Leaving Baichuan to start a business, eight people worked hard for over two months to create a popular Agent product. The founder said, "The Agent technology is somewhat mysterious."
“During my time at Baichuan Intelligence, my colleagues and I were always in a state of high excitement. Although we often worked late into the night, sometimes not leaving the office until one or two o'clock, we felt extremely fulfilled and happy inside.” When recalling that experience now, Xu Wenjian, the former head of Baichuan's toolchain, still has a sparkle in his eyes.
Xu Wenjian joined Baichuan when it was at the peak of its fame. More than half a year later, he chose to leave and embarked on his entrepreneurial journey once again.
Born in 1994, Xu Wenjian still retains the shadow of “technological idealism”. Even in his mid - thirties, he can say, “Entrepreneurship is a complex task. We need to make money without changing our original intention.”
From Xu Wenjian, we can see that those tech enthusiasts with ideals from the era of the “Six Big Model Tigers” still maintain their enthusiasm today. With the growth experience accumulated from the “Six Big Model Tigers” era, they are starting new entrepreneurial stories in the Agent era.
Growth Starts with Repeated “Disenchantment”
Xu Wenjian graduated from Nanjing Institute of Technology. In his early college days, he was introverted, so he deliberately forced himself to speak in public, even though his fingers would tremble with nervousness at that time. In addition, he actively participated in various college entrepreneurship activities. “Although the school is not among the top - tier ones, the entrepreneurship atmosphere is very strong. I'm very grateful for the help and inspiration from my teachers and classmates during this process. My alma mater shaped the prototype of the entrepreneur Xu Wenjian.”
Like many fresh graduates, Xu Wenjian also wanted to join a big company. After working at a startup for some time, he finally got his wish and joined Didi. At that time, he spent a year and a half in his spare time to reconstruct a technical architecture. At first, no one supported him, and some even questioned that he was reinventing the wheel. But as the project progressed, Xu Wenjian gradually won the recognition of his colleagues, leaders, and classmates from other business units. This experience also disenchanted him about big companies: they weren't as great as he thought.
However, the experience at the big company also planted the seed of entrepreneurship in him. At that time, his leader at Didi evaluated him as having great potential for entrepreneurship. So after leaving Didi, Xu Wenjian didn't rush to find a job but started exploring the path of entrepreneurship.
During that period, Xu Wenjian participated in two entrepreneurial projects simultaneously.
In the first project, he served as the core technical leader in a six - person small team and started developing a cloud Coding product. In hindsight, this was still a very forward - looking project. The team even received a $2 million investment based on just a PPT. But they encountered problems such as member attrition and no progress in the overseas market. The team was under great pressure, and Xu Wenjian even questioned himself every day whether he should continue. Finally, this project failed. Xu Wenjian still feels regretful when talking about it now. “Although there were many other problems during the process, the lack of perseverance was undoubtedly fatal.”
The other project was an AI education product that Xu Wenjian explored alone in his spare time. Initially, he just wanted to organize a learning exchange to get to know the cutting - edge world, but unexpectedly, he organized a team during this period, including four education doctors from Beijing Normal University and some R & D personnel. Finally, they developed this application together. Unfortunately, this project also failed after only four months.
“As the CEO at that time, I had obvious deficiencies in my understanding and experience of entrepreneurship. Although I could attract many excellent talents to join my team, the lack of continuous positive feedback and a clear strategic direction ultimately led to the failure of the project.” Xu Wenjian reflected.
“Baichuan Added the AI Label to Me”
After two setbacks, Xu Wenjian didn't immediately plunge into the next entrepreneurial venture. Instead, he chose to accumulate more experience in AI knowledge and practice first. At that time, Baichuan Intelligence had a good reputation and excellent technical strength in the AI field, so he sent his resume to Baichuan. It was the only job - seeking resume he sent at that time, and he finally got what he wanted.
Working in a large - model company, Xu Wenjian could clearly feel that the pressure and anxiety in these companies were much greater than those in application - oriented companies. “Because they always have to do benchmarks, and the competition is fierce.” But everyone's state became even more excited because of this competition.
“The first - batch practitioners I met who joined the Six Tigers all wanted to achieve something, or they were idealistic people. At that time, although these six companies had raised a lot of money, their scale was definitely not as large as that of big companies. So everyone's mindset at that time might have been relatively pure, just wanting to devote themselves to the AI cause.” Xu Wenjian said. “The biggest contribution of the Six Tigers to China is that they trained a large number of AI entrepreneurs.”
Nowadays, the “Six Big Model Tigers” are all seeking their own ways to survive beyond large - model R & D, each making their own choices. Looking back now, Xu Wenjian just sighs at the rapid changes in the large - model era: ChatGPT, which was once at the peak, quickly lost market share to its competitors; there were so many new technologies to learn every day; and various breakthroughs constantly refreshed people's understanding. Then he uttered that classic line: “I watched it rise to glory, and I watched it fall into ruin.”
