Recommendation
Explore the transformative potential of Agent and open a new chapter in human-machine collaboration.
Core content:
1. Multiple definitions and business insights of Agent
2. Innovative practices of AI Native companies in the field of Agent
3. Unique value of Atypica.ai and future prospects of human-machine relationship
Yang Fangxian
Founder of 53A/Most Valuable Expert of Tencent Cloud (TVP)
2025 is known as the "Year of Agents". From enterprise-level AI assistants to personal planning tools, various types of agents have sprung up. However, despite the high enthusiasm in the market, Agents have not yet formed a unified definition - is it a "next-generation app" or is it closer to an "intelligent collaborator"? Most people still regard it as an upgraded version of traditional tools, but the real potential for change may be far beyond imagination.
In this wave of agent exploration, AI Native companies are trying to break through the traditional framework and redefine its boundaries. They are no longer limited to the positioning of "efficiency tools", but are exploring the deep value of agents in business insights, creative generation, organizational change and other fields.In this interview, Dr. Fan Ling, founder of Tezan, will share his unique insights into Agents - simulating real user behaviors through large language models, allowing AI to not only answer questions, but also actively build user portraits, drive decision-making processes, and even expose blind spots in human thinking. This innovation not only challenges our understanding of Agents, but also heralds a new model of human-machine collaboration.【Core Insights】
- Product innovation: Compared with traditional AI, Atypica.ai's innovation is to simulate real people, use large language models to study typical users, and use multiple AI assistants to collaborate to conduct large-scale user interviews efficiently and at low cost.
- Divergence-first model: The divergence-first model is used at the reasoning layer. It is suitable for dealing with the non-consensus and artistic parts of business problems, which is different from the traditional convergence-first research method.
- The value of hallucinations: Allows AI to generate non-consensus opinions, broadens the boundaries of thinking, and is suitable for issues that require multiple perspectives (such as opinion polls).
- Organizational change: AI changes the way of working, from specialized division of labor to more versatile, and the company's organizational structure develops towards fewer positions and more compound skills.
- Employee potential: It’s not about AI replacing people, but that in a company like ours with 300 people, everyone can unleash unicorn-like potential.
- Observing the “mirror”: AI is not only a tool, but also a “mirror” for observing human society, which may reshape the way of work and life.
- Preset question bank questions: Simulated and preset question bank questions can enable AI to better understand user needs and ask appropriate questions.
- The value of multi-agent interaction: Multi-agent interaction can help us better understand people’s blind spots in decision-making, judgment or observation.
- Relationship between virtual agents and humans: The relationship with virtual agents in the future is hard to imagine. They can be observed as possible situations in the human world, or as tools for handling tasks.
What is the biggest difference between Atypica.ai and traditional agents?
Hello everyone, today we have invited Mr. Fan Ling, the founder of Tezan, to talk about a very interesting product they recently developed, Atypica.ai, and the thinking behind this product. This year is called the first year of AI agents by many people, and Atypica.ai is a very interesting AI agent product that I have paid attention to. So, what was the opportunity to create this product?Our company, TeZan, mainly helps big brands create content and AI-related products. We provide customers with tools for content management, distribution, and production. At work, customers often ask us: What kind of content should be produced? How to create this content? Which of the different content solutions is better? In addition to marketing, customers also involve content issues when understanding consumers and developing new products. These are typical business problems.In the past, we used traditional market research methods to solve these problems, such as qualitative research, quantitative analysis, data analysis, and consumer interviews. Now we are thinking: Can we use large language models to better solve these business problems? These problems are special because they usually have no standard answers, and solving one problem may lead to new problems.Traditionally, researchers have mainly used simulations to solve these complex problems. Previous simulations have focused on group behavior, like studying the overall trends of a population like studying a group of mice. But with the big language model, we can now better study and simulate individual behaviors. That's why we named our product "Atypical" - we want to use the big language model to study what a typical user would be like, so as to help solve various business problems.Let me give you a simple example: if we let AI finish reading Journey to the West, we can let it simulate the personality of Sun Wukong. For example, if we ask it "Would Sun Wukong like coffee or juice? What kind of movies would he like to watch?" AI can answer these questions based on its understanding of Sun Wukong's personality. On our website, we summarize our goal in one sentence: "Modeling the subjective world." We