Zhihu's AI Revolution: The Integration of Intelligent Search and Real-time Q&A

[baoshi.rao]

Against the backdrop of the information technology leap in generative artificial intelligence (Generative AI), there are three fundamental perspectives to consider regarding Zhihu's present and future:

It is one of the most important sources of Chinese language corpus for large language model pre-training. For instance, the recent phenomenon-level large model chat application Kimi Chat relies heavily on Zhihu as a key training data source (accounting for even over 80% of its data).

Every user who asks a question on Zhihu is essentially issuing a prompt; while every answerer on Zhihu essentially functions as a human language model based on personal knowledge and experience (corpus), values, and thought processes (algorithms and chains of thought). Based on this, 'digital answerers' powered by large language models will gradually emerge on Zhihu. Given that the 'question-answer' model is the foundational logic of Zhihu as a content community, as well as the core principle behind large language model tools and the driving force behind the 20-year evolution of search engines—Google, Baidu, and newer AI-powered Q&A aggregation tools like Perplexity—all recent advancements essentially revolve around the AI transformation of search engines. Therefore, Zhihu's search attributes—AI-based search—will be further enhanced.

From these three perspectives, Zhihu, amidst the clamor of AI, has the following potential ways to embrace the generative AI wave:

1. Become a better provider of Chinese pre-training datasets, acting as a perpetual source of high-quality Chinese content created by humans rather than machines. This ensures the content is more likely to be collected, utilized, and pre-trained. Participate in the construction of national-level Chinese language corpora, becoming a key contributor to dataset development. The entire community is undergoing a radical AI transformation. AI will pose questions and provide more responsible answers. The Zhihu community will see a surge of AI digital personas as independent IPs—historical, technological, medical, and literary vertical digital personas—answering user queries and delivering the answers users seek.

With search as the breakthrough and leveraging Zhihu's own large model capabilities (Zhihutu AI), the large model will participate in generating and summarizing answers to some questions, prompting users to ask follow-up questions and guiding them to engage in "multi-round dialogues" within the Zhihu community. This will attract more answer providers and make the large model a catalyst for boosting user activity.

The three paths outlined above—the first conservative, the second radical, and the third a combination of the two—will determine Zhihu's direction. Which path Zhihu chooses depends on what Zhihu is, its role in the large language model ecosystem, and what it excels at. At the Discovery Conference on March 20th, Zhihu introduced three AI-related features—

One is a search function that helps users find the 'consensus' within the community.

It essentially functions as Zhihu's version of Perplexity—users pose questions, and the Zhihai Tu AI model generates answers based on content within the community. Not all questions on Zhihu have answers provided by users. Often, a profound question requiring specialized knowledge may take a long time before an expert in the field can provide a high-quality answer. However, such questions can be referenced against high-quality answers scattered across the community in response to other questions. In such cases, if AI can search and generate answers, extracting valuable information from these responses and making effective inferences, it can provide an immediate answer to a newly posed professional question. The user asking the question can get a quick response while waiting for other professional users to arrive and contribute their answers. During the grayscale testing of this feature, Navis Li, a professional respondent in the consumer electronics field on Zhihu, raised a specialized question: Russia or the former Soviet Union seems to have a famous lens that can achieve rotating bokeh effects - what exactly is this lens? Frankly speaking, for such extremely niche professional trivia, waiting for human answers could take a long time or might never come. But based on AI search, other professional respondents and community creators on Zhihu were able to provide an accurate answer by refining, analyzing, and deducing from existing answers to other questions.

The second application is real-time Q&A for public editing purposes. It is equivalent to an AI-generated in-site Wikipedia. A high-quality question may have hundreds or even thousands of answers, with dozens of highly upvoted ones. Does everyone have the time and patience to read all the answers and then synthesize the knowledge they need? This is probably unrealistic. It requires a "best answer" based on the creators' responses—essentially a distilled and extracted Wikipedia-like function. In the early days, Zhihu had this feature, where users collaboratively summarized all answers through public editing, "crowdsourcing" a best response. The original intention was ideal, but the reality was somewhat harsh. Instead, it became one of Zhihu's most controversial features. Human subjectivity, bias, and tendencies are unavoidable and can easily create greater conflict, opposition, and controversy, leading Zhihu to eventually discontinue this feature. AI is truly the best fit for Wikipedia-style work. While AI may have biases, they can be constrained and controlled through enhanced training. AI's summarization and reasoning capabilities are improving exponentially. Therefore, based on the "best answers" provided by creators, we now have a better generator—large language models. This real-time Q&A is essentially AI-powered public editing, making the path for users to obtain simple, direct, and effective answers more straightforward and faster.

The third feature is the ability to continuously ask and follow up on questions, even without requiring prompts.

Zhihu has long been a relatively one-dimensional product: one question with multiple answers. Although there are algorithm-recommended related questions, they are not real-time queries from the same user and may not reflect the greatest curiosity or doubts in a questioner's mind. Now, after reading an answer to a question, users can follow up based on that answer and the extracted "best answer"—follow-up questions don’t even require active input of prompts, as the AI generates them automatically. This essentially encourages users to engage in "multi-turn dialogues" within the Zhihu community—just like what people do on ChatGPT and Kimi Chat. In response to user follow-ups, the AI can provide its answers, and human answerers will likely join the conversation over time. In this way, Zhihu, as a 'Q&A community,' has the potential to evolve from 'one question with multiple answers' to 'multiple questions with multiple answers.' In a content community based on Q&A, questions are the largest supply-side. In the past, questions mainly relied on people's curiosity and thirst for knowledge, but now they can be enhanced and inspired by AI. The benefits of this approach are obvious—generating more questions and, correspondingly, more answers.

