Go to the file. These are cumbersome, but Cursor does this very well.

A main line, of course, it is becoming more and more in line with people's habits, and its ability is becoming more and more powerful, and the power of this ability is overflowing with the LLM ability.

Cursor is a very early company. This company was founded in 2022 or even earlier. It did not start as a code editor; and even if it was a code editor, it was not popular at the beginning.

It is really known to everyone from July to August 2024, as Claude 3.5 Sonnet was released.

Zhang Xiaojun: It is because of the direct improvement of a model's capabilities that brings iteration of its product capabilities.

Xiao Hong: Yes, I have said so much before, what do I want to express the core?

Model capabilities are evolving rapidly, but "that shell" also needs evolution.

After each generation of model capabilities evolve, it may not even be the original manufacturer, but a third-party manufacturer presents its user-perceived value.

If there is no Cursor, I believe that Claude 3.5 Sonnet may be able to write code in Claude, but it will not be so smooth.

The law I have summarized myself is called: "Andy Beal's Law of the New Era".

(Andy and Bill’s Law, the original words are “Andy gives, Bill takes away”/“What Andy provides, what Bill takes away”.)

There is this law in the PC and semiconductor industries—regardless of what Andy Grove creates, he is the representative of Intel, and Bill Gates will eat it—Intel because Moore’s Law is, the cost may drop and computing power will rise after 18, but after 18, Windows will also eat its capabilities. Maybe Windows is more graphical, or provides more powerful capabilities.

LLM has been evolving. The models we see are getting cheaper and stronger, and even more powerful, they are even reflected in the fact that we think it is possible to write, answer questions, and search information. But next it will use tools, which can write code and call the API. We recently saw ChatGPT releases Operators that can call browsers - these are all model capabilities that are spilling out.

What is that "shell"?

The original factory is definitely defining, but it also needs to be defined with entrepreneurs. Claude is a very good example.

Everyone knows that Monica does not make its own large model base. During the Spring Festival holiday, I was reading some semiconductor-related books - Zhang Zhongmou's autobiography.

AMD founder Jerry Sanders said at the time: Real men must have a wafer factory!

It's just that those who diss only have design capabilities, but don'tA company with chip manufacturing capabilities - he said that such a company is too bad to compete with us.

But you see, in Zhang Zhongmou's words, TSMC has created two industries: one is a professional chip manufacturing company, which is TSMC; the other is a professional design company.

If TSMC had not appeared, there would be no division of labor between "design companies" and "production companies".

Zhang Xiaojun: You are a design company.

Xiao Hong: Yes. Of course, when we look at the industrial laws, generally speaking, vertical integration is initially, and there will be layers later.

From the beginning we held this view: the model should be ordinary commercialized.

You see this now, but it is a bit dangerous for people who do not have this ability to make a conclusion. Because the model is evolving rapidly. Theoretically, the state it presents does not seem to be ordinary commoditization, because there are always people who are more powerful - but from a longer-term perspective, I feel that is.

The choice of our company is that the model is developing rapidly and there are many companies. From a longer-term perspective, it should not be called stagnation, but several companies have entered a pretty good level. At this time, it may be easier to simply apply, and we don’t have to invest a lot of money in training the model.

Of course, we respect companies with models very much. This wave of major progress was brought about by their innovation and efforts. But this is not a conflict or zero-sum, it is called a question of "making applications" or "making models".

But even if there are model companies developing, some companies that mainly focus on the user's perspective and product perspective are needed to do some work. It's not either one or the other.

We have a good relationship with the model factory.

"Open source an entrepreneurial idea: What is your next ability? First do you do a good job in this part of the application"

Zhang Xiaojun: What level do you think today's product definition ability has reached? Who is stronger than the model?

Xiao Hong: According to the narrative just now, all breakthroughs are brought by the model, that is, the model runs further ahead. Basically, the model is driven first and the model is driven first.

After you will find that the original manufacturer is like ChatGPT, and I feel that it doesn’t feel that it will be so popular when it is released; including DeepSeek when it is released, I believe that all this was not what it expected - the original manufacturer is basically inadequately prepared.

Secondly, professional or application manufacturers always have to complete the PMF after the model is first issued. For example, Cursor. It was found earlier, and the OpenAI model might have been used at that time.

Zhang Xiaojun: It is equivalent to waiting there for the model to become stronger.

Xiao Hong: Ah, right!

You can open source an entrepreneurial idea: You predict what your next ability is, and do this part of the application first. Wait until this ability is available, and you will...

You will wait until then to do it, it will be too late, right?

Because there are always some people, whether they believe in this, or they know more about foundation model, or I do the programming field. There are always some people who will make this product.

This is also very interesting. You see, the difficulty of investing in VC has also become greater - because when you see that product, because the model is not ready, it seems not very work, and it feels stupid when it is used; but when the model is ready, it suddenly becomes very powerful. By then it will be commercialized better.

Its curve is a jump.

Zhang Xiaojun: But you see, the model capability is now first, and the original manufacturer may not be prepared enough when the model is released. As an application company, you can wait for it in front, as its model capabilities become stronger, bringing your explosive growth.

At this time, the original factory, the model is in your own hands, and it can follow up quickly. For example, use a follow-up strategy to create a Cursor-like product.

How will the competition between these two companies develop at this time?

Xiao Hong: Today we see that all domestic and foreign model companies have their own applications and open platforms - they basically have Chatbot and an open platform, and third parties can call it through the API.

There is one thing here: What does our understanding of AGI? It should still have public product attributes, and the original manufacturer should not do anything.

First, there are a lot of things, and it won’t do everything. For example, things in specific industries are generally not done.

Secondly, it may not do something particularly hard, or it may require many engineers.

I went to Google and saw people walking their dogs at 4:00pm. I think they shouldn't have to do 100 engineers and everyone's very specific applications.

Because for it, just hold the best part of the barrier in your hands, and it is better for it to leave other things to others.

