2018–2020: The Key Three Years When Douyin Overtook Kuaishou | A Conversation with Lessie AI Founder 于北川
Summary
Douyin overtook Kuaishou in 2018–2020 through a continuous positive loop of high-affinity seed users, efficient full-screen single-column distribution, category expansion, and commercialization-funded growth—not a single algorithm upgrade. Douyin refused to import seed users directly from Toutiao, using a black interface, a youth-oriented brand, and skateboarding, street dance, and technical content to break in, then relying on editorial curation to help the algorithm through cold start; by late 2018 to early 2019, its DAU had surpassed Kuaishou. “Without those people’s intervention in the early days, Douyin would never have taken off.”
Douyin’s long-running internal anxiety was not short-term DAU, but the ceiling of an entertainment-consumption product, which is why familiar-person social networking remained a strategic priority for 3 straight years. Even the earliest optimistic internal forecast was only about 12M DAU; even at 30M–40M, the team still benchmarked against the roughly 100M–150M DAU ceiling of products like Toutiao. The social experiment ultimately “almost certainly failed,” but in hindsight it did not stop Douyin from becoming an approximately 800M-DAU product, second only to WeChat. “The faster it grew, the more anxious we became, because we probably knew we were getting closer to the ceiling.”
Single-column versus double-column was not a UI preference, but a foundational system governing distribution fairness, creator supply, user psychology, and ad efficiency. Kuaishou’s double-column format let users choose content first, distributed traffic more evenly, and generated better returns for ordinary users and image-text creators; Douyin’s single-column format could force-feed content and ads efficiently. This two-sided ecosystem shift could not be validated by a short-term A/B test because “you can only test the consumption side, not the creation side.” 于北川’s view is that Kuaishou would have monetized better—and posed a greater threat to Douyin—if it had switched to single-column earlier.
The real gap opened during several extreme growth campaigns: an accidental Spring Festival resource allocation once doubled Douyin’s traffic, while the 2020 group-wide campaign against Kuaishou pushed peak DAU to 470M. Douyin Lite went from project approval to launch in just 2 weeks, with 7 key products across the group driving traffic through “Douyin Cards” and full-scale push notifications; Kuaishou distributed RMB10B in red packets, ByteDance RMB20B, and Zhang Yiming even asked, “Do we need to spend RMB100B?” LTV calculations showed positive ROI, and the pandemic subsequently helped retain the new users.
ByteDance could pay 2x competitors’ user-acquisition prices because Douyin had built strong monetization early and could recover the cost of new users within 1 year. Livestreaming, e-commerce, and local services were all tested early, but meaningful scale still depended on the platform first building a sufficiently large traffic pool. “Once traffic is large enough and the distribution container is efficient enough, many things happen naturally.”
ByteDance’s hardest early asset to replicate was not brute-force spending, but a flat organization that trusted young people and pursued strategic “extremity.” Roughly half of Douyin’s dozen-plus product managers were fresh graduates, and interns were expected to take direct ownership of problems. Extremity meant reframing “how do we grow 10%?” as “how do we grow 2x, 3x, or even 5x?”—forcing the team to rethink channels, resources, and product institutions instead of stacking five 2% optimizations.
Lacey AI’s startup bet is to first turn “one search into many” into a high-efficiency PLG tool through a People Search AI Agent, then convert one-sided usage into an intentional social network. The product targets sales, marketing, foreign trade, and recruiting in Europe and the US, using data sources such as ZoomInfo and Apollo to search, review candidates one by one, find contact details, and initiate communication; over time, it aims to have AI process 98 of 100 partnership requests and send only the 2 best matches to the user. 于北川’s view is that a pure efficiency tool may be absorbed by the model, so founders must answer what remains 3–5 years out rather than “putting their lifeline in the hands of a foundation model’s whim.”
Deep dive
1. Douyin Took Off Amid Massive Non-Consensus
Douyin was formally launched as a project around mid-to-late 2016 and began scaling in July or August 2017. When 于北川 joined as a fresh graduate and intern around March or April 2018, DAU was about 40M; during his interview the previous year, he had never even heard of the product.
