Interview with Xiaobai Founder Li Yan: AI Is a Brute-Force Aesthetic; Small Can't Be Beautiful
How to engineer and harness emergence is the defining thread of our era.
"How to engineer and harness emergence is the defining thread of our era." By Lili Yu
On a sweltering Sunday afternoon, in Tsinghua Science Park's QiDi Building in Beijing's Haidian district, I met Li Yan, founder of Wen Xiaobai, for the first time.
In conversation, he carried an almost excessive politeness. "I'm an introvert immersed in technology and product, so I'm not particularly adept at external dealings," he explained.
He said he'd prefer his product to speak for itself, while he stays behind the scenes. But clearly, that moment hasn't arrived. As a founder, he needs resources, and that means learning to persuade others.
Li Yan, born in the mid-1980s, studied at Huazhong University of Science and Technology's Software School and later at the Institute of Computing Technology at the Chinese Academy of Sciences. After graduating, he joined Tencent, then in 2015 moved to Kuaishou — then still operating out of a residential building with roughly 30 people.
He was a core architect of Kuaishou's AI system. With the support of then-CEO Su Hua, he built Kuaishou's first deep learning department, which later evolved into the MMU (Multimedia Content Understanding) division, responsible for multimodal model R&D spanning text, speech, and image.
In 2023, Li Yan joined the AGI startup wave. Among the many entrepreneurs at the time, some investors believed his company, Yuanshi Technology, was the only team capable of integrating the three core technical stacks: search, recommendation, and multimodal.
Over two years, Yuanshi raised tens of millions in seed funding from Kuaishou co-founder Su Hua, Redpoint Ventures, and Matrix Partners China, followed by a second round led by INCE Capital, totaling approximately $50 million.
In July 2024, the product Wen Xiaobai quietly launched. In form, it combines active Q&A chatbot functionality with passive feed scrolling, with a substantial portion of the feed content AI-generated. Consequently, some investors have likened Wen Xiaobai to "the Toutiao of the AI era."
In May 2025, Yuanshi released Agent Xiaobai Research Report, positioned as a counterpart to Deep Research.
By 2025, a consensus among most early-stage AI investors had formed: general-purpose agents or the domestic consumer market would inevitably become the main battlefield for tech giants, with startups struggling to break through.
In Li Yan's eyes, that day clearly hasn't come. "This market is large enough that the giants can't capture all the weight — there will inevitably be huge gaps left."
Notably, in the first half of 2025, Wen Xiaobai — after integrating DeepSeek — repeatedly appeared on lists such as QbitAI's AI assistant web traffic rankings for China and the AI Product Global Growth Rankings. Li Yan believes this wasn't merely DeepSeek's spillover effect. "There's a greater force at work."
In our conversation, he spoke more about this greater force. In a sense, it's Li Yan's obsession, the source of his entrepreneurial courage, and the miracle he witnessed at Kuaishou.
A persistent and fascinating question accompanying the AI era is re-examining the inertia left by the mobile internet era. In a sense, Wen Xiaobai — combining characteristics of both eras — serves as a compelling case study for this question.
The conversation follows.
Part 01
"The internet celebrities of the AI era definitely won't be the same people from the last era"
"Dark Surge": What's Wen Xiaobai's positioning? In some product rankings, it's categorized as a chatbot; in other descriptions, as AI search.
Li Yan: By our definition, it's neither an AGI story-telling general-purpose LLM company nor an upgraded AI search, but rather an AIGC content platform.
The previous generation of content platforms — Douyin, Kuaishou, Bilibili, Xiaohongshu — we call UGC. Wen Xiaobai is an AIGC content platform with large model creative capabilities added.
Users can actively acquire content through asking on our platform, or passively acquire content by scrolling feeds.
"Dark Surge": When actively seeking information, why emphasize "asking"?
Li Yan: We believe search is a product of the previous generation, while asking represents a more advanced one. The interaction between search and users is more like issuing a command to a computer — very flat. Asking emphasizes co-creating solutions with users; it's three-dimensional.
Search is also a high-threshold interaction. It requires users to learn to translate needs into computer language. Asking is human instinct — children ask questions before starting school. It's an information acquisition method closer to human nature.