Zhang Xiaojun: Anyway, you can collect money through the API.

Xiao Hong: Yes.

Zhang Xiaojun: So (as the founder of AI applications,) you cannot do the most fat business.

Xiao Hong: Yes, but of course some may have window periods - for example, they will do it in the long run, but they are not enough to take care of it in the short run - this kind of thing is very challenging for entrepreneurs today. Some people choose to do it, while others choose not to do it.

Would you like to do things that have window periods? Everyone gets different answers.

For example, in the last war, the app store, we later knew that we would be done by the original manufacturer, right? But in that moment, a third-party app store appeared, and got good results from the perspective of entrepreneurship during the window period, and some were acquired, right?

But how to do things with window period this time? What kind of achievements should be achieved during the window period? What kind of preparations should be made for the next step? These are more complex issues.

SummaryLet’s talk about it - API-based business, 1. I think the original manufacturer may not do it in vertical fields or specific fields; 2. The original manufacturer may not do it in dirty and tiring tasks; 3. Some original manufacturers may do it in the future, and there is a window period.

The so-called complexity of the window period is complicated - maybe you do it well, it won't do it - this is not logically derived.

You assume that it will definitely do it, or it will definitely not do it, not necessarily. It is possible that there is a leader who does a good job in the ecosystem, and the original manufacturer may not do it anymore.

Today is very flexible and there is no definite answer.

What is the opposite? That is what the original factory will do - Chatbot, it seems that everyone will do it.

"OpenAI is very sorry, DeepSeek's thinking display is the first time that the entire human race has seen it."

Zhang Xiaojun: Why do everyone have to make Chatbots? What is Chatbot fighting for?

Xiao Hong: I don’t need to compete for anything, and everyone doesn’t use the perspective of something. But it is quite in line with the imagination of AGI - a dialogue interface that can do anything.

Everyone seems to think that just making a model is not enough, at least there must be a Chatbot.

By the way, among all, I feel that DeepSeek is the most Buddha-like thing. But to this day, he has achieved the best results.

Zhang Xiaojun: Why?

Xiao Hong: First of all, I talk about Buddha. It only had its own app in December, and it had a web version earlier. If you look at its app, you will feel that it is really "shell-set" - just to show the model's ability with the simplest basic "shell".

Zhang Xiaojun: Put your own "shell".

Xiao Hong: Another perspective: If (DeepSeek) did not do this, the impact and dissemination this time would not be so great. Because many users use and see its app, see the thinking process, which has a huge improvement in user experience, and leads to a large number of dissemination - but all this is unexpected.

This matter itself must be very complicated, with various reasons, including the geopolitical background of China and the United States, with open source and closed source background.

What I want to say is: My biggest feeling is that this matter will give you great spiritual encouragement! ——From the perspective of bystanders, they may not want to encourage everyone.

Today, you may also think, open source is popular, should you open source be open? Should I build a technical brand? All kinds of thoughts. But most importantly, DeepSeek has been doing it at its own pace - Be Yourself!

He has been open source all the time, and he is also open source even if he is not popular before, but he is just doing things at his own pace.

I remember meWe were chatting on WeChat, and you asked me, what would I do if I were several other foundation model companies? I thought about it later - it is the most important thing to be yourself, not "stress".

Of course, the founders of foundation model have achieved much greater results than ours. I just assume that if I am running a company, the most important thing at this time is to follow my own pace. You will find that simple technology will also receive huge rewards. Although this is not replicable.

Zhang Xiaojun: Can you tell me about your feelings about the DeepSeek product?

Xiao Hong: Perplexity CEO posted a tweet saying that there are two huge experience innovations in the AI ​​era: First, Perplexity marks this sentence from which web page to quote to increase the credibility of the results.

The second is to show the LLM thinking process.

Put aside the various indicators of open source or technology - I talked to some friends in my hometown, and they could feel that the huge difference between DeepSeek is: DeepSeek demonstrates the process of how to think about this problem, which is an experience innovation.

Of course, OpenAI o1 also has thoughts, but OpenAI is very sorry - DeepSeek's thinking process display may be the first time that the entire human being has seen this, because OpenAI o1 has to be charged!

It has only recently shown the thinking process. It is because it is because others may see the thinking process and have this data to do their own training. So it shows a simplified version. To be honest, it doesn't seem to be very meaningful.

First of all, because OpenAI charges and has thresholds, many people don’t know about this at all.

Secondly, it did not fully demonstrate the thinking process, so o1 missed the experience innovation.

The DeepSeek thinking process is released, which is a huge improvement in experience. Also connected to the Internet. o1 was not connected to the Internet before.

And another point, indeed its model quality is very good. The baseline that everyone has experienced before is the average level, but the new DeepSeek model has reached the first echelon to experience it. The articles I wrote are much better, chatting has more emotional intelligence, the model ability itself has good evolution, and the experience is innovative. Innovation can be easily perceived by everyone, rather than being piled up by various thresholds.

Zhang Xiaojun: OpenAI defines the iteration of products - the first is Chatbot (chatbot). Through the ChatGPT product, we know the user form of Chatbot; the second is Reasoner (reasoner), which makes o1, but does not allow users to truly feel the link between technology and product, but instead feels the power of Reasoner from DeepSeek.

One more thing is that we did a podcast some time ago. Go to compare and read DeepSeek-R1 and Kimi K1.5 papers found that Kimi deliberately shortened the answers in its product experience, while DeepSeek's output is long.

In the past, people thought that it would be better to output short from user definition, but DeepSeek was counterintuitive and that it was better to grow. People like to think about it by looking at it.

Xiao Hong: Yes. We also have this feeling when we make Agent products recently.

We want to teach him: use bullet point less (summary of key points), don’t be a little bit. If you summarize too much, just write it out word by word. This is also amazing.

At the beginning, everyone thought it would be better to be more refined. When you find that Chatbot is a refined output, you will say, "Tell me more, I'll take a look."

