Zhou Yuan on AI Search, Human Conversation, and Zhihu’s Community Choices

Zhang Xiaojun interviews Zhihu founder Zhou Yuan about how asking machines and human-to-human discussion might coexist in a content community after the rise of large models. They also review Zhihu’s founding, competition, and post-listing adjustments.

Zhang has heard that Zhihu discussed large models before ChatGPT. Zhou corrects her: he remembers realizing the strategy needed revisiting after the late-November 2022 release, just days after a strategy meeting.

With Chinese technology companies focused on costs and efficiency amid the pandemic and other changes, the advance gave him a sense of being out of step. The team contacted people in the field to understand events firsthand, he says.

He recalls discussing collaboration with an entrepreneur and offering a possible $100 million commitment, which was not accepted. He does not confirm the person’s identity; the experience showed him that model development might require more resources than initially imagined.

Zhou says his early judgments included the growth of open models and the principle that AI should serve people. Unlike PCs or phones, large models arrived without an equivalent new medium, so their output would enter existing information environments.

Zhihu centers on people sharing and discussing with one another. Zhou says interaction with AI can be valuable, but if AI poses as community users without clear distinction, trust and incentives to participate may suffer.

He especially values people sharing problems they actually encountered and how they resolved them. Zhihu subsequently introduced a disclosure for AI-assisted creation and adjusted community governance to address possible confusion.

On investing in ModelBest, he says the organizations had a long-standing relationship and trust before the investment. ModelBest brought pretraining experience, while Zhihu contributed engineering and safety experience; their teams had worked together closely.

In fast-moving technical competition, he argues that identifying complementary talent, building trust, and forming a team are important to turning research capability into products.

Zhou says he felt more excitement than fear about AI’s opportunities. Zhihu began to help people share knowledge, experience, and insights and stimulate thought, a goal not identical to maximizing time spent consuming content.

He imagines AI as a medium that can help communication by organizing as well as transmitting information. Technical improvement, he argues, should support better exchange rather than replace relationships among people in a community.

Zhihu experimented with summaries of trending discussions and later removed the feature. Zhou explains that different viewpoints and the act of reading them can matter for open questions; compression into one summary can lose part of that value.

Zhihu Direct Answer, launched in 2024, explores AI search. He characterizes traditional search as opening multiple pages and synthesizing them oneself; AI search can provide an overview and sources first, then let users follow up or read further.

Early Zhihu combined search and asking in one button: users could search existing information or ask people when no answer existed. The team separated them after users found the interface confusing, but Zhou still sees a connection between the two.

He acknowledges that Direct Answer had not yet fit some users’ workflows. Conversations with creators revealed concrete needs that would require further product iteration.

Discussing Perplexity, Zhou sees it as more than a search tool people use and immediately leave. Users read, ask follow-ups, explore, and save or organize content, making it resemble a content product.

He argues that traditional search should be viewed together with the websites and services it connects as an ecosystem. Beyond information lookup, people need tasks such as ordering food or arranging travel, many of which vertical applications increasingly handle more conveniently.

So whether Perplexity disrupts search is not just a question of answer quality; it also depends on how new tools connect developers and services and complete tasks at lower cost. Zhou finds the idea promising but says it is too early for a conclusion.

He distinguishes search, chat, and documents: the model capabilities may overlap, but people have different purposes in each setting. Products should be designed around users’ situations, not only around model capabilities.

For content communities, Zhou maintains that AI should enhance human discussion rather than impersonate it. Depth, authenticity, and connection are qualities he thinks the product needs to protect.

Human conversation does not always put efficiency first. In consultations, Zhou says, users may not know what they really need to ask, so an expert may ask questions in return; seemingly casual conversation can also produce unexpected insight.

A short conclusion may not create enough confidence for a purchase decision. People may need to understand the issue themselves or rely on a trusted person; Zhou thinks AI answers still face questions of trust and accountability.

He wants to improve Direct Answer using actual user and creator feedback instead of chasing exciting concepts alone. He observes that subtle differences in model-output quality may not be apparent to every user.

Why insist on a community that encourages thought? Zhou says it follows from the team’s belief that everyone has valuable knowledge and experience in some area and that a platform can help others share, discover, and use it.

He recalls feeling that his programming work was far removed from a vision of technology changing business. Around 2005, an article in IT Manager World at a newsstand prompted him to try becoming a technology journalist.

He sent his belongings to Beijing before the magazine had formally hired him. It offered an internship, and an article about competition between browsers and desktop software later earned recognition, helping him stay.

Journalism exposed him to companies and new technology, but he also felt like an observer. Around 2008 he decided to leave media and create something himself in the technology wave.

He participated in Apple4us, a group blog where people with different backgrounds interpreted the same developments through design, engineering, and other perspectives. It made him imagine an editorial room that could keep expanding.

Twitter’s following relationships and Quora’s focus on questions and topics also prompted him to consider how people could discuss a question deeply. Zhihu later sought to connect that distributed knowledge through product design.

In his first startup, Meta Search, the team tried providing analytics to small e-commerce businesses, but many merchants at the time focused on expansion and would not pay. Zhou learned that founders need to understand genuine demand, not just spot an apparent opportunity.

When founding Zhihu, he wanted to do something meaningful that he could also use himself. He recalls bringing only a few pages to an early funding discussion and describing the product as knowledge-based social interaction when asked whether it was social.

In Zhihu’s first month, servers frequently failed, but users kept returning to discuss. The team saw this as a signal of demand and made service stability its highest priority.

He tells of a researcher who found an answer someone had posted years earlier and then began sharing work experience in hope of helping future readers. Zhou sees this mutual aid and connection across time as the community’s core.

Facing competitors such as Wukong Q&A, Zhou says similar question-and-answer forms do not necessarily make the same product. Business operations need efficiency, but discussion among people cannot always be measured by industrial output.

He acknowledges that broad platforms and recommendation systems have advantages in serving varied needs and that competition creates pressure. A community, he argues, needs a clear identity rather than changing direction solely under that pressure.

After listing, analysts and investors often assessed the company using similar scale metrics. Zhou acknowledges that this affected decisions and forced Zhihu to reconsider the relationship between community and scale goals.

Zhang notes that Zhihu pursued video and short content and later adjusted course. Zhou says the problem is not necessarily the format; excessive pursuit of numerical growth can distort the community ecosystem.

He treats creator experience, useful content, community atmosphere, and monetization compatible with the ecosystem as inputs, hoping healthy growth and profits follow. He admits that after listing he briefly thought more investment and scale would solve every problem, then found that was an illusion.

On criticism that Zhihu’s quality has declined and longtime users are leaving, he says the platform must listen and identify the underlying issue. Individual comments may not all be accurate, but that does not justify ignoring change users perceive.

On the falling share price, he admits that he checked it daily in the first months and struggled to accept it. He later decided the team should focus on sustainable operations and profitable growth; his confidence about the future share price is a personal view.

He does not think a community’s fate is fixed on day one. Products and media can evolve, but poor changes can also harm users; improvement depends on long-term building and judgment.

On being called “captain” and the company’s pirate culture, he describes the team’s sense of exploration. A CEO’s loneliness, he says, more often comes from decisions only that person can make, even when others are available to consult.

He recalls difficult and emotionally complicated conversations when the first cofounder retired. Over time he came to accept that people will join and leave as a company develops.

At the close, he says a community cannot stay vital simply by declaring that it represents a culture; its ecosystem must keep renewing itself. He hopes to answer what Zhihu becomes in the AI era through gradual product iteration.

Zhang closes the interview and invites listeners to suggest future topics.