“Looking back on my experience at Baichuan, I have to admit that there were some problems. Ultimately, these are all related to the organizational culture and values. We seemed to rely too much on luck and ignored the importance of our own efforts and team cohesion.” Xu Wenjian admitted frankly.
In Xu Wenjian's view, Baichuan Intelligence might have fallen into an operating rhythm similar to that of a big company too early. There were too many senior executives, and the work directions of each department were relatively scattered, making it difficult for the whole company to form a synergy, thus affecting the overall development of the company.
This was actually also mentioned in Wang Xiaochuan's public letter in April: “After two years of long - distance running, the front line was stretched too long and lacked focus.” And after focusing on the medical field strategically, “Each team didn't deeply think about ‘why’ and ‘how’ in creating medical value. As a result, the work goals of some teams wavered and deviated.”
Despite his sighs, Xu Wenjian is more grateful for his days at Baichuan. “Baichuan added the ‘AI’ label to me. This label not only represents a transformation of identity but also means that I have the opportunity to continue exploring deeply in the AI field and go further.” Xu Wenjian said.
Xu Wenjian initially went to Baichuan to understand how top - tier domestic AI companies understand AI and develop AI products, but his unexpected gain there was Agent.
Inside Baichuan, Xu Wenjian's team conducted a lot of research and experiments related to Agent, including the development of the first - generation Agents Workflow in China. “We were among the first teams to realize the value of Agents.” At that time, Xu Wenjian's team quickly produced a demo version internally, but due to various problems, this project was finally stopped.
Before that, there were only the LangChain and Microsoft's Prompt Flow frameworks in the market, and people didn't have enough understanding of the necessity of Agent engineering. Xu Wenjian also admitted that his previous understanding of Agent was relatively shallow. “But at Baichuan, I came into contact with the most cutting - edge Agent - related knowledge at that time, which completely overturned my previous perception.”
By the end of 2023, Xu Wenjian's attitude had changed to: Agent has the potential to reshape the whole world, and it is as important as large models. By the beginning of 2024, projects like Dify and Kouzi gradually emerged, and Agent also attracted wide attention in the industry. Xu Wenjian and another partner, Feng Lei, also founded Mars Radio Wave in December of that year.
Starting an Agent - related Entrepreneurship
Xu Wenjian and Feng Lei met during the entrepreneurial process. At that time, Feng Lei was introduced to help Xu Wenjian find investment. Later, they found that their ideas were in line, so they decided to do something together. Interestingly, they are complementary in terms of personality and experience: Xu Wenjian is more emotional and is good at injecting momentum into what they are going to do and taking the lead; Feng Lei is more rational and is responsible for holding back; Xu Wenjian has mainly worked in the To B field before and has relatively less C - end practice experience, while Feng Lei has a richer understanding and experience of the C - end.
When deciding on the entrepreneurial direction, they were constantly thinking about the question “How will AI become the biggest variable in the Internet era?” Finally, their answer was content consumption, which includes two dimensions: creators and consumers.
On the creator level, time, creative ability, and business knowledge can be called the three elements of creation in the Internet era. The change brought by AI is to make up for the lack of ability of creators who are short of one of these elements. For example, for those who have creative ability and business knowledge but no time, and for creators who meet the three elements, AI can bring a hundred - fold boost and expand the scale of creation.
On the consumer level, real - life people are more three - dimensional and multi - faceted than the “labels” given by traditional algorithms. AI can more intelligently analyze and extract people's memories and provide consumers with more personalized and customized content.
The personalized experience they hope to achieve is like this: AI can record important moments in people's lives, such as getting promoted or falling in love. These experiences form the user's personality, and AI will generate different suitable content according to the user's current experience.
For this purpose, Mars Radio Wave has planned three development stages: First, achieve “human - like flavor”, that is, make the expression of AI reach the level of human creation so that users can accept it; second, achieve “personalization” and truly make it different for each person; third, conduct in - depth exploration in vertical fields and achieve more in - depth customization. Currently, it is in the first stage.
After determining the general direction, they refined it to the AI audio direction, with the reason being “the technology is relatively mature and the cost is controllable”. In implementation, they first chose the popular AI podcast scenario, and the corresponding product is ListenHub.
However, the two didn't start product R & D at the beginning of the entrepreneurship. Instead, they spent a long time focusing on building the team. They hope that this team is sufficiently AI - native and highly self - rotating.
Team Building: Quality Is More Important Than Resume
Now, including Xu Wenjian and Feng Lei, Mars Radio Wave has a total of eight people. The responsibility boundaries among team members are not very clear. Everyone has their main responsibilities, but they can also participate in other work according to their interests.