"Dark Surge": How does this Q&A differ from products like Doubao, Yuanbao, or DeepSeek?
Li Yan: The so-called "model as product" approach of large models creates massive information hallucination. For ordinary users, you often don't know which parts of the information are true and which are false. But because we position ourselves as a content platform, we emphasize information accuracy, doing extensive retrieval and cross-validation on top of the model — it's not simple model usage.
On the asking front, we've also been quite aggressive. Recently, we're pushing an auto mode where users don't choose between deep thinking or web search. Just put your need in the box, and we'll intelligently match the optimal mode.
Because for ordinary users, it's like a cockpit full of buttons — overly complex. Whether a question is better suited for deep thinking or web search — even I get confused sometimes.
So much innovation isn't earth-shattering. Often, it's simply targeting different users, leading to different UI.
"Dark Surge": Is tilting toward ordinary users a deliberate design choice for Wen Xiaobai?
Li Yan: Look at the name we chose — it's designed for ordinary people. Wen Xiaobai, readable in one go, memorable, with zero psychological barrier to use. The little white (xiaobai) comes at your beck and call. We advocate a "xiaobai mindset" — no matter how powerful AI becomes, it must serve users, without pretension or showing off.
"Dark Surge": You mentioned that beyond asking, passive reading involves scrolling feeds. What's different between the mobile internet era and the AI era for feeds?
Li Yan: Compared to past UGC, our biggest difference is in business logic and product form.
Past UGC was organized more around producer-friendly interfaces, with many tabs. We organize around consumer-friendly interfaces. We consolidate information on the same theme better, letting users consume gossip more quickly. The information efficiency is different.
"Dark Surge": From producer to consumer — this represents a shift in discourse power.
Li Yan: In the past, producers were scarce, so platform business and product logic had to protect producer interests more, leaving them fixed positions. But this also burdened consumers.
The AI era blurs the relationship between producers and consumers. Producers are no longer so scarce, making it more appropriate for our new-generation platform to organize information around consumers.
"Dark Surge": How is your content currently produced? Mostly AI-completed?
Li Yan: You can understand our content production as a "multi-agent pipeline." We've pre-built complete instruction chains for scenarios like news quick commentary, deep research reports, and interactive Q&A. Core agents automatically assemble prompts, calling retrieval, reasoning, writing, visualization, and other capability components, completing multiple rounds of self-evaluation and correction internally.
Agents handle 90% or more of generation and quality control. Human experts intervene at critical nodes like topic selection, values verification, and style consistency, playing the role of orchestrator and gatekeeper. Additionally, user questions, feedback, and co-created content feed back into the system in real-time, becoming data soil for the next generation of content.
Our Xiaobai Research Report, launched in May, is an agent product benchmarked against Deep Research. Compared to many large models that return results after searching just a dozen web pages in a short time, we spend ten minutes searching hundreds of information sources before returning, finally generating a图文并茂 article through an HTML code engine.
"Dark Surge": What role does Xiaobai Research Report play in your product? Generally, research content isn't particularly oriented toward ordinary people.
Li Yan: In a sense, it's more like an engine that produces and creates for you.
It's generally assumed that only master's or PhD students or consulting firm experts need to write industry research reports. But ordinary people also have research needs.
For example, users often ask: I want to open a wellness studio, where in Chengdu would be suitable, what's the foot traffic like? We welcome elite users, but we haven't optimized much specifically for them.
Xiaobai Research Report, chat, and feed are all different modules of Wen Xiaobai. I feel we're building a product with a data flywheel, where each module also helps others improve.
"Dark Surge": Besides the producer-consumer relationship shift you mentioned, what else is different about feeds in the AI era?
Li Yan: Token-granularity personalization. Past recommendation systems were collaborative filtering based on user behavior. It didn't understand you or the article; it only saw whether you'd had cross-relationships. But large models deeply understand you and the article, using logical derivation to build connections between you, rather than based on past behavior. So there's new depth to personalized reading.
On creator-consumer interaction, there's currently another pain point: when co-occurrence samples are few, creator works at cold start can't find the most resonant consumers to interact with due to inaccurate distribution. This mechanism causes creators to gradually屈服于 traffic rather than staying loyal to good content itself. Large models' next-token prediction can do token-granularity causal inference, so distribution matching accuracy is significantly improved.