want to give AI specific personality traits through context to simulate real people.Now that we have the technology of multiple AI assistants working together, we can let one AI act as an expert to interview other simulated users. This will enable large-scale user interviews to be conducted efficiently and at a low cost. We have always wanted to do this, and we have finally achieved it in recent months, mainly because the capabilities of large language models have been improved, which can simulate more realistic consumer images.The second reason is that we have seen many forms of multi-agent interaction, which turns the work of the agent from a black box into an observable process. For example, Cloud Artifacts, Man or Devin, you can intuitively see how the agent works. This allows us to see a new way of interaction. Based on these observations, we quickly decided to put the ideas that have been in our minds into practice.I noticed that your product has a unique feature, which is to simulate real people. This idea is very innovative. Most AI assistant products on the market are mainly used for thinking and solving problems, such as OpenAI's Deep Research and Manus. These products allow us to see the most basic form of AI assistants. Many people still view AI in a traditional way, thinking that it is just a tool to solve complex problems. But your product focuses on simulating real people.Yes, there is a continuum in our thinking. I think that as humans develop, more and more things can be modeled and calculated. For example, physics is used to calculate and predict the objective world: for example, when an object falls, we can predict its trajectory, speed and time, and how it will bounce back.Now, the big language model helps us solve the problem of language communication, and then helps us understand the problem of thinking. This makes us realize that just as physics can model the objective world, we can also use the big language model to model the subjective world of people.I find the interview format particularly valuable. In the past, we always let AI answer questions, but now we can do the opposite: let AI ask questions, let people answer, and then let AI summarize and analyze. When there are multiple AI agents, the interaction between different AIs will generate many new possibilities. This is why we want to let more AI agents that simulate real people communicate and interact with each other.This hypothesis is very interesting. Google recently released the A2A protocol, which is mainly about the intercommunication between multiple agents. However, many people still understand it at one level: connecting agents focusing on different fields to form a workflow. In your context, the definition of Agent is actually more like a virtual individual.Take the example of Journey to the West that I just mentioned. By reading it, we can understand what kind of person Sun Wukong is. By understanding his personality traits, we can deduce what kind of decisions he might make, what kind of judgments he has, and even infer his views on certain things, rather than just understanding the knowledge he possesses.This is an interesting idea. I don’t know if there are other companies in the industry that define it this way. Many people may still simply regard Agent as the next stage of development of App.Everyone is very excited about the development of agents, especially the various functions achieved through MCP. Although we are also looking forward to it, we are currently focusing on simulation research using agents. As for the specific situation of other companies, I don’t know much. In fact, in the academic field, this is not a new topic, just like the word agent itself is not a new concept. In the past, simulation research focused on group behavior, such as theoretical research such as cellular automata. This is like studying ant colonies and observing the behavior of each individual. These studies have been ongoing and there have been many breakthroughs recently. For example, in the fields of marketing and information systems (IS) , many researchers are exploring the use of agents for market research. There is even an emerging discipline called "generative social science" that specializes in using agents to simulate social problems. Unlike the past simulation that focused mainly on group simulation, we can now study the individual level in more depth.Stanford Town is still quite popular for academics who have written papers before.The Stanford team also wrote an article in which they simulated 100 to 1,000 random Americans using an agent.What I just talked about was how the underlying logic is thought about. Maybe a more specific case is needed to explain to you how Atypica.ai works.The workflow is actually very intuitive. On the web page, you will see a dialog box, which is the same as common AI tools. You can raise business issues that need to be analyzed in the dialog box, and the system will ask three to five questions to clarify your specific purpose. For example, suppose I want to study user feedback on a product, the system will ask you: "What role are you in studying this issue? Do you want to know performance feedback or user experience feedback? After obtaining this feedback, do you plan to use it for new product development or competitive product research?" Through these questions, the system will understand your needs more clearly.In the second step, the system will organize the previous questions and answers into a series of specific work tasks. Then the third agent will conduct real-time searches on social media. We mainly cover platforms such as