中文精读
增长不是只把人带进来，还要让供给与反馈学会接住人
嘉宾：于北川，Lessie AI 创始人、抖音团队早期产品经理（当期身份）
主持：曲凯
原发布：2025 年 11 月 8 日｜时长：84 分 49 秒
整理：2026 年 10 月 18 日｜官方单集
从抖音早期增长讲到找人智能体 Lessie AI，这期看似跨了两个时代，核心问题却相近：带来更多用户以后，产品内部能不能形成下一轮价值？内容社区需要供给、消费与推荐共同运转，找人产品则要从单边效率走向可能的关系网络。
于北川在 2018 年加入抖音，主要亲历至 2020 年的消费端产品与增长。他也提醒，今天能讲出许多解释，不代表当时知道结局。本篇把这些解释当作有价值的内部回忆，保留非共识、失败和外部条件，不把一条成功史写成必胜配方。
一、把关系指标做高，不等于做成社交
回听：02:49—16:32。 于北川回忆，加入时抖音约四千万日活，团队仍担心娱乐产品的天花板。做熟人社交，是希望关系带来更多打开动机：今天不想看视频，也可能因为朋友而回来，提升长期留存并扩展人群。
难点是自然互动无法简单拆成好友数与互动频率。指标进入个人目标后，负责人可以用红包、积分或引导动作把它们做高，用户却未必真的有理由在这里继续联系。局部都“完成”，整体体验仍可能不连贯。
这不是数据无用，而是代理指标可能偏离真实目标。一个新增好友是关系的记录，不一定是有价值关系；一次受奖励驱动的互动，也不一定会再次发生。要观察奖励撤掉后、活动结束后，人还会为何而来。
他回忆 2019 年春节活动一度想建立社交，临上线前两周又改成拉新，两种目标都没有做好。活动能推动动作，却不能替代长期场景；战略摇摆还会让有限资源同时失去承接。
二、冷启动的调性，是供给与人群一起选出来的
回听：18:04—28:05。 于北川把早期差异归于三件事：没有直接从头条导入泛用户；黑色界面与年轻品牌；强内容运营。街舞、滑板、极限运动与技术运镜，服务一批具体年轻人。这些选择共同定义“谁会觉得这里属于自己”。
算法不是在第一天就知道答案。内容池小、行为数据少时，人工精选承担方向选择，并给代表性内容更多曝光。运营在这里不是算法之外的装饰，而是产生最早学习信号的方式。
泛化也不只是放宽人群。他描述一条链：及时引入新垂类创作者，给予初始曝光，让早期消费者产生反馈，推荐系统再学习，新的内容吸引新群体。若没有最初的探索机会，内容可能因“没有历史表现”始终得不到学习机会。
年轻人兴趣更广，是嘉宾的经验判断，不应变成人口规律；三项选择也没有对照实验。编辑推论： 小团队可借用的不是黑色 UI，而是让品牌、种子人群、供给与最初分发保持一致。杂乱导流可能增加人数，同时稀释学习信号。
三、单列与双列，改变的不只是观看效率
回听：28:16—31:41、32:53—43:32。 单列全屏让消费连续，平台容易插入下一条内容或广告。双列先让人选封面，可能给普通创作者和图文更多机会。于北川回忆，早期快手更平均的流量、突出的评论路径，给创作者不同的回报预期。
因此，一次观看更久，不等于整个生态更好。普通人是否得到首个反馈、愿不愿再投稿，决定未来供给。头部消费效率与普通创作者反馈，有时需要权衡，不能只拿消费侧时长裁决。
主持问为什么不做 A/B 实验。嘉宾指出，创作者同时面对实验内外的消费者，消费侧随机分组未必隔离创作侧长期影响。这里的问题是网络外溢与周期，不是所有生态变化都不可实验。需要考虑更合适的分群、阶段和长期证据。
容器也有题材边界。二三十分钟的知识或旅行视频，需要不同耐心；不能外放时，图文攻略可能更合适。连续短内容的反馈密集，不说明它能最好地分发所有东西。节目口语中的“无脑”，在这里理解为低认知负担的放松场景，不作为对用户或内容的贬低。
四、大额增长投入，依赖回收能力与跨组织配合
回听：34:42—37:03、48:55—58:14。 于北川对微视与视频号的比较，重点是入口动机：杂乱内容和大量导流，不一定给新社区清楚定位；好友点赞则可以先提供进入理由，再积累真实消费反馈。这是他的解释，未控制时机、组织和其他因素，不是独立的公司成败归因。
到成熟阶段，他回忆极速版、火山改名与集团多端春节导流。早期拒绝泛流量，与后来集团协同并不必然矛盾，因为产品承接能力已经不同。2020 年红包二十亿元、日活峰值 4.7 亿、按用户长期价值计算回报为正，均是其当期回忆，本文没有审计。疫情对后续留存的影响也不能删掉。
这条增长机制不是“敢花钱就赢”，而是新增、使用时长、变现能否支持持续投入。用户长期价值需要实际留存、收入和成本条件；小团队没有同等资源，也不能照搬集团补贴规模。
他还纠正“有 A/B 就随便上”的印象：大型产品先评审和论证，实验在后面验收。但谨慎也可能错过机会，例如他认为直播进入主流的尝试可以更早。可用的张力是保护既有体验，同时为新场景保留可控探索空间。
五、文化落在行动，极致首先要求改写问题
回听：58:40—01:06:03、01:18:27—01:21:47。 于北川强调弱化职位标签，让年轻人可以与核心负责人直接讨论业务，获得“我也可以承担”的感觉。他把文化理解为早期成员如何思考与做事，再影响后来者；扩张会改变这种传递，不能只靠墙上的词保持。