Also, the connection bandwidth between creators and consumers has become larger. In past products, producers couldn't interact with too many people. But now a creator can interact with 10,000, 100,000, or even 1 million consumers through AI assistants, knowledge bases, and prompts.
"Dark Surge": In your view, will AI-era creators differ greatly from short-video era creators?
Li Yan: The fundamental difference is that creation no longer requires a camera — you can directly use prompts to let AI produce. So users will also play with new verticals.
The short-video or live-streaming era basically relied on visual impact, amplifying people with more performative desire. Some people with genuine insight might not want to show their faces. The AI era may amplify and牵引 interesting souls more — interesting souls can be seen by more people.
Now users are also starting to create various bots to express themselves. Of course, it's still hard to determine exactly what kind of people will be internet celebrities in the AI era, but they definitely won't be the same people from the last era.
Part 02
"Giants can't capture all the weight, leaving huge gaps"
"Dark Surge": Some investors feel what you're doing is too同质化, and the consumer battlefield is clearly where giants must compete.
Li Yan: The同质化 perception may come from categorizing us as chat. The chat tech stack is already mature, so many do similar things. But what we're doing isn't chat — it's a reconstruction of information acquisition methods.
Looking back over decades, mainstream information acquisition has gone through several paradigm shifts: from portals to search engines to recommendation engines. Each technological transformation has spawned new giants.
In the mobile internet era, products all revolved around how people acquire information. Today, large models don't just find information — they generate it. This change is at the underlying paradigm level, not interface or feature level. What we truly want to do is close the loop of generation-distribution-consumption, making AI an understanding content partner that proactively provides content you want to know, should know, and that can spark interest.
Information acquisition is an ultra-large赛道 that can never be completely monopolized by one company. In the mobile internet era, Toutiao, Douyin, Kuaishou, and Xiaohongshu each had their strengths, showing the赛道's enormous capacity. Even a powerhouse like ByteDance couldn't capture everything. This is precisely why we believe startups have opportunities. The key is whether the angle you find is unique and whether the landing point addresses real needs.
So we're not "homogeneous" — we're carving out a new approach that most haven't seriously pursued yet, but that will certainly matter in the future.
"Dark Surge": What do you think is the root cause of why information acquisition platforms are hard to monopolize?
Li Yan: In the past, this producer-consumer two-sided network made supply-demand relationship adjustment difficult. Whether in supply or distribution mechanisms, many细分, diverse, long-tail, or mid-to-low-frequency needs went unmet. We always believe there are overlooked voices and overlooked perspectives in society. And AI, as a super-capability, will greatly alleviate this imbalance.
"Dark Surge": But if giants or model companies want to push in the same direction, it's not particularly difficult.
Li Yan: Why would giants change? They've already built core business models and ecosystems based on past producer and consumer habits. Others provide you content, you leave fixed positions for them. They have to look at profits, at financial reports, so they must pursue traffic efficiency. Even if they change, it's改良 without changing past product interaction forms — these are all baggage. We're completely new.
As for model companies, everyone sees different horizons and has different positioning. And even if everyone targets this direction, there are differences in talent density, effort invested, and iteration speed.
Today, there's no AI direction that giants can't pursue. But precisely because they have to manage dozens of directions, the force allocated to any specific direction is thin, and talent density may not exceed ours.
Our awkwardness today isn't that giants have already lined up battle formations — rather, it's that actually nobody is building content platforms.
"Dark Surge": What's the reason?
Li Yan: Content platform tech stack requires strong innovation. Many doing similar things are actually chat-based. Some giants use large models to改良 their existing recommendation systems.
But for feed + large models, we're the first company, and on the feed front, we still have a long road ahead.
"Dark Surge": Where do the difficulties lie?
Li Yan: The biggest difficulty is: when user interests are very dispersed and co-occurrence is very sparse, how do you reconstruct a clear interest chain of thought. Reinforcement learning, by bringing new emergent reasoning capabilities to models, can solve this predicament. This is actually a leap in model capability.
Second is the definition of domain problems — deciding what content gets priority recommendation on an AI platform.