Xiaohongshu, Douyin and Instagram. Instead of directly calling the platform's data interface, we search for content like a real researcher. After searching, we can see a lot of posts, including original text and comments. Based on these contexts, we will simulate the typical consumer portrait of the posting user, generally generating at least 5 typical user portraits, and some customers even require the generation of 100.Is Persona a professional term in the design field?Persona is equivalent to a portrait of a person's prototype. Then the next agent will be triggered to conduct an interview and ask the prototype people the corresponding questions. These questions will be related to the big problem you want to solve. After the questions are asked, these questions and answers are summarized into a paragraph. This paragraph will then generate a report with pictures and texts. This is probably the process.In product design, should we first work our way back from usage scenarios to deduce the current product form, or should we deduce its possible usage scenarios from simulated groups?Let me explain the four main business problems that our system can solve: The first is market insight. For example, we can analyze user feedback on a product. For example, when we recently studied the new energy vehicle market, we found that young families with more than two children may need MPVs ( multi-purpose commercial vehicles) . The second is product co-creation. We can invite target user groups to participate in product development. The third is product testing. Suppose you want to develop a chocolate for fitness enthusiasts, and you have three recipes A, B, and C. We can help you analyze which recipe is the most popular. The fourth is content planning. For example, many Xiaohongshu bloggers use our system to analyze their account positioning and plan future content directions. In addition to these intended uses, we also found that users are developing some interesting new uses. For example, some people use it to plan to study abroad, analyze their background, and get suitable school recommendations.Finally, let me share an interesting example: In the past, foreign companies had to rely on Chinese teams to conduct research to understand the Chinese market. Now, they only need to use Atypica to ask questions in French, and the system can analyze Chinese social media data and directly generate French reports, which greatly improves efficiency.Similar to "field research".Yes. We are now working with some authoritative media to integrate their unique data sources. These are accurate data at the macro level, different from hearsay on social media. We combine these rigorous data and opinions with the diverse voices on social media to form a complete analysis report. For example, when studying serious topics such as how China's new energy vehicles can expand in the Southeast Asian market, we are increasing the proportion of quantitative analysis.What new scenarios does the innovative design of Agent bring?
In the initial conception, this technology was mainly used for user research in the field of market research. This inspired me to think that this kind of simulation is particularly suitable for personal planning because it can help you explore different possibilities in life. Now, hallucination is often a big problem when AI is implemented, or its accuracy is a big problem. In your scenario, is accuracy a problem? In which scenarios will it be used better?I think illusion and accuracy have two sides for business research. Business problems are both scientific and unscientific, which is why I said earlier that it is both science and art. This is also why we need to discuss cooperation with some serious media and access their data sources. Although not every report needs access, for some special issues or advanced users, we need to obtain authoritative data. These authoritative data should take precedence over hearsay on social media to ensure the authenticity of the analysis.This is the Science part. As for the Art part, when we need to expand our thinking, we need to see more diverse perspectives, and social media data is particularly helpful at this time. When the language model directly answers questions, it will give a very structured answer, but when it is asked to analyze social media data, it will be found that there are many disputes and discussions, and the content will be richer and more diverse. This is the characteristic of business problems, and this is why we call it a complex problem. You need both a consensus based on facts and the exploration of possibilities beyond consensus. Sometimes you need to see multiple non-consensus states coexisting, and ultimately it is up to people to make decisions. The role of insight is not to help you make decisions, but to help you see more possibilities. So I think this system is particularly advantageous in dealing with this non-consensus, artistic part.In a way, Hallucination is a good thing because it makes people more open-minded. So we are developing a big model that we hope to file as soon as possible. At the reasoning level, we will not reinvent the wheel, but use the existing DeepSeek v3. But in terms of high-level reasoning, we want to make a divergence-first model rather than a convergence-first model.For deep research, you may want to find the answer rigorously. But what we want to do is a divergence-first model - before reaching an answer, let more people participate in the discussion and listen to more voices. I think this may be more suitable for dealing with problems that involve both art and science.Sounds perfect for the polls.I