Science and technology creator Zhihu CTO Li Dahai: Technical service content, commercialization-dependent content, it was difficult to scratch his head by "hehe"

Li Haizhihu Partner and CTO
In 2006, he graduated from the Department of Mathematics, School of Mathematical Sciences, Peking University, and worked for Google, Yunyunwang and Wandoujia successively. After joining Zhihu, Li Dahai has successively been responsible for the advertising technology team, data, algorithm and overall community business. Currently, he is in charge of Zhihu's big data team, content circulation and the development of new AI businesses.

—
Text | babayage
​Edit | Laugh Laugh​
2006-2010 Farewell to mathematics and joined Google at the beginning of growth Technical vision>Technical details
In 2006, Li Dahai, who was about to graduate from the Department of Mathematics of Peking University, fell into a long test that determined the direction of his life: stick to being a mathematician in the mathematics field, or leave the field of mathematics and switch to other professions completely.
Li Dahai's passion for mathematics can be called love, but learning is one thing, and he will face many practical problems as a career: in mathematics, solving a problem can be as short as seven or eight years, as long as decades or even a lifetime. He is well aware that he cannot sit firmly in the research laboratory and study a single subject in depth; apart from his personality, "I don't have so much confidence in my abilities. In the field of mathematics, I may only be a mediocre mathematician." .
"But if it's not math, what other options do I have?"
Since junior high school, computer has always been Li Dahai's interest, and he also established the Peking University Linux Club as an early initiator. Coincidentally, in 2006, Google took the first step to enter the Chinese market. Kai-Fu Lee started a talent localization project in mainland China, preached to famous schools across the country, and planned to recruit 50 closed-door disciples.
By chance, Li Dahai joined Google.
​
Tech creators: What was your biggest takeaway from your four years at Google?
Li Dahai: I have learned a lot in all aspects. The Google team is all great. There was a book that left a deep impression on me back then. Mr. Wang Yonggang wrote "Tao and the Law of Nature". Mr. Wang used a particularly lively language to clearly introduce the idea of ​​pattern design. I didn't expect that I would be in Mr. Wang's team from the beginning of the internship period. Another time I was in charge of receiving guests to visit the Great Wall, and the object of the reception was Guido van Rossum, the inventor of Python. In such an environment, all you need to do is open yourself up completely and absorb as much as you can.
In the four years at Google, it is not so much how much its technical strength has grown, the greater gain is to open up technical vision and strengthen technical thinking.

From 2010 to 2014, Yunyun.com, the unbalanced team of the first entrepreneur Li Dahai had a classmate. Hearing that Li Dahai was working at Google, he repeatedly lured him with food, instilling the concept and vision of entrepreneurship while eating. Three to five, Li Dahai, who had no idea of ​​starting a business, became more active.
It was not the straw that broke the camel, but a whole collapsed skyscraper - in 2010, Google withdrew from China, and Li Dahai decided to leave after careful consideration, "I still hope to be able to do things to serve Chinese users."
In terms of learning environment, Google is the best choice; once he decides to leave, only starting a business can make Li Dahai feel "worth doing". However, he did not cooperate with N Gu Maolu's classmates and friends - I don't know if this gentleman is distressed about food money - but joined Yunyun, founded by Liu Jun, the former vice president of Google's China Engineering Research Institute.
As soon as Yunyun.com was established, it attracted much attention because of its Google all-star technical team, which caused a lot of waves in the circle. There was no precedent for the business direction of Yunyun.com in those years. There was no such intuitive reference as "Sina-Yahoo", "Baidu-Google", "Renren-facebook", and even the founder Liu Jun himself said bluntly, "Yunyun.com is a brand new thing. It is completely different from what people know so far, and there is no analogy" - until many years later, it was abstractly described as "a social search engine".
At Yunyun.com, Li Dahai is responsible for the development of real-time search engines, basic libraries and some code frameworks.
​
Kechuang people: From the learning state to the entrepreneurial state, what kind of experience is it to make a hard switch between two hugely different modes?
Li Dahai: The pressure has increased sharply, but the influence has also increased. All aspects of the company are related to you, and it feels very challenging. I may still be quite suitable for starting a business, and I have a sense of freedom that the sky is high and the bird is flying.
There is an interesting thing, probably because many people in the team came out of Google, and many of the work habits at Google have been carried over, so it is not that difficult to adapt. People who come out of Google will have some "stinky problems". I have seen many useful tools in Google. After they come out, they find that the tools that are not in hand are particularly awkward, and they will say "Let's make tools first". I've also done a few things like Google Blaze's C++ Build System and so on.
​
Kechuang people: Recently there is a song "The Lonely Brave" that is very popular. There is a lyric in it called "Like me, the gaps are the same", a team with engineers as the absolute protagonist, whether everyone's shortcomings and defects are similar ?
Li Dahai: Absolutely. The unbalanced team is an important reason for the failure of the first startup. We also have product managers and very professional operation personnel, but these positions in the team are not strong enough to speak. As a technology-based company, we are making products. At the time, it did not grasp the pain points of users too much, resulting in a logic that looks good but it is difficult to run through. The social search we had to do in the past, taking buying a car as an example, was “I am friends with you—the content you searched will show my evaluation—directly promoting purchasing decisions and even purchasing behaviors.” In the end, we found that this logic requires a certain amount of information density as support, that is, it may rely on a set of products that increase information density. After a little thought, we know that this must be a large-scale product, and we can't make one directly. come out.
It is also because of this experience that I have a deep understanding of the density of information and the richness of information content.