To fresh graduates at the time, Toutiao looked “kind of dated,” and Douyin was not the first short-video product. Weishi, Xiaokaxiu, and others had already risen and fallen, leading the market to suspect that Douyin was merely “one wave.” Internally, too, no one knew what its long-term retention or ceiling would look like.
Douyin’s product team initially had only a dozen-plus people, roughly half of them fresh graduates. After 于北川 joined, 卷卷 told him: “Here, you don’t need to think of yourself as an intern, and you don’t need to think of yourself as a fresh graduate.” His strongest first impression was of a young, busy team taking direct ownership of important problems.
2. Two Accidental Traffic Surges Validated Young Users’ Retention
The first clear growth wave was tied to live quiz shows. ByteDance packaged the function as an SDK and embedded it in Toutiao, Xigua, Huoshan, and Douyin; university users were interested in the format but preferred downloading Douyin because Toutiao’s brand felt old, and then stayed for the content.
于北川 recalls that the 2017–2018 Spring Festival Gala resources may originally have been intended for Huoshan, but were transferred to Douyin for some reason, directly doubling the product’s traffic during the holiday. He called the acceleration “highly accidental,” but it gave the company its first clear view of the potential upside.
From late 2018 to early 2019, Douyin’s DAU surpassed Kuaishou; from 2019 to 2020, it continued widening the gap and became what 于北川 calls “the largest-DAU product in China apart from WeChat.” He views June 2017 through June or July 2020 as the full high-growth period.
3. The Closer the Team Got to the Old World’s Ceiling, the More Anxious It Became
于北川’s account of the early project logic is that Zhang Yiming mapped video length from “zero seconds” to 1–2 hours: mid- and long-form video and film already had platforms, but a structural gap remained between zero and 30 seconds, and zero and 1 minute. ByteDance therefore launched Douyin, Huoshan, and Xigua in parallel.
Douyin initially received very few group resources. Some estimated its ceiling at 6M DAU; the most optimistic forecast was about 12M. Even at 30M–40M, the team was still benchmarking against the roughly 100M–150M ceiling of products like Toutiao, while Kuaishou had already exceeded 100M.
That frame of reference created counterintuitive pressure: “The faster it grew, the more anxious we became.” In hindsight, one can explain how Douyin reached approximately 800M DAU, but 于北川 emphasizes that at the time almost no one dared believe a pure entertainment-consumption product could become China’s second-largest by DAU after WeChat.
4. Social Was the Main Plan for Breaking Through the Ceiling—but Not the Final Answer
于北川 distinguishes social networking from community by asking what comes first, relationships or content, before interaction occurs: in familiar-person social networking, people know each other first and consume the relationship; in a content community, content comes first and generates interaction. The two should not be conflated.
The team believed relationship density could increase the motivation to open the app: even without a strong entertainment need, users would return because of interactions with friends, improving open frequency, long-term retention, and user-base expansion. This is the mechanism that sustains DAU at WeChat, Facebook, and Instagram.
From 2018 to 2020, social networking remained an “absolute strategic priority, even the number-one priority” for Douyin. The company explored familiar-person social networking, the group launched Duoshan, and products such as Echo appeared in the market. 于北川’s conclusion is blunt: “It almost certainly failed.”
5. Data Can Be Hacked; Social Cannot Be Reduced to Local Metrics
曲凯 asked whether ByteDance’s inability to build social products was really a company problem, given that no new major social product had emerged in China during the same period. 于北川 accepted the broader context but still believed ByteDance was “not as good” at cultivating ecosystems between people.
If social networking is decomposed into median friend count, interaction frequency, and penetration, then written into OKRs level by level, teams will pursue the metrics “by any means necessary”: points, red packets, and forced prompts to add friends can all manufacture data without showing natural interaction in a real use case.
于北川 summarizes it this way: “Data can be hacked.” Social requires a complete context, coherent flow, and an instinct for people; local optimization can easily damage the whole. That is precisely where it conflicts with ByteDance’s strengths in rational decomposition and rapid experimentation.
