Conversation with Xue Tianlu: Before the Second Enlightenment

Welcome to Dine Podcast, I am Du Xiao, this is our third episode. Dine is a design studio in Beijing. Since its establishment in 2014, we have provided design services to hundreds of enterprises and startups. You may have used the products we designed. Nowadays, we use a lot of AI to help us complete our design work. This program has no fixed theme, but it is currently focusing on the stories created by ordinary people in the AI era. The guest in this episode is Xue Tianlu, who is a colleague I call directly and is also the first intern I recruited. He later gave up the opportunity to further his studies at Indiana University and became a regular employee of Zhizhu. Tianlu later also briefly joined Dine as a designer. Later he joined the early Xiaohongshu and was responsible for the withdrawal product and later subsection jumping. For 25 years, he worked as an EIR for a period of time at Sequoia. He is currently the Whole Founder and CEO of Second Enlightenment. Their latest product, Clawly, is an OpenClaw client built for ordinary people and runs in the cloud. You only need to download an APP. You can let your lobster do things for you without the need to deploy the environment locally. Let’s start with Tianlu’s Webcode experience. So when did you start getting in touch with Webcode? I was probably in 203 or 204 when I saw the product Cursor. When I saw it at that time, the slogan was very clear. It was AI coding or something like that. But I didn’t get the full meaning of it, but I thought this product was cool. Then later on, I posted a post on Geek, which was probably about 2004. In the future, people will miss the time when people were still writing code, similar to such a pretentious description. Then when I saw this post later, I thought how could I be so prophetic? But when I saw Cursor, I didn’t try the Cursor product very much. Then I didn’t use AI coding or Webcode in depth, but there was no word Webcode yet. The next time I came into contact with it should be. I am a little specific now, but I have to look for it. It was probably at the end of 2004 or the beginning of 205, when I was still myself. I guess it should be that a certain big model has been updated. It is not the Replet that came out. Oh, I also used that one. Yes, it should be. I remembered that it was one. It also made an advertisement at the time and said that everyone can use it. I don’t know that it should be V level because each version of the CodingAgent of Replet has V V level VRV3 and the like. Then it was the first time I started using it myself. I just tried to do something at the beginning, but I used the Web version. Then I wrote two things. I wrote two or three things. At that time, I was working in Byte Routine in those two days because if I was responsible for it, As far as my position or that matter, you basically have ten hours of meetings a day, and it feels like this. But at the time, I remember that I had about a day and a half, and I didn’t have much, or some meetings, I didn’t hear much or anything. Then I just kept coding this thing. There should be no ethics committee. Then I coded several things, and then some simple things, such as notifications of Manchester United games, and then coding a website, monitoring Tesla stocks, some very simple things, and then I felt very forced when Reple Agent was first released. It is the first one that is equivalent to integrating the entire site, which means you can have a database, you can have a front-end, you can have a back-end. I remember that he was still writing Python at the time. I don’t know what code he wrote, but he didn’t need to setup anything. Yes, because I remember that after I pulled the code, I studied for a long time how to run this Python thing locally, because all I had been exposed to before were things that could be GS. I didn’t pull it down because you said this, which reminded me of the cursor from earlier at that time. I actually downloaded cursor and then, because cursor is about setting up an environment, right? Because I wrote some very simple code a long time ago, but it was very troublesome to need a setup environment, so I gave up. I suddenly remembered this thing, so replit is because of it. You can do everything in one stop on the web, so I started to rely on this thing. And today, two years later, replit has claimed to be able to put your iOS app on the shelves with one click. Yes, and then later, I should never do it again. I have never used replit again. After that, I probably paid for two months and bought it for me. I also bought it. I just said that the first thing you did was all related to personal interests. It was not related to work. I have a very clear impression. It was the first one called the results of the Slow Game. It was similar to this but it didn’t seem to be implemented. The second one was because I probably mainly bought a few stocks at that time, Tesla Meta and so on. Then I wanted to make a web page, which is to put the latest information of these four stocks on a dashboard. I can share the first thing I made. I made the app very unscrupulously, but I wanted to use it to open a link, that is, the webhook API database, and then this kind of long running task. But one of the things I did at the time was to monitor everyone’s working time in slack, because slack has an online and offline status, right? I asked the app to record the person’s online status every once in a while and then turned it into a table. You are awesome. You can now sell this product as a SaaS product. But I remember that the model at that time was sonit3.5. I am talking nonsense now, but it is very early or 3.7. How do you get pictures? How do you monitor online? Do you want screenshots? There is still this. This has an API, that is, it has an API. Yes, yes, so I used it to run a web app and open up such a small thing in our internal office. At that time, the model reminded me and warned me that you need to obtain the consent of the parties involved. I said that it was already a pretty awesome model. Then I made this thing and it ran, but it was very difficult because it was written in python. I didn’t know very much how to change some things, layout, including some deliveries. But this was my first replet. I tied a domain name to it, paid for it, and then ran it for a period of time. OK. In fact, it is still very complete. The complete concept of Sass is. What follows is that after I resigned, it was probably around May of 2025, or around that time. In fact, I can start to do some things of my own. During that time, I was actually watching various startups on X. New ideas or what our products look like. Then I started to make some ideas. It was not very utilitarian. I wanted to make this product. That is to say, I could have some time to experiment with some things at first, so I started to do more at that time. Then I started to use cursor to do some things. First of all, I am a person who likes history very much. I like history very much. I like Chinese history and foreign history. Then, what I was familiar with before was actually Chinese history. For example, I read a lot of history books when I was a child. Then there was a period of time when I felt that my history should be more refined. This history thing, this topic may not have anything to do with vibe code, but history thing, in fact, there are some big subjects. For example, Westerners like that historical topic very much, which is the Roman empire. There are some very funny ones, for example, on the So I spent some time studying the Roman Empire, and then another one was really, really, and then another one, another very representative historical topic, was the Warring States Period in Japan, and this one was something I actually didn’t know much about before. I played a lot of Three Kingdoms games when I was a kid, and then Koei, a company, usually made one with the theme of the Japanese Warring States Period, and then another Three Kingdoms version. Usually the Japanese games of the same generation would be a little more advanced than the Three Kingdoms version, with similar gameplay. But when I was a kid, I couldn’t play Japanese stuff. I didn’t understand the culture, so I spent a lot of time studying the Warring States Period in Japan, and spent a lot of time studying it. Then during the research process, I discovered a very interesting topic, which is Japan. There is something called 100 Castles, which is the entire country of Japan. There are 100 castles scattered throughout Japan, usually built from 1400 to 1660, which is before the Warring States Period, two or three hundred years of castles. Then there is an activity, which was mainly for tourists at that time. They often They like some people who go to Japan very much. They will check in the 100 famous cities. Then I discovered this matter. In fact, I knew this matter earlier, maybe three or four years ago. Then when I knew this matter, I thought this was a good thing and an opportunity to learn about history. But I had never had time or much time to control to see these things. Then last year, for example, when I gradually became more rational, I began to have more time to really visit. Make some plans to visit these famous cities and then add them. This is also another kind of collection batch, but the cost of this collection batch will be higher because you really want to go to this city, and the doorway is very deep. You have to collect a symbol for this, right? He has several methods. One method is that you go there and stamp a seal. There is a seal under the city. Then, they are all different. They are all different manufacturers. Then I went and started to do this. I probably started in the first half of last year, starting from the Spring Festival, and then started to have a purpose. I went to several cities at a time. After that, and in the process of doing this, I found that this thing is to meet my own needs. I wanted to have a good way to track. I went to those cities and saw a place on the map of this city. Even I might be able to satisfy some simple feelings of achievement, including this city. Some simple ones, where to buy Tamacheng Seal, are very specific, where to buy Tamacheng Seal, transportation, and even what time it closes, complicated details like this, and it would be best if it is the details of an app, right? Then I found that there are several apps on the app store, three or four of which are individually made as giant blues. It is possible in the iOS7 era. I don’t know, it has not been updated in 100 years. So I found that I can make such an app. I started using cursor and made this app. Have you visited all of them in the past 100 years? I am now 51. My current track on the app is that I have visited 51. My app also made a very beautiful one, which has a seal collection function. Because you were also a designer before, it was also the original one. So how did the materials or materials in your app come about? First of all, my app has nothing. I didn’t have the most complete plan at the beginning. I did it wherever I thought. My first idea is my first pront, which is literally to tell the cursor. Give me a tiktok style one that slides up and down, right, because I directly thought that my product looks like this, right. Then immediately after this page is finished, the second page is to make a map. There is no material left here. Then I answer your question about materials. The problem with materials is that there are two kinds of materials. One material is because I made the seal. There is now a batch collection system. You can check in when you go and collect the batch. This is also a paid function. If you collect more than three, you have to buy pro. Something like this, and then the batch function was something I thought of later. Then I thought I could add this thing, and at the same time I could also make a payment, because I didn’t want this app to start. It was very annoying to pay, so I came up with this idea. In fact, there may have been a period of time before. It may have been the first version of nadovnr in Maybe there was no nadovnr pro. I used O3 at the time, so I used O3 for the base model. Then the tripe image model used a certain image model in tripe. Then I let it generate all my small 3D things. Then it was very manual. I didn’t write one at the time, such as the display of calling the API, because I still used the tripe app or the web to copy prompts to make it constantly updated and generated. Then there may be changes in the middle, and some of the most annoying things in the middle. It means you do image processing. For example, when you generate something, you need a transparent background or a slight change of the background. This is very good. You should know that there is no particularly good tool for this until now. When we talked today, it happened to be GPT2 GPT image2. I haven’t tried one of our products yet. You took this thing and then I used GPT to do this and it was very easy to hit its limit. I was a member of 20 dollars at the time. Then for this, I upgraded to 200 dollars, so that part of the seal It is completely generated by GPT. Do you use its image model to generate 100 seals? 