Kimi Engineer Goes Viral in Female Attire: But He Prefers Discussing the Essence of AI

Kimi Engineer Goes Viral in Female Attire: But He Prefers Discussing the Essence of AI

2026-11-19 12:06

Introduction: Open-source multi-pathway evolution and fundamental benchmarking are the true drivers of model advancement. On October 18, Kimi engineer Tang Feihu appeared at the 2026 CloudBridge Conference hosted by Jiepei Chain, dressed in a full maid cosplay outfit.

October 18, Kimi engineer Tang Feihu made a striking appearance at the 2026 CloudBridge Conference organized by Jiepei Chain, donning a full maid cosplay ensemble. During a panel discussion on "Round Table π," he sported a long, cobalt-blue bang hairstyle, paired with a ruffled dress and meticulously crafted makeup.

Tang Feihu exudes a peculiar contradiction. While his name is Tang Feihu, colleagues and friends more commonly address him as Xiao Dao or Dao Niang. Months ago, he was photographed in women’s clothing during an interview. Recently, this video resurfaced online, catapulting him to viral fame and introducing many to Kimi’s “cosplay maestro” persona.

The benefits and drawbacks of sudden fame are both evident. He has gained increased visibility, frequently appearing at industry events, effectively becoming a walking ambassador for Kimi culture. In private conversations and online discussions, people constantly speculate about what kind of company allows such free self-expression in public settings.

The cosplay outfit Tang Feihu wore at the event was DeepSeek’s iconic whale costume. He explained that Kimi’s official cosplay gear had not yet been completed, so he couldn’t wear it. His home wardrobe also includes a DeepSeek whale suit and a Claude saintess outfit.

His rise to fame has positively influenced others. For instance, young fans who share his passion for female cosplay now have a stronger argument when negotiating with parents: “Living like him can still be successful.”

Of course, much of the attention stems from outsiders’ curiosity. Some Kimi users even told Tang Feihu their neighbors—a Chinese herbal medicine merchant—had started asking about him. Similar private messages arrive daily, averaging three to four per day.

This has become somewhat bothersome for Tang Feihu, who simply wishes to remain within his inner circle, “self-indulgent in his own niche.”

For over a decade, Tang Feihu has consistently worn women’s clothing and participated in cosplay, while also being deeply passionate about algorithms and competitive programming. He has taken part in or mentored dozens of international computer science competitions. When human-written code could no longer keep pace with AI, he stopped coding altogether and transitioned into a developer advocate role at Kimi.

Before the CloudBridge Conference, Tang Feihu spent nearly two hours speaking with us. He addressed some identity-related controversies and shared several unconventional insights about the large model industry: Harnesses aren’t as crucial as they’re made out to be; models should target deeper, more fundamental benchmarks instead of chasing scores in office-use scenarios. The most important aspect of open-source ecosystems is preserving diverse technical pathways.

In an era where everyone is racing to catch up with model iterations and secure market positioning, Tang Feihu seems to have preserved a part of himself that refuses to be tamed by work.

01

I’ve Been Biased by Survivorship

Baijing Lab: You recently went viral as a male coder in drag. Last year, Zhou Yu, an engineer at Dify, received widespread attention. There were earlier discussions around programmers wearing women’s clothes during the early days of mobile internet. Why has the topic suddenly gained so much traction now?

Tang Feihu: I’d love to know myself. The video was recorded months ago, and I have no idea why it suddenly went viral. It’s actually quite troubling for me.

Baijing Lab: Why is it troubling?

Tang Feihu: I don’t want to be seen by irrelevant audiences. This outfit has nothing to do with people outside my community. For example, (he shows us a WeChat chat log from a Kimi user whose neighbor—a Chinese herbal medicine vendor—asked about him). That kind of intrusion really bothers me.

I receive dozens of similar messages daily. It feels like I’m being stared at. Can you understand?

Most of the media coverage I don’t even want to open because I find it distorted and misleading. It’s hard to feel anything but powerless. Maybe that’s just how the internet works.

Baijing Lab: Looking at public news, you’ve been in women’s clothing for speeches since 2017, but lately it seems less frequent?

Tang Feihu: Not at all. I’ve always worn women’s clothing to events. For a trans person, isn’t this normal? In overseas communities, I never faced such issues. People don’t view it through a sensationalist lens.

Baijing Lab: Did you read any of the comments after your recent video went viral?