Zhang Xiaojun: To sum up, DeepSeek is not only strong in technology, but also has a good product definition, right?

Xiao Hong: I think it should be said that - it displays the technological innovations in a user-perceptible way, and this display is not only a product-level display, such as free. You may generally not regard it as a product strategy, but a commercial or business strategy.

OpenAI does not release the o1 thinking process, and may not be so user-oriented - it is thinking that others can train the model if they get this - there will be a little competition (thinking).

So I would say DeepSeek is simple.

I made it and took something for everyone to feel - great. There aren't so many things.

Zhang Xiaojun: What do you think of DeepSeek's longer-term strategic position?

Xiao Hong: I don’t know. Maybe DeepSeek is still considering - mainly focusing on the open source model, or using DeepSeek as an important consumer product like OpenAI and commercialize it? I'm not sure how they'll think about this, but today it's really in a very good position.

I'm not sure if DeepSeek will make a super app the organization's goal? Many times it is chosen.

OpenAI would not have turned ChatGPT into what it is today if it weren't Sam Altman but Ilya.

This is still very dynamic.

But no matter what. Today, DeepSeek's model has improved, and the response of the DeepSeek app to the world - although I don't think it can be copied, it is also exciting.

"So I thought: OK, it's not enough, there should be a virtual machine"

Zhang Xiaojun: We were talking about what the original factory would do just now. The original factory looks like they will make Chatbots.

Xiao Hong: Of course, Chatbot itselfIt is still evolving rapidly, including looking at the human imagination of the future. The so-called AI Assistant is to say something to him, he helps you do things and check things.

This is in line with human imagination, and the original factory will definitely make this thing.

Zhang Xiaojun: Where is the application company?

Xiao Hong: Of course there is controversy. Some investors or founders hold this view: OK, Chatbot is something the original manufacturer will definitely do, so we won’t do it and avoid it.

I am not that pessimistic myself.

Today's technology is still developing rapidly. Do you directly regard it as if it cannot be done? A little early.

Frankly speaking, our recent Agent product should look like a Chatbot, which is very in line with everyone's imagination.

But what it does on the application side is very complicated, and this complexity is not like Monica, it does a lot, such as "function". It would be quite complicated to use these models well - I think it's worth a try.

A little like what I just said about the third category of products in the open platform: even if there is a window period, it is worth a try.

Zhang Xiaojun: The definition of Agent is "a large language model mobilizes the external world". Is it essentially different from the products we have seen before?

Xiao Hong: The concept of Agent will be available in 2023, but there has never been a product form that is truly felt by everyone.

Maybe Zhu Xiaohu will say that I don’t believe in Agent, just like he said I don’t believe in AGI. Maybe this time next year he will say I believe the Agent (laughs) - hope so.

Agent is to be able to perceive the environment and perform tasks independently. It's very abstract to give it a sentence, it can handle this task.

The core reason why the Agent field has fallen below expectations in the past few years is that the model is not smart enough to automatically solve many things.

Second, before, including the process of doing Monica, we said: OK, it can perform tasks, okay, we will connect it with some APIs, and we will connect it one by one.

The language model alone is not enough. If it wants to search online, I will send it a search API; it needs to be able to read the knowledge base. OK, let's make a knowledge base so that users can upload files, search the knowledge base, and give you the answer; it needs to do PPT and draw pictures...

In the process of developing Monica, we made a lot of APIs for it and connected all kinds of APIs upwards. But have you discovered that this is very similar to making a feature machine - the feature machine is, hey, users need this? good! At that time, mobile phone manufacturers specially made a feature phone.

The feeling of a feature phone is that the phone that adds up one feature after another.

Zhang Xiaojun: Stacked.

Xiao Hong: So in the process of making Monica, frankly speaking, although Monica has integrated a lot of things and it does take a long time, it feels like it is making a feature phone, feature phone - you are connecting it to APIs one by one.

But the real Agent should be able to write code, call the API and execute it. It can handle many long-tail tasks and does not require developers to write them themselves. This is in line with everyone's imagination of Agent.

I remember a senior, Bai Ya (founder of Youzan), who told him that Monica has done the best and integrated a lot of things into it.

He said: Red, the ultimate is not enough, personalization is enough - to achieve the ultimate, you are hao123, and to achieve personalization, you are Google. This is a very inspiring sentence, and we spent a lot of time studying it.

When I saw Cursor, all kinds of code-writing companies emerged, becoming more and more popular, rather than being used by engineers. Today you interview many engineers, and they can find many problems. The more senior engineers will find the bigger problems, but the more they use them, the more they feel, the more they feel.

And I remember that after Cursor came out, there was another competitor called Windsurf. Windsurf has a different place from Cursor. But Cursor later followed up with the YOLO mode, the name is magical, Y-O-L-O. Much like the younger generation can speak.

But it means You Only Look Once—I can handle it as long as you watch the process once.

What is the difference between it and the original one? You originally wrote a piece of code and ran it with Python, but your computer may not have a library installed, and it will report an error if you click. With the YOLO mode, it will automatically put it in the LLM to help you solve it.

One day I was using Windsurf's similar YOLO mode. I found that I told it a question, and it said, OK, I will go to Github to download this code and do something, and then write it again - I felt like I was hit by lightning at that moment! ——It is actually using tools, and it can use human tools. (For example, on Github, in theory, there are all kinds of codes, right? He can use them when creating all tools.)

At that time, I thought the Agent era was really here.

Zhang Xiaojun: Will the original manufacturer do this? Or will the application company do it?

Xiao Hong: The original factory is worth doing, but it seems that the original factory is not doing well enough. (Laughs)

So I think that the so-called Agent should be able to solve long-tail needs and call various tools. The best tool to be called is of course the existing code and it writes its own code, and it can handle things through the API.

Of course, these are not enough. Because humans still have a lot of knowledge, orMany services are not called through APIs, but through Web calls. Overseas, I think I have to bring a browser.