think we need more diverse perspectives on public opinion polls, or on business issues, for example.I didn't know you defined it like this at first. I used it as deep research, so my first question was to ask him about the current comparison of AI development in China and the United States, and what policy suggestions he had. He generated five different Personas for me, including employees of large companies, Silicon Valley engineers, and so on. I got a lot of different answers from their perspectives, which was very interesting. This gave me an idea: Isn't this just like a good expert interview? Traditional expert interviews are all paid. If in the future, as I just said, professional data sets are connected to these professional fields, can it also become a form of future Agent?We have some clients now. Because many of the companies we work with are consumer product companies, they invest a lot of energy and resources in consumer research every year. These companies have accumulated a lot of research documents. Although you may only read a simplified report in the end, the actual process contains a lot of interview records. We will collect these interviews - there may be thousands of them - and then convert them into interactive exclusive agents. In this way, if you are a brand, for example, you don't have to rely solely on the voice on social media when doing research.You may have your own unique Agent for your own employee library.I think this is a very good thing. In the past, you would simplify things and only read a final report. In the end, there are obviously hundreds or thousands of very different people, but in the end, they are all turned into a quantitative report? But now these people seem to have returned to being individuals, interacting with each other. I can't say that they are completely the same as consumers, but they are more like people than these cold numbers and labels.When I explain to clients and users, I always say that we all want to drink fresh juice, but sometimes we choose instant juice due to economic or environmental constraints. Although instant juice is low-cost and available at any time, it simulates the color, aroma and taste of fresh juice as much as possible. Traditional research methods are like telling you the composition of the juice, but knowing the ingredients alone does not allow you to experience the real taste. No matter how cheap instant juice is, it at least restores part of the real experience. That's why we say that what we are doing is like an "instant" version of research.You may be more familiar with databases built on a rational and scientific level, but yours is an agent at the artistic level. So what dimensions of materials does it need to collect to make this agent more like a real person?First of all, the ability of the large language model is the foundation. The large language model is like giving a person a personality. What we need to do is to tell it which personality to stimulate. So we are not inventing personalities, but stimulating personality traits that the large model itself has already understood. For example, we stimulate a certain personality through context. On social platforms, context may be relatively short, that is, users' posts and replies. If it is an interview, it may be a conversation of one or two hours, which is relatively long. If it is a work like the complete works of "Harry Potter", it contains all the relevant contexts, the effect will be better, and it can show richer personality traits.How many of these Personas do you foresee creating in the future?We are rebuilding the Persona library behind every survey now. So you can understand it as a process of accumulation, that is, if we have it now, it is not enough, anyway, this number is huge, this number is larger than I thought. Multiply it by five, at least multiply it by five is our Paku. But I think what is more important is that these are some of them that you can understand as not of the highest quality. It is to solve a certain problem. I just interview you for maybe ten minutes. So we are now doing something new, that is, AI intelligently invents questions for a business problem you want to solve, and people answer them. For example, I recently launched a question that I launched yesterday. If you have any suggestions for user experience, you can send it to our core users, and AI will automatically target this question. Ask questions about the big problem I want to solve, and let people answer them by voice or text typing, and then such AI can summarize your entire reply and give me one.Asking good questions is an important skill. Because I often do interviews, I also ask AI to make some interview outlines for me, but I feel that they are usually quite monotonous.There are two methods here. The first is to simulate (Impersonate) - you need to let AI understand what kind of person you are so that it can ask the right questions. I will demonstrate this product to you later. It is important to let AI understand that the context of the question is not just a prompt. For example, "I want to do an interview about AI agent, give me some questions" is not enough. AI needs to understand why you ask these questions, and as an interview designer, you also have to spend time interacting with it. The second method is to preset the question bank. AI can select and adjust questions from the question bank, and can also grasp the timing of asking questions. For example, there are 200 questions in the question bank, but not all questions need to be asked every time.How is Agent reshaping organizational structures and ways of working?