Tang Feihu: I haven’t dared to look. I just want to stay within my small circle, “self-indulging in my own corner.”

Baijing Lab: I checked comments on X—they’re generally positive. Some say it’s voyeuristic, others say they didn’t expect Kimi’s culture to be so inclusive. No wonder Kimi calls itself Moonshot.

Tang Feihu: In China, if someone has income and produces tangible results, society tends to be more tolerant. Recently, I’ve even seen kids with similar situations using my example to convince their parents. But these correlations aren’t particularly strong.

Baijing Lab: What kind of weak correlation are you referring to?

Tang Feihu: There’s no strong link between being transgender and working in computer science. I’m a victim of survivorship bias—can you see what I mean?

Baijing Lab: You once discussed the topic of yuri in a podcast with a former Google colleague. You just mentioned survivorship bias—does this mean there are many others like you?

Tang Feihu: We were talking about appreciating yuri-themed content. That has absolutely nothing to do with one’s sexual orientation or gender identity.

Recently, mathematician Deng Yu, who just won the Fields Medal, also enjoys yuri content. These are completely unrelated phenomena.

02

Every Large Model Has Its Own AI Maid

Baijing Lab: Actually, I’d like to ask—now DeepSeek’s community is popularizing “whale girls,” and AI maid narratives are trending. You and Zhou Yu may not have intended it, but this phenomenon clearly emerged. One interpretation is that the competition in the large model industry has become so fierce that companies resort to gimmicks to stand out?

Tang Feihu: That’s an interesting question to answer.

First, this is actually two separate issues: Zhou Yu’s case, mine, and the whale girl trope are entirely independent.

The “whale girl” phenomenon belongs to the broader trend of anthropomorphization in anime culture, which has existed for over a decade. Artists often transform abstract concepts into character designs based on their traits—could be female, male, or non-gendered. Anything can be anthropomorphized. For example, guns in games like *Azur Lane* or *Girls' Frontline* are portrayed as girls. Similarly, *Sword Art Online* features anthropomorphized swords. In China, many Ayou (members of the anime community) draw food items or cultural relics as characters. Even programming languages like C++ and Python have their own anthropomorphic avatars.

Anthropomorphizing large models fits perfectly within the established norms of anime culture. Last year at a convention, I even saw a male version of DeepSeek—but it didn’t go viral.

Turning formless large models into characters is creatively intriguing. Today, almost every major large model—except Dou Bao—has its own anime-style avatar. Merchandise like costumes and plushies based on these avatars are already being produced. In some cases, these fan-made versions spread wider than official branding.

This reflects another cultural trend: meme culture. The “whale girl” image has become a meme. So has Huang Renxun.

Baijing Lab: Why do they become memes? Why hasn’t Kimi become one?

Tang Feihu: Kimi has memes too—plenty exist on Bilibili. It’s just that DeepSeek is more mainstream, and their Flash version is affordable and widely adopted. More users = broader meme propagation.

Baijing Lab: At the CloudBridge Conference, you wore a whale suit. As Kimi’s Developer Relations lead, why wear a DeepSeek whale suit? (laugh)

Tang Feihu: Because Kimi’s outfit isn’t ready yet—it’ll take another month. I don’t just have the DeepSeek whale suit; I also own a Claude saintess outfit.

Baijing Lab: So back to the original question—do you think this is a result of intense competition in the large model industry, forcing companies to innovate just to gain visibility?

Tang Feihu: Not at all. I was just being myself.

Baijing Lab: Why is it going viral now?

Tang Feihu: It’s not my doing. I’ve always been like this—nothing’s changed. Recently, someone I met years ago at a convention reached out again after seeing my viral moment. He said, “I can’t believe it’s been over ten years, and you’re still active in anime culture.”

03

Harness Isn’t That Important

Baijing Lab: This is another question I find fascinating. Many people change as they age, sacrificing personal interests and hobbies to prioritize work—especially in such a fiercely competitive field as large models. Yet you seem remarkably carefree.

Tang Feihu: The core reason is that since transitioning into DevRel, I no longer touch production. My stress level is significantly lower than my colleagues’. However, if you're an engineer aiming for faster, better growth, touching production is still essential. Engaging with live systems inevitably leads to mistakes and pressure from potential losses to the company.

Baijing Lab: You joined Kimi around 2024. Before that, you were deeply involved in blockchain applications. Why choose Kimi?