100 seals. Then you need to insert a point here. In addition to the 100 name, Japan also has a 100 continuation name. Those people were not satisfied, so they made 100, and then they made 100 more. I originally wanted to add 100 continuation, but later it was too annoying and there were too many things, so I cut off 100 continuation. Because in the process of going to some cities, I found that many of the 100 cities were in ruins. They were just residential buildings. In fact, later I felt that even if I gave myself an excuse, it would be nothing good. What about this APP for you? It is a very Japanese culture, a kind of collection of punch cards, especially stamps, a kind of digitization of gameplay. And I want to talk about this. This is very funny. I did it about two months ago because I have a support email in the APP. I should not have received it in history. The support email of this APP. Because this APP is not promoted very much, there should not be many downloads. The only one I received is a support email. There is a user. He asked when the 100th support will be renewed. Is he Japanese? He should be Chinese, Chinese, but I forgot it is Fandi or Jianli. Back to the material generation here. Material generation. This one is generated and grown with Tradby. Then I will do some control on the style, right? I will give him some prom and let him go to 3D and so on. Lovely squares and the like, and then that, and then if you want to talk about more details, because because because of each city, if you listen or try my APP, because each city will have some details, some elements, different, each city is different, so what I actually do is I will tell him some actual pictures and history of each city, and then he will generate this. He sometimes generates it very well, because each city, that is, each city does have some culture or a general there, so it should be different. Not all cities have this treatment. Because of the characteristics of some cities here, maybe 6-70% of the cities will have something of their own. Some very famous cities, such as Daman City and Record City. Obviously, there will be many, that is, Record City will have cherry blossoms, etc. Then another material is because I am a. The homepage of my APP is a card for each city. It will have a picture of a castle, and then there will be some content for the report. What I just wanted to ask is these pictures. Did you use photos or generated them? For me, I didn’t use photos because there was a copyright issue with the photos, but it wasn’t just a matter of mysophobia. I circumvented the copyright issue. I used AI to compile it, basically feeding it a lot of material and telling it the history of the city, so that it could generate as many prongs as possible. At that time, in fact, the pronging style was not that easy because today’s pronging is very prong-y. But about a year ago, at that time, pronging required a lot of prongs, so it was a bit laborious. So you really spent some time. And when you look at it now, You should still be able to see the traces of the picture on AI. If it is fresher, then the material will be completely written by AI and then collected. Then it may click some, add some links, jump to some official websites of this city, and micro-encyclopedia. All the content is probably written, and there is no CMS. It is very simple, just 100. It is hard code written in the code. Then this app is probably possible. The basic functions are written. In fact, it only took a week or two. I think it should have taken a week or two. For example, looking at the cursor now. You should be able to see the lock. It is very fast to write. But then you need to adjust some details. It took a long time to adjust. It is similar to making the card slide up and down. How to make the card stick to your hand. It is such a very detailed problem. At that time, even the current AI should not have a good solution. It is not 100%. Then the AI at that time was even worse. It took a lot of time to solve some very detailed design level and interaction level problems. A hundred people came behind you to officially start a company, build a product, and find people. What happened in the middle? What prompted you to start this idea? First of all, I started from my own rationality. I just wanted to start a business. That is to say, I wanted to have some time to explore on my own, including the direction of the team, and then rest for a short period of time, so I didn’t start right away. I had never been to Italy before, and I particularly wanted to go. This has something to do with history, right? I especially want to go to Italy. I want to go to Italy. That is, I think traveling is not necessary for me. But I think there are some places that I really want to go to once. Then I don’t have to go again. It doesn’t matter. But I have never been to Italy before. And I like the history of Rome very much. Of course, if you want to have West Burke talk to me about the history of Rome, I am very welcome, but I probably went to Italy in July or August. It should be three weeks, three weeks or four weeks. Then I basically went to all the places in Italy that I could go to in such a short period of time. Then after I came back, I went to a company of my friend, a friend we should all know together, to experience it and get back to a kind of work or what it feels like to work. The specific thing I did was actually building a product. Then I really worked in a team. Then when I was coding with the consultant of a friend's company, I was really working in a product, in front of the user product. Although my responsibility was a product, what I actually did was that I would do AI coding to do all my things, and then do it from the front to the back, and then do it all the way. Then I probably did this thing for about a month or two, and then I found some feelings. Another reason at that time was that maybe I wanted to return to some of this kind of routine state because I took a break, so my friends happened to confirm it and then I did it for a while. But after this experience, Around September last year, I made it clear that I was going to start, I was going to start doing my own thing, I was going to start building a team, and then I started to start in the direction I was optimistic about and started to invest in making products. In fact, I think I can insert a paragraph here about your previous work experience, such as when you graduated or started on Zhihu. I think you should start on Zhihu. I was in 11 years, when I was still studying business in the United States, and it was a two-year program, supposed to be from 10 to 12 years. Then in 11 years, even if it was the vacation between the two academic years, I went back to China for a trial. When I returned to China for a trial, I was submitted to Innovation Factory. Then I was a product, and I worked as an interaction designer in Innovation Factory, which is to do practical work. Its mechanism is to arrange you into some of the projects that Innovation Factory invested at that time. At that time, I was assigned to a project that he could force out later. I found it very boring, but I didn’t understand the product, and their team was also very small. There were maybe a dozen of them, and I was assigned to it. Then I felt that this project was very boring. I might have done something, but it was not very important. Then I joined a group of interns who wrote a bit like students of the same level. There was another girl who went there. At that time, she also went to a project that she didn't like. Then one day she told me that she went to Dot.com. Dot.com was the star project of the Innovation Factory at the time, the star project. Yes, yes, she particularly liked Dot.com. I was right to Dot.com at that time because she was a bit dot.com. How should I put it? I felt a strange feeling. I didn’t like it that much. Then she especially liked Diandian. Then one day she applied for her own transfer. Then she let me find out that I could transfer to another project by myself. Then I went to find the person who was in charge of our design team at the factory at that time. Then I asked her if I could start some other projects. Then she looked confused. Then the following is like this. I think the latter story is a coincidence. I actually told you before that we were on that Twitter. It should be that I want to follow you, or I want to follow each other and forget. I must know you because you should be a relatively famous designer in the design circle at the time. Yes, I think from an objective point of view. Then I think no one should object to my statement. Then, I then because you were on Twitter, which was still called Twitter at the time, and posted a selfie. I didn’t know the story. I probably didn’t. I hadn’t seen you post a selfie on Twitter, so although I had never met, I knew what you looked like. And then, because we didn’t have any videos before. During that time, I should have used a real person avatar. I forgot. Anyway, I just saw yours. Don’t say that I posted a selfie. It was so disgusting. I saw your photo on Twitter. Yes, I was in the factory at that time because the factory should not have announced the investment at that time. It was a secret thing. Then it was in innovation at that time. It was not long after I moved there. I had just moved. I had just moved there. At that time, I remember it very clearly. It was in the section of the third phase of the building. Zhong'an Chong was there. The factory should be on the 17th and 18th floors. Then One day at noon or midnight, I was making lunch in a very simple place in the middle. When I was going to make lunch, I came across us. We were right outside the area of the canteen. There should be no canteen. No. There is a canteen, it's just a thing, it's a corridor, it's actually a corridor, right, and then there's a pool, and then I asked you, who are you? You should remember, is that so? How do I remember that there was someone who brought you over and said yes? I started Zhihu when I was studying. I should have started Zhihu at the beginning of 2011. Zhihu might have been launched at the end of 2010. I liked it a long time ago. I used Cora very much. The story may be boring and can be solved later. Then I actually mean that I am already a Zhihu user who attaches great importance to me. And I applied for registration by myself using a Google Sheet. I filled in one. What am I? I filled in one. What am I in HTML5? What expertise, information, etc. Then he invited me in, and then I was so I had more than 3,000 accounts at the time. I should be a registered user with more than 3,000 accounts, so I was already using Zhihu at that time, and then I knew that Zhihu was already in the factory. So I joined Zhihu at that time. I was an intern at the beginning and then when I started school, I felt that entrepreneurship in Zhihu was still relatively early. There was a typical kind of entrepreneurship, very passionate, everything was possible, and the development was very fast. The people there were very, very very. People in the early Zhihu were very interesting, and the users were also very interesting, so I felt that what I was doing was much more interesting than reading. Reading is to learn something from a book, which is as long as reading a paper. For me, this is something I particularly want to do. Then in the schools in the United States, it is actually possible. They are more casual. You can just send an email and you can study. Then I sent an email saying that I want to study. Can I just postpone it and postpone the next school year? Then they agreed. So I was working in the post at that time, that is, becoming a full-timer, that is, continuing to work. It was about 11 years. If I joined the job, I may have joined the job in July or August. I probably forgot the specific details. Then officially, it may be 9. It doesn’t matter. I stayed in the job for almost three years. So this is my first official job. Do you remember the influential things you did at Zhihu at that time or the more important products? I think it may be the most influential thing now, but it doesn’t feel particularly influential when I say it. In fact, it was me who left the Zhihu column. You may not know it. I always thought at the time that you should have already left. Then of course I felt that because there was a more important project in the middle, Zhihu was going to open registration. When the registration was opened, I was the one who set it up. At that time, Brother Chong may have already left. I don’t know what’s going on when he might make an APP. Then I came to design that one. What does the open registration homepage look like? What does the unlogged homepage look like? Because before the revision, the version you designed was very minimalist. That homepage on Earth. Then this design itself. Designing this homepage is not a particularly awesome project. But during the design process, I need to discuss more with Zhou Yuan about what the core value of Zhihu’s product is. Because there may be a sentence or a message to convey to the user. During this process, I will think a lot by myself. What is the core value of Zhihu? And in the end, we may finally have a version. The version of publish data at that time was to share your knowledge, experience and insights with the world. That's it. But this has a greater impact on me. It makes me think about what the core intelligence of Zhihu's product should be. It was possible at that time, around 2013 or 14, and it must be long-term. I think what should the positioning of this product be? Then I finally said that my answer to this matter is not necessarily the same as Zhou Yuan or others. But my own answer is that I think Zhihu is a platform for Gao Zhilang content on the Chinese Internet, and there may not be something like UG CPUC at that time. I just think that we must provide all high-quality content. This is my own answer. If it is high-quality content, if the vision is this, then our products may be. For example, the product is to serve the vision. Zhihu only had Q&A at the beginning. It is obvious that you should not only have Q&A. Maybe Cora may have published something similar to article at that time. I kind of forgot that Cora published a blog very early. Then I felt that Zhihu must have a form that is similar to an article or a format where you can publish content without asking questions. Then I always wanted to promote this thing. Zhou Yuan’s impression at the beginning was that he was not very excited about it. But because of the early days of Zhihu, everyone was actually relatively open and free. You push a project. At some point, you can also push it online, or you can promote it as long as you find enough people. So at that time, I basically said that it was one of the more important projects before I left Zhihu. The last project was one of the last projects. Just make this thing completely and push it online. Then I still remember that after I made this thing, I struggled with the name for a long time because it was called a Zhihu article. This thing was very boring. Then I went to find Ji Xin, the CEO Cofounder of Zhihu at the time, and then I asked him for his advice. It was his name. He didn't think about it at first, but he was very clear about this challenge. I was very excited. He met me one day, and he suddenly said two words to me, which is to turn the column, which is called column. I think the