曲凯 notes that the recent “keep the streak alive” feature seems effective. 于北川 acknowledges its value but sees it as a shallower form of social networking that no longer carries the strategic mission of breaking through the company’s ceiling.
6. Douyin Used Distinctive Positioning to Avoid a Direct War with Kuaishou and Huoshan
In early 2018, Huoshan already had more than 30M DAU, at one point exceeding Douyin, but shared similar seed users and everyday, lower-tier content with Kuaishou, whose DAU was approaching 100M. Kuaishou was larger and had a more complete creator ecosystem, giving Huoshan a structural ceiling.
Douyin firmly refused to import seed users from Toutiao. 于北川 sees this as one of 卷卷’s important principles: even if moving from zero to one became harder, the team wanted to prevent Toutiao’s existing audience from dictating the new community’s content demand.
On the brand side, Douyin used an all-black interface—rare among major apps at the time—along with an independent logo and youth-oriented visual identity. On the content side, it concentrated on skateboarding, street dance, young men and women, and technical camera work. 于北川 stresses that this was not about superior taste, but about being “distinctive and independent enough in its positioning.”
7. In Cold Start, Editorial Judgment Mattered More Than Recommendation Algorithms
When the user and content pools were still small, machine learning had too little behavior to learn from. Early Douyin therefore created an “Editor’s Picks” label, with operations staff selecting content aligned with the community’s direction and product rules ensuring it received more exposure. In substance, this manual selection functioned as a media editor-in-chief.
曲凯 challenged the premise: editorial taste usually implies a niche product, so how could it support a mass-market platform? 于北川’s answer was that Douyin’s young users were inherently more generalizable—they could watch street dance, anime culture, and remix videos, but might also watch “tacky videos.” Their range of interests was much broader than that of older users.
Early positioning was therefore not meant to permanently seal off content, but to select the user group with the greatest capacity to expand. “The starting point is still choosing the right people,” after which categories must be added quickly enough to prevent the product from remaining trapped in its initial niche.
8. Category Expansion Was a Three-Leg Relay Between Operations, Algorithms, and New Users
When a new category first entered Douyin, the recommendation system had not yet learned to recognize its value, so its metrics often looked worse than those of mature content. Operations had to provide the initial traffic, giving creators sufficiently positive early feedback to keep supplying content.
As supply increased, the recommendation system gradually learned to match the content with existing users. Young seed users also had the capacity to consume across interests, allowing the new category to survive within the original community rather than being eliminated quickly because of weak initial metrics.
Once the category reached a certain scale, the content itself became a growth hook, attracting new users who genuinely preferred it. 于北川 describes not a one-time “content expansion,” but a loop of “creators → algorithm learning → user generalization → new users.”
9. Single-Column Maximized Distribution Efficiency; Double-Column Protected the Creator Ecosystem
Douyin’s full-screen single-column format reduced user pre-screening, allowing the platform to decide the next piece of content directly. After 5 videos, an ad might appear; users could watch 3 seconds or even the whole thing before realizing it was an ad. The container improved both content generalization and monetization efficiency.
Kuaishou’s double-column format let users inspect a thumbnail before choosing. Early on, it also hid like counts, and after entering a video, the first downward swipe showed the comments section. Traffic was therefore more evenly distributed, average feedback for long-tail creators was higher, and posting rates and interaction were naturally stronger than on Douyin.
Double-column was also friendlier to images: ordinary users could receive distribution without producing video. 于北川 recalls that image-text content accounted for a high share of Kuaishou in 2018, leading some people to predict it could become “China’s Instagram.”
曲凯 asked why the formats could not simply be A/B tested. 于北川 explained that creators could not be split across two ecosystems simultaneously; an experiment could measure the consumption side, but not long-term supply, trust, or changes in user psychology. “Ecosystems and new changes cannot be A/B tested.”
10. Douyin Truly Broke Through the Ceiling Because Consumption Was Simple Enough
Looking back at approximately 800M DAU, 于北川’s core explanation is not social networking but “being simple enough—simple and mindless enough to let you consume.” Douyin competed for every idle moment, and its biggest competitor was arguably not Kuaishou but gaming.