There is another essential difference: I was using Windsurf at the time, and it ran on my computer, but sometimes it asked me to confirm whether to install this library; or it performed a command line operation and asked me to fill in yes or no, because it might really ruin my computer, or something conflicted - but in fact, it asked me to fill in yes, which was quite a buck.

If I were a very novice, how would I know yes or no? But it asked me to fill in yes, as if the responsibility was on me, and it was not its business if it was broken.

Many years ago, someone asked Bill Gates: Why did Windows talk to me about this thing, saying that opening this thing may harm the computer, yes or no? Depend on! You and Microsoft don’t know, how could I, an ordinary user, know?

So when I saw it, I felt that there seemed to be no You Only Look Once, and you still have to do it. Besides, if you are really a novice and an ordinary user, you really can't understand it.

So I thought: OK, it's not enough, there should be a virtual machine - Chatbot should have a computer on the cloud, and execute all the code it writes and the things it wants to check through the browser on that computer.

Because it is a virtual server, it doesn’t matter if it breaks, it can have another one. It can even release that server after the current task is executed.

So I think that the architecture is called a virtual server or a browser. If you can write code and call the API, it can handle various long-tail tasks - this is what we are doing.

Zhang Xiaojun: Is the model capable of ready now?

Xiao Hong: Today is just right. We only realized late last year that it was almost the same.

Just just now, everything was connected - for example, to observe Cursor, to see it using Github code, and back to the development process I mentioned at the beginning, from Jasper to ChatGPT to Monica to Cursor, what is the next one?

You will find that it consumes more and more tokens, and it is even a little advanced, right? For example, you have exceeded the ability of ordinary people to write code, but AI is OK.

And there is another point, it needs to run multiple iterations on its own. If you set a task for it, it will encounter setbacks and may report an error. It will try and keep trying to solve it.

Today, the model is good enough, so we have developed the most basic version and run it out. Of course not smart enough. We saw ChatGPT released Operators and other things, and many tasks are not completed at high enough, and the model still needs to be smarter.

So I still look forward to the manufacturer's model being smarter! The cost is lower!

Application vendors are not availableThe consumption of love consumes a huge amount of tokens, swallowing up a large amount of tokens.

"It should be asynchronous"

Zhang Xiaojun: Is your product similar to Cursor?

Xiao Hong: It should be made unlike it.

Zhang Xiaojun: Is it a programming product?

Xiao Hong: No, it’s not programming, consumer-grade products should make everyone feel like Chatbot.

Consumer grade is very important.

Zhang Xiaojun: Will it evolve along Monica?

Xiao Hong: New products will be released separately. Because you can't imagine that hao123 and Baidu are the same product. (Laughs)

Zhang Xiaojun: It's still a Chatbot-like product, but have you done a lot of iterations in the product?

Xiao Hong: Yes.

Zhang Xiaojun: Is it similar to DeepSeek?

Xiao Hong: Reasoning is necessary to release the thinking process. Of course we may call it Planner.

For example, if you ask a question, it will break it into many steps, and each step will solve what is going on, and finally solve the problem.

There are two very important things here: one is about the model and the other is about the experience.

I will say first about experience, it should be asynchronous.

All Chatbot experiences today are all about you sending a sentence, and it responds to you. For example, in the process of its reply, what you say to it may interrupt the above sentence.

So today's Chatbots are all synchronized. Some friends will use a description called "waterfall-like", which is A-B-A-B. But human chats are not like this.

You send me a message, maybe I will not reply after a long time. During the long time, you send me two or three more messages, and I will reply to your topic together. Or a colleague said he wanted to do something, but after halfway through, I found out that he had made a mistake, and I told him that he had made a mistake and I would start over.

The human communication process is not like Chatbot now, it is A-B-A-B. It has a lot of forks, and even a certain task needs to be executed for a while.

So, if all Chatbots are issued today, they must be done within one step, which has many limitations. Many times, for example, if you ask it to do something, it will eventually get the task, but you have to do something first, get the result, and use that result as part of the input before doing the second thing, which takes time.

We cannot imagine that you handed it to any intern who is smart, and he will answer everything you ask him in seconds - this is unrealistic. He needs to take some time to inquire and think.

Zhang Xiaojun: Your product is probably, I will give it a request, and it will reply to me after a while?

Xiao Hong: It will tell you what it wants to do, and the next thing it will do. If there is progress, it will tell you - this is our fantasy of the best interns.

If you just said to get this done, he said: OK, I will do this, and I will give you synchronously after each step has its results. And you can say to him halfway through: Oh, my demand was wrong at that time. Can you do this again? He said: OK, I will change it, and I will give it to you if I finally complete the output.

This is what we are doing, and it will be more like humans.

Zhang Xiaojun: Nowadays products can be replied in seconds, so this product must meet my more complex needs before I will give it this time. How can it meet more complex requirements?

Xiao Hong: That's what I want to say. Today's model capabilities are able to complete some complex and multi-step tasks. It’s just that there is no such product, so everyone can’t feel it.

I would like to give an example. For example, ChatGPT released Deep Research, and there is a test set called GAIA. We are doing this question internally.

One of the problems is that there are several animals on a video on YouTube at a certain time. We were surprised to find that our agent opened YouTube and was watching it. He would use YouTube shortcut keys to accurately locate it to that second, and then tell us that it is these animals.

You see, this process is different from Chatbot in the traditional sense.

First, it can watch YouTube pictures, not subtitles.

Secondly, we even found it using YouTube shortcuts, and was shocked that it answered this question.

Zhang Xiaojun: From the perspective of a novice user, what kind of task instructions can I give it to it that can satisfy me? And how long does it take?

Xiao Hong: For example, if you want to analyze what rules does Elon Musk’s Twitter have, you can let it go.

It may call the Twitter API, grab everything, analyze it in semantics, and give you a pretty good output result. These tasks are a bit advanced.

We use Chatbot, and we generally don’t use these tasks, but in fact, many tasks can be solved today, but information retrieval and query are definitely the highest frequency tasks.