The right logic and wrong timing for joining Wandoujia in 2014 "This is the era of mobile search, and there is a time window for this matter. If you still want to do search, come to us." "Auntie Cui" Cui Jin, founder of Wandoujia His remarks moved Li Dahai, and he joined Wandoujia with unpaid search ambitions.
Turning the clock back to 2014, the combination of Wandoujia, the head app store, and search technology has almost unlimited imagination: users only need to enter the restaurant they want to query in the app store, and the store can directly call up the review APP And jump to the page of the hotel; as long as the user searches for the reservation, the store can call up the reservation APP and directly present the reservation page...
But when you get started, problems come one after another: First of all, the technical difficulty threshold is quite high. "The Internet in the PC era is based on HTML, and crawlers can efficiently process all pages, but in the mobile era, the APP side is structured information. The cost of calling is very high, which in turn affects the efficiency of search engine construction”; if the technical problem is still at the methodological level, the real devastating blow comes from the overall change of the mobile Internet ecology, the rapid rise of mobile phone manufacturers, and the formation of head effects. The practice of self-made app stores and pre-built applications not only makes the APP ecology also form a head effect, but also almost destroys the survival soil of third-party app stores.
It was not until a few years later that the WeChat applet used another methodology to run through the business model that Wandoujia had envisioned. At this time, Li Dahai had also found a dream worth devoting all his energy to.
​
From 2014 to now, Zhihu’s value-matched content treasure mine wrote this, which basically summarizes Li Dahai’s workplace migration model: from technology to products, or from obsessed with technical details to more comprehensive industry insights. In 2015, Li Dahai was invited by Zhou Yuan to join Zhihu. There are two important reasons for joining Zhihu: First, in Li Dahai’s view, Zhihu has core resources that were scarce in his previous entrepreneurial projects— - Effective information with both value and quantity, that is, high-quality questions and answers, "After the first two lessons, I realized the value of content, and the combination of these content and technology can detonate greater energy, and at that time Zhihu At the technical level, it has not yet begun to be recommended by machine learning, and it has great potential.” Second, the concept of Zhihu fits his values ​​very well: high-value knowledge should be carried by vivid stories, not boring lectures. "When I was studying mathematics, I felt very painful. It was a system of axioms. Axioms, Theorems 2 and 3 lead to conjecture 4. The stories behind these theories, how they came about, and what problems were solved at the time were not reflected in the textbooks. These sections allow learners to digest knowledge more efficiently.”
"Helping more people understand the world more comprehensively in a simpler way is worth doing all the time." If the previous entrepreneurial goal was only to make a great product and a career that would bring both fame and fortune, in Zhihu, Li Dahai really experienced a unique sonorous sound of idealism, which is the reverberation of the brand that changes the world every step of the way.
At first, Li Dahai was only in charge of algorithms and big data, but he took over the homepage soon after. The homepage recommendation system was the most important product of Zhihu, but the current situation was not ideal at that time. Not only did a request have to wait for two seconds, but even the homepage The push logic is that each user has a mailbox, and Zhihu puts 200 messages in it.
Li Dahai quickly abandoned this "old algorithm", cut Zhihu's homepage into a recommendation system, and used machine learning technology to quickly improve the user experience. Key indicators such as user duration, second retention, and DAU have been greatly improved.

