HomeArticle

Tang Feihu, an engineer of Kimi, appeared at the Cloud Bridge Conference: Wearing women's clothing is just my daily routine.

白鲸实验室2026-11-19 11:26
AI has widened the capability gap between ordinary people and professionals.

On October 18, Tang Feihu, an engineer at Kimi, appeared at the 2026 Cloud Bridge Conference hosted by Jiepei Chain in a maid outfit cosplay look. During the multi-person discussion in the Round Table π session, he wore long cobalt-blue curly hair with straight bangs, a ruffled long dress, and had very delicate makeup.

There is a wonderful contrast in Tang Feihu. Compared with the name Tang Feihu, his friends and colleagues are more used to calling him Xiaodao or Daoniang. A few months ago, Tang Feihu was interviewed wearing women's clothing. Recently, this video was unexpectedly dug up, making him quickly go viral online, and letting more people know this "cross-dressing guru" at Kimi.

Both the benefits and troubles of rising to fame are very obvious. He has obtained more exposure opportunities, and his frequent appearances at industry conferences make him like a walking cultural spokesperson for Kimi. In private occasions and online discussions, many people can't help but be curious about what kind of company would allow employees to show themselves so freely at public events.

The cosplay look Tang Feihu wore today is DeepSeek's iconic whale costume. He told us that Kimi's cosplay costume was not ready yet, so he didn't have time to wear it. At home, he not only has the DeepSeek whale costume, but also the saint costume for Claude.

His rise to fame has also had a positive impact on some people. For example, teenagers who are as obsessed with women's clothing cosplay as him have an extra bargaining chip to negotiate with their parents, that "you can also live a good life if you live like him".

Of course, most of the extra attention comes from the novelty-seeking psychology of people outside the circle. Even a Kimi user told Tang Feihu that his neighbor, a traditional Chinese medicine merchant, also began to inquire about his news. He receives three or four similar private messages every day.

This is a little troubling for Tang Feihu, who only wants to "enjoy his hobbies in his small circle".

Tang Feihu has been wearing women's clothing and doing cosplay for more than ten years, and he is also obsessed with algorithms and competitions. He has participated in and coached no less than dozens of international computer competitions. After the code handwritten by human programmers can no longer catch up with AI, he stopped writing code and started working at Kimi to serve developers.

On the eve of the Cloud Bridge Conference, Tang Feihu talked with us for nearly two hours. He responded to some identity controversies, and also had many non-mainstream judgments about the large model industry: Harness is not that important, models should challenge more essential Benchmarks instead of indulging in score competitions in office scenarios. The most important thing for the open source ecosystem is to retain enough different technical routes.

In an era when everyone is busy catching up with model iterations and competing for industry positions, Tang Feihu seems to always retain a part of himself that is unwilling to be domesticated by work.

01

I am affected by survivorship bias

White Whale Lab: Recently you became famous as a cross-dressing guru. Last year, Zhou Yu, an engineer at dify, attracted widespread attention. In fact, the topic of programmers wearing women's clothing was discussed back in the early days of the mobile Internet. Why has the discussion suddenly become so high again now?

Tang Feihu: I also want to know. Because that video was recorded a few months ago, I have no idea why it was suddenly forwarded everywhere. Actually, I am very troubled.

White Whale Lab: Why are you troubled?

Tang Feihu: I don't want to be seen by irrelevant circles. This look has nothing to do with people outside the circle. For example, look at this (he showed us the WeChat chat records on his mobile phone, which was the chat between a Kimi user and him. It turned out that the user's neighbor, a traditional Chinese medicine merchant, asked the user for information about him), this kind of thing disturbs me a little, which makes me very troubled.

I receive many similar messages every day, and I feel a sense of being stared at, can you understand that?

I don't want to open most of the reported content, because I think they are all distorted, and I feel a little powerless. Maybe the Internet is just like this.

White Whale Lab: I read in public news that as early as 2017 you gave a speech wearing women's clothing, but it seems that you have attended events in women's clothing less in recent years?

Tang Feihu: No, I have been wearing women's clothing to attend events all the time. For a trans person, isn't this (dressing) very normal? When I was overseas, I didn't encounter so many troubles. People there won't look at it with a novelty-seeking perspective.

White Whale Lab: Did you read the comments from netizens on the video that made you famous later?

Tang Feihu: I dare not look at it. I just want to enjoy my hobbies in my small circle.