Today, whether it is used as a search or as a chat tool, that part of the demand will not disappear. It won't be gone just because the Agent appears. This part of the demand still exists.

But, I want to say, because the Agent product form appears, humans have further expanded their use of Chatbot boundaries, and this is in line with everyone's imagination.

Zhang Xiaojun: Assuming that Chatbot continues to evolve, it also has such a function, we will need a second entrance and a third entrance toIs it satisfying us? Do we need a special asynchronous one?

Xiao Hong: The reason we use a new product is because we feel it is necessary.

Monica is also a product of many users. It is inevitable that there will be the original user habits in it, and you cannot start all of these again.

A newer product, a product without burden, is better adapted to these assumptions.

Zhang Xiaojun: How high do you expect this product to be?

Xiao Hong: I don’t know. Not sent out yet, haha.

Zhang Xiaojun: How did you think of this product first?

Xiao Hong: The whole process is to summarize from Jasper to ChatGPT to Monica to Cursor to Devin. Devin is very consistent with the architecture I just mentioned - Devin exactly is this architecture.

It directly selects the hardest-core engineer, but I prefer to choose general rather than to go to a specific industry. For example, for engineers.

This architecture is very consistent with what I imagined with Agent, and it should be for ordinary users. It shouldn't be like Devin targeting the engineer industry and priced at $500 - is this a bit like OpenAI's 200, hahaha?

Zhang Xiaojun: When this company starts to charge fees, you can always pay it at a low price.

Xiao Hong: Price should be part of the positioning.

What is the positioning? We think it is a consumer-grade product or a mass product, and you should price it in a mass way. At least the price for entry, right?

But with more usage, because the cost is there, you can get users to pay more. But your basic pricing determines whether you are a consumer-grade product or an enterprise-grade product.

It is also very important for model manufacturers. Today we test that only Claude Sonnet 3.5 in the world can run the architecture we just mentioned - we call this Agentic's ability internally.

The traditional Chatbot alignment method, during the training process, is to assume that a round of dialogue should solve your problem as much as possible. So everyone teaches the model according to this. But after testing, only Anthropic's Claude Sonnet 3.5 has long-term planning and gradual problem solving capabilities.

Long-range planning is a problem. After you give it, it says OK. I will make this plan well. In the 1234 steps, I will first solve this problem in each step. After this matter is done, I will get input and then solve the next thing. These two abilities are combined. And traditional Chatbots try to get everything done in one round.

We believe that it is caused by different training methods. Therefore, the model manufacturer should train it specifically for Agentic or AgentModeled by itself.

Zhang Xiaojun: Before the New Year, everyone said Agent, Agent, Agent. After a New Year, I didn’t expect that the one that became popular was Reasoner. (Laughs)

Xiao Hong: Yes, but it will be too late to chase after you when you are popular.

You still have to be yourself.

You have to have your own rhythm - if you can't be stressed, it's too late to be stressed.

Zhang Xiaojun: I imagine that you can see your product in Q1 this year. It is an app, right?

Xiao Hong: There is a Web and an App. We may do some small-scale tests first.

A very core issue, or a very important responsibility of a product manager, is to control user expectations.

Suppose it can do everything in the world, such as: How do I make $1 million? This is not something that should be performed by an agent.

But if we can give more specific examples to make everyone's expectations more reasonable, everyone will use them smoother.

"Don't worry about me! Keep moving forward quickly!"

Zhang Xiaojun: I feel that you are in good shape today (compared to before), why?

Xiao Hong: Because the examples I just gave you, such as: Agent's cool example, really happened in these few days.

It was unexpected. Of course we were doing it before the New Year, but after all, in the process, we tested it these days - we found: Damn, it's so awesome.

I just read DeepSeek's paper that said, there was an A-ha moment, they said "A-ha, wait" - we are experiencing this feeling, these days. (Laughs)

Zhang Xiaojun: Suddenly it will happen, and you don’t know why it will happen, and do something beyond your expectations.

Xiao Hong: Even your craziest imagination, you don’t know.

Damn it is actually watching YouTube videos, it tells me: Hold down the shortcut key to watch.

You really feel that you are creating something like life.

Zhang Xiaojun: Do you expect to see many such Agent products this year?

Xiao Hong: Everyone can do it.

Zhang Xiaojun: In the AI ​​entrepreneurship ecosystem, what kind of mentality should entrepreneurs maintain?

You see, many AI companies nowadays do not only make one product, but they continuously release products. At the same time, you don’t know what the next change in model capabilities will be, or whether the products you make are transitional products, and there is also a fierce competitive environment with giants. What kind of mentality should you start a business?

Xiao Hong: Technology progress is the biggest dividend. There are always new technologies happening, so today is still a good era for entrepreneurship, so everyone can be more optimistic.

I repeatedly said that I should have my own rhythm. Maybe you have less gains and losses, and according to your own love - I rarely use the word "love". In an internal sharing, I would like to share with my classmates what love is?

One day I got off work late and drove my colleague's car home. I think, what if I accidentally get into a car accident when I drive tiredly? Or what to do if you bump into someone?

At that moment, my own answer: If I could only make one phone call, I would tell my partner: Don’t worry about me! Keep moving forward quickly! Don't be distracted by me. I handle my own affairs.

At that moment, I was no longer important, but that thing was even more important.

When you think like this, you seem to seriously think that this matter is more important than you.

How did all this be established? It’s still the thing you need to pay attention to - make yourself very excited and proud. To some extent, it is not necessarily directly related to business - I think DeepSeek may have been doing this all the time.

We can't say it, the business was successful today, but it is indeed in a relatively good position.

So, first, optimistic. Because I believe that technology is progressing rapidly, there are still many things in progress.

Second, find the thing you like or believe in. Once you find it, you don’t need to say anything else.

Zhang Xiaojun: Did you find it today?

Xiao Hong: I think I have found it, and I can tell my classmates: Don’t worry about me.