name may not be that perfect, but it may be better than all the names I thought at the time. It has a sense of elitism, that is, before the column, can’t everyone write and post a column? It is in a magazine. The column in the newspaper is a privilege that only a few people have. There is also a name, which will also affect the decision-making of a product, that is, you let users post articles. Is it in a blog? Because I didn’t want to call it a blog at the time, or you can just post it in one place, for example, like the current exit post, but the column may be limited. You have to have a column, or in general users will post like this. But in the end, I didn’t have a good name, so I just used the column. Then we just used the column. In fact, it was just me at the time, plus maybe one or two front paragraphs, plus a back paragraph. I still remember their names. Now there is a former master who is with me. I am now in the company, and then we made this product. I think this is a project that has the greatest impact on me individually, and it still has a certain impact, although I think it may be okay. So your next, more interesting career experience is Xiaohongshu, and it is the early Xiaohongshu. I think my parents want to say it is one. It is also a bit like my parents eating. There are also some coincidences and some factors in it. I was in 2016, it should be 2016. In the summer of 2016, I took my parents to the United States. After playing for a while, I took them to build citizens. Anyway, I took them to the classic route of Highway 1 self-driving, and then detoured to Vegas and went to Yosemite. It was a loop, and it was mainly self-driving. During that experience, I was in the back. When we circled back to San Francisco, we had a mutual friend and he was in Airbnb. Because we met a long time ago, when I passed by Jinshan, I still remember that day. I should have gone to see him in the afternoon, and I also went to watch the Warriors game in the evening. And anyway, because I had a good relationship with him, I was working at Airbnb at the time. I went to him and asked him to visit. Airbnb was in a very fancy office that was supposed to be very famous. It was a garden-like office at that time. Now the bank heard that the lease was canceled. So we were in the Airbnb office on the top floor. We went to the top floor. He had some beautiful places and a table tennis table. I still remember that we played table tennis there. At that time, I was probably a little bit confused about my career. I can’t say that I was confused. But there is one thing I want to consider about my next career, or what kind of things I should do, and what kind of choices I should have. Then he was obviously working in a company like Airbnb, and we also talked about this topic at the time, but it was not particularly developed. I came back to China. It should be less than a week or two after I returned to China. He called me and said, very hurriedly. He said that I was going back to China. I wanted to return to China to join. It was something like this. Then he was very surprised by the news. Then he asked me if I wanted to come with him. Xiaohongshu. In fact, I knew Xiaohongshu's product at the time because it was for women. In the early days, it was actually used by women for shopping. In fact, I was definitely not the target user. And the product design at that time could not be said to be, and there was not much customer appeal, so I didn't pay special attention to this product. Then as soon as I remembered it, I wanted to say that the logo was very long and drawn at the beginning. It was an open book. The color, which was very dark red, was a particularly good color because I had just moved to Shanghai and had been in Beijing before. I had just moved to Shanghai for a while, so Xiaohongshu happened to be in Shanghai. At that time, there was basically nothing worthwhile in Shanghai. The Internet companies I went to Bilibili were also relatively small. Many of them had not yet started. It was not like today. In fact, there were not many Internet companies in Shanghai at that time. So I went to Xiaohongshu to learn about it. After going to Xiaohongshu to chat, I found that the team was actually pretty good. In the early days, Xiaohongshu was relatively elite. In the early days, Xiaohongshu included Mao Wenchao and some early founders or executives or some more important people. In fact, they are all in the United States. They have been to some pretty big companies. Then maybe Xiaohongshu has a special feature. Many of them have a background in consulting companies and come back to work as executives or e-commerce operations. For example, when I first joined Xiaohongshu, I went because I was working on community products. Then I went to sit down with the operations team. Then I found that a very basic girl who was chatting with me was a girl who was doing content operations. She graduated from Stanford GSB, and she was a basic person who did simple content note review. She graduated from Stanford Business School. This is the error. So in the early days of Xiaohongshu, everyone actually had a relatively strong elite atmosphere, and everyone wanted to do awesome things. I really liked the team banner of Xiaohongshu at that time, and I didn’t think too much about it. I thought the company was quite interesting, so I joined Xiaohongshu. I joined Xiaohongshu in October 2016. At that time, Xiaohong The DAO of the book should be relatively good. I can say that it is relatively small anyway. Basically, when you open it, it is a very small, small-scale product. Then I joined Xiaohongshu. Until I left, it was basically four full years. I was doing community products in Xiaohongshu. The so-called community products are actually this APP except the mall tab, because it may be community. If you are small, for example, if the message tab and my tab, there is nothing you can do. If the product is small, In fact, I was working on the content of this product, from its style to discussing with early recommendations and how to distribute this content. Later, maybe around 2017 or 2018, I started to focus on making cameras and making cameras. At that time, Douyin came up, so everything was included in the category of the community. I have done all these things, and then I want to say that I have done all these things. When I left, I should say that it is already a relatively large-scale company. Maybe the more life it has, the more life it has, it has reached hundreds of millions, right? Probably a relatively complete experience of the process of Xiaowangshu from small to medium size. I think that when I was making Xiaowangshu, when I started to make this product, I actually felt that I didn’t know anything, just that I didn’t know what a content platform was. Because at that time, we looked back to 2016, when Douyin may have just started or not yet started. At that time, everyone thought that the largest and most powerful so-called imbalanced products in the market were Weibo and WeChat. But today we all see that TikTok and Xiaowangshu are actually completely Even Kuaishou may not have been that big at that time, including video, so I look back at the experience today. In fact, these platforms are essentially based on UGC and maybe in the future. After this product has developed to a later stage, PGCPUGC has built content platforms or content communities based on the content generated by these users. Then the above personalized recommendation distribution is a very important part. Judging from the results, we have constructed such a product form. So there is no doubt that Xiaowangshu is in China today. I think it is in China. We can think of the platform in China today. There are only four influential ones, or add five at most. Douyin Xiaowangshu Kuaishou Station B, video account, the others are not important, it is Xiaowangshu. I think from today's point of view, it is definitely very valuable. Even depending on different people's views, it may be the most valuable, the most special one. Although it uses a lot of materials, there is no doubt that it is not money, and it is not as good as video accounts and Kuaishou. But the scene and value of its products include it because it has a strong search, including its graphics and text, so it has a strong differentiated value. But when I was doing it, I didn’t know anything, just what any of these things were. And the people who were doing these things with us at that time, including the early community operations at that time, such as Mao Wenchao or Conan, were the management of Xiaowangshu. I think everyone didn’t have a very clear sense of what kind of product we were making. In the middle, everyone would think that we were going to become an Instagram. But if you look at today’s results, it was just a passing cloud. In fact, in the end, the product form that Xiaowangshu constructed This is a very powerful product with a long-term moat based on UGC search and information flow distribution. Of course, if we are in AI today, this is another matter. But when we were doing it, we actually did it in a step-by-step manner without knowing what we were doing. Then I think the most interesting part at the time was that in fact, for example, when we were looking for jobs today, what strategies might everyone divide into C-end products? But at that time, In fact, there is no telling what a product can do or cannot do, so it is just to recommend. I think it is basically my job. In fact, it is to work with the few people who are most compatible with the recommendation team, including the most comprehensive few people from the operation team. I often sit with them, and I am more exclusive. I stay in a room with a product by myself, and only talk to the product. I often communicate with many users, including at that time, to solve the product step by step and encounter some problems, which impressed me deeply. Because when you were confused, you also liked to hang out with engineers, just chatting about something together. Because a designer like me might prefer to spend a lot of time thinking by myself instead of seeking this kind of communication from outside. I think this may also be your characteristics or abilities, because many of the Xiaolong data board products are actually very simple in nature, such as some recommendations. But I think it may be difficult for us to talk about special details today. Essentially, if you don’t understand some of the basics of recommendations, such as how to recall For example, for example, there are steps like gold and rough. It may be difficult to know where it is. For example, some steps will affect the product experience and which will not. And if you don’t know, when engineers are recommended to make these product decisions, they will actually do some, and they will actually do it. It is not necessarily to make the best composite user product experience, or how to talk about your product value data. So I think for someone who makes, for example, you are making this kind of content platform product, you have to have a very good understanding of how your distribution is done, and then give this Your role in judging your product, including operations and recommendations, may be further apart, right? Many times, for example, in the early days, people who worked in content communities or platforms, everyone has more or less the same feeling. Operations feel that their content cannot be distributed. This is a very classic problem. Of course, in this problem, every case deserves to be looked at very carefully, looking at the entire link to see why it cannot be distributed. What are the problems in it that are worthy of being discussed? It is very valuable to apply it to all mechanisms. In fact, collecting cases one by one, debugging, and modifying the possible problems. I think this cumulatively makes the product continuously iterative and better. But because for example, the problem of content distribution, users do not have a strong perception. It is not like the C-side problem of a product. I clicked this button and the button did not work. In fact, distribution itself is a kind of product design, right? Or product experience. Many times if a piece of content that should have traffic has poor traffic, this is a bad product execution. At this time, it is difficult for users to report this problem, so you have to collect it from many aspects, whether it is operations or KOL. Of course, there were no KOLs in Xiaoshu in the early days. They were all people who grew up with Xiaoshu. From their feedback on their own traffic, you can tell that some of them are indeed very valuable. This is because this thing has great experience. Why does it not work? But I think there are a lot of them in the middle, and there is no way to summarize them into a big, detailed matter of a thing. Especially the product itself is a product that has grown from small to large. This company has grown from small to large. It is not built on the premise of a very complete system like the current Byte. So there is a lot of practical, very interesting, but very fragmented, and very diverse work that needs to be done. For example, another very interesting thing is in Xiaohongshu. I actually spent a lot of time on it. To build the AB test platform within Xiaohongshu. In fact, the product itself can be considered. I am the product manager of this product. Of course, there is no such title. It was me and another colleague of mine at Xiaohongshu at the time. It was the first to do data. It is the data of data engineering. We also have some products. We basically worked together because no one said at the time that we wanted to build an AB test platform. At that time in 2018, everyone was top. In fact, we did not have such a strong perception that we must have such a strong thing. Then, if you make recommendations, AB test is very critical. This should not need to be explained. Everyone knows everything about the situation. But if your entire team adopts this data-driven system, you need a good tool. So at that time, I actually spent a long time to make this internal tool to make this internal tool easy to use, including more functions in the future. In fact, more product functions will use this thing platform. This matter has a great sense of accomplishment for me, including myself, as I said in my