Tang Feihu: I wanted a place with abundant compute power for large model training. I interviewed at several companies, but only Kimi responded quickly—so I joined.

Baijing Lab: Why shift to Developer Relations later?

Tang Feihu: Partly due to the immense pressure on engineers. When we launched our open-source large model, we needed someone to manage developer relations—a role no one else was filling. I stepped in.

Baijing Lab: How intense was the pressure?

Tang Feihu: You can reference the recent viral ByteDance poisoning incident. Our situation differs, but the parallels are instructive.

Baijing Lab: When did you start doing Developer Relations?

Tang Feihu: Around the time K2 was released.

I’ve been learning while doing it. Large models evolve rapidly, bringing new technologies and countless community issues. Answering questions means facing new problems daily. It’s a continuous process of learning and adapting alongside the community.

Our community hosts many exceptionally talented users. Their questions and solutions built with large models offer me constant inspiration. Interacting with such skilled individuals makes me grow stronger too.

Baijing Lab: Can you give an example?

Tang Feihu: One user ran out of tokens and tried every trick to conserve them. Using edge devices and deploying a Subagent via K3, he achieved performance close to fine-tuning a large model directly. He implemented numerous inference optimizations.

Baijing Lab: How do you measure the success of Developer Relations? Is DeepSeek’s developer ecosystem more mature?

Tang Feihu: Not necessarily. Measuring developer ecosystems via KPIs is difficult. Developer Relations is inherently intangible—focused more on service quality and reputation. Domestically, Shenwang stands out. Internationally, Amazon leads.

For smaller domestic firms, investing in Developer Relations is often seen as extravagant. High cost, delayed returns. Few companies offer such roles. Only after K3’s release, with growing user base and community foundation, did we begin systematic developer engagement—like the recent Kimi Community Ambassador initiatives in Shenzhen and other cities.

Baijing Lab: As Kimi’s evangelist, have you noticed any new trends in interactions with developers?

Tang Feihu: After K3’s launch, user queries and needs surged. With model capabilities evolving monthly, the nature of questions changes constantly—including shifts in how Harness is used.

Even before DeepSeek launched Harness, our community had users leveraging Codis architecture and similar plugin-based approaches. You see how usage patterns evolve, yet the core goals remain consistent.

Baijing Lab: When DeepSeek Harness first launched, it sparked controversy. Many didn’t grasp its business model. Some thought it would empower niche geeks to build vertical agents. Now, most users are content creators building plugins. What’s your take?

Tang Feihu: Plugin-based models aren’t unusual—most software industries operate this way.

From a community developer perspective, my core belief is that open-source diversity fosters more product forms and categories—benefiting everyone in the ecosystem.

Baijing Lab: From a model vendor standpoint, Kimi launched managed agents on November 5, essentially offering enterprises a pre-built Harness foundation. This seems like a different path from DeepSeek’s Harness.

Tang Feihu: In the large model era, you need specific plugins or skills? Just ask the model—it can install them. There’s no need for elaborate frameworks. What you build today might be obsolete tomorrow. The real challenge is simplification. For example, in Codex’s latest update, they removed many previously included components.

Building a Harness is extremely difficult because you’re fighting against a dynamic system. Designs valid today may fail tomorrow. Take Kouzi—it’s a form of Harness. Many of its design choices now seem redundant, so it’s being phased out.

Baijing Lab: Do you favor DeepSeek’s Harness approach?

Tang Feihu: Offering users choice is good enough. If all tech paths converge toward homogeneity, it’s not a healthy ecosystem.

Different technical routes aren’t inherently superior or inferior. Even if differences exist, they’re not yet discernible at this stage.

Baijing Lab: DSH seems more focused on serving developers, almost indifferent to commercialization.

Tang Feihu: Expecting Harness to generate profit? Harness is pure public good.

Baijing Lab: How does Kimi’s Harness differ from DeepSeek’s?

Tang Feihu: I don’t think we should compare ourselves directly with DeepSeek. Each company has unique circumstances. Current differences are merely technical route choices—not yet indicative of superiority. Harness evolves rapidly. Many users now build custom Harnesses using Kimi Code, tailoring them to their needs. Customized Harnesses likely better meet user demands.

But is a one-size-fits-all solution truly optimal? Kouzi has many plugins, yet it’s nearing removal. Judging such tools is hard. A single model update can break multiple plugins—far from ideal.