Games require interaction and investment. Douyin could be “mindlessly opened” in a moment of boredom. 曲凯 compared it to a short-drama comment saying, “I’ll leave my brain here for a while”; 于北川 sees this as a universal human need: everyone needs to relax every day.
曲凯 further summarized ByteDance’s strength as distributing “mindless subjects.” 于北川 did not accept that absolute formulation, but acknowledged that recommendation systems have an advantage with short, fast content: the higher the consumption frequency, the denser the feedback data, and the faster the next round of personalization learning.
11. Even the Strongest Recommendation System Cannot Make One Container Distribute Everything
A 30-minute knowledge-sharing video or travel vlog may be the best content of its kind in Bilibili’s selection-oriented environment, where the slower pace is acceptable. Put it into a continuous short-video feed, however, and users lack the same patience. Traffic feedback will push creators to keep Bilibili as their primary platform.
The same applies to image-text content. Guides and gossip do not require sound but do require careful reading; they are naturally constrained in a dynamic video feed and better suited to Xiaohongshu. 于北川 uses the question “What kind of power bank cannot be taken on a plane?” to illustrate how a search-oriented guide mindset is formed.
Xiaohongshu and Douyin will therefore coexist for the long term: the former offers denser image-text information and is better for search and quiet reading, while the latter has lower consumption friction. Both can expand their subject matter at scale, but the primary container still determines which needs each platform serves best.
Douyin and Kuaishou, which has also shifted to single-column, compete more directly. 于北川 believes Kuaishou would have monetized better and made Douyin more vulnerable if it had switched earlier. But both have accumulated enormous habit assets, making a “7:2:1” outcome more likely than one platform completely killing the other.
12. Weishi Never Built a Cold-Start Loop; WeChat Channels Found a Misaligned Entry Point
于北川 believes Weishi’s problem was not merely algorithmic technology: it imported creators from Douyin and Kuaishou all at once, accepted traffic from external sources including WeChat, and ended up with no clear seed category and too little historical behavior to train its recommendation system.
曲凯 compared it with early Huoshan, but 于北川 sees a key difference: Huoshan or Xigua already had established users and traffic, giving them a learning base from which to generalize. Weishi started from zero while facing growth pressure as a strategic project, making it difficult to spend 1 year building positioning first.
WeChat Channels used a more clever approach: it did not create a standalone app or compete directly with Douyin through the main feed, but used “videos your friends have liked” to satisfy gossip and relationship-based endorsement. Users entered video consumption, their behavioral data fed back into the main feed, and the content community entered a positive loop.
13. The 2019 Spring Festival Social Gamble Changed Its Objective 2 Weeks Before Launch
Alipay’s Five Blessings campaign, WeChat’s shake-to-win mechanic, and red packets had created Spring Festival growth legends. During the 2018–2019 Spring Festival, ByteDance bid for the CCTV Spring Festival Gala; Baidu won the prime spoken-ad placement, while ByteDance secured a secondary sponsorship slot but still expected to generate enormous traffic.
The original plan was not simple acquisition. Red packets would guide users to add friends, then enter an Instagram Story-like environment for interaction. Duoshan also launched around the same time, though 于北川 says it may not have been closely related to the Spring Festival campaign.
About 2 weeks before launch, after intense debate, the team changed the objective to new-user acquisition. 于北川 sees this as a lesson in strategic indecision: ultimately, neither objective was executed particularly well.
14. The 2020 Group-Wide Campaign Took the Contest to RMB20B in Red Packets and 470M DAU
To defend its lead and widen the gap with Kuaishou, Douyin launched Douyin Lite in 2019, taking only 2 weeks from project approval to launch and quickly reaching tens of millions of DAU. From late that year into early 2020, Huoshan was also renamed Douyin Huoshan Edition.
Kuaishou wanted to secure both core naming rights for the 2020 Spring Festival Gala. It planned to reshape the short-video market with red packets and, according to 于北川, a large-screen up-and-down-swiping format, adding tens of millions of DAU on the peak day. ByteDance’s response was that both the red-packet amount and market noise had to exceed Kuaishou’s.