speech, I am the person who conducts the most experiments. But they may have changed the platform now. I don’t know. We even once did it because I spent too much time. He made a ranking and gave me an award. It wasn’t very important. He just gave me a random one. Then Mao Wenchang took a photo with me. A story like this, just because, but that was very special. It was done by myself. It was not a big one because the platform itself was mainly done by me and another colleague. It had nothing to do with the product, but for me, doing such a thing in a company For me, it gives me a great sense of accomplishment. It just makes me feel that I haven’t done anything before. I haven’t spent any time at Facebook, right. I don’t know what it is to be truly data-driven. I haven’t been there before, and I haven’t been there before the novel. So at that time, it was just based on my own feelings. I felt that I figured this thing out by myself. Then I also learned some of these theories of data analysis. It’s also quite interesting. What is significance and non-significantness? This is a very basic thing. There may not be any step-by-step tutorials. So I started to make it. I think this part is a very interesting part of my countless experiences in retrospect. Yes, I was a designer, a product manager who was a designer. What was your title at the time? Was it a product manager? Or was it a product leader or something? At that time, I thought this was another interesting point. Because I was talking directly before, including when I was in some of my later experiences, including when I was at Dine, I actually did a lot of design, right? Of course we all thought that we also had to do some product work. But in fact, we will do the delivery and execution of the design. Then when we went to Xiaowuwu, I didn’t preset what I had to do, but Xiaowuwu actually had designers and product managers in his company. So when I went for the interview, after the interview, I might go and they were ready to hire me. Then I talked to Deng Chao. Deng Chao is the product leader of Xiaowuwu. He is responsible for the design of the product. Then he asked me whether I wanted to be a product manager or design, and he gave me a choice. I didn’t really want to be a product manager at the time. Maybe I don’t know if this can be broadcast. In the early days, we all hated product managers. They only gave orders but didn’t do anything, right? Then later I thought about the product manager’s right to speak. It seems that the product manager’s right to speak is bigger. It can decide what to do. If you become a designer, it becomes possible that you have to take the demand. So I chose to be a product manager. But I also told Deng Chao at the time that I wanted to do design. The title you enter is the product manager. I actually made the first product, not on the web, not on the app. I made it for the web, and made a book for enjoying. An early function for KIO, a long note, similar to the one in the book for enjoying, an editor for the public account function. When I was making that product, I actually made the design of that product myself. I didn’t want to. It was just a habit. I didn’t want to use other designs. Including me at that time, I also designed an official website for the overseas version of Enjoy the Book. At that time, it was very beautiful but it was not launched online. Then gradually I discovered that when I was doing some very interesting things, I didn’t necessarily have to do the design myself, so it changed to maybe a main component. Then for example, I started to participate in some apps. When doing these things, I probably did it very quickly. I probably did it for a month or two, and then I immediately found a photographer to extract it, etc., and entered into this more traditional workflow. So I want to say that the representative is actually the product manager. And then why did you join ByteDance? The core reason is that Byte at the time, for example, or the experience of making C-end products, was because he made products like TikTok. Of course, TikTok was made before, for example, Musically, which is globalization. What is global is that no company in China has been so successful in globalizing C-end software products, and TikTok is such a product. It may be the same as Xiaoshu, or it may be original. Musically is not called TikTok, but it is also original and does very well. It is a product with many very awesome features. At that time, I thought Alex was mainly Alex. It does not mean that people here may not know how to offend people. It is the management of Alex and Yimin. They are a very good person. I really wanted to learn from them at that time. Then I really feel that they have a lot for me to learn and want to know why this company can be so awesome. This has created such a motivation for me. So for example, it is not for, not simply, not to say OK company, I want to go to better opportunities, but for example, when I went, for example, if I go to other companies, I have no, not even a second thought that I want to go to another company, that is, another big company, because Uncle Xiaolong is also very valuable to me, or an experience that I like very much. I actually don’t like going to a big company. But Byte is a place that I think is very unique, and its products, its brand, and the way it does it are undoubtedly very unique, so I just have a learning mentality. One night I read all the immigration Weibo. He may have deleted it later, but I read all the Weibo one night. I still think it is very interesting, right? Then Alex also did some of his interviews online, which is one of the few interviews he has on the Internet. I mainly work on the Shanghai part of Douyin. Well, then on Douyin, I think if I look at the results, I think I should have done some awesome things. Do I always use the word "awesome" to make some more powerful things? Um, for example, the Xiao Huo Man may be considered one, but I don't think it is the same. First of all, Xiao Huo Man and Sparks should be very successful trial products now. I think according to some people may be very, very successful, but it was indeed me who was in charge of the team at the time. Even my influence on these products should be very, very large. Uh, as for other products, I think I have done a lot. If you look at it from a product perspective, for example, if you carry it on your resume, it is not a good analogy. You can say how much data has been improved and how much scale it has reached. There are many, many products like this, but I think the ones worth mentioning may be Xiaohuoren, Spark, and we have made the scale of Douyin IM to a relatively large scale. What this number does not say is that we actually do it in Douyin anyway. There is still ineffective progress in socializing with acquaintances. I think the results may be like this. Others, I have done a lot of big impact in Byte, which is a very large amount of data. Including later, I started using a lot of AI to do a lot of things in 2032 or 24, including AI special effects. Yes, because I was responsible for AI special effects at that time. I was responsible for special effects. Uh, so I also did a lot of AI things. I think I have some comments on Byte. That is, my expectations for Byte were relatively high at the beginning. It should be said that I think Byte is a very awesome place. And the recommendations in it, uh, include the way they do things. Everyone, uh, the products they make are actually very good, and the data is also great. So, but my overall feeling, I think there are some stages that I have experienced. When I first entered, I felt that there was a very strong sense of freshness. For example, when I was a small clerk, I may feel that the recommendation team I worked with may be a top recommendation engineer. Maybe not more than five, but when I went to Byte, I found that there might be twenty people like this, and I felt like wow, so the talent is really a lot, but gradually, um, as time goes by, I will find Byte itself, the configuration of his team or the way he does things, I will see more problems. But I feel that it is difficult for me today. I still see a lot of problems that are difficult to solve, especially when they get bigger. After getting bigger, getting bigger is a big factor. Then later on, I also made a lot of efforts. I want to make a lot of changes, right? Changes in the directions that I can reach, but overall in the end, I still find that any company is still There is still a very long version of him, and there is also a short version of him. I think that in the later period, it must be in my own. I gradually feel that I can learn less and less from this company. At least I personally may be my problem. Of course, I think that in the past, for example, Alex and I may have had more or less contact with Alex. There must be some cooperation. I think Alex is still very powerful. He is very powerful in terms of his judgment of product benefits, method judgment, etc. But later on some things about the company as a whole, I think in the long run. I don’t want to be a manager myself. This thing is not very attractive to me. It seems to be in charge of a lot of people and very big things. But in fact, you can’t do a lot of innovation in what you do all day long. Then, you are doing some management things and doing some communication things. There are a lot of processes, so this is what I don’t like. I think it is definitely necessary. But this is also liked by some people, and some people are very comfortable with it. But this is the main reason why I don’t want to stay on my own for a long time. I still prefer to do something. Interesting and ground-breaking, try something new Speaking of this, interesting and innovative things, would you like to talk about your new product, one of your new products? In fact, recently we are actually making several products. And we are, as Du Xiao just said, you said this should be, it is a product that we have recently invested more experience in, called Clauli, and we may also change the name. It doesn't matter. The origin of this product is also very interesting. It was when we were in January. At that time, we were making an agent product. The agent product we made at that time was more aimed at productivity and strong productivity. For example, we explored some directions of making a video agent. Then we have made a very complete one. We made a video agent, made a PPT, and also had some more basic general products for research. At that time, we made a product based on the Cloud Agent SDK. During the process, we found that we wanted to add this product. We want to give it some attributes. For example, we want it to have a lower threshold for ordinary people to use it, but at the same time it may have very powerful capabilities. For example, it is more common sense, such as better memory, etc. This is something that everyone often mentions when doing everything now. And then, from the end of October last year, the agent actually has a lot of ecological aspects, and some technical things gradually came out. Of course, we may not have talked about it today. For example, in November and December last year, things like Opus actually had a very big change in Vibeship. Then we If we make such a product at the end of October or January, we definitely want to take advantage of the booming ecosystem, such as many and many skills, right? We don’t have to build our own skills or similar things separately for every scenario. So we found that the Open Cloud framework is actually very satisfying for what we want to do with the previous product halfway through. For example, it has better memory, it is well compatible with the skill ecosystem, and it has many designs, such as basic System Prompt and so on. He is more like an ordinary person, easier to get started with, an agent experience, more like a human being, etc. Such basic things, so at that time, we started as a relatively simple and fast project. We thought that we were using Open Cloud ourselves. I don’t know if you have used Open Cloud, but today Open Cloud, everyone should have listened to you. We started based on ourselves. We installed some Open Cloud ourselves and then included it locally and in the cloud. We also deployed some. We found that there are many, many problems with this thing, right? Everyone knows Deploy it as fast as possible, and then you have to configure the token yourself. If you want to get some capabilities, you have to do a lot of things yourself. So we made a very simple idea at that time, that is, we wanted to make a product that ordinary users can easily use our Cloud. And this is not what Tencent is doing downstairs. Isn't this what Tencent is doing downstairs? Yes, yes, the possible results are compared with the results. I think we originally wanted to release it before the Spring Festival, but we found that it took a lot of time to develop a certain product. But I think today I think this product We have done a better job. In fact, I think if a user really wants to use Open Cloud, I think our product is a better product that can offer a personal experience of Open Cloud. But I think that as Open Cloud makes this product, we have also discovered the direction of Agent. I think some of its challenges are whether it is product positioning or whether it really goes to market or commercialization. Of course, there are many in China, such as installation, one-click installation, etc. This is something that I think is definitely valuable, but it may be relatively short-term. So now we are actually also rethinking some of the positioning of this product. We also have some relatively big things, and we may gradually improve this product. We are now basically in a. We have conducted internal testing for about a month, more than a month, and then there is probably One or two hundred internal beta users are all friends of friends. And the internal beta users are not so many who are mainly engaged in the Internet, but they may all be white-collar workers, which are more white-collar and knowledge-based jobs. Then we found out from the internal beta that ordinary people can actually be very capable. They still need the most advanced agents. There are many capabilities offered by various agents. But there are several huge gaps here. The first one is that they