Baijing Lab: Why did Harness suddenly go viral? What does it mean for model vendors?

Tang Feihu: It enables better alignment with proprietary models, unlocking superior performance.

Frankly, I don’t think Harness is as pivotal as the community suggests. It’s a transient, evolving transitional solution.

Baijing Lab: Models are increasingly enhancing long-horizon task handling and multi-agent collaboration. Does this reduce the space for Harness?

Tang Feihu: Yes. I’ve made several predictions, all correct. First, I don’t consider Harness critically important—people overestimate its significance. Second, Harness will persist because it can always improve end-user experience, whether by saving tokens or better constraining behavior. Third, Harness must continuously evolve alongside advancing models. Heavy-handed design is unwise. Developers must exercise restraint and embrace minimalism.

From an end-user perspective, more choices are better. Today you face Scenario A, tomorrow Scenario B. Having ready-made options lets you immediately evaluate and select the right Harness via an evaluation framework. More choices are always preferable.

04

Focusing on Product-Level Benchmarks Is Short-Sighted

Baijing Lab: Currently, model capabilities are evolving toward Agent-driven long-horizon tasks. What trends do you anticipate?

Tang Feihu: We focus more on foundational benchmarks. Improving these deep-level metrics naturally enhances agent performance as a side effect.

Baijing Lab: What do you mean by foundational benchmarks?

Tang Feihu: Making models smarter. OpenAI tests unsolved mathematical problems to demonstrate intelligence. A-factory experiments with drug discovery. These benchmarks already surpass human capability. Only achievements here carry real meaning for models. Product-level benchmarks—like those in office automation—are short-sighted.

Baijing Lab: Are we now competing on AI-for-Science benchmarks?

Tang Feihu: Yes. We should seek truly significant benchmarks, not confine ourselves to product-level metrics. If you excel at these fundamental benchmarks, the product-level ones will naturally improve—indicating genuine model advancement. That’s true generalization across benchmarks.

You’ll notice most benchmarks plateau at 63–64 points, revealing no meaningful gaps. The real core differences in models aren’t well captured by current benchmarks.

Baijing Lab: What about the AA leaderboard, widely cited when models launch?

Tang Feihu: I think these benchmarks are too weak.

Baijing Lab: Now each model vendor has internal benchmarks in new scenarios, though they’re not publicly disclosed?

Tang Feihu: Each vendor likely has better benchmarks, though I don’t know specifics. If Kimi’s benchmark improves from 30 to 40 points, and all other benchmarks simultaneously rise—that would indicate a more fundamental benchmark, as it impacts the entire ecosystem.

05

Global Models Are in Cold War Heat

Baijing Lab: Since part of your role involves training enterprise clients, what do you observe as key factors influencing enterprise model selection this year?

Tang Feihu: Primarily model capability ceiling, cost, and inference speed.

Baijing Lab: Some argue most jobs don’t require high-capacity models. This year, many domestic models launched “Flash” variants. Why hasn’t Kimi released a Flash model?

Tang Feihu: Almost ready. Kimi emphasizes output capacity, but that’s ultimately a cost-performance trade-off. A cheap token price doesn’t matter if the model keeps failing, requiring repeated inference. Eventually, costs accumulate. Once enterprise procurement reaches scale, cost accounting becomes unavoidable.

Baijing Lab: This seems to validate your 2023 April prediction on X: future AI will split into two types—one controlled by big corporations (like OpenAI, Google), the other open-source and locally deployed. You predicted a war between them. Where are we now?

Tang Feihu: The conflict is now boiling over—so much so it brought down Huang Renxun. (laugh) This is cold war heat. The downfall wasn’t just media narrative; fundamentally, Huang created a Twitter account supporting open-source models and turned a blind eye to chip sales to China—leading to his fall.

Huang Renxun is now frontline in this cold war heat.

There’s a board game called *Cold War Heat*. Players represent the US and USSR, not fighting directly but deploying proxies across influence zones. This mirrors efforts to embed influence within companies like NVIDIA—identical dynamics.

You could even use large models to simulate this board game, designing a chip war strategy game—possibly even generating it instantly for others to play.

(Supplementary context: After our conversation, Huang Renxun’s official X verification was revoked. Elon Musk, who supported Huang, promptly restored it for free.)

Baijing Lab: How did the Huang Renxun incident affect domestic model vendors?