“怎样增长百分之十”与“怎样增长两到五倍”会产生不同答案。前者容易累加功能优化，后者迫使团队考虑新渠道、新资源与战略变化。这个提问技巧的价值是打破既有约束，不是无条件要求更多预算或更长工时。
到创业阶段，负责人还要从熟练的业务动作里抽身。写文档、测 bug 可以给即时反馈，却可能回避更难的问题：产品定义是什么，关键负责人在哪里，未来六至十二个月资源怎样准备。不是细节不重要，而是创始人还须组织公司继续存在所需的条件。
编辑推论： 每周同时检查“把当前动作做好”和“改变关键约束”两类工作。后者也要有具体产物，不能只写战略思考；例如明确一项招聘责任、一个现金假设或一次用户范围收缩。
六、Lessie 的双边愿景，先从一边愿意付费开始
回听：01:07:07—01:17:54。 Lessie AI 当时定位为找人的智能体：替销售、营销、外贸和招聘等角色检索数据库，逐个审阅画像，寻找联系方式并辅助联系。它切的是“一找多”的劳动密集流程，效率与准确性是嘉宾产品陈述，节目没有独立成功率验证。
长期设想是，AI 提高联系密度后，收件人也需要 AI 筛选；双方可能逐渐形成平台内的关注、沟通与资料认领。但第一天直接做双边网络，容易因为两边都太少，谁都没有价值。因此先借邮件等已有渠道，让单边工具有用、有人愿付费，再观察关系是否真的长出来。
这里有两道不能跳的门：找到人不等于对方愿意被联系；发送更多不等于合作更多。收件人意愿、信息准确性、拒收与质量控制，需要独立设计。未来双边网络是愿景，不能拿当期找人功能证明已经建立。
他回答“大厂也来做怎么办”时，串起三个问题：长期留下什么，为什么现在有窗口，怎样在窗口里起量。用户数据与关系网络也必须有合法使用边界；巨头暂时没关注，并不是永远的护城河。
应用练习：给增长链补一条反证
挑一个内容新垂类，记录引入供给、初始曝光、有效消费、作者再创作与新用户进入五个环节。每个环节写一个失败信号：例如点击升高但无人看完，消费满意但作者没有反馈，新增上涨却只有奖励存在时才回来。
再挑一项找人或沟通工具，分别写发送者所得、收件人所得、双方同意的条件，以及单边价值尚未形成前可借用什么现有渠道。最后只选一个主要增长目标，明确活动结束后的承接。目标不是凑更多动作，而是检查下一轮价值从哪里来。
来源边界
依据官方简介与完整转写。2018—2020 年经营数字、资源安排和公司比较均为于北川回忆；更早立项故事为他后来听来，未作亲历事实。Lessie 的效率、网络与发展路径属于 2025 年产品自述和愿景。片尾音乐不采作主持观点，转写损坏数字不补写。编辑推论与练习另行标识。
回听导航
- 10:51—16:32：关系目标为何会被指标替代。
- 18:04—28:05：调性、精选与泛化学习链。
- 28:16—31:41：容器、创作者与实验外溢。
- 37:24—43:32：分发题材的场景边界。
- 46:23—58:14：活动目标、集团增长与论证。
- 58:40—01:06:03：行动习惯与改写增长问题。
- 01:07:07—01:21:47：Lessie 的单边起点与创业责任。
相关主题：内容冷启动、增长证据、单边到双边。延伸收听：FlowGPT 的创作者生态、Mercor 的专家匹配与履约。
English edition
Growth needs more than arrivals: supply and feedback must learn to serve them
Guest: Yu Beichuan, founder of Lessie AI and an early product manager on Douyin; roles at the time
Host: Qu Kai
Originally published: November 8, 2025 | Duration: 84 minutes 49 seconds
Edited: October 18, 2026 | Official episode
Douyin's early growth and the people-search agent Lessie AI seem to belong to different eras. They share a question: after users arrive, can the product create another cycle of value? A content community needs supply, consumption, and recommendation to work together. A people-search tool must establish useful individual productivity before a possible relationship network.