actually don’t know that agents can do these things. This is a A very big problem is that we often do this product because we are still relatively simple and have not done much education and the like. We may have done it. We may have told users that it can be used in this way. Then users know and say OK. It turned out to be so awesome, so today everyone still doesn’t know that such a product form can do such a thing. This is the first one. The second one is of course. I found that this discussion has entered a comparative environment, as if talking to investors. The second one is the agent. If he can exert his maximum ability, he needs to be better. In addition to a good base model, everyone knows that it actually needs to obtain your complete or more accurate context. This problem is relatively simple for programmers, but for many white-collar workers, especially non-programmers, or white-collar workers who work in transactional tasks or mainly workers, their context is basically scattered, or there is no good way to feed it to the agent. Today, many companies and many products are solving this problem, but so far, there should be no company or any product form that has solved this problem well. So this is a big gap. Another big gap. And apart from this, there is no doubt the price, because today’s best model is actually different from slightly better models. When it comes to complex tasks, especially agent-related tasks, the difference is still relatively big. Some models are actually very cost-effective, but they are okay for general tasks. But once we enter complex scenarios, especially ours. In fact, basically it is draining a lot of the agent's capabilities, so if you always use the best model to let users do things, it will be very, very expensive. We have not done a good job in this psychological construction. We can spend so much money. Although in this X or the narrative of VC or the narrative of Silicon Valley, everyone says that you are actually hiring an employee. Why am I here? But now, except for companies in Silicon Valley, most companies do not have such a strong mentality. Just me, I spend thousands of dollars or thousands of RMB a month to hire an agent. Doing these things is actually replacing my employees. Speaking of this, we have an internal project management software called MICE. I have always believed that this form of Kanban requires a lot of manpower to maintain, or linear. First, you have to create a task. Of course, you can use some methods to automatically create tasks. But for a design team like ours, many tasks still need to be written and created manually. Although we have done some simplified work, for example, you can click on a conversation in Slack. It will automatically extract the content and help you create a task. But you still have to create a task. I have been thinking about the future of work or the form of future project management software. I made an attempt. That is to say, I was actually inspired by the notion of agent. They have been pushing that function some time ago. Then I went to use it. After about a day, I explored and figured out how it works. After so many years of development, it has a strong ecosystem of connectors. The principle is that there is a very, very detailed system prompt to control this. These connectors are constantly pushing or grabbing the context, and then using a loop or trigger to do something. In fact, when I first saw it, it turned out to be such a simple thing. I was a little surprised, but after using it for a day, I found that the effect is very, very good. It can be implemented in slack. To add it, you can choose to add it or not to add it. He will read all the conversations you send in Slack and will be punished once. This agent is wrong, so I just moved all the tasks on our message in one day and asked him to serve as an office assistant to help us manage these tasks. At the end of the day, it cost about 3000 credits. I forgot that it is really valuable because there is no formal fee for it. I calculated it and it should be about 50 US dollars. I haven’t officially used it for one day. I just set up this workflow and it already cost 50 US dollars. So I felt that this thing was definitely not sustainable, so although it was very useful, I stopped it after using it for a day. Then I tried to use the same SLU in mess to make a similar product. We called it sure, which is called the sous chef. The second chef is the chef. All of us are in the catering category. I also added some connectors to him because notion. Although the connector is very powerful, there are only so many commonly used ones, about five or six. For example, we put slack, figma dropbox github, the most important github, and then use an agent loop to passively accept or actively capture some context in the project. For example, he can analyze who has touched this file, who has opened it, who has saved it, saved several versions, and what comments were posted. Then combined with the tasks in our task board, he can roughly guess the current progress of the task, so this task board has now become a semi-automatic form. I think this It makes us happy to use your quality. What is the agent loop that we wrote ourselves? What is the use of the open clause? What is not useful? We use the convex workflow. Because it is a very long workflow, we use a tool that can be persistently orchestrated so that it can run for a long time. Now, I think this is what I think it would be interesting to talk about this. But our current company has about 9 formal employees. Then we only use linear now. We also use heavy and linear, but I, including everyone, should open linear very few times. We basically use cloud code and sometimes use clody, which is similar. For the principle, it is to manipulate linear and include all. For example, when you just said it, we thought of some specific details. For example, the dashboard of all our data. For example, we use some log services. There are some very useful ones, such as axiom and post hoc services. We actually use them. Clarly's agent then includes Clarco, because it does not have scheduled tasks and does not do well, so we often use Clarly's agent to regularly look at these boards. And I don't know what he is looking at. Anyway, he may be looking at the board. It may be looking at specific data, or it may be looking at this reple, this code reple, and yesterday's submission. For example, I have an agent. This agent will also do it. In fact, every day, he will review the latest new submissions every three hours and then every 24 hours. I will announce the new PRs of the past day, the merging PRs, and any new features. Then he may have about four or five time points in a day, and broadcast them once at each time point. We may have six or seven major post hoc dashboards, and broadcast this dashboard once, and then the recent trends. Then he will find some things, and anyway, very, very, all are timed characters and driven. Then one of the things we are doing now is to drive him to take the next step. Now we may basically find the first step of the problem, including review. But maybe the next step what we are doing now is to allow him to do this. This means that we are not a form product. That is to say, we are using qually in Heavily's work building so that he can do more, such as fixing a problem by himself. But we actually have many companies that may say they are doing it. I am not sure, but we are not actually at this step yet. We are just problem discovery. Basically, we don't look at any of this kind of Sanmao software. We all use the agent method to talk about each other's working methods. I think especially for small teams. In fact, it’s quite interesting. It’s just about using the power of AI to accomplish more things with fewer people. This is what you mentioned before we started recording, because from my understanding, you have also made a lot of products, although some may not be disclosed yet, but you have made so many products based on the scale of nine people, and the time is also very short. I think this is quite powerful. Yesterday, I talked with a friend of mine about Tai’s entrepreneurship, and he asked us what kind of products we made. Then I said that we have made so many products, and he said wow. He asked if you were all-staff programming. I had never thought of this word before, and I felt like I didn’t know if it was a buzzword. I asked him if it was a fashionable buzzword or something like that. But in fact, we are indeed full-staff programming, and what we are now, and we are not touting this thing. I just describe our way and what we are now. For a new employee who wants to join our team, my expectation for him is that he will be agentic or coding-oriented and then manage various things in an agentic way. As a supplement, we intervene in the work in this way. So now, in fact, it will be a bit like the question you just asked, which is how we allocate it now. In fact, first of all, in terms of how everyone works, all of us actually need so-called programming. But in fact, we are using various agents to do various woollish coding, analysis and other tasks. Then in addition to others, it is us. We will have some loose bodies for each project. There may be a person in charge, and then one or two persons in charge. Then there are some members who may be involved in multiple projects together, but roughly these 9 people may be distinguished. We may have three main projects now, and the number of people involved in each project may be about a dozen. But the number of people involved in each project alone may not exceed four or five people at most, which is roughly the same form. But I think this is actually a very interesting one, but maybe now because of everything Things are changing too fast, so there is no way to discuss this topic in an organized way, which is how an organization should evolve in the AI era. I think the problem now is that everyone is trying different ways, and many bragging people think their own way is the best. And many people may still use very ancient ways, etc. I thought of an entry point, that is, they can't do anything now. We exist now. What is the value of our existence before the emergence of UBI? As a design company, what we deliver is no longer design. Unless it's something like marketing, the projects we deliver are actually already code, and design MD. I think the value we have now as a design company means that there is such an information gap before AI completely masters the design. This time difference or information gap also has some information gaps, that is, design MD has really become like I don't know what is more popular now like scale, right? In fact, a design company seems to have no value anymore, right? My feeling is to respond to your question just now, that is, who is there? Is there anything that AI can't do well? I was thinking about it when you were talking about it. And I think you said a lot of it. I basically agree with it. One of my big feelings is that the more you use it, the more you will find that there are actually many things that AI can't do, or it is very difficult to do well. It is a relatively contradictory experience. But in fact, this is how I feel. When I do some things, such as coding or writing and reading information, I will find that the efficiency is greatly improved. This is what you do, and you will not use it in the future. You're right, but in some things, the more you use it, you will still find that there are some things that it just can't do well. Then you have to, or don't have to, and you will realize that you really should invest more energy in the things that it can't do well, and think about how it can do it well. It can give some basic good things, but you can’t really make a good design. Then for the product, you want to make a product. It may include the design, including the product itself. Its definition, differentiation, entry into the crowd, and what is the positioning? This is not done well at all, right, but I feel that many things cannot be done well, so the more I use it, the more I find that it does not do many things well, so there is a kind of contrast or something, that is, many people say AI doom or something. It emphasizes a lot of work. I think it is true in the long-term trend. But in fact, what AI can do today I think is still relatively narrow, but it may be a deep feeling that people are good and people are irreplaceable. Of course, I think the opposite is true. If you don’t have AI, then today’s people Maybe if there is no AI, indeed he is very good. Except for a few, except for some special professions, I think his profession does not need AI. But most people, I think, should use AI. Because in the process of doing this, we may have some comparisons and comparisons with some technical aspects of Mondane. We actually spent a lot of energy because Clawley is a very complete one. What we actually do is very similar. Later, we found out that it is very similar to the perplexity and the personal computer, but we don’t have it. We don’t have such a market yet. That is, we are a very complete agent. He has his own computer. The configuration of the computer is actually not low. It is a cloud computer and it can also control your mobile phone. It can also control various things on your computer. In the process of doing this, we discovered that there are some fashionable sayings that you have to make products for agents, right? But in fact, the problem we encountered is that the agent's infra is indeed a very barren one. Basically, today, one by one, for example, the field is still, but this matter, I think it may be a little bit into a special and more segmented field. But our experience is that if you want to make an agent today, it will not be built on the shoulders of some giants, it may be built on the shoulders of some giants. Only the model is a giant, but everything else has not yet been built. All other necessary foundations and implementations are still in the process of development. Then the agent has his hands and feet tied by many things, including data that users cannot use, including the lack of good tools. Of course, recently everyone has done a lot of cli and other such things, right, so that he can have a lot of access. But until today, in fact, we all know that the Chinese Internet or