This requires you to both understand models and have clear认知 of content赛道, making us a very rare team in the market with both capabilities. Over the past year, we've also accumulated highly organized data in new scenarios, using it for reinforcement learning to help birth more refined chains of thought.
Facts have also proven that with almost less money, we've reached the front of many AI product rankings this year. So sometimes, I actually hope our positioning isn't very clear, because that's where startup opportunity lies. Running in a straight line, startups have no way out.
"Dark Surge": But the surge in many AI products early this year seemed more like spillover from DeepSeek traffic.
Li Yan: Q1 was definitely so. But our feed business has also been continuously growing — we can feel a greater force at work.
Look at Douyin and Xiaohongshu — their DAUs are also growing. Such homogeneous two-sided networks have this much traffic, let alone a new form. It's like the revolution of electric vehicles over gasoline cars.
Xiaomi entered the EV market quite late, but selling 30,000 units monthly generates nearly 10 billion RMB in revenue. The scariest thing is: you're first in this market with no competitors, but the entire relevant user base in China is only 300,000.
In my view, going to a large market at least offers survival possibility. Small battlefields have theoretical ceilings. You do everything right, but no matter how hard you try, you can't break through.
"Dark Surge": Is building a small but beautiful company not attractive to you?
Li Yan: AI companies can't be beautiful if small. AI is violent aesthetics — its first principle is scaling law. It needs more data, more compute. How can small have power?
Those application companies under 10 people not touching large models are another matter. But people like us with strong technical backgrounds wouldn't make that choice.
"Dark Surge": Would going into a specific vertical to build an agent not be an option for you?
Li Yan: This relates to the maturity stage of technology. Currently, acquiring traffic through large model underlying technology is highly efficient. Additionally, we believe making vertical choices through data-driven approaches on top of general technology has higher success probability than arbitrarily picking a vertical. Because this is choice based on user interaction data, not on subjective认知.
And we're not experts in any vertical. We're better at technology, infrastructure, containers — users from various verticals can all come.
We're certainly surviving in the gaps between giants. But precisely because this is such a large赛道, I believe giants can't capture all the weight, leaving huge gaps.
Part 03
"We're lifted up by a greater force"
"Dark Surge": Is the执着 of building a platform a common dream for executives who came from giants?
Li Yan: I'm not that kind of空降 executive from a giant. When I joined Kuaishou, it was still in a residential building.
Half of Kuaishou's AI-related tech stack was built with my participation. So I was a core builder of Kuaishou AI, not a simple participating observer.
After growing the company large, we came out again. So I'm someone more adapted to small companies.
"Dark Surge": So you enjoy the entrepreneurial stage more.
Li Yan: Many mature organizations, when doing things, are more like carving —微弱改良. Startups are a violent aesthetics. They have no money, no resources, but have an engine. This engine naturally carries power, letting us grow. In our bones, we're builder-type people.
"Dark Surge": How would you understand Kuaishou's success?
Li Yan: Many people ask how we rose up. To be凡尔赛, we don't know either.
Compared to macro strategy, what we felt more was from the grassroots, from the technology engine, from底层力量. We were lifted up by these forces, and after being lifted up, we weren't adapted to it — like, how did we become a giant?
"Dark Surge": So you believe more in that greater force.
Li Yan: During the first funding round, an investor also told me they'd invested in many small, beautiful B2B scenarios like what you described.
Then he said something that particularly moved me. He said, maybe because you came from Kuaishou, you believe more in consumer success. One general's success comes with ten thousand bones. From a societal perspective, many projects die. But because you've succeeded before, you feel it's not particularly hard.
"Dark Surge": How would you understand this force?
Li Yan: Many people's analysis, logic, and deduction seem like very small efforts to me. I've recently been reading a book called Discovered, Not Designed. It says complex systems must emerge, not be designed. Emergence is the greatest force, not logic, deduction, or analysis.
I've told many investors: we've passed the era where a great theorem could support world progress. There's no logical deduction that directly leads to success. What great theorems or laws have there been in recent decades? The defining thread of this era is how to engineer and harness emergence.
So I also feel wronged — it's not that I deliberately want to build a platform. Rather, we believe the driving force of world progress is a complex system, not one or two deduced laws.
"Dark Surge": What's key in this complex system?