Tang Feihu: Impact remains unclear, but it definitely benefits domestic chips. Some proposed restricting NVIDIA chips to closed-source models—a ridiculous idea. To developers, it’s akin to banning cryptography algorithms.

Baijing Lab: Have you met Huang Renxun?

Tang Feihu: I’ve attended several of his live talks. He’s like Steve Jobs—an visionary leader. Without his relentless pursuit of GPUs and computing centers, none of today’s breakthroughs would exist.

For example, Anthropic helped researchers accelerate cancer drug development from 10 years to 3. OpenAI proved over 300 mathematical theorems. Shandong University published a paper proving Goldbach’s conjecture under Riemann hypothesis assumptions. Behind these advances lies Huang Renxun’s foresight. We need more leaders like him—visionaries driving progress.

06

Learn to Use Models as Levers

Baijing Lab: You and Yao Shunyu both graduated from Hefei No.1 High School. Have you met since?

Tang Feihu: Probably not. I was in the science experimental class at Hefei No.1. Now, all my classmates span the entire large model ecosystem—from investment and compute infrastructure to tokens, models, even OpenAI.

Baijing Lab: You were admitted via recommendation to Harbin Institute of Technology, unlike Yao Shunyu, who entered Tsinghua through direct exams.

Tang Feihu: I didn’t want to attend school—I was rebellious. Now I regret it. I should’ve studied harder.

Baijing Lab: You were indeed rebellious—you dropped out after three months at HIT. But you carved a non-mainstream path and succeeded. Today’s youth have few upward mobility channels, making your journey seem luxurious.

Tang Feihu: Back then, I rode many waves of opportunity. Job hunting was easy—before graduation, I had five or six offers. Now, the rise of large models has reduced entry-level positions. Fresh graduates struggle to land their first job. But this group is precisely the most dependent on that first role—creating a steep career curve. It’s a very unhealthy trend. I don’t know how to fix it.

Baijing Lab: This trend began last year.

Tang Feihu: It’s worse this year.

Baijing Lab: Recently, we reported that many top-tier university CS students with large model training or infrastructure experience are receiving fewer offers this year. How should today’s youth respond to the AI-driven job market squeeze?

Tang Feihu: The core skill now is leveraging large models as levers, regardless of your field. Don’t let models dominate your thinking—don’t surrender decision-making to them. Even in model training, you must master how to use large models: optimizing operators, building evaluation frameworks, judging model capabilities.

Specifically, actively use AI to identify tasks AI cannot replace—downplaying execution skills. Most problems aren’t unsolvable by AI; they’re just expensive. This is a temporary phase. Future models will make tokens cheaper.

Second, domain expertise remains vital. Can you, like Terence Tao, use GPT-6 to discover and prove mathematical theorems? Probably not—without specialized knowledge. You’ll see models act as levers, widening the gap between experts and novices.

Finally, explore widely and build rich knowledge structures. I competed in many contests. I once thought the algorithms I learned would never apply to my career. But during a model training bottleneck, I applied contest-specific algorithms and data structures—solving it instantly. Helping colleagues with what I knew brought me great joy. Ironically, the knowledge I use most now came from those competitions.

Source: Baijing Lab

#Large Model

Disclaimer: Contains third-party opinions, does not constitute financial advice

Share To
X
Telegram
WeChat
QQ
Link
Recommended Reading

The AlphaGo Moment in Cybersecurity, or Starting from These 466 Agent Matchups

14 days ago
The AlphaGo Moment in Cybersecurity, or Starting from These 466 Agent Matchups

AI is not a search bar—it's a supercomputer costing just a few dozen dollars

08-05
AI is not a search bar—it's a supercomputer costing just a few dozen dollars

Anthropic Locks in 6 Years of Compute with $10 Billion Commitment

08-05
Anthropic Locks in 6 Years of Compute with $10 Billion Commitment

Not building robots—why is it worth $14 billion?

07-31
Not building robots—why is it worth $14 billion?

AI is no longer competing on benchmark scores, but on profitability

07-25
AI is no longer competing on benchmark scores, but on profitability

ByteDance's Douyin Beans priced at 68 RMB—worth it for 382 million monthly active users?

07-25
ByteDance's Douyin Beans priced at 68 RMB—worth it for 382 million monthly active users?

Fields Medalist Concerned About AI Extinction Heads to OpenAI

07-25
Fields Medalist Concerned About AI Extinction Heads to OpenAI