whether it is the global Internet, the agent is inaccessible to many things. So, I To sum up, we made a very full-featured agent with a complete execution environment for quality control, and then we discovered that the infrastructure work we had to make up, or the infrastructure work we had to do, was very huge. And one thing here is that even if we are halfway through, we want to build these infrastructures. Of course, this is not necessarily what our team wants to do, etc. This does not seem to be an interesting thing to say. It is just a business perspective. I think we have discovered a point, and I think we can mention another point. This may not be so interesting, but I think it is very important. For example, we are actually doing Clark, so we are mainly oriented to users. When we do gold market in the future, we will mainly target overseas, including this product. We are also an English product, etc. But during our internal testing, we found a lot of domestic friends and friends, and because of these opportunities, we came into contact with some domestic, some common ones like Shanghai, some doing advertising, and doing this kind of mainstream companies. These small companies can't set up that kind of thing. Then I will follow Because they are very coincidental, because their friends or other friends happened to use our products, and some company bosses thought this was very interesting and started using it. Then I talked to some companies and they were using it. They didn't even have a clear concept of what agent is. Their AI may not be very useful. Just understand their needs and their daily scenarios. One of my basic findings is that many of today's small and medium-sized companies, non-interconnected, and even larger companies, they can get very great value from agents. I'm absolutely 1000% sure about this, but they have no idea how to start today, and how they should go about doing this. Even Tencent can't do this. Isn't it integrated in WeChat? No. Just today. Just today. For example, lobster is so popular, right? There is no doubt that my mother deserves lobster. Then that one. I went to visit the boss of a company the day before yesterday. The boss of a small company is not an Internet company. He sat down and asked me. The first sentence he said was, I don't know how to pronounce that. Now that everyone should be Hermes. The summary reading is Hermes. I don’t know. It was the first sentence he asked me. Do you know that Hermes is the agent? But he, for example, is definitely not a person who does not understand technology at all, but he is definitely a very ordinary person who is not in the Internet circle. That’s why he has become so popular. But in fact, it is because of his popularity in the market, in self-media or on social media. There is a huge gap with everyone’s actual adoption at work. Yes, I can only say that, definitely Tencent. Some of Tencent’s products are doing well. In other words, Doubao may also be doing well, but in fact we think that the agent can really bring out the greatest value. For example, some of the superficial use of agents in our work that you and I just described can actually do a lot of huge improvements in efficiency. If we talk about many of the current agent products, in fact, the cloud co-work is actually a product direction that I think is more suitable. Or in my opinion, it is a product direction that I think is most suitable for truly bringing agents into the work of the general public, because it is actually equivalent to By using the user's local, you can do more diverse manipulations of various document-based and transaction-based tasks. Computer use here will definitely strengthen its capabilities. But in fact, co-work is a place in addition to code, that is, its slogan is literally claw for anything besides coding. Like I went to that one yesterday, I visited that company where I use our internal testing products. He showed me a case that impressed me very much. It was one of my users. My user was a person from that company. Then he talked to his boss. And others at the same time, he was very proud to show that he used Claw for example to make a form, which is still a very complicated form. He said that he made it in ten minutes, but usually he might have to do a like this, because Claw for example can do almost the same thing as co-work. But I, for example, as a developer of Claw for example, I don’t have heavy use, and I don’t need to make a form, but because in fact, we have made these basic capabilities. In fact, these basic models have this ability, which is the Sing tool given to him, so he can do this. And this is actually, I think, maybe back to the point just now, that is, I think that today, everyone's demand is actually very strong. But this change is because it is too huge, and everyone's boundaries for its capabilities are very vague, and the product form, everyone is still there. Many products, most of the product experience is very poor, so I think today is still in the very early stages of adoption. So I think today, if any company that is doing agent, I think there is a very big opportunity, but there are different ways of doing it, you can have a very The product definition of Touchtap, or something like that, of course we didn’t say that we must do work. I just said that we actually didn’t want to make a work-oriented product at the beginning. We wanted to make a more personal product that ordinary people can use. This is the reason why we made an app in the first place. But today, if we directly Take a look at the scenes that are in strong demand and must be work-related. Although our internal users actually find that they use a lot of them. They help them manage their fitness, tell their fortunes, and have a superior user to help them fight landlords and count cards. He is essentially a user. He is essentially a smarter friend or AI. But these possible work scenarios are definitely still the most valuable today. So your final product form is very similar to Poke. I actually like the form of the Poke product. Then it is actually open cloud, but it was released in September last year, and he met his challenge today. I think his challenge is definitely the same. If you are not focusing on this kind of productivity scenario, because it is difficult for me to pay, this commercial aspect is very related to what I just said. If you can have a free product, it doesn’t matter. A very popular product called Mobby allows ordinary people to make some Apps, and then he also made an App that other people use to make money. It has a bit strange function like this, but it also encounters some problems with you. And today I feel a bit like I read one that day. I read something that day, but this one. I should have read some early books on computers and the history of computer development. It was a bit confusing. I suddenly remembered which one it was, but when I read that book, I felt that it was very similar to the era we are experiencing today. In the beginning, many things were focused on productivity. Especially last week, I watched a video of Steve Jobs, probably in the 1990s, when he was at Next. He was introducing the products he had at that time, and he drew some people they were going to target on his whiteboard, such as lawyers, doctors, and so on. Basically, you were selling a SaaS, or there may not have been a concept of SaaS at the time. You basically sold a product that was very productive. But later on, in fact, the entire computer Everyone knows that computers, TikTok, and mobile phones have become very, very entertaining. I think that they may all go through such a process. In fact, today, you and I may be dealing with AI and don’t know what to say. How to put it, people who understand this kind of related dynamics better will know that the capabilities of AI are already very strong, but what ordinary people are exposed to or experienced is actually very large. With this, there is a big gap with the ceiling, and then maybe the demand for high productivity will be used first, and then slowly return to entertainment in the long term. There will be such a process. When I talk about this, it seems a bit like a mentor. That's not what I mean. I just want to have some opinions. I suddenly thought that I actually thought of that. You should have seen it. Before that, it should be last week. Kapathy Andra Kipathy posted a tweet, and then Andra Kipathy had a little bit. Now it feels like this. How should I put it? I made an inappropriate analogy the other day. I said he was Confucius in the AI world or something like that. This is not it. It sounds a bit disrespectful to Confucius. Can it be cut? But what I mean is that he actually said a lot of very influential views and concepts in the AI era. Starting last year, for example, the word Vibe coding is Coin, right. Then later, he said, he said, in that speech at YC, he said this is not Agent, this is not The Year of Agents, this is The Decade of Agents. He said some very powerful concepts and then, at the end of December last year, the famous Obers During this period, everyone realized the magic of Vibe coding, and then he also put this. It should be him. I remember it was another person who said Obers 4.6 under his post. At that time, 4.5 was still Clocko. He compared it to the Gutenberg printing press. I can’t remember clearly. I remember it was another person who said it and then he forwarded it. But I am a little bit wrong. I mean, Andrea Capati from Shang Shang Zhong posted a Tweet. In fact, if you want to find it, you can look it up. He posted a Tweet, and he discovered that in fact, the AI that most ordinary people use and have the most contact with is actually TraderBT, the most common model, and especially the Voice mode of TraderBT, which is a more legacy model. Then, they actually have a lot of problems and even make fun of AI. This matter has become a mean thing on TikTok or Instagram, and people will make fun of AI. Why is it so stupid? But in fact, the real cutting edge of AI is 4.6, 4.7 or something like this. This new thing is actually very, very capable of doing very powerful things. He is describing this very huge gap. It is this body feeling that is very similar to my body feeling. So I feel that no matter what product or model you make today, I think it is all the same. You are standing at the starting point of this journey. If you look at it from a utilitarian point of view, you have a lot of opportunities and can do a lot of great things or eventually. You have a large-scale product. You are not particularly active on social media now. So if anyone wants to follow you, where can they go to find you? Or some of your products, official website, etc., which can be shared later and shared in this shownotes. First of all, I am also looking for one. How about myself? Let me talk about the state of things. So I am in the process of my own transition. If I want to pay attention now, I think it is okay. Mainly pay attention to me first. I am on axe and geek, mainly on these two platforms. In fact, today's episode is our first podcast. The previous episode with Wang Xiaoguang. The episode with Wang Xiaoguang is actually an audio interview transcript. I just finished writing the article and felt that it was a pity to throw it away, so I released it. At the beginning, because CBVV is a very professional anchor. When he started our video conference, he asked me if I wanted to record like what is called N2N, which is split-track recording. I said no, it should be Famous last word, and then we started. It turned out that both of our audio tracks were one track, and they were all mixed together. And when cutting, although it can be cut, there will still be some. I also prefer to steal people to talk like this, so there are many places that are broken. Thank you everyone for your testimony. Someone has said it in the comments, so we also spent a lot of money to purchase microphones. Okay. This is the program for this episode. Thank you Tianlu. Then you can follow Tianlu on social media and the future product releases of their company. Okay, thank you everyone. I am very honored to be able to participate in the first episode of the podcast. Thank you everyone. Bye, bye.

