Zhihu's enemy isn't Douyin, it's a scale-driven version of itself
Zhou Yuan looks back: after going public, video and short content were two strategic retreats, and the root cause was treating a community as a scale business. A community can only be ecosystem-driven, growing by overflow, not by expanding the roster.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
When GPT3.5 came out, the first reaction was disconnection
Zhou Yuan says the timing wasn't a month before the release, but at the end of October 2022 when GPT3.5 came out, just days after a strategy meeting had ended, and the first reaction was "this strategy needs to be updated again." He describes a sense of disconnection: for the past two or three years the theme for Chinese tech internet companies had been cost cutting and efficiency, while across the ocean technology suddenly took a step forward. Zhihu then contacted OpenAI engineers, and quickly found that "the things you can think of, actually at that time everyone would think of," and many Chinese companies were contacting the same people; in the end most of the engineers they reached had no intention of changing jobs.
— Zhou YuanOffering $100 million to push large models was thinking too small
Zhou Yuan's offer to a friend whose background suited building large models was: Zhihu has many scenarios, and can put up $100 million to push this. The other party politely declined, and then in the first half of 2023 his team quickly raised a few hundred million dollars. Zhou Yuan says, "Looking back now, I was still thinking too small." This is one of the few moments in the whole episode with a specific dollar figure and self-correction.
— Zhou YuanMaking AI pose as users is a dead end
Zhou Yuan says that in early conversations with big companies, some judged that AI doesn't need sleep and doesn't get tired, so you could turn AI into NPCs in the community to talk with people and pose as users, and the data would go up. He felt at the time this definitely wouldn't work: communication between people requires trust; users simply talking with AI is fine, but mixing them together so you don't know the other side is AI and think it's a user would challenge the incentive mechanism. His analogy is a digital virtual singer performing on the same stage as a real singer—the virtual singer isn't a question of singing for 24 minutes, it can sing for 24 years, and can hit all the high notes, so then it's meaningless.
— Zhou YuanAI isn't just a medium, it changes information in transit
Zhou Yuan gives an abstract but useful framework: understand AI as a medium; information goes from point A to point B, and what's different about AI is that it can change the information during transmission. So if you treat the whole community as a big medium, AI can play more roles inside it and assist people in improving communication. He also acknowledges that in the BERT year (2018 or 2019), what small models could do in the community was valuable but very limited.
— Zhou YuanHot list summaries went offline: compression kills native value
Zhihu made hot list summaries, many users used them, but later took them offline. Zhou Yuan explains: the hot list has a lot of open-ended discussion, and the value lies in the discussion itself and reading different viewpoints; after summarizing and compressing, a lot of the native value subject is lost—you don't necessarily have to read 10 or 100 answers to reach one consistent conclusion; you might have several different conclusions, and that's more valuable. His characterization of this kind of product experiment is "not necessarily in the small building, and not necessarily reaching 5%."
— Zhou YuanWhere search engines should be disrupted has already been disrupted several times over
Zhou Yuan sees search engines and the web pages and sites they can index as one overall ecosystem, with two types of needs: finding information, and completing various tasks (booking flights, finding images, ordering takeout, borrowing books). The latter type isn't done by search engines, it's done by going to websites. He says, "Where search engines should be disrupted, at present they've already been disrupted several times over." Takeout is search with a pull-down, but it carries supply chain, delivery people, and review information, and traditional search engines find indexing and interaction very troublesome. So the search market is a highly fragmented market; as long as the commercial flywheel of production and consumption can close the loop at lower cost, there's an opportunity to become an independent vertical search.
— Zhou YuanWhat Perplexity lacks is where the new developers are
Zhou Yuan maps Perplexity to "Perplexity plus the model behind it that compresses all web pages," and in between there's still a piece left: the services that traditional search engines can index. Some services have already become closed-loop companies like takeout; the remaining part he thinks may need to be solved through more agents. He sees the small and medium website operators behind the original big flywheel as developers, and the entire open web like an algorithm-driven big app store. So the question becomes: where are the new developers, and what are they. If the two together can solve consumer problems at lower cost and higher efficiency, then it's very imaginative.
— Zhou YuanA community is ecosystem-driven, not roster-expansion growth
Zhou Yuan attributes the post-IPO wavering to one essential difference: community growth is ecosystem-driven, overflow growth, not roster-expansion growth. The ecosystem must be very good and healthy to overflow and attract more users; scale-driven is more number-oriented, roughly equivalent to budget-driven, which is to say it wants a numerical result. Ecosystem-driven requires a longer time to build, like land value that needs cultivation. So using a scale-driven approach to do things on community soil, whether video or short content, can deform. He says video and short content today should both be very organic supplements to community content; the problem isn't the media form itself.