Keep an external perspective and avoid the information cocoon room Kechuang people: You may be the person who knows the bottom layer of Zhihu best, from the logic principle to the technical details, how do you maintain the "user perspective" "external" under the influence of such a complex internal information perspective"?
Li Dahai: Good question. First of all, everyone in the technical team must be a player in the community. I ask everyone to have a certain amount of active time every day. You must experience your own products first, and only when you play can you know what the user's pain points are and why they hurt, otherwise you will not be able to empathize. ;Secondly, we have established a user experience center to continuously listen to users' feedback through various methods and channels, including offline interviews, collection of effective user feedback, data analysis of user behavior traces, etc. - Zhihu There is a strong emphasis on data science, with a strong emphasis on the scientific nature of experiments.
​
Kechuang people: The information cocoon room caused by the recommendation algorithm and the head effect caused by the human information node are the challenges that the content-based community is bound to face. How can Zhihu avoid this phenomenon?
Li Dahai: Zhihu has always been committed to improving and optimizing the content layering, evaluating the content in each field. In addition to recommending high-quality content in the field of interest to the user, I also want to recommend cross-field high-quality content to him. In addition to paying attention to the accuracy of the recommendation system, we will also pay attention to its surprise, so that the content that the user sees can continuously expand his horizons and avoid the information cocoon.
As for the head effect of creators, it is necessary to mention Zhihu's vision: to allow everyone to share their knowledge, insights, and find the answers they need. In this scenario, there is obviously an objective fact: the answer you need is likely to be from someone you can't imagine at all. He may come from a KOL or a newly registered user, so Zhihu's product logic must be Focus on premium content.
Of course, Zhihu will also be able to give KOLs who continue to output high-quality content the influence they deserve, including commercial returns. After all, content is the core value.

High-quality content is the soil of commercialization. The biggest challenge: Semantic recognition of content communities. Tech creators: high-quality content and commercial content have their own value for Zhihu. As a heavy user of Zhihu, I feel that Zhihu can control business information in one place. Within a comfortable range, what principles and concepts are based on Zhihu's sense of balance between commercial content and high-quality content?
Li Dahai: The realization of commercialization does not necessarily affect the quality, professionalism and neutrality of the content. There is no inherent contradiction between the two, but it depends on how commercialization is done. In the past few years, Zhihu has made many attempts and explorations at the commercialization level. Judging from its achievements, Zhihu has formed a relatively stable business model and diversified income sources. However, the premise of these commercial explorations is still to maintain the high-quality production and operation of the community and content. Zhihu has always insisted on not exchanging low-quality content for traffic, sacrificing community atmosphere in exchange for income, and not destroying the network ecology by rubbing hot spots and gaining attention. It's still the same to this day. When earning income collides with our fundamental values, we choose to give up our income and defend our values. Reasonable income promotes content creation, because commercialization encourages the production of professional and neutral high-quality content, and maintains the professional characteristics of Zhihu, rather than changing and shaking.
​
Kechuang people: In Zhihu, which is the biggest problem and challenge you encountered? Can you describe the specific events, problem-solving ideas and solutions in relative detail?
Li Dahai: Zhihu is a content community. How to better drive the production and circulation of content through technical means is of great significance to us. We have been trying to broaden the application boundaries of AI algorithms in the content community. So far, there have been many application results. Zhihu AI algorithm application has also run through multiple scenarios from content production, consumption and community governance.
There are many technical challenges. I think the biggest challenge is the precise identification of semantics, such as the identification of the yin and yang strange content of "Wali". The core difficulty lies in the complexity of network language. For example, the classic "hehe" will have very different meanings due to different relationships between the two parties, different scenes and times of speech. Even manual judgment is difficult to standardize, and the training challenge of algorithmic models is even more difficult. Based on the cumulative observation of users' practical experience, Zhihu, together with the algorithm team, has established a content emotion model, a user intimacy model and a text recognition model from the three aspects of emotional tendency, intimacy, and text characteristics. Iterative perfection. The combination of the three models not only gets rid of the limitations of a single algorithm model, but also makes "Wali"'s yin and yang strange gas recognition accuracy surpassing most human judgments.

Technical team management experience and performance, selection of scientific and technological talents: As the manager of a technical team of more than 800 people on Zhihu, can you share your management experience, such as how to keep the same frequency with the strategy and how to select talents?
Li Dahai: My management framework is a point, a line, and a point. The point is the key position. What skills are needed and what requirements must be met must be accurate and effective; the line refers to the extraction of the most important things in the large team. This may require Many teams can only do well with cooperation. How these teams break down their goals, how to cooperate and divide labor, and how to set up processes requires in-depth consideration; the aspect is the system, norms, and culture, which can affect the team's engineers in all aspects as much as possible to improve cohesion.
Zhihu's R&D team attaches great importance to aligning the business, and everything is aimed at achieving business goals and supporting business value. Therefore, we hand over the performance appraisal rights of the business R&D team to the business team to ensure that the R&D team gives the greatest support to the business.
As for the selection of materials, there is no perfect person. Once a person is over 30 years old, a concept is formed in his heart, and it is very difficult to change Ta. But on the other hand, people have great flexibility and flexibility, so we must use people's strengths to stimulate his strengths, so that his heart is really willing to do this.
​
Kechuangren: Your personal hobby, future personal life vision?
Li Dahai: I usually like to play football and poker with my friends. The future life vision and Zhihu's business have a high degree of fit. In the long run, I hope to help more people through content services, knowledge services, interesting education, etc., and to lead Zhihu with my colleagues to create more value, whether it is an enterprise Management or technological innovation can solve more problems and enter a new situation.
​
-- End--