White Whale Lab: I read the comments on X, and I think the overall evaluation is not bad. Some people say it's not just novelty-seeking, others say they didn't expect Kimi's culture to be so inclusive, no wonder Kimi is called moonshot.

Tang Feihu: In China, if a person has an income and can deliver results, people may be more tolerant of him. Recently I also saw that children in similar situations to me will use my example to persuade their parents. But in fact, these things do not have a strong correlation.

White Whale Lab: What did you say does not have a strong correlation?

Tang Feihu: There is no strong correlation between being trans and working in the computer industry. I am affected by survivorship bias, can you understand that?

White Whale Lab: You and your former Google colleague also recorded a podcast about lily topics before. You just said you are affected by survivorship bias, are there many people in your situation?

Tang Feihu: We talked about appreciating lily-themed works, which has nothing to do with your sexual orientation or gender identity.

Deng Yu, the mathematician who just won the Fields Medal recently, also likes to watch lily-themed works. These are completely unrelated things.

02

Every large model has its own AI girl

White Whale Lab: Actually, I also want to ask that now the Deepseek community is also popular with the whale girl narrative, and the AI girl narrative is also becoming popular. Including you and Zhou Yu, it may not be deliberate, but on the other hand, it did happen. There is an interpretation that the competition in the large model industry is so fierce that people can only rely on these fancy tricks?

Tang Feihu: This question is very interesting to answer.

First of all, this proposition contains two separate things: Zhou Yu's case, my case, and the whale girl are two completely independent matters.

The whale girl belongs to the anthropomorphism phenomenon in the two-dimensional culture, which has existed for more than ten years. When illustrators see abstract things, they will draw them into anime characters according to their characteristics, which can be girls or men. Everything can be anthropomorphized, for example, guns, the domestic games Azur Lane and Girls' Frontline are anthropomorphism of guns. And Touken Ranbu is the anthropomorphism of cold weapons. There are also many otakus (the name for members in two-dimensional communities) in China who draw food or cultural relics into anthropomorphic images. Even programming languages such as C++ and Python have their own anthropomorphic images.

Turning large models into girlish anthropomorphic images is very consistent with the inherent habits of the two-dimensional community. Last year at a comic exhibition, I even saw someone cosplaying the male version of Deepseek, but this image did not go viral widely.

Anthropomorphizing large models, which have no specific form, is very interesting from the perspective of creation. Now almost all large models, except Doubao, have their own two-dimensional images. Some stores are also making costumes and dolls of these images, which may spread more widely than the official images.

This in turn reflects another cultural phenomenon, the meme culture. The whale girl image has become a meme. Jensen Huang has also become a meme.

White Whale Lab: Why did they become memes? Why didn't Kimi become a meme?

Tang Feihu: Kimi has its own memes too, you can find them in many videos on Bilibili. It's just that DeepSeek is more well-known to the public, and its Flash version is cheap so it has more users. The more people use it, the wider the corresponding meme spreads.

White Whale Lab: You also wore a whale suit at the Cloud Bridge Conference. As Kimi's head of developer relations, are you wearing DeepSeek's whale suit? (laugh)

Tang Feihu: Because Kimi's costume is not ready yet, it will take another month. I not only have the DeepSeek whale suit, but also Claude's (saint costume).

White Whale Lab: So back to the original question, do you think this is because the competition in the large model industry is so fierce that everyone needs to play with new things to get a sense of presence?

Tang Feihu: I'm not doing that, I'm just being my true self.

White Whale Lab: Why did you become popular recently?

Tang Feihu: It's not my problem. I've always been like this and never changed. A very interesting thing happened recently: a person I met at a comic exhibition more than ten years ago, who I haven't talked to since then, saw me after I became famous, added me as a QQ friend again, and sighed that he didn't expect that after more than ten years, I am still active in the two-dimensional culture.

03

Harness is not that important

White Whale Lab: This is also a question I am very curious about. I think many people around me will change as they grow older, giving up their interests, lives and hobbies to work, especially in a highly competitive industry like large models, but you seem very free and easy.

Tang Feihu: I think the core point is that after I switched to DevRel, I don't need to touch production, so my pressure is actually less than that of my colleagues. But if you are an engineer and want to grow faster and better, you still need to touch production. Because when you are exposed to the production environment, you may make mistakes, and face the pressure of causing losses to the company.