— Zhou YuanThe first month after IPO was unhappy, then inflation set in
Zhou Yuan admits he was a bit inflated in the period right after going public. The first month was unhappy because the day of the IPO happened to be the day a certain Chinese concept stock blew up, all Chinese concept stocks avalanched, and Zhihu went public that day; he felt at the time his luck was especially bad and knowledge value was undervalued. Then because they raised a lot of cash, and after 2021 commercialization growth and user growth were relatively fast, from a model perspective he began to inflate internally, feeling that as long as you add investment and keep expanding scale, all problems could be solved. He says this was an illusion; the more real thing is still the very real experience of each creator in the community.
— Zhou YuanCouldn't make peace with the stock price for months, then split it into two layers
Asked how he made peace with the stock price acting like a plunging roller coaster, Zhou Yuan says for the first few months he couldn't really make peace with it; every night he'd take a look, and seeing it fall again he'd think "who's buying our stock again." Later he entered another state, splitting the problem into two layers: one is the overall problem of Chinese concept stocks over the past period, and the other is not being too affected by external changes, not catering to the market with numerical disclosures or strategic moves, but needing to be very firm in being yourself, with the core being sustainable development and sustained profitable growth.
— Zhou YuanA CEO's loneliness has nothing to do with whether anyone is with them
Zhou Yuan raises a paradox: many people say a company CEO is a relatively lonely person, but this has nothing to do with "whether anyone is with them." A CEO's loneliness shows in that some decisions only you can make; even if someone can discuss it with you, the part they can share is relatively small. It doesn't show in daily communication, nor in whether the people who started the company with you are still on the same boat—even if they are, this CEO may still be lonely when he should be lonely.
— Zhou YuanIn their own words · checked verbatim
In my memory it should have been October 2022, the end of October, GPT3.5. At that time we had probably just finished a strategy meeting, only a few days before, and after it ended the first reaction was that this strategy needs to be updated again.
我印象里面应该是2022年10月份 10月底吧 GPT3.5 我们在那个时候应该是刚开完战略会 没几天 然后开完以后第一反应是 这个战略需要重新再更新一下
Zhou Yuan3:08
It's like at that moment the rocket is launching, and suddenly a company next to you tells you, should we go build a rocket? Then the people in the middle of the launch would definitely say, I'd rather wait and see.
就好比说在那一刻 就是这个火箭正在发射 突然之间旁边有家公司告诉你说 我们要不要去造个火箭 那你在发射过程中的人肯定就说 那我还是等着在看吧
Zhou Yuan5:10
It's like you have a digital virtual singer, and you arrange for him to sing with one of our real singers. It's actually meaningless. The virtual singer isn't a question of singing for 24 minutes, he can sing for 24 years, he can hit all the high notes, then it's meaningless.
就好比说你有一个数字虚拟歌手 你安排他去跟我们的一个真实的歌手去唱歌 他其实没有什么意义 虚拟的歌手 他不是唱24分钟的问题 他可以唱24年 他所有的歌运都可以唱得上去 他就没有意义了
Zhou Yuan10:13
I think up to now, where search engines should be disrupted, at present they've already been disrupted several times over.
我觉得截止到目前为止 搜索引擎 该被颠覆的地方 目前都被颠覆了好几轮了
Zhou Yuan26:22
Ecosystem-driven needs a longer time, needs building, it's like land value, it needs cultivation.
生态驱动 它需要更长的时间 需要建设 它就像一块土地价值一样 需要培育
Zhou Yuan1:03:48
Even if they are, this CEO may still be lonely when he should be lonely.
即使在 这个CEO他可能还会在 他该孤独的时候 依然很孤独
Zhou Yuan1:16:51
Figures
| Funds Zhou Yuan offered to a large model partner | $100 million | 5:10 |
| Amount the other team raised in the first half of 2023 | A few hundred million dollars | 6:11 |
| User penetration of hot list summary-type product experiments | Not necessarily over 5% | 20:18 |
| Share of the core population most empowered in the AI era | Possibly 0.2% | 1:00:46 |
| Time from Zhihu's launch to switching domains | A month later | 53:35 |
| Zhou Yuan's internship salary as a reporter | About 2,000 yuan | 40:31 |
Glossary
- PMF / Product-Market Fit
- The state where a product matches market demand and users keep using it.
- LTV / Lifetime Value
- The total revenue a user brings over their entire lifetime.
- NPC / Non-Player Character
- A character in a game controlled by the program, not played by a real person.
- Perplexity / AI search product
- A search product that generates answers with AI and attaches sources.
How to listen
Founders and product leads of content communities and AI search products, and investors who care about strategic wavering at community companies after they go public.
0:00-2:06 opening and chapter preview, and a segment with garbled transcription at 34:28-35:30.