欢迎收听Dine Podcast,我是杜肖,这是我们的第三期节目 Dine是一家北京的设计工作室,自2014年成立以来 我们为数百家企业和创业公司提供设计服务 你可能用过我们设计的产品 如今我们大量的应用AI帮助我们完成设计工作 这档节目没有固定主题 但目前关注的是普通人在AI时代创造的故事 本期的嘉宾是薛天禄,天禄是我在直呼的同事 也是我招募的第一个实习生 他后来放弃了在印第安纳大学深造的机会 成为了直呼的正式员工 天禄后来也曾短暂加入Dine作为设计师 后来他加入了早期的小红书 负责撤区产品,以及后来的子节跳动 25年,他在红杉当过一段时间EIR 目前是Second Enlightenment的Whole Founder和CEO 他们的最新产品Clawly 是一款为普通人打造的 运行在云端的OpenClaw客户端 你只需要下载一个APP 就可以让你的龙虾帮你做事 不需要在本地部署环境 我们从天路的Webcode经历开始聊起 所以你是什么时候开始接触Webcode的? 我大概在,我应该是在二三年还是二四年的时候 我看到了Cursor这款产品 然后当时看到的时候 就它就是那个slogan写的非常明确嘛 就是什么AI coding之类的 但是没有get到它整个是什么意思 但是我觉得这款产品很酷 然后我后面的时候 我在极客上发了一个帖子 大概就是在说应该是二四年的帖子 就在未来就是人们会怀念当初人们还在写代码的时候 类似于这么一个装逼的描述 然后当我后面看到这篇帖子的时候 我觉得我怎么这么有预见性 但是我当时在看到Cursor的时候 没有怎么去试Cursor这个产品 然后我也没有深度的用AI coding或Webcode 但是还没有Webcode这个词 然后下一次接触到应该是 我现在有点具体的但我要去找可以找到 就大概是在二四年底或者二五年初 还是自己的时候 我猜应该是某个大模型有更新 不是它是那个Replet出了 哦那个我也用了 对哦它应该我想起来它是一个 它当时还做了一个广告说 所有人可以在 它应该我不知道是V级 因为它Replet的每个版本的CodingAgent有个V V级VRV3之类的 然后当时是我第一次我自己开始用了 就是开始就试着搞一些东西 但是但是我是用它Web的版本 然后就写了两个东西 写了应该是两三个东西 当时那两天因为在字节的工作 Routine就是因为如果在我负责 就我那个这什么位置或者说那个事情 你基本上一天可能要开十个小时的会 类似这种感觉 但当时我记得我有大概一天半的时间 我就没什么或者说 有些会我就没怎么采见或怎么样 然后我就在一直Coding这个事情 这个应该不会有什么道德委员会 然后我就Coding了好几个事情 然后就是一些简单的 比如说什么曼联比赛的通知 然后什么Coding了一个网站 什么监控特斯拉股票 就一些非常简单的事情 然后当时觉得非常有逼 Reple Agent刚发布的时候 是第一个相当于把全站给集成起来的 就相当于你可以有数据库 你可以有前端 你可以有后端 我记得他当时写的还是Python好像是 我不知道写的什么代码 但反而就是不用setup任何东西 对 对因为我记得我把代码拉下来之后 我还研究了好久 怎么让在本地把这个Python东西跑起来 因为我之前接触的都是可能GS的东西 对我就没有没有拉下来 因为你说这个让我想到了 就是那时候更早之前那个cursor 我其实下载了cursor 然后但是呢 因为cursor就是要setup一个环境 对吧 因为我之前很久很久之前 我也写过一些非常简单的代码 但是需要setup环境就很麻烦 所以我就放弃了我突然想起来这个事情 所以replit就是因为它 你在web里面可以一站式的完成所有东西 所以我当时开始靠这个事情 然后两年后的今天 replit已经声称可以让你的iOS app一键上架了 是然后后面 后面我应该再也 再也没用过replit 在那之后 我应该是付了两个月的费 对我也买了 我也买了 刚才说你第一个做的都是一些个人兴趣相关的东西 对 没有任何跟工作有关的吗 我印象非常清楚 就是第一个叫做慢点比赛的结果 发布许类似这个 但这个好像没有实现 第二个就是 因为我当时可能主要就买了几只股票 Tesla Meta 什么之类的 然后我就想做一个网页 就是把这四个股票的最新的信息放在一个dashboard上 我可以分享一下我做的第一个东西 我做那个app非常无良 但是我想用它打通一个链路 就是说webhook API 数据库 然后就是这种long running task 但我当时做的一件事情是监测大家 在slack的工作时间 因为slack不是有一个在线和不在线状态吗 对 我就让那个app去隔一段时间去记录一下 这个人的在线状态 然后变成一个表格 你这个很牛逼 你这个现在都可以拿出去卖的一个产品 SaaS产品 但这个我记得当时的模型是sonit3.5 我现在有点乱说了 但是非常早 或者3.7 那你怎么拿图片呢 怎么监控的在线 你要截图吗 还是有这个 这个是有API的呀 就是它是有API的 对对对 所以我当时就用它跑通了一个web app 然后打通我们内部办公的这么一个小东西嘛 当时那个模型就提醒我说 警告我说 你要取得当事人的同意啊 就是我说 那已经是蛮牛逼的模型了 然后我就做了这个东西 跑起来了 但是跑得非常艰难 因为是python写的嘛 我也不是很懂 怎么去改一些东西 布局啊 包括一些交货什么的 但是这个是我的第一个replet 我还给它绑了个域名 我还给它付了费 然后就跑了一段时间 OK 那其实还是非常完整 完整的一个sass的概念 是 后面的话就是 就是我从自己的离职之后 大概是25年的5月份左右吧 或者说那段时间前后 其实就是可以开始干一些自己的事情了 然后那段时间 其实就是在看各种创业公司 X上的各种新的想法 或者大家产品长什么样 然后自己开始做一些想法 没有 也不是很功利的时候 我要做这个产品 就是说 开始可以有些时间去experiment一些东西 所以那时候开始更多的 然后那时候开始用cursor 做一些事情 首先我是一个非常喜欢历史的人 我很喜欢历史 我中国历史 外国历史都很喜欢 然后 然后我之前比较熟悉的 其实是中国历史 比如小时候看很多什么史书 然后后面有一段时间 就是我觉得我的历史要更精进一些 这个历史这个东西 这个话题可能跟vibe code也没什么关系 但是历史这个东西 其实是有一些大的subject 比如说西方人很喜欢那个历史话题 就是罗马 Roman empire 就是有一些很搞笑的 就比如说the x上 应该你也知道 就是什么 什么一个中年男子 什么 他的兴趣会 会坍缩成要不就是讨论烤肉 要不讨论罗马帝国 像那个什么Elon Musk那样的人 他们肯定就比较偏右派的人 他们都会比较喜欢这个话题 所以我有段时间去研究罗马帝国 然后另外一个很 真的 然后另外一个 另外一个很有代表性的历史话题 就是日本的战国 然后这个是一个 我之前其实非常不了解的 小时候我玩非常多 三国制的 然后呢 光荣这个公司 他做一般会做一款 那个日本战国的题材 然后再做一款三国的就交替 然后通常日本同代的游戏 会比三国的那个版游戏 好像更先进一点 类似玩法上 然后但我小时候就 玩不进去日本的东西 又不懂里面的这个文化 所以我就有一段时间 去研究了日本战国 就花非常多时间去研究 然后在研究的过程中呢 我就发现了一个 一个很有意思的话题 就是日本的 有个叫100名城的东西 就是日本整个全国这个地方 他又散布着日本全国的100个 通常是建立在 1400到1660 就是战国之前的 两三百年期间的城 然后 然后就是有一种 有个活动 当时主要是以游客为主 他们经常 他们很喜欢去日本玩的一些人 他们会去打卡 100名城 然后我就发现这个事情 其实我是在更早之前 可能在三四年前 我就知道这个事情 然后我知道这个事情的时候 我就觉得 这个是一个很好的 去了解历史的一个东西 一个机会 但是我一直之前 没有没有时间 没有太多时间去控 去去看这些东西 然后在去年开始 比如说我从自己 逐渐要理智的时候 我就开始有比较多时间去 真正的去 visit 做一些计划 去看这些名城 然后就加 这也是另外一种收集批 只不过这收集批的成本会更高一些 因为你真的要去这个城 并且 这要说起来 门道很深了 他这个你收集得有个象征 对吧 他有几种做法 就是一种做法是 你去那里盖个章 他那个城下面有个章 然后另外一种做法 就是你去买那个城 他有个叫玉城印的东西 知道的听众 应该知道我在做什么 就发现某一种 就每个城都会出一个玉城印 然后在都不一样 都是不同的生产商 然后我就去 我就开始去做这个事情了 我大概从去年的上半年开始了 就春节开始 然后就开始有目的的 我一次去去好几个城 类似之后 然后在做这个过程中了 我就发现 就满足 这个东西就是存在一个 我自己需求的空缺 我就想有一个很好的方式去track 我去了那些城 然后看到这个城地图上面有个地方 甚至我可能可以满足一些简单的成就的感觉 然后包括这个城 一些简单的一些玉城印在哪买 就是非常具体的 玉城印在哪买 交通 然后甚至他几点关门 类似这种繁杂的细节信息 然后它如果是一个app的情细就最好 对吧 然后我就发现app store上有一些这么几个app 有三四个单独做巨蓝 就是那种可能iOS7时代 我不知道就是100年没有更新的 然后所以我就发现我可以做这么一个app 我就用cursor开始 就把这个app做出来 你最终这100年程都访问了吗 我现在51个 我现在在app上来的track 就是我访问了51个 我的app还做了一个非常精美的 这个有个印章收集的功能 因为你之前也是设计师嘛 最初的也是 所以你这个app里面的这种物料或者素材 是怎么来的 首先我这个app没有什么 一开始没有最完整的规划 我是想到哪做到哪 我第一个idea是说 就是我的第一个pront literally就是告诉cursor说 给我一个tiktokstyle的一个上下滑的一个 对 就是因为我直接想到 我这个产品就是长这样 对吧 然后马上这个页面做完之后 第二个页面就是把一个map做出来 然后到这里都还没有剩下任何素材 然后我回答你素材的问题 素材的问题是 有两种素材 一种素材是 因为我做了那个印章 就是现在有个batch收集的系统 你去过就可以打卡 然后把这个batch收集了 这也是个付费功能 你超过收集超过超三个要买pro 类似这种 然后就是这个batch功能是后面想到的 然后后面想到说可以加一个这东西 然后同时也可以做一个付费 因为不想这个app一开始付费就很烦 所以然后想着这个idea 其实就是之前可能跟当时有一段时间 那个应该是在X 当时有可能是nadovnr第一版 一一可能刚出 还是之类的 当时就很多人做一些3D的小东西 然后就是模型可以生成一个3D的小玩意 然后觉得这个style很好 当时可能就看到这个style 然后觉得说 那我做一个印章的话 我就可以用这种方式去生成 所以我就基本上是用 当时是用 所以现在想来那时候 可能没有nadovnr pro 当时用的是O3 就是我机座模型用是O3 然后tripe图片模型用的是 tripe里面某个图片模型 然后当时就让它去生成 我所有的这种3D的小的东西 然后就非常manually的 当时也没有去写一个 比如说调用API的显示 因为我还是用tripe的app 或者web去复制prompt 去让它不断的更新 不断的生成 然后中间可能要改之类的 然后中间一些最烦然的地方 就是你去做图片处理 比如说你生成出来的东西 你要透明的背景 或者说那个背景 稍微有点改动 就非常 这个你应该知道 就是这个 到现在这个应该也没有 特别好的工具 今天我们聊的时候 正好是GPT2 GPT image2 我还没有试 我们的一个产品 你接了这个东西 然后所以当时我就用 GPT做这个事情 然后非常容易 hit到它的limit 我当时是20刀的一个会员 然后为了这个 我就升级到200刀 所以那个印章那个部分 完全就是靠GPT生成的 用它的图像模型来生成 100个印章吗 100个印章 然后这里要插入一个点 就是日本除了那个100名称之外 它还有个续100名称 就它那些人不知满足 就搞了100个 它后面还再搞100个 我一开始本来想把续100 也弄上去 后来就是太烦了 就是太多东西了 所以我就把续100砍掉了 因为我在去一些城的过程中 发现去100名城 很多都是废墟 都没有 都就是什么居民楼在那里 其实后来我觉得 就是我给了自己一个借口 就也没什么好 那个什么 对你这个APP 还是非常日系文化中的 一种收集打卡的 尤其是盖章这种 对 玩法的一种数字化 而且这个我要讲到 这个很搞笑就是 我大概在两个月前 因为我在APP内有一个 那个support的一个邮箱 我应该历史上没有收到过 这个APP的support邮件 因为没怎么推广这个APP 应该没多少下载 我唯一收到一个 有一个support的邮件过来 就有个用户 他问什么时候 把续100名的支持 是日本人吗 他应该是中文 中文文的 但我忘了是凡底字 还简离字 回到这边的素材生成 素材生成 这个是一个用Tradby生成长的 然后那个style 就我做一些控制 对吧 我就给他一些prom 让他去3D什么什么之类的 可爱方什么之类的 然后那个 然后这里面有如果要谈到更细节的话 因为因为因为 因为他每个城 如果听众 或者说去试一下我的APP的话 因为每个城会有一些细节的一些 一些元素 不一样的 每个城都不一样 所以我其实做法是 我把每个城的一些 实际的图以及历史告诉他 然后他就会去生成这个 他有时候生成出来挺牛逼的 因为每个城 就是每个城确实是有些 文化 或者有个将军在那里是不一样的 所以应该是 不是所有城都都有这个待遇 因为有些城市在这边什么特点 可能大概6-70%的城会有一些自己的东西 一些很有名的城 比如大满城 记录城 Obviously就是会有一些很多的 就是记录城会有樱花等等 然后另外一个素材就是 因为我是一个 我那个APP的首页 是一个每个城的卡片 它会上一个有个城堡的图片 然后下面会有承报的一些内容 我刚才想问的就是这些图片 你是用的照片吗 还是生成的 对我就没有用照片 因为照片有一个版权的问题 但也不就是有个洁癖的问题 就是我就 规避了版权的问题 我就相当于用AI重汇 就是基本上给它喂很多素材 告诉它这个城的历史 让它尽量生成拟针风格 然后在那个时候 其实拟针风格没有那么容易 因为今天拟针是很拟针了 但在大概在可能将近一年前 那个时候拟针就还要很多的prong 所以有点费力 所以你真花了一些时间 然后 当你现在看的话 应该还是能看到 它是AI上的痕迹那图片 如果新鲜一点的话 然后那个素材就完全让AI写 然后就去收集 然后它可能会点一些 加上一些link 跳到这个城的一些官网 微级百科 所有内容大概就是写 也没有搞一些个CMS 很简单 就是100个 它就hard code写在代码里面 然后这个app就是大概可能 基础的功能写出来 其实就花了 我觉得应该花了一两周 比如说现在看那个cursor的 那个lock应该都可以看出来 就是非常快写出来 但是后面就是调一些细节 调了非常久 就是类似那种 比如说让这个卡片上下滑 怎么粘手 就是这种非常细节的问题 当时就是就现在的AI 应该我觉得也没有 很好的解决 没有百分之百完美的解决 然后当时的AI就更惨一些 就是花非常多时间去解决 一些非常细节的设计层面 交互层面那些问题 一百名城到你后面 正式创业 做公司 做产品 找人 这中间 发生了什么 是什么促成了你这个 创业的想法 首先我从自己的理智 我就是想要创业 就是说我想稍微有一些时间自己探索一下 包括方向上团队上 然后休息一小段时间 所以没有马上开始 我之前没去过意大利 我特别想去 这个跟一个历史有关 对吧 我特别想去意大利 我想去意大利 就是我觉得旅游对我来说 我并不是一定要去旅游 但是我觉得有些地方 我是很想去一次 然后之后我就不用再去 无所谓了 但意大利我之前已经没有去过 然后我非常喜欢罗马的历史 当然如果要有西伯克 跟我聊罗马历史 我是非常欢迎 但我就大概有去 大概在七八月份的时候 去意大利去了 应该有三周的时间吧 三周四周的时间 然后基本上把意大利能去的地方都去完了 大概中间这么短时间 然后回来之后 我就去我朋友一个公司 就我们应该都共同认识的朋友 去体验了一下 重新回到一种上班 或者工作的感觉是什么 具体做的事情 其实就是在build的一个产品 然后我就真正的在一个 OK 这个我觉得值得讲 是我那段经历是 真正在一个团队里面 把vibe coding 或者AI coding用进去 在这之前 比如说字节 小红书 很显然我没有用任何的AI coding 然后在跟朋友公司 那个顾问那段时间呢 就是真正的在一个产品里面 面前用户产品里面 我虽然职责是一个产品 但我实际上做的事情是 我会AI coding 做我所有的东西 然后从前段后段 然后一整条连路去做 然后这个事情可能做了大概 在那边一两个月 然后找到一些感觉 当时另外一个原因是可能自己 因为休息一段时间 想回到一些这种 就是routine的状态 所以就是当时我朋友 正好他们也确认 然后就去做了一段时间 但这个经历之后 到去年大概9月份左右 我就比较明确的 说我要开始 我要开始做自己的事情了 我要开始去 组建团队 然后开始在自己 看好的方向开始 就投入去做产品 其实这里我觉得可以插一段 就你之前的工作经历 比如说从什么时候 毕业或知乎开始 对我觉得你从知乎开始讲吧 我是11年的时候 那时候我还在美国读研究商 然后那是一个两年的program 是从10年 supposed to到12年 然后11年的时候 就算是两个学年之间的放假 然后我就当时回国试习 然后当时回国试去创新工厂 试习 就是我投递的是创新工厂 然后我是一个产品 就是以一个交互设计师的身份 在创新工厂 就是做事实上 然后他的机制是会把你安排到 创新工厂当时投资的一些项目里面 然后当时我被分配到了一个 他后面可以逼掉 就我觉得非常无聊的一个项目 但是我就也不懂这个产品 也他们的那个团队 但是也很小 他们可能就是十几个人 我是被分配进去的 然后我觉得这个项目非常没意思 可能也做了些事情 但是不是很重要 然后跟我进去了一批实习生 就是写的有点像同级的学生一样的 有另外一个女生 她是去了 当时也她也去了一个她不喜欢的项目 然后她有一天她跟我说 她去了点点网 点点网当时是创新工厂的明星项目 明星项目 对对对 她特别喜欢点点网 我当时就对点网 因为她点网是有点 怎么说呢 我觉得异样的感觉 就我没有那么的喜欢 中性 然后她特别喜欢点点网 然后她有一天就申请自己transfer过去了 然后她就让我发现 原来可以自己去 transfer到另外的项目的 然后我就去找那个当时工厂 负责这个我们设计团队的一个人 然后我就跟她提出说 我是不是可以开一些别的项目 然后她就面露南色 然后后面是这样的 后面这个故事我觉得是比较巧的 我其实之前跟你 我们俩是在那个推特上 应该是我要关注你 或者是我要互相关注忘了 我肯定是知道你 因为当时你应该是在设计圈 比较有名那个设计师 是的 我觉得从客观角度是 然后我觉得我这个statement 应该没有人会反对 然后呢 我然后因为你在推特上 当时还叫推特 发过自拍 我不知道这个故事 我应该没 我没看过 你在推特上发过自拍 所以我 虽然我没见过面 但我知道你长什么样 然后 然后 因为我们没有 之前没有任何的视讯 那段时间我应该用了真人头像 我忘记了 反正就是我看过你的 不要说我发自拍 这么恶心的 我就在推特上 看到过你的照片 是 然后我当时在工厂 因为当时应该工厂 还没有公布投资了之后 是一个秘密的一个事情 然后在创新 在那个当时应该也是 搬过去不久 刚刚搬 刚刚搬过去 当时是 我记得挺清楚的 就是在第三期那段楼 中安冲那里的段楼 应该工厂 就在17 18楼 然后 我有一天中午 还是半晚的时候 在打饭 在中间那种很 简陋的一个地方 去打饭的时候 我就碰了 对我们就在 我们就在那个 饭堂的那个区域的 外面 应该没有饭堂 没有一个饭堂 就是一个东西 就是一个走廊嘛 其实就是一个走廊 对 然后还有个水池 然后我就 我就问你 你是 那谁吗 对你应该记得 是这样吗 我怎么记得是有个人 领你过来说要 没有 没有 那我真的是 应该是我就直接问你 反正我就给你打上了照面 然后我就说了 我是 然后基于这个之后 我就跟你 我就知道 原来知乎也在这 因为知乎是一个 这个要说一下 知乎是我在读书的时候 我就开始 我应该在11年初的时候 应该 知乎有可能是 10年底上线 我很早以前 我就特别喜欢 我当时刷Cora 刷的特别多 这样故事可能比较无聊 后面可以解掉 那我就其实 我的意思是 我已经是一个知乎的 非常重视 对我 并且我是自己 用一个Google Sheet 申请注册的 我是填了一个 我是什么 我填了一个 我是什么HTML5的一个 什么专长 信息什么之类的 然后他就把我邀请进去了 然后我就 所以我当时是3000多号 我应该是3000多号注册用户 所以当时我已经在用知乎 然后我就知道知乎已经在工厂 这件事情让我很兴奋 反正这个都是这个结论 就是我发现知乎招式上 然后后面我就去找工厂 那个人就说 我想transfer到知乎去 然后他就非常为难的 负面就为难的同意了 所以当时我就后面到加入知乎 最开始就是实习生 然后后面到开学的时候 我就觉得在知乎这种比较创业 当时还是比较早期 就是有一种典型的那种创业 很激情 然后所有事情都是可能的 然后发展也很快 然后里面的人都非常非常 早期知乎的人都非常有意思 然后用户也很有意思 所以感觉做的事情就比读书 有意思非常多 读书就是学一些那种书本上 长得看论文 就对我来说 这是一个我特别想做的事情 然后在美国那边的学校 其实就是可以 他们就是比较随便的 你就可以发一封邮件 就可以修学 然后我就发一封邮件 就是我要修学 我能不能就是延后 延后后面的学年 然后他们就同意了 所以我就当时后面在职务 就是转正 就是继续工作下去了 也是11年大概 如果入职的话 有可能是7、8月份入职 大概具体忘了 然后正式的话 可能是9 不重要 就在职务待了 差不多三年的时间 所以这是我第一份正式工作 你还记得当时在知乎做的那些 有影响力的事情吗 或者比较重要的产品 我觉得现在看来 可能最有影响力的事情 但说出来感觉好像也不是特别有影响力 其实就是那个知乎专栏 是我 我走的 大家可能也不知道 就是当时我一直特别想 当时应该你已经走了 然后当然我觉得 因为中间有一个比较重要的项目 是知乎要开放注册 开放注册的时候 我就是我来设当时 赶哥可能已经 我不知道他可能去做APP 怎么回事 然后当时我来设计那款 开放注册的首页长什么样 就未登录的首页长什么样 因为就要改版之前 你设计的那版非常极简 那个地球那个首页 然后这个设计 这个本身设计这个首页 不是什么特别牛逼的项目 但在设计的过程中 我就需要跟周元讨论比较多的 知乎这个产品的核心价值 是什么 因为可能有一句话 或怎么样 有个message跟用户传达 那在这个过程中 我就会自己想很多 就知乎的核心价值到底是什么 然后最终呢 我们可能最终有一个版本 当时的publish数据的版本 是与世界分享你的知识经验和见解 是这样 但这个对我来想比较影响比较大 就是让我去思考 知乎这款产品的核心智能力 应该是什么 当时是可能 13年14年左右的时候 肯定是面向长期的 我觉得这款产品定位 应该是什么 那我最终 我对这个事情的答案呢 跟周元 或者跟其他人不一定一样 但我自己的答案就是 我觉得知乎是中文 中文互联网上 高智朗内容的平台 然后当时也可能没有UG CPUC这种东西 就是觉得我们要提供所有高质量的内容 这是我自己的答案 那么如果是高质量内容的话 如果愿景是这个的话 那么我们的产品就可能 比如产品是服务愿景的 知乎最早只有问答 那很明显的想到 你不应该只有问答 那可能 Cora当时可能也出了类似于 article之类的 我有点忘了 对Cora很早就出了blog 对 然后我觉得知乎一定要有一个 类似于article 或者说是不需要问题 也可以去发布内容的形态 然后我就一直想推这个事情 然后周园一开始 我印象中是他不太 可能大家对这个没有那么兴奋 但是因为早期知乎 大家其实也是比较开放自由的 就是你推个项目 某种时候你也可以去推上线 或者说你把它推动 只要你找到足够的人 所以 所以我当时就基本上 我在离开知乎之前 比较重要的一个 最后一个项目 最后一个项目之一 就是把这个东西完整做出来 推上线 然后我还记得我把这个东西做出来之后 然后我为这个名字苦难了很久 因为叫知乎文章 这件事情就很无聊 然后我就去找季欣 对当时知乎的那个CEO Cofounder 然后我就问他的建议 就是他的名字 他一开始也没有想好 但是他对这个挑战 还显然是很兴奋 他有一天遇到我 就是他突然跟我说了两个字 就是转栏 就是叫专栏 这个名字我觉得可能也没有那么的完美 但是反正比我当时想的所有名字都可能要 有一种精英化的感觉 就是专栏之前的话 不是每个人都能写 都能发专栏吗 就是在杂志上的 在报纸上的那个专栏 是只有少数人才能拥有的特权 另外也还有一个 这个名字 他也会影响到一个产品的决策 就是你让用户去发文章 到底是在一个博客里面 因为当时还不想叫博客 在一个place 还是说你可以偏偏发 比如说像现在的exit post一样 但专栏可能就限定了 你得有个专栏 或者说大体上用户会这么下 但最后反正 我就没有看好的名字 就直接用专栏 然后我们就 我其实当时就我 再加上可能一两个前段 再加上一个后段 我现在都还记得他们的名字 现在有一个前大师 跟我在一起 我现在在公司 然后我们就把这个产品做了出来 我觉得这个是一个 如果说单个的 我对这个项目影响最大 然后到现在 它的影响仍然有一定 虽然我觉得可能还好 的一个项目 所以你下一段 比较有意思的职业经历 是小红书 而且是早期的小红书 是这个是我觉得 我家长想要说是一个 也是有点像是我家长吃乎一样 也是有些机缘巧合的一些 一些因素在里面 我是在 16年 应该是16年 夏天的时候 我带我爸妈去美国 玩了一段时间 就是带他们去 建设市民 反正是带他们去 经典的路线 一号公路自驾 然后绕到Vegas 去优胜美地 就一个环线 然后都是自驾为主 然后那段经历里面 我在后面 绕回来到旧金山的时候 当时我们有共同的朋友 然后他在Airbnb 然后因为我们很早以前认识 然后我路过 金山的时候 我还记得当天 我应该是下午去找他 然后晚上我还去看了 勇士队的比赛 气候赛 然后anyway就是我 因为跟他关系比较好的 当时在Airbnb工作 我去找他玩 找他去visit一下 Airbnb当时在一个 非常fancy的 应该是非常有名的 当时的花园一样的办公室里面 现在银行听说退租了 然后我们在Airbnb的办公室里面 顶楼 我们去到顶楼那里 他有个很漂亮的一些地方 然后有个乒乓球台 我还记得我们在那打乒乓球 当时我可能对职业上 其实就有一点 不能说是迷茫 但是有一点在想考虑说 我接下来在职业上 或者说我应该做些什么样的事情 应该有些什么样的选择 然后他显然是在一家比如说 标语那公司Airbnb 然后当时我们也聊了这个话题 但是没有特别展开 然后他可能也没有能给出什么 特别有见地的东西 然后他看起来 他所在路线 就是他已经在美国 可能已经在排队绿卡 在一些科技公司 就一些期权的即将到手之类的 这么一个状态 很舒服的状态 然后我在美国那个旅程之后 我就回国 回国之后 应该不到一两个星期 就打电话给我 就说 就非常的急匆匆 什么 他说我要回国了 我要回国加入 就是大概类似这种的话 然后对非常吃惊的一个消息 然后他就问我要不要一起来 他应该没有这么 没有这么有感染力的描述 但大概的意思就是这个 然后让我就非常吃惊说 他本人在美国 这么向上的一个环境 或怎么样 他也可以留在美国 但是突然回国 加入一间 加入一间 小红书 其实我当时知道 小红书的产品 因为它是女性 早期是其实女性购物用的 其实我肯定不是目标用户 然后当时的产品设计 也不能说是 也没有什么客取之处 也没有太多的客取之处 所以没有特别特别的关注 特别多关注这个产品 然后我一记得就想说 一开始那个logo 就长挺抽的 就是一个打开的书 那个颜色 这个很深红色的 是特别好的颜色 因为我当时已经 刚搬到上海 就之前在北京 刚搬到上海一段时间 所以小红书正好在上海 那时候上海其实 基本上没有什么值得 去的互联网公司 B站也比较小 拼的多个还没起来 不像今天一样 其实当时上海 没有什么互联网公司的选择 我就去小红书了解了一下 然后去小红书聊完之后 发现这个团队 其实还是挺 就非常的 早期小红书 还是比较精英的 然后早期小红书的 包括毛文超 以及包括一些早期的 Founder或者高管 或者说一些比较重要的人 其实都是在美国 他们去过比较牛逼的 一些大公司 然后或者小红书有个特色 是他们很多是 咨询公司背景的 回来做高管 或做电商做运营 然后当时比如说我去 刚开始加入小红书的时候 我去 因为我是做社区产品 然后我去跟运营团队 那里一坐下来 然后发现跟我聊天的 一个非常基础的就是做内容运营的一个女生 她是Stanford GSB毕业的 就是基础的做简单的内容笔记审核的一个人 她是Stanford商学院毕业的 就是这种误差 所以小红书早期其实是大家是有一种比较强的精英氛围 然后大家都非常想做牛逼的事情 我蛮喜欢小红书当时的团队旗帜的 然后也没有想太多 就觉得这公司挺有趣 就加入小红书 当时加入小红书是在2016年的10月份 当时小红书的DAO应该还是比较 这个应该能说吧 就反正比较少吧 就是基本上打开一看是一个非常小的 小规模的一个产品 然后我加入小红书 就是直到我离开的时候 基本上是四个完整年的时间 我在小红书就是做社区的产品 所谓社区的产品呢 其实就是这个APP除了商城Tab以外的 因为可能就是社区的 因为如果你小 比如说如果消息Tab和我Tab 也没有什么东西可以做嘛 如果是产品小的话 其实就在做这款产品的内容的 从它的样式到 跟早期的去推荐去讨论 怎么去分发这些内容 到可能后来在17年 18年左右开始重点要做 camera要做相机了 当时抖音起来了 所以所有事情都包含在社区上面的范畴里 这些所有事情我都做过 然后就要想说做这些所有事情 然后从我离开的时候 应该想说 已经是一个相对来讲规模比较大的 可能越活 应该越活已经上亿规模 对吧 大概比较完整的经验小王树 从小变中等大小的过程 我觉得从当时做小王树的时候 开始去做这款产品的时候 其实我觉得是什么都不知道的 就是不知道什么叫做一个内容平台 因为当时我们回想到16年的时候 抖音其实可能刚开始或者还没开始 然后当时大家觉得这个市场上最大 最牛逼的所谓的失调产品 社区产品就是微博和微信 但今天我们都看到 抖音TikTok 小王树其实完全 甚至当时快手可能是也没有那么大 包括视频 所以我今天回看当时的经历 其实这些好几个平台 其实本质上它都是基于UGC 和可能未来 和这个产品发展到更后期之后 有PGCPUGC 基于这些用户产生内容 构建出来的一个个的内容平台 或内容社区 然后上面的个性化 推荐分发是非常重要的一环节 就结果上来看 我们构造出了这么一种产品形态 那么小王树在今天毫无疑问 我觉得它在中国的 我们可以认为中国今天那种平台 就有影响力的就四个 或者说最多加五个吧 抖音小王树快手B站 视频号 其他不重要了 就是小王树 我觉得从今天来讲 肯定是这里面非常有价值的 甚至是取决于不同人看 可能最有价值的 最特别的一个 虽然它用物量毫无疑问 也不是钱 还不如视频号和快手 但它的产品的场景和价值 包括它因为它有很强的搜索 包括它图文做得特别好 所以它有很强的差异化价值 但在当时做的时候 什么都不知道 就是不知道 这任何一个东西是什么 然后当时跟我们一起 做这些事情的人 包括比如当时早期的社区运营 比如说 或者说毛文超 或者是柯南 就是小王树的管理层 我觉得大家都没有一种非常清晰的sense 我们在做一个什么样的产品 在中间 大家会想着说 我们要成为一个Instagram 但你从今天的结果上来看 那也都是 过眼云烟 实际上最终这个小王树构造出来的产品形态 这个以UGC搜索 信息流分发 构建出来的一个很牛逼的一个长期护城河 还不低的一个产品 当然在今天如果在AI的话 这是另外一回事 但我们当时做的时候 其实是抱着一种 不知道自己在做什么 一步一步的摸索的形态去做的 然后我觉得当时最有意思的部分是 其实我们 比如说今天我们今天产品去找工作的时候 大家可能会分什么策略产品 C端产品 这些区分的分法 但是在那个时间间 其实完全不存在说一个产品 能做什么事情 不能做什么事情 所以就是跟推荐 我觉得基本上是我的工作 其实就是跟推荐团队的最合现在那么几个人 包括运营团队的最合计那么几个人 我经常跟他们坐到一起 我就比较排斥 我就自己待在一个产品的一个屋子里面 只跟产品说话 我就经常是跟 包括当时我也跟很多用户有很多交流 一步步的去解决这个产品 遇到一些问题 这个让我印象很深刻 因为你当时在吃糊的时候 也是很喜欢跟工程师混在一起的 就是一起在聊一些东西 因为我像我这种设计师 可能更喜欢自己用大块的时间去思考 而不是去向外寻求这种沟通 我觉得这个也是可能你的特质或者是能力吧 因为很多比如说小龙数据板产品 它其实本质上也很简单 比如说一些推荐的 但这个我觉得可能今天我们很难讲到特别细节 就是本质上你如果不了解一些推荐基础的 比如说怎么召回 怎么比如说有金牌粗牌这种步骤 你可能很难知道这里面有哪些地方 比如说有些步骤是影响到产品体验的 有哪些不会 然后因为如果你不知道的话 其实推荐工程师去做这些产品决策的时候 他其实就会做一些 他其实就会做到就会就会 不一定是做出最佳的复合用户产品体验 或者说是怎么说你的产品价值数据的一些决策 所以我觉得对于一个做 比如说你是做这种内容平台产品的人 你得非常了解你的分发是怎么做的 然后给予这个 你去做你的产品判断的角色 包括运营跟推荐 他们可能隔得更远 对吧 很多时候比如说早期 这个做过内容社区或平台的人 大家都有或多或少都有感觉 就运营觉得自己的内容分发不出去 这是非常经典的问题 当然这个问题里面 其实每一个case都值得去非常仔细的 全链路的去看为什么分发不出去 这里面它有哪些问题是值得被 非常有价值的应用到所有的机制上 其实一个一个case的去收集 去debug 去修改这里面可能存在的问题 我觉得这累积起来 让这个产品不断的在迭代变得更好 但因为像那个比如说内容分发的问题 用户是没有很强的感知 它不像一个产品的C端问题一样 我点这个按钮 这个按钮不工作 其实分发本身是一种产品的设计 对吧 或者产品的体验 很多时候如果一个该有流量的内容 它流量不好 这就是一个不好的产品的执行 那这个时候用户很难反馈这个问题 所以你就得从很多方面去收集 无论是运营的或是KOL的 当然早期小树也没有什么KOL 都是跟小树一起成长起来的人 他们对于自己流量反馈 从中你就辨别出来 有些确实是非常有价值的 就是这个东西经验牛逼 它为什么不行 那这里面其实不断的 我觉得累积起来 就让这个推幻系统 变得一步更robust 当然后续你整体的基础的技术 比如基础的这些模型的提升 当然也很重要 但是我觉得中间有很多这样的 没有办法把它归纳成一个大的 一个东西的一些细碎的事项 特别是这款产品本身是一个 从小到大 这个公司从小到大 不是建立在一个像 比如说现在的字节一样 一个很完整的一个体系 的一个前提下做出来 所以这里面有非常多的实际 很有意思 但很零碎 很多样化的工作需要去做 比如说另外一个很有意思的事情是 在小红书 其实我花了很多时间去 去build小红书内部的 AB test的平台 就其实这个产品本身可以认为 我就是这个产品的产品经理 就当然也没有这个title 就是我跟当时我另外一个 在小红书的同事 它是最早是做数据 它是数据工程的数据 大家也有些产品 就是我们一起基本上 因为当时也没有人说 要做一个AB test平台 就是在170 18年那个时候 大家从 就是top 其实是没有那么强的感知 我们一定要有这么强的一个东西的 然后但是你做推荐的话 AB test是非常关键 这个应该不用解释 所有做情况事情都知道 但是如果你怎么样整个团队都 都adopt这种数据驱动的体系 你就需要一个好的工具 所以当时其实我花了很长时间去 去做这个内部的工具 让这个内部工具变得好用 包括后面更多的 其实更多产品的功能 会用到这个东西 平台 然后这个事情对我来讲成就感很大的 包括我自己就在讲话说里面 我自己是开开就是开实验最多的人 我但他们现在可能平台换了 我都不知道 就是我们甚至一度 因为我开了过于多时间 他搞过一个排行 然后还有给我发了一个奖 然后那个不是很重要 就是随便那种发了一个 然后毛文昌还跟我合个照 类似这种故事 就是就因为 但那个就是非常 就是自己搞的 也不是什么很盛大的 因为那个平台本身 就是我和那另外一个同事主要在搞 像这个事情 就跟产品也没什么关系 但是对我来讲 在一个公司里面做这样的事情 对我来讲 有很大的成就感 就是让我觉得 而且我也可能之前也没有 也不是那么什么 我也没有在Facebook什么待过 对吧 也不知道怎么真正数据驱动是怎么回事 也没有在之前 在小说之前也没有在自己待过 所以当时就是从自己感觉 感觉是自己把这个东西摸索出来了 然后中间也会学习一些 这些数据分析的理论 也挺有意思的 什么显著性不显著性 这非常基本的东西 可能一步一步也没有什么教程 就开始把它做起来 我觉得这部分是对我回顾来讲 小无数经历里面很有意思的一部分 对啊 一个设计师 设计师出身的产品经理 你当时的title是什么 是产品经理吗 还是说产品负责人或者是什么的 当时我 这个是另外比较有意思的一点 就当 因为之前我在直呼 包括我在后面一些经历的时候 包括在Dine的时候 其实我做了很多设计 对吧 当然我们都认为自己 我们也要做一些产品的工作 但实际上我们会做设计的交付和执行 然后去小无数的时候呢 我并没有预设我一定要做什么 但小无数其实他公司里面 他就分了有设计师和产品经理 所以我去面试的时候 面完了之后 我可能就要去了 他们也准备要我了 然后我就跟邓超聊 邓超就是小无数那个产品负责人 他负责产品的设计 然后他就问我要做产品经理 还是做设计 然后他给我去选择 我当时就不太想做产品经理 可能这个我不知道能不能播 早期我们都比较讨厌产品经理 只会指挥不干事对吧 然后后面我就想产品经理的话语权 就是讨论到产品经理的话语权 好像更大一些 是能决定能做 如果做设计师的话 就变成可能你要接需求 所以我就选说我要当产品经理 但我当时还跟邓超说 我说我要做设计设计的时候 我还是可以做的 就是那还怀出来一种 我还是喜欢做设计的 然后当然邓超也OK 对吧 然后所以我的title 进去的title就是产品经理 我其实做了第一款产品 不是在web 不是在app给你做的 我是做了一个web的一个 做一个享用书 早期的一个给KIO的功能 长笔记 类似于享用书里面的一个 公众号功能的一个编辑器 在做那个产品的时候 其实我从那个产品的设计 都是我自己做的 我就不想 就是有种习惯性的 我不想用别的设计设计 包括我当时 当时我还自己设计了一个 享用书的一个海外版的一个官网 我当时现象现在都做得挺漂亮 但没有上线 然后但是逐渐的就发现 好像其实在做一些很有意思的事情的时候 也不一定要自己去做这个设计 所以就是它转变可能是一个主件 