Zhang Xiaojun: Did you find a better position in the AI ​​ecosystem today?

Xiao Hong: This is something I believe. I'm not sure if there is a map from God's perspective, is it in a good or a bad location?

I don't have this perspective.

But I have a self-centered perspective: I am very happy to do this.

"When you realize you are leading, more radical, super radical"

Zhang Xiaojun: I feel that in recent years, something has happened every Spring Festival and is brewing.

Xiao Hong: Yes! As I told you before the Spring Festival, it is very similar to the Spring Festival in 23.

At that time, we had already established a project to make Monica before the release of ChatGPT, and we immediately saw ChatGPT release, but that Spring Festival was frankly speaking, there was no popularity in China, and there was no super hot abroad. Until Sam Altman tweeted that 1 million users, China and the United States became popular in all respects.

I was very nervous at that time and had seen many independent developers innovate in GPT-based APIs. So, I didn’t have a break during the 23rd Spring Festival. On New Year's Eve, there is also an online call meeting.

25 years are very similar, but 25 years are different. One thing is called "There will always be accidents." Originally it was Agent, everyone worked hard during the Spring Festival, but they didn’t expect DeepSeek to appear. There is Sora in 24 years - AI will not miss any Spring Festival, hahaha.

25 years are more like 23 years. In 24 years, the entire industry has been linearly extrapolated for 23 years, which is conceivable, such as multimodality, and after writing and making pictures. But the video was a little surprised. No one expected it to be so fast, but immediately digested the information.

In 23, Monica's innovation mainly adds context and uses a good product form, but in 24 years, you find that major manufacturers have followed suit, including bean buns, and it is possible that everyone also makes plug-ins and browsers.

The founders still have to innovate quickly - big companies can understand that when you innovate, it is very dangerous; when you are compared with big companies, it is also very dangerous.

Zhang Xiaojun: For example, compared with bean bread?

Xiao Hong: Haha, we don’t have the mentality of comparing with bean bread, it’s a different market. But obviously Doubao has understood this matter and will do it.

You still have to give full play to your organizational flexibility, look at the technological trends, and run faster.

First, be ahead;

Second, when you realize you are ahead, you are more radical and super radical.

Today we reviewed it and felt that Monica was not radical enough in 23 years.

Zhang Xiaojun: What do you say? What should you do?

Xiao Hong: Users should be done more. My second time starting a business, I will have a slightly better operation at the beginning, such as having more experience in recruiting, but I may be recruiting on a large scale and grouping the team more radically and quickly, rather than understanding this matter with good management.

It depends on where you are – if you know you are innovating and you are leading, you should be radical.

Zhang Xiaojun: I think this is your change.

Xiao Hong: What may this change come from? When we chatted in 24 years, you may think I should be more well-managed and more conservative.

Zhang Xiaojun: And for a while in 24 years, you should be in a bad state.

Xiao Hong: This is normal. If you take this part out and share it with you: When you encounter setbacks or difficulties, you will definitely feel that it is very difficult - you will linearly extrapolate the setbacks, and you will always be very difficult - you will be worse and you will feel more frustrated inside.

Recalling, because technological progress should have many opportunities for innovation, we should be more radical when there are opportunities for innovation - this is a summary and review of 23 and 24 years and a prospect for the next.

Zhang Xiaojun: Do application entrepreneurs need to spend a lot of money? Do we need to continue financing, or should we make revenue quickly?

Xiao Hong: The people have a scale in their hearts. Make the product well.

It may sound a bit grand, which means "human welfare", and you can get enough rewards if you do it. Instead, it doesn't need itThere are so many tricks, so there is no need to burn so much money.

You don’t even have to buy KOLs, and you don’t have to spend money on DeepSeek, and you don’t give any KOLs 1 cent. There is no TikTok that prohibits it from advertising on it.

This is exciting. You have fantasized about countless miracles - as an App entrepreneur or experienced the mobile Internet, you will think: Wow, it’s so great to make a product that is the first in the overall list! If you are the first in the list of more than 100 countries in the world, wow, it is simply the craziest fantasy!

It happened.

It is in the era you are in, as if the company in the industry is not far from you, has done it.

This is of course very exciting.

Zhang Xiaojun: What are your organizational goals? Is it a super app?

Xiao Hong: This description of Super App is very accomplished and oriented.

I shared yesterday with a group of high school and college entrepreneurs, a bunch of young hackers: "Think with the age of the times, not with physiological age." How many years is this year?

You will find that some founders may drop out of school and start a business. A while ago, VCs said that we were looking for these people in the 1990s. Because they may have missed the mobile Internet, and now they are in their battle year - I think this description is quite correct.

Let me look back, I graduated in 15 years. During my college years 2011-2015, I was getting better every day, so I chose to start a business. But I didn't realize: Damn, the mobile Internet was over when I graduated. But the senior brothers who may be a few years older than me started their business. The achievements, achievements, and things in the secular sense are all better.

It is because I started a business in the first year/second year/three years of mobile Internet. Share with young hackers: Your physiological age is also very good this year. If we say that the popularity of ChatGPT in the end of 22 or 23 years is beginning, this year may be the beginning of two years or three years, which is very early.

With a group of friends, follow the dividends of technology, and do some things, this is our goal. I hope everyone enjoys this process. As for business returns, there are too many elements that you can't control.

What else did I do during the Spring Festival? Study and read papers. You learn it, and this positive feedback is very strong. No one can stop you from doing this. But business competition is not like this.

Zhang Xiaojun: Look at the products with the largest number of users now, whether they are ChatGPT or DeepSeek, they all have their own models. What do you think of models?

Xiao Hong: First of all, it is very appropriate.

This time, the technological progress and the driving force for technological dividends all come from the model. Model companies do this and it gets rewards that match the job, that's that simple.

Zhang Xiaojun: Why did you decide not to make a model since day 1? Have been entangledNo?