Yu joined Douyin in 2018 and mainly recalls consumer product and growth work through 2020. He acknowledges that explanations available today were not clear then. This reading preserves uncertainty, failures, and external conditions rather than turning success into a guaranteed recipe.
1. Improving relationship metrics does not establish social interaction
Listen: 02:49–16:32. Yu recalls roughly forty million daily users when he joined, alongside anxiety about an entertainment product's ceiling. Familiar-person social features might create reasons to return even without wanting video, strengthening retention and broadening adoption.
Natural interaction cannot simply be decomposed into friend counts and frequency. Once those measures become individual targets, rewards and prompts can raise them without giving people a reason to maintain contact. Locally completed goals can coexist with an incoherent experience.
This is a problem with proxy measures, not evidence that data is useless. A recorded friend does not necessarily represent a valuable relationship, and rewarded interaction may never recur. Observe why people return after rewards and campaigns end.
Yu recalls a 2019 holiday campaign intended to build social behavior, then redirected toward acquisition two weeks before release. Neither objective went well. Campaigns can induce actions without creating lasting settings, while shifting strategy weakens the follow-through for both objectives.
2. Early identity comes from choosing supply and people together
Listen: 18:04–28:05. Yu attributes early differentiation to avoiding broad traffic from Toutiao, a youthful brand and dark interface, and strong content operations. Dance, skateboarding, extreme sports, and technical camera work served a specific young audience. Together they communicated who would feel at home.
Algorithms did not know the answer on day one. With limited content and behavior, human curation selected a direction and supplied representative work with exposure. Operations helped generate initial learning signals rather than merely decorating an algorithm.
Expansion required more than a broader audience declaration. Yu describes introducing new creator categories, providing initial exposure, receiving consumption feedback, letting recommendations learn, and attracting new audiences through that supply. Without exploration traffic, a category can remain trapped by its lack of historical performance.
Young people's broader interests are his experience-based interpretation, not a demographic law. The three early choices lack a controlled comparison. Editorial inference: Small teams should borrow alignment among brand, seed audience, supply, and distribution rather than copy a black interface. Chaotic traffic can raise headcount while diluting learning signals.
3. One or two columns change more than viewing efficiency
Listen: 28:16–31:41 and 32:53–43:32. Full-screen single-column feeds enable continuous viewing and easy insertion of the next item or advertisement. Two columns ask people to choose a cover and may create more room for ordinary creators and image posts. Yu recalls more evenly distributed traffic and prominent comments on early Kuaishou producing different creative incentives.
Longer viewing does not necessarily mean a healthier ecosystem. Whether ordinary creators receive initial feedback and publish again affects future supply. Efficient head content consumption and feedback for ordinary contributors can require tradeoffs beyond watch time.
Asked why not run an A/B test, Yu notes that creators serve consumers inside and outside experimental groups. Randomizing consumers may not isolate long-term supply effects. The problem is spillover and time, not a ban on experimenting with ecosystems. Grouping, stages, and longer observation need consideration.
Containers also favor different material. Twenty- or thirty-minute knowledge and travel videos require another kind of patience. Written guides can fit situations where sound is unavailable. Dense feedback from short content does not establish universal distribution superiority. The episode's colloquial “mindless” describes low-effort relaxation here, rather than a judgment about users or content.
4. Large acquisition spending needs recovery economics and coordination
Listen: 34:42–37:03 and 48:55–58:14. Yu compares entry motivations in Weishi and WeChat Channels: mixed supply and heavy traffic may fail to define a new community, while friends' likes can provide a reason to enter and generate real consumption signals. This is his explanation, without controlling timing, organization, or other causes.