然后比如说我开始参与到一些app 这些事情的时候 其实很快 我可能就做了前面做了一两个月 然后就马上就可能 就找摄影师去提取取后等等 就进入到这个比较传统的一个工作流程里面 所以在想说其实代表就是title是产品经理 然后对 那你后来加入了字节跳动 是因为什么原因呢 核心的理由就是字节在当时 在我比如说或者说做C端产品的这个严历 就是因为他做出了TikTok这样的产品 当然TikTok之前是比如说Musically做的 就是在全球化 全球化的就是中国没有任何一家公司 在做C端软件产品的方向上 做全球化做的如此成功 并且TikTok是如此一款 可能是也跟小树一样 也可能是原创的 Musically不是叫TikTok 也是原创的 然后做的非常好 然后有很多非常牛逼的地方的一款产品 然后当时就觉得Alex 就是以Alex为主吧 并不是说 这里可能会不知道怎么得罪人 就是Alex以及伊敏这些字节的管理层 就是是一个非常 我当时非常想要学习 然后非常觉得他们有非常多可以让我学习 想知道这家公司为什么可以这么牛逼的 一个对我来说产生了这么一个动机 所以比如说他并不是为了 并不是单纯 并不是为了说OK 公司我要去更好机会 而是说比如说当时我去 比如说如果去别的公司 我根本没有任何 没有一秒想过我要去认为一家别的公司 就是别的大公司 因为小龙叔对我来讲 也是非常非常有价值 或者说我很喜欢的一个经历 我其实不太喜欢去大公司 但是字节是我认为是非常独特的一个地方 然后他的产品 他的品牌 他做的方式毫无疑问 非常独特 所以我就是抱着一种学习的心态 我有一天晚上 还把移民的微博全部看完 他可能后面把它删掉了 但是一天晚上还把微博全部看完了 还是觉得很有意思 对吧 然后Alex也是他在网上的一些访谈 就是他在网上有为数不多的几个访谈 我都看完了 然后就当时我觉得核心目的就这个 就学习的心态 然后有一个机会去上海 就是自己主要是抖音的上海的部分 嗯 然后在抖音呢 我觉得如果说结果来看 我觉得应该还是做了一些牛逼的东西的 我是不是老是做牛逼这个词 做了一些比较厉害的东西的 呃 比如说那个小火人 可能算是一个吧 但我好像也不是 就首先小火人跟火花 应该是现在非常成功的试掉产品 我觉得 According to 一些人可能是非常非常成功的 但确实就是我当时负责团队 甚至是我对这些产品的影响应该非常非常大 呃 其他呢 我觉得也做了很多 你如果从产品的角度 比如说你如果携带 简历上 不是一个很好的类比啊 你可以说是什么数据提升了多少 达到多少规模 就这样的产品是非常非常多的 但我觉得值得说的可能就是小火人 火花 然后以及我们把抖音的IM的规模做到了一个比较大的规模 这个数字不说的 反正就是我们其实在抖音里面做了 在做熟人社交这个事情 还是有了不效的进展 就我觉得结果上可能是这样 其他呢 在字节里面做了很多大影响的 就是数据量特别大的 包括后面 二三年二四年开始 用很多AI做很多事情 就AI特效啊 对 因为当时我负责AI特效 我负责特效嘛 呃 所以我也我也做了非常非常多AI的事情 我觉得我对 如果但这个可能是对字节 也不是那么不是那么比较 比较富项的一点评价 就是我一开始对字节的期望值是比较高的 应该这么说 就是我觉得字节是一个很牛逼的地方 然后里面的推荐 呃 包括他们做事的方式 大家呃 大家做成做出来的产品 实际上也做得很好 然后数据也很棒 所以但是 但我的整个体感呢 我觉得有一个 有经历的一些阶段 就刚进去的时候 我觉得是有非常强的新鲜感 比如说 在小文书的时候 我可能觉得跟我合作的推荐团队里面 可能 顶尖的推荐工程师 可能不超过五个 但一去到字节呢 我发现这样的人可能有二十个 就有种感觉 就哇 所以人才真的是 非常多 但逐渐随着 嗯 事情时间的推移呢 我会发现字节本身 他的团队的 配置或他做事的方式 就会看到比较多的问题 但我觉得今天我很难 我也没有一个很好的position 我也不是什么可以评判自己的问题 的一个position 对吧 但我的 作为一个个体的体感 就我觉得这家公司 在能力 效率 在边界上 我觉得 我还是看到很多很难解决的问题 特别是变大 变大以后 变大是一个很大的一个因素 然后在后面 我也做了很多努力 我想去做很多改变 对吧 在我能够触及的方位内改变 但整体 in the end 就还是发现 任何一家公司 他还是都 还是有他非常长版 还是有他的短版 我觉得后期的话 就肯定在自己的里面 我就逐渐感觉到 我从这家公司 能够 至少我个体 有可能是我的问题 从能学到的东西 变得越来越少 当然我觉得 之前比如说Alex 我也跟Alex 可能碰多或少 肯定有一些合作 我觉得Alex还是很厉害的 他无论对产品的判断利益 还是对方式的判断等等 但后面对整体公司的一些事情 我觉得长期来看 我不想在自己做一个管理者 这个事情对我来讲 没有特别大的吸引力 好像管了很多人 管特别大的东西 但实际上 整天做的事情 其实你不能做特别多的创新 然后 就是在做一些怎么说 management的一些东西 然后做一些沟通的东西 然后有很多的流程 所以这个是我不喜欢的 我觉得肯定有必要性 但是这个是也有些人喜欢 也有些人很舒服 但这是我不想长期 待在自己的最主要的原因 我还是更想做一些 有意思的突破性的 尝试一些新的东西 说到这个 有意思的创新性的东西 要不要聊一下你的新产品 新产品之一吧 我们其实最近 我们其实在做好几款产品 然后我们就是 刚刚杜潇说的这个 你说这个应该是 就是我们最近投入比较多经历的一款产品 叫Clauli 然后这个名字我们可能也会改 不重要 这款产品的起源也是 挺有意思的一点 就是我们在一月份的时候 我们在一月份的时候 当时我们在做一个agent的产品 我们当时做的agent产品 是比较瞄准生产力 强生产力的方向 然后我们当时比如说在 我们探索了一些 做视频agent的方向 然后呢 我们已经做了一个 非常完整的一个 做了一个做视频agent 做了一个做PPT的 然后也有一些比较基础的 general的 做research的这种产品 都做出来了 然后当时我们是基于 Cloud Agent SDK 做出一款产品 然后做的过程中呢 就发现我们对于这款产品 想加的 我们想赋予它的一些属性 比如说我们想让它 普通人的使用门槛更低 但是同时可能有很强大的能力 然后比如说它 又更常识的 比如记忆更好等等 这种现在大家做一切 都会经常提到的东西 然后以及当时 从去年十月底的时候 其实agent有很多生态的 一些技术的东西就逐渐出来 当然我们今天可能没有 还没聊到 比如去年11月12月 当时Opus这些东西 其实是有非常大的 对于Vibeship的一个转变 然后当时我们 如果我们在十月底 一月份的时候 做这么一款产品 我们肯定想利用 正在蓬勃兴起的生态上 很多优势 比如说很多很多的skill 对吧 我们不用每一个场景 都去单独做一遍自己的skill 或者类似的东西 所以我们前面那个产品 做到一半 就发现Open Cloud 这个框架 其实是非常满足我们 想做的很多的点 比如他有更好的记忆 他能够很好的兼容 skill的生态 以及他有很多 他在比如说基础的 System Prompt等等 这些设计上 他更像一个普通人 更简易可以开始上手 一种agent的体验 更像人等等 这种基础的东西 所以当时我们就是一开始 是一个比较简单快速的项目 我们想 我们自己在用Open Cloud 我不知道你有没有用Open Cloud 但今天Open Cloud 大家应该都已经 怎么听你了 我们就基于自己 我们自己开始 自己装了一些Open Cloud 然后包括在本地 包括在云端 我们也部署了一些 我们发现这东西 有非常非常多问题 对吧 大家都知道 部署起来尽量挂 然后你要自己配Token 然后你如果要弄一些能力的话 你又得自己做很多事情 所以当时我们就做了一个非常简单的想法 就是我们想 我们想做一个让普通用户可以很简单的用到 我们 Cloud 的一个产品 然后这不是腾讯楼下 这不是腾讯楼下做的事吗 对对对 可能结果上 结果上我们做的比较 我觉得我们本来是想春节前发布 但是发现打某产品花了非常多的时间 但我觉得今天 我觉得这款产品 我们做到了一个比较好的 其实我觉得是 如果一个用户 他真的要用Open Cloud 我觉得我们产品是一个比较好的 能够offer一个Open Cloud 的一个 personal体验的一个产品 但是我觉得Open Cloud 随着做这款产品 我们也发现了Agent这个方向 我觉得无论是在产品定位 还是说他真正到后面 go to market 或者商业化上 他的一些挑战 就是你单纯一个 跟好用的Open Cloud 这个我觉得是 就很难让用户去买单 或者让公司去买单 除非你就是 当然中国很多 比如可以装机 什么一键装机 什么之类的 这个是我觉得肯定有价值 但是可能比较短期的东西 所以现在我们其实也在重新思考 这款产品的一些定位 我们也有一些比较大的东西 可能再逐渐去改 这款产品 我们现在 我们现在基本上是在一个 我们大概内测了 可能一个月 一个多月 然后有 有大概一两百个内测用户 都是朋友的朋友 然后内测用户呢 就不是以 没有那么多的 以互联合为主 但是都可能以白领 这种 就是偏白领 偏知识型工作为主 然后我们发现 我们从内测中发现 其实普通的 普通人 其实还是能够非常 就他们还是非常需要 agent 现在最先进的这种 各种agent offer出来的很多能力 但这里存在 几个很巨大的鸿沟 第一个是 他们其实不知道 agent能做这些事情 就是这是一个 非常大的问题 就是我们经常是 因为我们现在 这款产品 还是做的比较简单 没有做太多的 教育之类的事情 经常是我们 我们可能做了 我们可能跟用户说了 他可以做这么用 他可以干这个事情 然后用户才知道说 OK 他原来可以这么牛逼 所以今天大家还 不知道 这样的产品形态 可以做这样的事情 这是第一个 第二个就是 当然这个 我发现这个讨论 进入到了一个比较 好像是在跟投资人 讲话的一个环境 就是第二个就是 第二个就是 agent如果 他能够发挥出 他的最大能力 他需要 除了需要更好 除了需要很好的 机座模型之外 这个大家都知道 它上面其实需要 获取你很完整的 或者说更准确的 上下文 那这个问题 在程序员是比较简单的 但在很多白领 特别是这种 非程序员 或者说是这种 事务性 工作的白领 或者工作者为主 他们的 上下文是 基本上 比较分散 或者说没有一个 很好的方式去 可以喂给agent 那今天很多公司 很多产品 在解决这个事情 但到目前为止 应该还没有任何一个公司 或任何一个产品形态 很好地解决这个问题 所以 这个是一个很大的 另外一个很大的 轰沟 然后除了这个之外 毫无疑问就是价格 因为今天最好的模型 事实上 跟稍微好一点的模型 其实差别 真正进入到复杂任务 特别是agent 相关任务的差别 还是比较大的 就有些模型 其实是 性价比很高 但是做一般的任务 还可以 但一进入到复杂的 特别是我们这种场景 其实基本上 它在榨干agent的很多能力 所以 如果用 一直用最好的模型 去让用户做事情的话 非常非常贵 大家也没有做好 这个心理建设 能够花这么多钱 虽然说 这个 X上 或者说 这个VC的叙事 或者硅谷的叙事 就是大家说 你其实是在 hire一个员工 我怎么在这里 但是现在除了 除了硅谷的公司 应该大部分的公司 还没有这种很强的心态 就我 我一个月花个 几千美金 或几千人民币 去hire一个agent to do 这些事情 其实是在替代我的员工 说到这个 我们内部有一个叫 MICE的项目管理软件 我一直认为这种 看板的形式 需要大量的人力去维护 或者说linear这种 都是需要一些 首先你要创建一个任务 当然可以通过一些方式 让它自动去创建任务 但是对我们这种设计团队来说 很多任务 还是需要去手动 去写 去创建 虽然我们已经做了一些 简化的工作 比方说 你可以在slack里一键 去根据一段对话 它会自动提取里面的内容 帮你创建一个任务 但你还是要创建一个任务 我就一直在思考 这个未来工作 或者说未来项目管理软件的形态 我就做了一个尝试 就是说 我其实是受到那个 notion的agent的启发 他们不是前段时间 一直在推那个功能吗 然后我就去用了一下 用了大概一天的时间 我就摸索 算是摸索透了一些 它的工作原理 它无非就是通过一些非常 经过这么多年的发展 它有很强的那个connectors 生态 原理就是有一个非常非常详尽的系统prompt 去控制这个 这些connectors去不停的 要么去推 要么去抓context 然后呢 去以一个这种loop的形式 去或是trigger吧 去做一件事情 我其实一开始看到它是 原来是这么简陋的一个事情 我还是有点惊讶的 但是我用了一天之后 发现效果非常非常好 它可以实现在slack里面 去add他 你可以选 就是add他或者是不add他 不add他的话 他所有的你在slack里发的对话 他都会读 然后都会触罚一次 这个agent wrong 所以我就一天 我把我们mess上 所有的任务全搬进去 然后让他去作为一个办公室助手 去帮我们管理这些任务 然后一天下来大概花了3000credit 我忘记它真是价值 因为它那个还没有正式的收费 大概算了一下 应该是50美金左右吧 就一天我都没有正式去用 我只是把这套workflow set up起来 就已经花了50美金 所以我觉得这个东西 肯定不可持续 所以虽然它很好用 但是我用了一天之后 我就把它停掉了 然后我就在mess里面 试图去用同样的slu 去做一个类似的产品 我们把它叫sure 就是叫什么副厨 二厨 大厨是chef 对我们所有都是这个餐饮类的 我也是给他加了一些connector 因为notion虽然connector很强大 但我们常用的就那么多 大概就五六个 比方说我们把slack 把figma dropbox github 这个最重要 github放进去 然后也是用一个agentic loop的方式 去被动接受 或者主动去抓取一些 项目中的context 比如说他可以分析 这个文件被谁碰过 然后谁打开过 谁保存过 保存了几个版本 然后发了什么评论 然后再结合我们任务板里的任务 他就能大概的猜出来 这个任务目前的进展 所以这个任务板 现在已经变成了一个半自动的形式 我觉得这个 就让我们用的很爽 你那个素质 你们是 什么是 用什么自己写的agent loop 有用什么 open clause 没有用什么 用的是convex的workflow 就是一个 因为它是个很长的workflow 所以我们把它用了一个 可以persistent的编排的工具 让它可以很长时间的run 我们现在 就是我觉得这个 如果要聊这个 我觉得应该也有意思 但我们现在公司大概 正式员工的话 9个人 然后我们现在只用linear 我们也用heavily用linear 但是应该我自己 包括我猜大家 应该打开linear的次数 其实很少 我们基本上都是通过 cloud code 然后有时候用clody 就是类似吧 对于原理吧 就是去操纵linear 然后包括所有的 比如说 刚刚你说的时候 想到一些具体 我们那些细节 比如说 我们的所有的数据的看板 比如说我们用一些日志服务 有一些很好用的 比如说axiom 像post hoc 这种打点的服务 对 我们其实都是通过 clarly的agent 然后包括clarco 因为他没有定时任务 做得不好 所以我们很多时候 用clarly的agent 定时的去看这些看板 然后 我也不知道他看的是啥 反正他有可能是看看板 有可能是看具体的数据 然后也有可能是看 这个reple 这个代码reple 昨天一天的提交 比如说我有一个agent 这个agent也会 他其实就是每个 他每天会 每三小时 把最近的新提交review一遍 然后每24个小时 会把过去一天的新的PR 就merging的PR 去announce一遍 就有什么新feature 然后他一天可能大概有四五个时间点 每个时间点 播报一次 我们可能有六七个比较主要的 post hoc的dashboard 播报一次这个dashboard 然后最近的趋势 然后他会有些发现 然后反正非常非常 就全部都是定时人物 然后驱动的 然后我们现在 接下来在正在做的一件事情 就是驱动他做下一步 就现在可能我们基本上 第一步的问题发现 包括review他都有 但可能下一步 我们现在在做 就是让他可以 这个说我们现在都只是 不是form产品 就是说我们在heavily的 把qually用到我们的 这个工作楼里面 就让他可以更多的 比如说自己把一个问题修了 然后但这个我们其实 有很多公司可能说他们在做了 我也不确定 但我们其实还没有到这一步 我们现在只是问题发现 基本上我们都不看任何的 这种三毛软件 我们都是用agent的方式来做 聊对方的工作方式 我觉得尤其是小团队的 其实蛮有意思 就是无非就是借AI的力 去用更少的人 完成更多的事情 这个也是咱们在开录之前 你提到的 因为我的理解里面 你们也做了很多的产品 虽然有些可能还没有想公开 但是你们依九个人的规模 做了这么多产品 而且时间也很短 我觉得这个还是蛮厉害的 我昨天跟一个我的朋友 聊太的创业 然后他问我们做的产品 是什么之类 然后我就说我们做了这么几款 然后他就说哇 他问你们是全员编程吗 然后我之前没有想过这个词 然后我感觉 我不知道是不是一个流行词 我问他这个是个时髦的流行词 什么之类的 但是实际上我们确实是全员编程 以及我们现在的 以及我们并不是要吹捧这个事情 我只是说描述我们的方式 然后以及我们现在 我对于一个新的员工 如果要加入我们团队的话 我对他的预期就是他要以这种 agentic或者说以coding为主 然后以通过agentic的方式去管理各种事物 为辅 以这种方式来介入工作 所以现在事实上 就会有点像是你刚刚问的问题 就是我们现在怎么分配 我们实际上首先大家的工作方式上 我们所有人都其实需要所谓的编程 但事实上其实就在用各种的agent去做 各种的woolish coding 还是分析等等各种工作 然后除了之外的话 就是我们 我们会有一些松散的每个项目 可能有个负责人 然后或者一两个负责人 然后有一些obviously有些成员 可能是多个项目都一起在 但大致这个9个人会可能区分在 我们现在可能主要的三个项目 可能每个项目叠加起来的人数大概十几人 但是可能每个项目单独参与的人数 可能最多不超过四到五个人 大概这么一个形式 但这个我觉得如果 我感觉这个其实是一个很有意思 但是可能现在 因为一切变化的都太快了 所以没有办法很好的 有组织性的去聊 这一个话题 就是一个组织在AI时代下 应该怎么去演进 我觉得现在的问题是 每个人都在尝试不同的方式 然后很多吹牛逼的人 就觉得自己的方式最牛逼 然后很多人可能还 还用很古老的方式等等 我想到了一个切入点 就是说他现在做不了什么 我们现在存在 在UBI出现之前 我们的存在价值是什么 作为一个设计公司 我们交付的已经不是设计了 除非是一些那种marketing的东西 我们交付的项目 其实已经是代码了 还有design MD 我觉得我们现在存在的价值 作为一家设计公司存在价值 就是说在AI完全掌握设计之前 打这么一个信息差 这个时间差的感觉 或者信息差 信息差也有一些 就是design MD真的变成像 我不知道现在有什么更 像scale这种流行程度的话 对吧 那其实一个设计公司 好像也没什么价值存在了 对吧 我的感受是就回应你刚刚的问题 就是有什么人 有什么AI做不好的事情 我刚刚你在说的时候 我也在想 然后我觉得你说很多 我都是基本上认同的 然后我的一个大的一个感受 就是用的越多 反而是会发现 其实有很多事情是AI明确 做不了的 就是或者说很难做好 就是这是一种比较矛盾的体验 但是事实上我的感觉就是这样 就是会在一些事情的时候 比如说很喜欢的coding 或者写东西看信息 就会发现效率大幅提升 你就是做这个事情 以后不用它 你就就是对吧 但是呢 在一些事情 然后越用 你就还是会发现 有些事情它就是做不好 然后你不得不 也不是不得不吧 你就会意识到 你确实就应该 在这些它做不好的事情上 投入更多精力 以及要思考它怎么能够做好 这有一种很矛盾的体验是这样的 我觉得它做不好的东西 很显然 比如说像设计 一些产品的定义 它是做得很差的 不光是设计 就是可能设计 我觉得它也做的大概率是不好的 它可以给一些基础的好的东西 但是你真正要做好的设计 做不出来 然后产品的 你要做一款产品 它可能包含着设计 包含着产品本身 它的定义 差异化 切入人群 切入的定位是什么 这个完全就基本上是不行 你只能说 你有些东西让它选 或者你跟它讨论 非常棒 但它没有办法提供 还有很显然的 就比如说 你如果要去做 go the market 这个跟产品的定位 是非常紧密相连 甚至可以说是整体是一个事情的一个事情 这个也完全做不好 对吧 但我觉得很多很多事情是做不好的 所以说越用就越发现 它确实很多事情做不好 所以是发 就有一种 有一种 或者说反差或怎么样 就是有很多人说AI doom 或什么 它强调很多的工作 我觉得确实从长远的趋势上是的 但是实际上 今天AI能做的事情 我觉得还是比较窄 但是可能很深的这么一个种感觉 就是人是好 人是不可替代 当然我觉得反过来 你如果没有AI 那个现在 今天的人 可能如果没有AI的话 确实他也很 在很多 除了一些少数的 除了一些特别的职业之外 我觉得他的职业也不需要AI 但是大部分人我觉得就应该用AI 不用AI的话 就是相当于电脑出来 你没有电脑对吧 但是反过来 如果你 只有AI的一个 就是有很多人 只有AI去做的话 很多基本的事情 都没有办法做得非常好 我可能就重复 但因为我的想法是这样 OK Clawley还有什么想聊的吗 比较有意思的东西 或者说 Clawley 因为我们在做这个过程中 可能有些比较 比较 比较 Mondane的一些技术层面的东西 我们其实花了非常大的精力 就是我们 因为Clawley是一个非常完整的一个 我们其实做的东西很像 后来发现很像 那个perplexity的 那个personal computer 只是我们没有 我们暂时还没有这么market 就是 我们是一个非常完整的一个agent 他有带自己的电脑 这个电脑的配置其实还不低 然后它是一个云端的电脑 然后它也可以操纵你手机 也可以操纵你电脑上各种东西 我们在做这个过程中 发现 给 就是有一些时髦的说法 叫做你要给agent做产品 对吧 但是事实上 我们遇到的问题就是 agent的infra 确实是一个 基本上今天 还非常贫瘠的一个 一个 一个一块 比如说领域 还是 但这个事情 我觉得可能有一点略微 进入到一个特别 比较细分的领域 但我们的体感就是 今天如果要做一个agent 他不是建立在一些巨人的肩膀上 他可能建立在一个 只有model是巨人 但其他所有东西都还没有建起来 其他所有的必要的基础 实施都还在发展的过程中 然后就agent 他被很多东西绑住了他的手脚 包括用户拿不到用的数据 包括没有很好的tool 他也 当然最近大家做很多cli等等 这种东西 对吧 让他可以有很多的access 但直到今天 其实我们都知道 中国互联网 还是无论是 全球互联网 他agent是很多东西是拿不到的 所以 所以我们 总结一下的话 就是做quality control 我们做了一个非常全功能的 具有完整执行环境的一个agent 然后在这里面发现了 我们要补的或者说我们要做的基础设施的工作 非常的巨大 然后这里面有一点 就是 甚至我们做到一半我们都想去做这些基础设施 当然这不一定是我们团队上想做的事情等等 这个说出来好像不是有意思的事情 只是说只是商业上我觉得 是我们发现了一个点 然后另外一个点我觉得可以提一下 就这个也可能不是那么的有趣 但是我觉得是很重要的 对于agent这个事情 比如说我们其实在做clark 所以我们是主要的面向用户 我们未来做gold market的时候 我们去主要面向海外 包括这个产品 我们本身也是一个英文等等的产品 但因为我们内测期间 我们找了很多国内身边的朋友朋友 然后因此一些机会 就接触到一些国内 一些上海这种普通的一些白做 有些做广告 做当道的这种公司 这些小公司不设不了那种 然后呢我就去跟 因为他们是非常机缘巧合的 因为他们的朋友什么的朋友 正好用了我们产品 然后有些公司的老板就 这个很有意思 就开始用 然后我就有跟一些公司去聊 他们在用 他们都不是甚至 没有一个很清晰的概念 agent是什么 他们的AI可能没怎么用 就了解一下他们的需求 和他们日常的场景 我的一个基本的发现 就是今天很多广大的这种中小公司 非互联的 甚至可能到大一点的公司 他们能从agent上获得非常大的价值 这是我绝对1000%肯定 但是他们今天完全不知道怎么开始 以及他应该怎么去切入 做这个事情 即便是以腾讯这种 这种推法都没办法吗 在微信里面都集成 没有 就今天 就今天龙虾 比如说 龙虾比如说像这么火 对吧 毫无疑问 我妈都值得龙虾 然后那个 我前天去拜访一个公司的老板 小公司老板 不是互联网公司 他坐下问我 第一句话是 我不知道那个怎么读 现在那个大家 Hermes 应该是 总结读法是Hermes 不知道 就是他第一句问我 就是Hermes 这个agent你知道吗 但是他 但比如说 他肯定不是一个完全不懂技术的人 但是他肯定是一个非常普通的 非互联网圈子的人 就是他已经火到这个程度 但是其实 就是他在声量 在市场上 在自媒体 或者在社交媒体上 火的这个程度 跟大家实际在工作中的 adoption差距非常大 是 就只能这么说 肯定腾讯 腾讯有些产品还做得不错 或者说 豆包可能也还做得不错 但实际上我们认为agent 这个事情 你真正发挥出最大价值 比如说你我刚刚描述的 我们工作中一些粗浅的 使用agent的场景 其实他可以做非常非常多的 大效率的巨大提升 如果说众多的现在这种agent产品 其实那个cloud co-work 其实是一个在产品方向上 我觉得比较契合 或者说是在我认为 真正让普罗大众agent带入工作上 我觉得最契合的一个产品方向 因为他实际上就是相当于 用用户本地的 然后可以比较多元的操纵 各种各样的文档性 事物性的任务 就这里computer use 肯定会加强它的能力 但事实上co-work 就是一个除了代码之外 就是他那个slogan literally就是 claw for什么什么 besides coding 什么之类的 像我昨天去那个 我那个 我拜访那个 我用我们内测产品的公司 他给我展示一个case 让我印象特别深刻 就是我的一个用户 就是我的用户是 那个公司的一个人 然后他跟他的老板 以及他同时其他人 他非常骄傲的展示 他用claw for example 做了一个表格 which还是一个 很复杂的表格 他说他做了十分钟 但平时可能要做一个这样 就是因为claw for example 可以做到 co-work差不多的事情 然后但是我都 比如说我作为 开发claw for example 我都没有heavy的用到 我也不需要做表格 但因为事实上 我们做了这些基础的能力 其实这些基础的模型 有这个能力 就是给他的sing tool 所以他可以做这个事情 然后这个本质上 其实就是 就我觉得 可能回到刚刚那个点 就是我觉得 那个今天 大家的需求 其实我觉得很强 但是这个改变 因为他过于巨大 以及大家对于他能力的边界 是非常模糊的 以及产品形态 大家都还在 很多产品 绝大部分的产品体验 是很差的 所以我觉得今天还在 adoption的非常早期 所以我觉得今天 如果任何一个做agent的公司 都我觉得有非常大的机会 但是不同做 有不同做的方式 你可以有一个非常 触跳的产品定义 或怎么样 就是 当然我们也没有说 我们一定是做工作的 我只是说 我们其实一开始 并不是想做一个 以工作为主的 我们想做一个更personal 普通人能够用的 我们的一个产品 这是为什么 我们一开始 要做一个app的原因 但是今天如果直接看 强需求的 肯定还是工作相关的场景 虽然我们的内侧户 其实发现 他们用的是非常多 他们个人 帮他们管理健身 帮他们算命 还有个优户 帮他斗地主 帮他算牌 就是用户他会 他本质上就是一个 就把他当成一个 更聪明的一个朋友 或者说AI在用 但这些可能工作场景 在今天肯定还是 最有价值的 所以你们最后的产品形态 就很像Poke 我其实蛮喜欢Poke 这个产品的形态 然后它其实是 open cloud 是 但是去年 9月份发了 然后他今天也 他遇到他的挑战 就是 我认为他的挑战 肯定也是 如果你不是主打 这种生产力场景 因为我很难去付费 这个商业上 跟我刚刚说的那一点 是很相关的 如果你可以免费产品 无所谓 如果你可以烧钱也可以 但是无论这种产品 都不可持续 所以他最近也在做一些转型 他做了一个 Recipe的功能 其实就是 某种程度上 让普通人 有点像是 有一款还在内侧的 声势很大的产品 叫Mobby 就是让普通人 可以做一些Apps 然后他还做了一个 别人用的Apps 可以赚钱 类似这种 有点奇怪的功能 但是跟你也遇到一些问题 然后今天 我觉得有点像是我那天看一个 我那天看一个什么 但这个 我应该看了一些 早期的关于电脑 这个计算机发展历史的一些书 有点的转 我一下子想起来是哪一本 但就是 但跟我 就我看那本书的时候 就感觉到 跟我们今天在经历的时代很相似 就是很多事情 最开始的时候 都是以工作强生产力为主 特别是 我前上个星期 我还看了一个乔布斯在 应该是90年代 他在Next的时候的一个视频 他在介绍他当时的产品 然后他当时在他的白板上 画了一些他们要打的人群 什么律师 医生 等等等等 就是基本上 你就是在卖一个SaaS 或者说当时可能没有SaaS 这个概念 就基本上卖一个 非常productivity的一个产品 但到后面 其实整个电脑 大家都知道这个电脑 TikTok 手机 变得非常非常娱乐 我觉得就是 可能都会 经历这么一个历程 就今天其实 你我或者说 在跟AI 打交道 不会说什么 怎么说呢 就是更了解这种 相关的动态的人 他会知道 AI的capability 已经非常强了 但是普通人 接触到的 或者普通人体验到的 其实是非常大的 跟这个跟着天花板 有很大的gap 然后可能是 生产力强的需求 会先用上 然后长期可能 慢慢的回到娱乐 会有这么一个过程 我这个讲 这个好像有点像 什么导师 不是这个意思 我只是想一些看法 那个 我突然想 我其实想到 那个 你应该有看到 那个之前 应该是上上周 Kapathy Andra Kipathy 发了一个tweet 然后 就是 Andra Kipathy 他有一点 现在感觉变成了这种 怎么说 我那天打了一个 不太恰当的比方 我说他是 AI界的孔子 什么之类的 就是 这不是 听起来好像有点 不尊敬孔子 要不可以剪掉 但我的意思是 就他其实 说出了很多 非常有影响力的 观点和观念 在AI时代 从去年开始 比如说 Vibe coding 这个词 就是 Coin 对吧 然后到后面 他说 他说 他说 在那个YC的 那个演讲上 他说 这个不是 Agent 这个不是 The Year of Agents 这个是 The Decade of Agents 就他说了一些 非常有强有力的概念 然后 然后 然后在去年 12月底 那段著名的 这个Obers 这个所有人都体会到了 这个Vibe coding 的 魔力 的那段时期 然后他还把 这个 应该是他 我记得是另外一个人 在他的 Post下面 说 Obers 4.6 当时4.5 还是Clocko 是把它比做了 Gutenberg 那个印刷机 我有点记不清楚 我记得是另外一个人说 然后他转发了 就是 但是我有点差题了 我的意思是 上上中 Andrea Capati 发了一个Tweet 其实如果要找 可以找一下 他发了个Tweet 就是他发现 其实大部分的 普通人 他们用的 接触最多的AI 其实是TraderBT 最普通的模型 以及 Especially 就是TraderBT 那个Voice mode which是一个 比较 Legacy的模型 然后呢 但他们其实 就会有很多问题 甚至 取笑AI的问题 这个事情 在TikTok 或者说Instagram 上变成了一个mean 就大家会取笑AI 怎么那么蠢 但是实际上 真正AI的前沿 或者说4.6 4.7 或者这种 这新的东西 其实已经非常非常的 能够做非常厉害的事情 就是他在描述 这个非常巨大的gap 就是这个体感 是跟我的体感 非常非常像 所以我觉得今天 无论你做任何产品 或者做模型 我觉得都是一个 你站在一个 这个旅程的起点 就是有 如果从功利的角度看 你有非常多的机会 可以做非常 多伟大 或者是最终 最终有很大规模的产品 你现在在社交媒体上 也不是特别活跃 所以有什么 想关注你的人 可以去哪里找到你吗 或者是你们的一些产品 官网之类的 可以待会儿分享 分享到这个 shownotes里面 首先我也在寻找一个 我自己怎么样 因为我现在毕竟 我需要为我自己的产品 或公司去 说一些事情 所以我要从一个之前 那个 叫什么 隐世的 hermit的状态 变成一个 这种要说一些 要在社交媒体上 说一些事情的状态 所以我现在在自己 transition的过程中 然后我现在如果要关注 我觉得可以 主要关注 我先关注我 我在axe上和极客上 主要是这两个平台 其实今天这一期算是我们第一期播客 上一期跟王小光那一期 对跟王小光那一期 其实是一个音频的采访稿 我只是在写完文章之后觉得 好像扔了有点可惜 就放出来了 就一开始 因为CBVV是非常专业的主播 他一开始我们俩视频会议的时候 他就问我说 要不要就是叫什么 N2N这样录制 就是分轨录制 我说不用吧 应该 Famous last word 然后就开始了 结果我们俩的音轨 都是一条 都是混在一起的 而且剪的时候 虽然能剪 但是还是会有一些 我又比较喜欢 去抢人说话这样子 所以有很多地方是破碎的 也感谢大家的指证 在评论里面已经有人说出来了 所以我们也赤巨资购入了话筒 好吧 这就是本期的节目 感谢天路 然后大家可以在社交媒体上 关注天路 以及他们公司未来的产品发布 好 谢谢大家 非常荣幸能够参与到第一期播客 谢谢大家 拜拜 拜拜