White Whale Lab: You joined Kimi around 2024, right? Before that, your research enthusiasm was always on blockchain applications. Why did you choose Kimi?

Tang Feihu: Because I wanted to find a place with computing power to do large model training. I interviewed several companies at that time, and only Kimi gave me a reply quickly, so I joined.

White Whale Lab: Why did you later switch to developer relations?

Tang Feihu: On the one hand, the development pressure on engineers was too high. At that time, we were working on open source large models, and we needed a developer relations maintenance role to do this work. It just so happened that no one was doing it, so I took the job.

White Whale Lab: How high was the pressure?

Tang Feihu: You can refer to the very popular ByteDance poisoning case some time ago. The situation we encountered is different, but you can use it as a reference.

White Whale Lab: When did you start doing developer relations?

Tang Feihu: Probably when K2 was released.

In fact, I was learning while doing developer relations. Large models change very fast, and it is new technology, so the community accumulates a lot of problems. You have to answer their questions, and new problems keep emerging. Some problems you see for the first time make you think "it can be done this way", it is a process of constantly learning and growing with the community.

Many users in our community are really very excellent. The problems they raise and the things they make with large model methods bring me a lot of inspiration. When you communicate with these excellent users, you will also become very good.

White Whale Lab: Can you give an example?

Tang Feihu: For example, one of our users didn't have enough tokens, so he tried every means to save tokens. He used his own end-side device and the K3 model to build a Subagent, which performed very well, almost as good as the performance he tuned for the large model himself. He did a lot of reasoning optimizations in it.

White Whale Lab: How to measure whether developer relations is done well? Is DeepSeek's developer ecosystem more mature?

Tang Feihu: Not really. It is very difficult to measure the KPI of the developer ecosystem, and developer relations is very difficult to quantify, because it is more about services and word of mouth. Currently, DeepWeb is doing well in China, and Amazon may be doing the best abroad.

In fact, in China, especially for small companies, doing developer relations is a luxury. It requires a lot of investment, and it is difficult to see immediate results. In fact, not many companies in China provide such positions. It was not until after the release of the K3 model that we had more users and community foundation, and began to systematically carry out developer activities, including the recent Kimi community ambassador activities in cities like Shenzhen.

White Whale Lab: As Kimi's evangelist, have you found any new changes in the process of contacting developers?

Tang Feihu: After K3 was released, users' problems and demands have increased. And because the model capabilities are changing, users' problems have new changes every month, including the way Harness is used is also changing.

In fact, before DeepSeek released Harness, some people in our community used the Codis architecture it depends on, and used the model in a similar plugin way. You will also find that the way people use the model is changing, but the problems they want to solve are the same.

White Whale Lab: When Deepseek Harness first came out, it was very controversial. Many people couldn't understand its business model, thinking that if it was used by geeks, they would build their own vertical Agent. Now it may only be some content bloggers making some plugins. What do you think of this?

Tang Feihu: The form of making plugins is not uncommon, many software industries do it this way.

From the perspective of community developers, my core idea is that the open source community should be more diverse, and having more product forms and categories will benefit everyone in the entire community.

White Whale Lab: From the perspective of model manufacturers, Kimi launched hosted agents on November 5, which is equivalent to directly providing enterprises with a set of Harness base. It and Deepseek Harness seem to be two different routes.

Tang Feihu: In the large model era, for what plugins and skills you need, you can directly ask the model to install them. You don't need to make a big fuss around plugins, the things you make today may be eliminated tomorrow. I think it is more important to learn to subtract. For example, in a recent update of the Codex version, a lot of original content was actually deleted.

So making Harness is actually very difficult. Because you are fighting against something that changes dynamically. Many things you design today may become invalid tomorrow. For example, Coze, which is also a kind of Harness, many of its designs now look a bit redundant, so it is on the way to be taken off the shelves.

White Whale Lab: Are you more optimistic about DeepSeek Harness's approach?

Tang Feihu: I think it's good to provide users with different choices. If everyone's technical routes tend to be homogeneous, it is not a good ecosystem.

These technical routes are different, and I don't think any one is better or worse. Even if there is, we can't tell at this point in time.

White Whale Lab: DSH seems to prioritize serving developers first, and doesn't seem to care about commercialization.

Tang Feihu: Do you still expect Harness to make money? Harness is purely for public welfare.

White Whale Lab: What is the difference between Kimi's Harness and DeepSeek's Harness?

Tang Feihu: I don't think we can compare us with Deep