When recruiting, Xu Wenjian pays more attention to people's quality. “Quality is the top priority, even more important than the resume.” For Xu Wenjian, a person's growth potential, awareness, and self - motivation are more precious. Therefore, although the educational backgrounds of the Mars Radio Wave team members vary, from junior college to master's degrees from prestigious schools, and there is even an intern from Tsinghua University, “but these people all show extremely high growth potential.”
Judging from the screening results, there are mostly young people in the team. “Some experienced job - seekers may be reluctant to change because of the heavy burden of the mobile Internet era. Their experience has become a constraint on their growth. This is actually ironic.” Xu Wenjian sighed.
Mars Radio Wave has its own screening process to find talents that meet the requirements. Mars Radio Wave has set up three rounds of interviews, and the quality interview for candidates actually starts from the first round. For example, for the same question, different people will show different solutions and attitudes. Xu Wenjian will also spend a lot of time communicating and confirming with candidates later.
Xu Wenjian hopes that team members gather together because of a common motivation. “Entrepreneurship is not only a technological competition but also a competition of organizational culture and values.” The experience at Baichuan made Xu Wenjian deeply realize the importance of organizational culture and values, so he has invested a lot of time and energy in continuously updating ideas and aligning values with everyone.
Xu Wenjian and his team will think about things in the next few months in advance and give the team a big goal periodically and timely synchronize background information with the members. “We believe that as long as team members understand the goal and the motivation behind it, they can arrange their work independently and promote the project forward.”
Currently, this model is running smoothly. “What I'm most proud of is that as long as we set a clear big goal for the team, even if Feng Lei and I leave for one or two months, the team can still operate efficiently. They will actively improve the goal and direction and present the results to users. In this process, we can completely let go and let them make decisions and execute independently.” Xu Wenjian said happily. “The concept behind this team operation model is the unity of culture and values.”
This management method enables everyone in the team to grow very rapidly. “Looking back over the past three to six months, I can clearly see the changes in each person. They have gradually grown from initially being only good at a single skill in a certain field to becoming compound - type talents with multi - dimensional abilities. This rapid growth and change are also a major feature of our team.”
When making internal decisions, the team more often judges the direction through full discussions, common - sense judgments, and data analysis rather than relying on authority. “Because the whole team has such a consensus: everyone needs to keep improving, and no one is always right, including me.” Xu Wenjian said.
Developing ListenHub in More Than Two Months
The team really started investing in the R & D of ListenHub in March this year. The entire R & D cycle was actually only a little over two months, and it was officially launched in May.
Currently, there are three engines inside ListenHub, each responsible for different tasks: The first engine is responsible for analyzing the user's input intention and will expand different - structured and in - depth - analyzed articles for different questions; the second engine is responsible for generating human - like and highly personalized content; the third engine is responsible for converting text content into various forms of audio.
Currently, there is no unified standard or framework for Agent, and everyone is groping forward. Although from a more macroscopic perspective, the processes of each team may be similar, such as analysis, planning, execution, attribution, and reflection, in terms of specific implementation methods, the concepts and practices of each team are very different, and this difference brings completely different effects.
“There is indeed an element of mystery in Agent technology. We studied many open - source AI podcast generation tools and found that their structures are different. Through continuous attempts and explorations, we found a more effective implementation method than the open - source structure. This may be a stroke of luck or a talent, just like finding the best combination among 50 materials.” Xu Wenjian said.
The team also tried several different models and called different models to complete different tasks in different scenarios. Some concepts and architectures also refer to Xu Wenjian's previous product development experience in entrepreneurship.
The team also uses AI tools to improve efficiency. Xu Wenjian also uses AI programming tools to develop products. “In the past, an excellent engineer might be ten times better than an ordinary person; now, they might be a hundred times better.”
However, under the current architecture, from pre - processing to content generation and then to multi - modal conversion, there are a large number of details to be handled in each link. During the R & D period, the team members were not fully in place, so an important task for the team was to repeatedly consider which functions were necessary and could be implemented first and launch a minimum viable version with limited resources.
For example, the team postponed the introduction of the reflection mechanism. The consideration behind this is that the reflection mechanism is necessary in some scenarios, especially in general Agents that emphasize action accuracy; in vertical fields, the generated content is already relatively in line with user expectations and relatively accurate, so there is no need to rush to launch the reflection mechanism now.
Another important but not fully implemented function is RAG technology and a more intelligent mechanism. RAG mainly focuses on finding the information closest to the user's query through retrieval and cannot achieve memory extraction and understanding. The memory of an intelligent agent is more complex and depends on a more complex reasoning and analysis process. Therefore, the team needs to build a more intelligent information analysis and extraction mechanism, which will be a very crucial part of the product and can effectively sort out high - value information for users in different scenarios.
Regarding the popular MCP, the R & D team has reserved various interfaces in the architecture design, but currently, this is not the most important thing. “Users don't care whether they are using MCP, Coding server, or a specific protocol. What they really care about is whether they can use our product.”
In Xu Wenjian's view, the essence of an AI product is to