7 key products across the group jointly collected cards, with the rare card designed as the “Douyin Card”: users of Toutiao, Xigua, Pipi Xia, and others had to go to Douyin to claim it, while full-scale push notifications across all products on the Gala night also routed users directly to Douyin. 于北川 believes this kind of resource coordination “could happen only at ByteDance.”
Kuaishou distributed RMB10B, ByteDance RMB20B. In the 1–2 weeks before the campaign, 于北川 recalls that Zhang Yiming, who was presumably in the UK, asked, “Is the money enough? Do we need to spend RMB100B?” Douyin ultimately reached peak DAU of 470M; LTV-based ROI was positive, and the pandemic unexpectedly helped retain those users.
15. High Ad Efficiency Let Growth, Livestreaming, and E-Commerce Share One Economic Model
Douyin’s 2018 ad load may already have reached “1 ad every 6 videos” or “1 every 10.” Because the platform could recover user-acquisition costs within 1 year, ByteDance could keep spending aggressively, even buying traffic at 2x competitors’ prices.
E-commerce and local services were tested as early as 2018, but true scaling depended on a much larger traffic base. Livestreaming scaled first from 2019 to 2020, while e-commerce accelerated further around July 2020; 曲凯 regards 罗永浩’s first livestream as a landmark moment for Douyin e-commerce.
The team later found that the feed could not only “sell products, people, and merchants through video,” but also directly distribute People You May Know, e-commerce products, and store cards. If placement, timing, and user matching were accurate enough, large traffic could turn cards into transactions.
于北川 therefore views monetization as a consequence of traffic scale: early weakness in livestreaming, e-commerce, or local services did not necessarily mean demand was absent; the platform may simply not yet have been large enough to continuously fund new use cases.
16. Douyin Did Not Simply “Launch First and A/B Test”; It Even Missed the Livestreaming Window Through Caution
于北川 corrects the outside impression: at the time, features first went through rigorous review, extensive debate, and persuasion of key stakeholders. Only after the team decided they were worth building did it open experimental traffic. A/B testing came late in the decision process; it was not a button that replaced product judgment.
High-speed growth gave the team and organization room to be patient and think carefully, but in hindsight he believes he should have been “more aggressive.” His clearest regret is that in 2018 Douyin did not directly add livestream distribution to the Feed, offering only a Following entry point and leaving livestreaming without enough traffic for an extended period.
Kuaishou already had a large base of mid- and long-tail hosts and a mature live-show business. Douyin’s excessive caution over Feed content meant livestreaming did not receive sufficient traffic for a long time, showing that “nothing is a problem under growth” can also conceal opportunities being deferred.
17. A Flat Organization Gave Young People a Real Battlefield
ByteDance deliberately weakened hierarchy in its early years: internal cards mainly displayed departments, OKRs, and contact information rather than highlighting who was a VP. Ordinary employees could discuss business directly with Zhang Yiming, Zhang Nan, and others. The effect of that proximity was “de-mystification”—employees came to believe, “I can do many of these things too.”
曲凯 wondered whether Douyin’s heavy use of fresh graduates was simply a result of limited early resources. 于北川 insists the bigger reason was that “ByteDance really trusted young people at the time,” while rapid growth kept creating critical assignments for them to take on.
于北川 sees his generation as beneficiaries of the last wave of mobile-internet opportunity: “Talented young people need more opportunities and dividends.” 曲凯 adds that they “need to go to war.” After the mobile-internet dividend ended, 于北川 became anxious for a time, until AI created a new growth battlefield. “In a sense, AI saved everyone’s soul.”
18. ByteDance’s “Extremity” Meant Strategic Reinvention, Not Simply Throwing Resources at a Problem
于北川 believes ByteDance’s expansion made it inevitable that the company would recruit large numbers of experienced people externally. New teams had not experienced the transmission of the original culture, so habits of candor, clarity, and giving young people opportunities would be diluted. This is an inevitable consequence of becoming a large company: it can be delayed, but not reversed.