Xiao Hong: To be honest, I don’t have it. It should be different at different times.

In day 1, I didn’t have this resource. At that time, I felt that I was very pragmatic and started to work with the best model. At that time, I might be "doing business" or "making products". I'm not sure if I will do it in the future - this is very honest.

In other words, it may have a reduction in costs, or it may have user needs, or you have this ability.

If you add a more comprehensive perspective, it is: Are there anyone else in this market doing a good job? Can you find manufacturers that have strong or weak relationships with you and help you get this done?

I generally believe that the industry will be stratified after maturity. This is a very dynamic thing.

Zhang Xiaojun: Are you one step ahead of others in the matter of Agent?

Xiao Hong: I don’t know other people’s progress, but there are not many teams who have the opportunity to try to make the Agent product I just mentioned. Because, it requires a lot of compounding abilities.

He wants to do Chatbot, some AI programming related, and browser related, because he needs to call the browser and has a good perception of the boundaries of LLM - what level will it develop today and what level will it develop next.

First of all, there are not so many companies at the same time, and companies with these capabilities may be doing a very specific business at hand. Some of our classmates just happened to have time to do these things together.

First, we are relatively lucky.

Second, if everyone is stressed and doing Reasoning today, will they leave some time for startups?

Zhang Xiaojun: How far can the model expect the capability spillover to go?

Xiao Hong: If you are asking about the United States’ Stargate stack, how can it improve its capabilities? I have no idea.

But I don't think I can short it.

You can't assume it will be slower in the next few years, and you should understand this with the craziest fantasy.

I sent an instant during the Chinese New Year, quoting a great man's sentence: "Civilize the spirit, barbarize the body, and leave the rest to AI." What are the others?

The future may be "spiritual civilization" and "physical health". As for intelligence, we will leave it to AI. I believe I should be able to see it in my lifetime.

Peak (Chief Scientist Ji Yichao) met Huang Renxun at this year's Nvidia customer thank-you meeting. He asked him a question and said: What will happen in the next few years surprise you?

Huang Renxun’s answer is: Basically nothing. He thinks nothing is possible.

I think I should watch this wave with this mentality.

Looking back, in the past few years, I just took the Spring Festival as a matter of the past few years.All shocked?

So, it's better not to short it.

Zhang Xiaojun: This is just a step toward Reasoner.

Xiao Hong: Reasoner, you are still thinking more, but the agent also needs to execute and feedback, be able to accept feedback from the environment, use tools, and use feedback from the environment as input - more like humans.

Of course we are working as software agents, working in white-collar workers. To a certain extent, robots and intelligent driving are "a large-scale used agents".

Zhang Xiaojun: You just said that there are some different knowledge-hows in training Agent and training large language models. Is there any knowledge-how that you can share?

Xiao Hong: We don’t have knowledge-how, this is the knowledge-how of foundation model, this is not our knowledge-how.

We only know that the alignment method should be different, such as the process of training and instruction SFT in the end - which one I am not an expert.

Our team believes that different data should be needed to specifically align them.

"Think in game mode", rather than "think in logical reasoning"

Zhang Xiaojun: Do you have a more vivid metaphor to describe your current entrepreneurial state?

Xiao Hong: It is difficult to summarize the entire process in one word. At this moment, at this time of 2025, I think the goals are clear, there are specific things to do, and I am happy.

Zhang Xiaojun: Do you pay more attention to experience this time?

Xiao Hong: You can't say that, but I know that many things are not something you can control yourself. You should do some of the things you can control yourself. There are too many things that cannot be controlled, such as geopolitics, you can only treat them as an input, but you have no way to control them.

It is more important that you do what you want to do well.

For the first time, I have more gains and losses.

I asked DeepSeek yesterday and translated the three words "greed, anger, and ignorance". He is very good at explaining that "greed" means attachment to good situations, "anger" means dissatisfaction with bad situations, and "ignorance" means ignorance of the truth of the world.

Zhang Xiaojun: We must have the craziest imagination of the AI ​​era and implement it based on the craziest imagination. Do you have any particularly crazy imaginations? For the world in the next 5 years.

Xiao Hong: The white-collar lifestyle may be a detour for human beings.

Longing the history of human beings for ten thousand years, it is rare to sit in a place with high-intensity mental strength and not exercise much, maybe 100 years. In ancient times, we needed spiritual civilization construction and physical labor to make our bodies better.

In the past 100 years, people have started to have diabetes and hypertension because you work like this.

If AI does this, everyone should live more like in the past - to improve their spiritual civilization and improve their health.

Zhang Xiaojun: What are humans doing?

Xiao Hong: Of course, white-collar workers will definitely not die, because they always need to sign and take responsibility. (Laughs)

Zhang Xiaojun: Recalling whether there was any wavering time in the past three years of starting a business?

Xiao Hong: I will definitely do it when I encounter difficulties, but I have no choice but to finder - you won’t say that I am in difficulties now, so I will quit.

It’s not that it’s a bad thing to not choose, it’s this mechanism that makes you more resilient.

Zhang Xiaojun: When is it swaying?

Xiao Hong: It can’t be called swaying, I must have a relatively low time.

When the new dimension has not yet appeared and the old dimension seems to be imaginable. When you can make linear speculation, you will be a little less excited, or will be more likely to be disappointed when you encounter setbacks. Because it is all linear and conjecture.

It is still in the early stages of AI technology, and there will always be new things coming out, which is the biggest change in 24 to 25 years.

Zhang Xiaojun: In the past 23 years, this wave of AI entrepreneurship feels like a giant's opportunity, and everywhere, it feels like a model company's opportunity. Many people will ask: What are the opportunities for AI application companies? ——Do you still have such doubts now?

Xiao Hong: Don’t confront application companies and model companies.

First, the model company will also do the application itself. Application companies should respect model companies very much. Because most of the capabilities or progress are provided by model companies.

We are in the elevator and enjoy all this. It is obviously wrong for you to steal someone’s business. The relationship between application and model should be viewed reasonably.