How receptive are your customers to these AI agents?That's right, we don't just want to serve existing customers. We want to help everyone who has business analysis needs, because the concept of "Business" is not just B2B, everyone has their own business needs.So it is not a pure ToB product.Yes. Now our user base has grown far beyond its original size. Originally, we mainly served large enterprises. They are very interested in our products, but they also raised several questions. The first issue is data security - for example, if they want to ask users for feedback on new products, this will expose new product information, so we need to solve this problem. The second question is whether we can integrate their private data, such as internal interviews and surveys. Third, they hope that the report will not only provide general answers, but also go deeper into specific issues. Fourth, they are concerned about how to turn these generated results into practical actions, such as turning them into new product development plans or the next round of project briefings, which are all very interesting.It felt like it killed four employees at once.Or I think that we originally divided our employees based on their skills.Now we may need more people with more complex skills. So originally there may be people doing research, development, design, and other things. Now we may need four times as many people to do these four jobs together, instead of everyone doing one job. That means they are all three floors, but this is the case, and then the feedback from our customers is probably like this.You just talked about the changes that AI may bring to organizations. We all feel that with more and more AI products, whether for personal use or for commercial use, everyone's work will be re-evaluated, such as the need for more complex skills as you just mentioned. As a business manager, how do you think you should organize your company in the future?I think any change is ultimately good, but the process of change is always uncomfortable. Whether a person wants to change or not, everyone knows that change is necessary. But human nature makes us not want to change too much.Speaking of the opportunities brought by AI, although our company mainly focuses on AI applications, we are also gradually promoting the application of AI. In this regard, large companies like Tencent also do the same.I think AI is bringing about a major shift: it is changing traditional industrial thinking. In the past, we divided job responsibilities into very fine categories, and everyone focused on their own professional field, repeating the same work over and over again, becoming so-called "tool people." But the emergence of AI makes this high degree of specialization less necessary. For example, if you are a writer or illustrator, you may worry: Is AI helping me or threatening me?I think the AI era is more like a return to the Renaissance, where everyone can become a versatile all-round talent. What companies need in the future is not fewer employees, but fewer fixed positions. I like this saying: everyone can become a "unicorn". This does not mean that a company only needs one person, but that a company like ours with 300 people hopes that everyone can realize their unicorn potential.Become a large "one-person company".Ownership for everyone can be more complete. Originally, I was only responsible for my own little piece of land. Now, you can be responsible for something more end-to-end, and everyone will have a greater sense of meaning. Originally, you were only responsible for this part, and you were over-optimizing a part every day. Therefore, our way of working has changed. We used to have iterations every two weeks, so that only those who continued to do things could really participate. Because there was a lack of rhythm in the two weeks, only those who did this thing every day could control the overall situation. Now we usually have one iteration or API card a day, and sometimes we can do three iterations a day. We now have short meetings of half an hour or 20 minutes, and we can see the results that evening and continue to discuss the next day. Each meeting has become shorter, but changes and progress accumulate faster, unlike before when we had to wait two weeks to hold a lengthy meeting.This working state is very good. In the past, it took at least ten people, including front-end, back-end, and product managers, to form a team to make a product, and that was not enough. One plan would take 300 man-days. Now we can probably do a project with two or three people, and these two or three people are fully responsible for the project results. If a task takes more than ten man-days, we will question whether we are going in the wrong direction.Will the company become an “incubator” in the future?I think these are probably just concepts, like incubation technology or something. I think the ultimate function is very important. You will feel that you have control over some things, and your tools are more AI, and you will not feel that you are a tool of AI. It is true that everyone is more like a CEO. It does not mean that everyone has to bear liability, but everyone's work is more complete, which is good.Yes. In the process of AI transformation, there are actually two paths. One is that everyone learns to use various AI tools by themselves. There are many domestic and foreign tools on the market. The other is to promote some unified tools from the company level, and integrate them into the previous workflow. What do you think of these two methods now?I don’t think there is an absolute right or wrong in using new technologies. The key lies in how people do it. Although people who are willing to try new tools will always be exposed to new technologies earlier, latecomers may also have their own advantages.Let me share an unexpected story with you. According to the original plan, we should have focused on improving existing products rather than developing new ones. But one day, one of my students made a usable demo version. After seeing it, I discussed with the CTO whether to give it a try. Surprisingly, our CTO completed the product development in just one week.During the development process, we encountered some challenges. For example, we found that the interaction method of the product was very similar to Manus, and there were already ready-made development frameworks on the market. It might take 3-5 weeks to complete using these frameworks, but our CTO chose to rewrite the code himself. This made us understand a truth: only by doing it yourself can you truly understand how to do it. Although we all know that we should use AI to write code and use tools such as Cursor, practice is the best teacher. Now our engineers choose to use the simple Next.js framework for development.Practice makes perfect. Many people only discuss what should be done, but may not have actually made a product. We found the right direction step by step through practice. Although only my CTO and I have mastered this development method, we plan to let the other five project teams learn it. Some colleagues find it difficult to transform old code and old products, but we want to encourage everyone to try. We hope that everyone will use AI to develop products from the beginning of the project.What kind of product innovation do we need in this round of AI wave?