He prefers to call the company’s culture a “code of action.” How the first few dozen people thought, collaborated, and persuaded others through logic influenced the second and third waves of employees. People who did not fit a high-consistency culture also left faster, making early high attrition and cultural intensity two sides of the same coin.
If he had to summarize ByteDance in one word, he would choose “extremity”: in action, “solve the problem at all costs”; in thought, keep asking how to maximize returns and expansion. It was not “brute force creates miracles,” but a refusal to reduce strategic questions to local optimization.
When the target is merely 10% growth, a team can stack five features worth 2% each. If the question becomes “How do we grow 2x or 3x?” Zhang Yiming might even ask about 5x. The team then has to reorganize channels, resources, and product form. “Only by seriously thinking through that question can you come up with methods that break the old structure.”
19. AI Founders Must Answer “What Will Remain in 3–5 Years?” from Day One
After leaving ByteDance, 于北川 found it difficult to readjust to slower growth and more reporting: “I was a little spoiled.” He first started a company with former colleagues, then founded one independently, gradually shifting his focus toward a purer working environment and long-term assets.
AI founders broadly face the question, “Will OpenAI build this?” 于北川 remains wary of single-point AI image-generation, AI presentation, and other pure efficiency tools: once the model internalizes the capability, the product may lose its position, and he “cannot imagine what it grows into.”
His bottom line is: “We cannot put our own lifeline in the hands of a foundation model’s whim.” Products can evolve, but the starting point must be able to become a defensible business 3–5 years out, rather than arbitraging a temporary capability gap.
His abstraction from the previous internet generation is that major products usually have ecosystems: Pinduoduo and Taobao have merchants and products; Douyin and Xiaohongshu have content and creators; Tantan and similar products have people. Lacey AI is therefore exploring whether “AI can connect certain people.”
20. Lacey Started with a “One-to-Many Search” Tool, Then Tried to Rebuild Intentional Social Networking
Lacey AI positions itself as a People Search AI Agent serving marketing, sales, foreign trade, headhunting, and HR in Europe and the US. After users describe a target profile, the Agent calls professional SaaS and databases such as ZoomInfo and Apollo to search for candidates and review them one by one.
Previously, users had to manually filter and verify candidates, find contact details, and reach out individually; the relevant SaaS tools could cost tens of thousands of dollars per year. Lacey aims to compress the process to 3–5 minutes and continuously improve search, outreach, and conversion through memory and context.
于北川 describes the end state as “a sender and a filter”: AI will cause public, intentional connections between people to explode, while the other side will also need AI to process dense inbound requests. If 100 people want to collaborate, AI can speak with 98 first and arrange meetings only with the 2 best matches.
Cold start will still begin with one side: the tool experience must be 10x, 20x, or even 100x better than existing products, first accumulating hundreds of thousands to 1M PLG users. Email can then evolve into platform IM, subscriptions, and following; people being searched can also return to claim their profiles, gradually creating new relationship and data moats.
21. The Founder’s Challenge Is Not Doing More Business, but Resisting Cognitive Inertia
于北川 believes large companies train people to analyze and solve concrete problems; startups require strategic choices, talent organization, and resource acquisition, including hiring key leaders, establishing operating principles, and preparing financing and ammunition for the next 6–12 months.
Founders can easily use writing documents, testing bugs, and solving familiar business problems to prove they are busy, using immediate feedback to relieve anxiety while avoiding the truly difficult work of defining the product or hiring a CTO or head of growth. 曲凯 calls this muscle memory; 于北川 summarizes it as “cognitive inertia.”
Resisting cognitive inertia does not mean abandoning the front line. Founders must understand concrete details while putting most of their attention on strategy and resources. “Hard but right” questions often produce no answer after 1 or 2 days, but solving them can transform the company.
Faced with “What if ByteDance builds this too?” his checklist has 3 layers: what remains long term—is it ARR, or a unique ecosystem and user data; why is the giant not competing directly now, and how long is the window; and how can the Day One entry point scale quickly? If none of these can be answered, “whether ByteDance does it today, next year, or the year after makes no difference.” What remains is luck.