Secondly, the emergence of DeepSeek and the emergence of open source have had such a big impact on everyone, so it should be more optimistic for application companies.

All application companies will say that the model will become a commercialized, or the open source model will catch up, and everyone will do the use case and solve specific problems.

In the past, this sentence was useless because you still had to adjust a third-party model, and the model manufacturer had at least the right to charge a very expensive price.

Look at the American API, all company APIs are more expensive than Chinese manufacturers. Because of the domestic competitive environment, API manufacturers have a more intense competitive environment, so APIs can also have higher gross profits, in the past.

But DeepSeek appears, making it really happen. You are really paying attention to your application, use case, users, and growth, and don’t worry about one day the model manufacturer wants a very high gross profit, and one day the model manufacturer decides to do what you want to do, forming a competitive relationship with you.

Today is hereDon’t worry about this matter, this is the second point.

Third, make good technology and products, and the market will reward you.

In fact, there are many independent developers who think that if the model does not appear, they will not be able to achieve such a large scale. Therefore, it should be more optimistic, not more pessimistic.

Is it as big as the mobile Internet? I think it may not be that big, because that change is also accompanied by billions of people from not using smartphones to smartphones, and this change is huge. But it's already big enough.

Don’t be obsessed with being the “next byte”. (Laughs)

Zhang Xiaojun: In the past two months, has DeepSeek brought about a qualitative change in the underlying soil environment of global AI entrepreneurs?

Xiao Hong: It will take some time to see the specific changes and impacts, but everyone will rethink their strategies. For example, Sam Altman publicly admitted that there is no open source may be wrong.

I think founder also has a thinking model called "thinking in the form of game". Game does not mean zero sum or competition, but rather "thinking in the form of logical reasoning."

Zhang Xiaojun: What does it mean?

Xiao Hong: Logical reasoning means that, for example, Baidu has the best algorithm engineer, Baidu will definitely do the recommendation, which is called logical deduction.

If you have this, you will get this. This is called logical thinking.

Think about the game method is that your appearance and other players may make the entire environment different.

Zhang Xiaojun: "You" refers to yourself.

Xiao Hong: Yes. If there was no DeepSeek wave, everyone might not think about open source, but because of the emergence of DeepSeek, everyone should seriously think about it: Should we open source?

This is something you could not logically deduce in the past.

It is a change that occurs because of the addition of a certain player. Sometimes this player is a third party, and sometimes it may be yourself.

For example, I even thought that if OpenAI did ChatGPT at the beginning, it was a third-party company that made ChatGPT, but it belonged to a third-party company. Perhaps OpenAI chose to be a pure platform company, rather than a consumer app company.

Founder should think from this perspective.

Zhang Xiaojun: This is what I want to ask you. In the case of linear derivation, everything seems desperate. What kind of mentality should be used to deal with a dynamic environment at any time? ——Everything is changing.

Xiao Hong: Logical deduction you will be desperate. If the world is logical derivation, today's biggest recommendationThe engine company is Baidu.

First, it is not logically deduced. First of all, it is important for everyone to know this.

Second, try to make yourself an important variable.

Third, how to be as good as possible? There is no way, so you can only do your own thing.

Zhang Xiaojun: Tell me about your life. What is life like for a founder? This year is exactly 10 years for you to start your business.

Xiao Hong: My first company was registered on January 20, 2015. I was still in my senior year at that time.

Because I didn’t know what the founder is special.

But first of all, the work intensity is very high, and the mood often fluctuates. Hahahahahaha.

Zhang Xiaojun: What do you think of this world today?

Xiao Hong: First, it is the AI ​​era; second, I think it is in a good era of globalization. I am not a geopolitical expert, and it sounds like everyone has their own problems, so the overall situation is more conservative and isolated.

However, everyone does not want others to be isolated, they only want to be isolated. Everyone will hope that their entrepreneurs will think more globally - today's Chinese entrepreneurs should be more radically global.

Everyone should go to the international market to experience it. They need to participate in global competition, rather than compete in the market we are used to.

When I was working in this company, none of our founders had lived overseas for a long time. Our English level was probably high school at its peak, because the English level of everyone in college declined. (Laughs)

I joked at the time that if a founder who lived in the United States was placed with me at the same time, I would have chosen to vote for that founder myself.

But first of all, it shouldn't be like this comparison, but you do your own thing.

The idea at that time was also very simple, that the global market should be larger, and the market would give tuition fees to the founder to go to school.

The last few quick questions and answers

Zhang Xiaojun: The last few quick questions and answers. A favorite food worldwide.

Xiao Hong: Hot dry noodles, welcome to Wuhan to eat.

Zhang Xiaojun: A favorite place.

Xiao Hong: Company Office.

Zhang Xiaojun: A knowledge point that few people know but must be understood.

Xiao Hong: Snoring is very bad for your body, it will affect your sleep and make your brain hypoxia. If you snore, go to the hospital to have a look early.

Zhang Xiaojun: Recommend two must-read books.

Xiao Hong: "When the Human Stars Shine" and "The Road to Happiness".

Why?

BeforeThe latter will make you very motivated and also make you very anxious; the latter will make you less anxious.

Zhang Xiaojun: Tell me about a venture capital opportunity with potential that you have discovered now.

Xiao Hong: When choosing a vertical field, the business process in that field should be a little complicated, and today's model cannot be solved well; but the magic of your prediction model will definitely come, such as explaining the year.

You just do this business process well and wait for the model to become stronger.

Zhang Xiaojun: What is your very ideal day like?

Xiao Hong: Actually, this is the case during the Spring Festival holiday, perfectly fitting this day.

When I wake up in the morning, I read some books first, and when my colleagues get up, I handle the work, and when I read some books at night, these are the days during the Spring Festival. Feel very happy.

Because reading is something you can control, everyone works together, and why is it better to work during the Spring Festival? There will be many unexpected things missing.

It's more about you doing something you can control.