The “old products” that Mr. Fan mentioned also have many AI functions. So what is the biggest difference between current new products and previous ones?I think it's like this. We first did a lot of AI understanding and generation functions. That was supported by the previous generation of machine learning technology. Now we need to reconstruct these things, not only the cost is lower, but the effect will be better. I have also been thinking about our products. As an enterprise software company, many enterprise software now have to add AI functions. But in fact, this makes the cost higher because you have to pay token fees, and customers don't have extra budget. I say the worst practice is "new wine in old bottles" - AI is new, but the product is still old. The result is that the product experience has deteriorated because the continuous addition of new functions has led to increased token consumption. It was originally a simple software, but now it costs more. Customers think the product has become more complicated, and the company advertises outside that "we have connected to GPT" and so on. This kind of daring pursuit of AI trend does not actually create business value.I think a better approach is "new wine in new bottles" or "old wine in new bottles". "New wine in new bottles" means to truly develop new products through AI, which is the direction we are trying. "Old wine in new bottles" means to use new forms to meet existing needs, such as our content management system. Later, I thought that instead of completely reconstructing it, it would be better to use AI to redesign the interaction method. In this way, it can be turned into a content database, but only the interaction form is changed at the upper level. I think this approach will be better, and customers will find it valuable and refreshing.Second, it is also easier for customers to understand why we should start charging token fees instead of just tenant fees as before.If so, does it mean that for customers, a smoother transformation method is to put new products in new wine or new products in old wine to establish a new workflow? It is not like many people are trying to use AI to improve productivity, but find that the cost is very high and cannot be reduced at all, or much.I don't know what you have observed, but everyone is not very satisfied with the so-called Copilot. The issue is cost, or the output of this part of the investment is not very satisfactory, so many may really need AI-based capabilities, and rethink whether some work processes can be redesigned based on the capabilities of the big meta-model?Are there any industries or scenarios in your mind that can be designed first?We have been making some attempts recently. For example, we only knew that brand companies needed more content, but later we gradually understood that because the media is more fragmented now, there is more demand for content. Many companies are operating accounts, such as Xiaohongshu and Douyin. At the beginning, we just wanted to help customers solve the large amount of content needs of the accounts, that is, to be able to produce a variety of content that can pass the review and is low-cost. But then we gradually realized that each account, especially device accounts, needs to have its own personality . This personality is like differentiated personality traits. Each account needs to have its own characteristics, and the content should also be produced according to this personality. So we don’t just use the company’s original data sources to simply edit, piece together, and adapt (Re-adaptation) , but to recreate (Recreate) .Based on the existing 0 to 1 content, create some graphic and text content that suits your personality.The video content reminded me of another point. This function can also be used in AI companionship. Because I feel that one thing that AI companionship lacks is personality. The reason why the previous egg was so popular was because its personality was very distinct. Most so-called AI companions are too obedient and submissive, which is actually not what everyone wants.Yes, so I think in fact we also see that the personality of the large language model is challenging, because the model is very submissive and very easy to tame. Very obedient, sometimes we want it to have more objections, it doesn't like it. So I think this is also a possible result of over-alignment, so when we want to add context to the large model to increase its opinions, we will also make some attempts in this area.Yes. If this is not the case, for ordinary people, if my own feelings interact with the big language model for a long time, your thoughts will be very solidified because it keeps praising you.It is very interesting to see the initial attempts to add personality to large language models, which brings rich species and personalities to the virtual world. In the future, will we live with these virtual agents? What form will they exist in? Will we only use them as tools?"Partner" is a question that is easy to answer in concept but difficult to imagine in form. I really like to observe the Replay function in our ATP. I often watch the interviews between agents, which is very interesting. Sometimes I will deliberately write in the prompts that instead of doing an interview, let them argue, so that the discussion between agents will become very intense. I don't treat them as real people, but observe them as possible situations in the human world. Through this observation, I can find some of my blind spots, which is why I like this kind of product so much now. Of course, agents like MCP that help you handle various tasks are another logic, just like there are many minions helping you work. But I prefer this kind of conceptual challenge. I think that through the interaction of multiple agents, we can better understand people's blind spots in decision-making, judgment or observation, which will be very valuable. For example, I sometimes ask it about the situation of some companies: "This salesperson is working hard, but the effect is not good. It has been three months. What should I do?" Although the agent's answer may not cite real data, this way of discussion can indeed bring me a lot of new perspectives.Thank you, Mr. Fan, for introducing this new product to us today and sharing the innovative and pioneering ideas behind it. We will continue to pay attention to the development of this product to see what changes it will bring to our lives.I hope everyone will try to "break" it in various ways.