At a later stage he recalls a lightweight app, Huoshan's renaming, and traffic coordinated across the group. Early avoidance of broad traffic and later group promotion need not conflict because the product's capacity to serve arrivals had changed. His recollections of two billion renminbi in 2020 holiday rewards, a 470 million daily-user peak, and positive returns based on lifetime value are not audited here. Pandemic effects on subsequent retention also matter.
The mechanism is not simply spending boldly. Acquisition, usage, and monetization must support continued investment. Lifetime value requires real retention, revenue, and cost assumptions. Small teams cannot copy a large group's subsidy scale.
Yu also challenges the idea that A/B testing allowed indiscriminate launches. Large product changes required argument and review before experimental acceptance. Yet caution could miss opportunities; he would have tested live streams in the main feed earlier. Preserve existing experience while allowing controlled exploration of new settings.
5. Culture is action; ambition begins by rewriting the question
Listen: 58:40–01:06:03 and 01:18:27–01:21:47. Yu emphasizes reducing status labels and allowing young staff to discuss work directly with senior decision makers. That encourages responsibility. Culture consists of early people's habits influencing later hires; expansion changes this transmission, and slogans alone cannot preserve it.
“How can this grow ten percent?” differs from “How could it grow two to five times?” The first invites accumulated feature optimizations; the second forces consideration of channels, resources, and strategy. The technique breaks inherited assumptions rather than automatically demanding more budget or longer hours.
Founders must also step outside familiar operating tasks. Documents and bug checks provide immediate feedback while potentially avoiding harder questions: product definition, critical leaders, and resources for the next six to twelve months. Detail remains important, but organizing continued company existence becomes part of the job.
Editorial inference: Review both improving current actions and changing crucial constraints each week. The latter needs concrete outputs too, such as a hiring responsibility, a cash assumption, or a narrower audience—not merely an entry called strategy.
6. Lessie's two-sided vision starts with one side willing to pay
Listen: 01:07:07–01:17:54. Lessie AI then positioned itself as an assistant for sales, marketing, export businesses, and recruitment: search databases, review profiles, find contact information, and assist outreach. Its entry point is labor-intensive one-to-many search. Efficiency and accuracy remain the founder's product claims without independently demonstrated success rates.
The longer vision is that more AI-assisted outreach creates demand for recipient-side filtering. Platform follows, conversations, and profile claims might emerge. Starting directly with both sides risks too few participants giving either side value. Instead, existing email channels support a useful, paid one-sided tool before testing whether relationships grow.
Two gaps remain: finding someone does not establish willingness to be contacted, and sending more does not mean more cooperation. Consent, information accuracy, rejection, and quality need independent design. A prospective network cannot be treated as already established by search features.
Asked about large companies entering, Yu connects three questions: what endures, why a window exists now, and how growth occurs within it. User information and networks also require lawful use boundaries. A giant's temporary inattention is not permanent defensibility.
Exercise: add a falsifying signal to the growth chain
Choose a new content category. Record supply introduction, initial exposure, useful consumption, repeat creation, and incoming users. Add a failure signal to each: rising clicks without completion, satisfied viewers without creator feedback, or users returning only while rewards persist.
For a people-search or communication tool, separately describe sender value, recipient value, mutual agreement, and existing channels available before both sides form a network. Choose one main growth goal and specify what happens after the campaign. The point is to identify the next cycle of value rather than accumulate actions.
Source boundary
This reading uses the official introduction and complete transcript. Operating figures, resources, and company comparisons for 2018–2020 are Yu's recollections. Earlier origin stories were heard later, not personally witnessed. Lessie's efficiency, network, and development path are product statements and aspirations from 2025. Closing music is not treated as host commentary, and damaged numbers are not reconstructed. Editorial inferences and the exercise are identified separately.
Listening navigation
- 10:51–16:32: Relationship goals displaced by metrics.
- 18:04–28:05: Identity, curation, and learning through expansion.
- 28:16–31:41: Containers, creators, and experimental spillover.
- 37:24–43:32: Contextual limits on content distribution.
- 46:23–58:14: Campaign objectives, group growth, and review.
- 58:40–01:06:03: Habits and reframing growth questions.
- 01:07:07–01:21:47: Lessie's entry point and founder responsibilities.
Related themes: content cold starts, growth evidence, one-sided tools becoming networks. Further listening: FlowGPT's creator ecosystem and Mercor's matching and fulfillment.