Li Yan: Evolutionary speed. Starting point actually doesn't matter. And speed often depends on the engine.
"Dark Surge": What's Wen Xiaobai system's core engine? Where specifically does its sense of power come from?
Li Yan: The core engine is of course large models. But its true sense of power comes not just from compute and parameters, but more from an "interest estimation + chain of thought" system we've built on top of models. We hope it can not only understand language, but understand people — know what you're thinking, what you care about — then respond in a natural, resonant way. This leap from "capability" to "resonance" is a direction we particularly value.
Past content production, including short video, was actually quite handicraft. Much of editors' labor was枯燥 and repetitive. I believe AI changing the world should start from these places.
"Dark Surge": Looking back, which decisions in this entrepreneurship were overly conservative?
Li Yan: On PR and fundraising, we were both quite conservative.
Entrepreneurs have no resources and need to persuade others. But I used to resist this, spending too little energy. For entrepreneurship this is very bad — a CEO still needs to maximize company interests.
"Dark Surge": A persistent reflection in the AI industry is re-examining the inertia brought by the mobile internet era.
Li Yan: The past definitely had much luck. Currently, mobile internet's demographic dividend is gone, capital dividend is also gone. You need to be more pragmatic and efficient to have a glimmer of hope to catch up.
"Dark Surge": Given model capabilities and payment habits, many AI application companies doing consumer products chose to go overseas from day one. How does Wen Xiaobai consider its commercialization? Also, with limited resources, how do you allocate weight between models and product?
Li Yan: We're currently polishing our overseas product form. While overseas large model infrastructure is more mature, products that truly combine "active asking + passive feed" into a closed loop are still very scarce. We hope to replicate the "generation-distribution-consumption" chain validated in the domestic market to overseas, making AI not just a tool but truly a high-frequency information外援 and content partner.
Commercialization won't be one-size-fits-all. We'll切入 from high-frequency, high-value usage scenarios, penetrating around content services users truly need. AI product business models may not be traditional advertising, e-commerce, or subscription, but rather new ways of realizing "content connection value."
As for resource allocation, we've consistently insisted on deep cultivation where we should self-develop, and not重复造轮子 where we can leverage others. Embracing open source is an important way to improve efficiency — standing on giants' shoulders. But we're absolutely not just套壳. Xiaobai Research Report and the feed engine are both self-developed. User interest modeling, especially cold start and long-tail interest chain reasoning, can't be completed by general-purpose large models alone — we must build our own content赛道 models. This is Wen Xiaobai's differentiated core, and where we're most willing to place heavy bets.
"Dark Surge": Are you someone who easily accepts reality and failure?
Li Yan: Entrepreneurship is修行. Like playing ball — just swing every shot with full effort.
When I first started my company, I attended an Alibaba-organized Xixi Forum in Hangzhou. They asked me to share. I said after doing AI for so many years, I really can't claim expertise, because I've lost more than won.
The AI赛道 — hot in the 1980s, dead in the 1990s, up again in 2010, down in 2018, up again in 2022. Life and death is the basic law of this赛道.
"Dark Surge": What's the biggest change in the AI you've worked with across two eras?
Li Yan: Before 2017, the entire AI technology paradigm was supervised learning, training models based on user-labeled samples. Then the amount of labeled data directly affected model success or failure.
There was talk of small AI and big AI. CV, NLP — these were human-labeled, with limited labeled data, called small AI. Search and recommendation were equivalent to all netizens labeling for you, so called big AI.
With continuous breakthroughs in reinforcement learning, plus milestone achievements like AlphaGo, AI could play with itself. Entering the large model era, it can generate infinite data to train with. So today's AI is not only not smaller than search and recommendation — it's even bigger.
Look at many company founders — they tend to believe that without users, they can generate data themselves. They feel training the large model well is enough, not caring about user participation. This is indeed a major difference in the past year or two.
"Dark Surge": After DeepSeek, many indeed believed that so-called user flywheels and data flywheels had limited impact on intelligence improvement, so many model-training companies adjusted their focus.
Li Yan: I believe things will change again later — user participation will also become important.
AI has its different stages. At different stages, the greatest source of power manifests differently. And finding the greatest source of power is what people doing technology are better at.
Recommended Reading
Where Money Flows, People Rise and Fall