A search, a real-time Q&A, and a follow-up question—all are powered by Zhihu's AI. Zhihu calls this three-in-one AI feature 'Discover · AI Search,' represented by a 'four-pointed star' logo on the left side of the homepage.

At first glance, it has search functionality but not as a dedicated search box. It has large-model dialogue capabilities but doesn’t present them in a chat interface. It allows follow-up questions, but AI isn’t the sole responder. It still looks like the Zhihu we know, requiring many curious questioners and high-quality professional answerers. And now, AI stands by their side. In the three potential paths for Zhihu to embrace AI, the platform has chosen the third one. This path ultimately points to Zhihu's most critical lifeline as a knowledge Q&A community—active users, a continuous stream of high-quality questions, and trustworthy answers.

Zhihu believes that "Discovery·AI Search" could become the lifeblood of its community, but it remains highly vigilant against radical AI adoption across the platform—where large numbers of AI-generated questions and AI-provided answers dominate. The community would be flooded with robotic and digital questioners and respondents, with most questions and answers being AI-generated.

In a recent conversation between SiliconAngle and Zhihu's founder and CEO Zhou Yuan, Zhou expressed extreme caution toward "NPC respondents": "The platform shouldn't actively pursue this direction. Why would a platform want to populate its ecosystem with so many NPCs?" He believes Zhihu's "AI Search" is a feature that "doesn't rely on traditional information feeds, but instead adopts a new approach based on large model capabilities and interaction methods, while providing extremely direct data feedback." This "direct data feedback" refers to real "people," not AI. As questioners, users can obtain desired answers faster and more directly; as answerers, their responses can be more frequently searched, retrieved, indexed, and regenerated into new content. For regular browsing users, it offers a different interface and interaction method, allowing them to "search" and "discover" questions and answers that previously required scrolling through endless information feeds – a crucial factor for community engagement. This process generates more content, which also serves as training data for large language models.

"AI itself is a language model that doesn't encounter problems. Only humans face issues in society. When you experience heartbreak or job loss, these pains and desires are uniquely human. The problems and corresponding content generated by people are what's authentic. AI can assist you, but if we remove this human layer, AI just becomes self-referential and loses significant value," Zhou Yuan told Silicon Star.

Based on this understanding, he opposes presenting "AI Search" generated answers and questions under robot IP identities. Instead, he insists they only appear within the "four-pointed star" badge indicator, delivered through private chats – exclusively for individual users without appearing in information feeds created by actual "content creators" like questioners and answerers. Facing AI, the fundamental question Zhihu must answer is: How can humans share wisdom with AI while maintaining respect and encouragement for human originality? This isn't a new problem, but it requires a novel solution. Currently, their answer is: Through AI, help people discover the broader world created by humans and more human-generated content, rather than using AI to find more AI-generated content.

It points to a more critical issue: As increasingly more content worldwide becomes AI-generated, could human experience and human-created content become an increasingly marginalized scarce resource? The premise and foundation of AI-generated content are corpora derived from human experience—content created through historical, social, economic, and technological practices—whether articles, data, images, sounds, videos, or code. But if human-created content becomes a scarce resource under the "squeeze" of AI's exponential growth, where will the experience and corpora for future AI-generated content come from?

This matters to everyone, and for Zhihu—which has effectively become a vital repository of Chinese-language corpora worldwide—it's an unavoidable question. "The development of artificial intelligence relies on three key factors: chips, models, and data. Chips are indeed a bottleneck, but they are an engineering challenge. With sufficient scale, iteration becomes possible—our engineers are quite capable. Models benefit from open-source advantages and won’t evolve slowly. However, data is a non-renewable resource, and it can’t be acquired immediately. Whether you invest $1 billion or $10 billion, you can’t build a UGC community overnight. Language data is a non-renewable resource," Zhou Yuan believes.

"If you treat language data as static, even if it appears so, you must view it as a piece of land. You can’t harvest corn from one plot and sell it elsewhere without considering whether the land might face pest infestations tomorrow," Zhou Yuan argues. He suggests that global language data faces the risk of depletion or impoverishment, much like oil.

Therefore, ensuring the flywheel of data supply and consumption keeps turning—where more data becomes training material for large models, consumed by AI while also being replenished and inspiring humans to create better content—becomes the most valuable solution for Zhihu in the generative AI wave. This is especially true given its pivotal role in the global Chinese internet language data ecosystem. What is the value of becoming the world's largest renewable and continuously supplied Chinese language corpus? The answer from the United States is: Google has paid over $60 million cumulatively to Reddit, the most important content community in the US, to purchase high-quality corpus data.

What is the answer from China? With the data trading market maturing under the dual drive of government and market forces, Chinese language corpus will become an important tradable data resource. Identifying the most significant Chinese corpus will emerge as another scarce yet vital 'water' resource following chips and computing power. Everyone knows that in AI development, it's often the 'water sellers' who make the first profits.