Vol.200 Exclusive Interview with 林凡 of Maimai: What New Changes Can AI Bring to Social Networking and Recruitment

Summary

• China never produced a LinkedIn because of supply and demand, not product shortcomings. During the decade of mobile internet, jobs were so easy to find that “Why would a boss spend time talking to a candidate? I only do it when candidates are hard to hire,” which gave Boss Zhipin’s rapid matching model the edge; now that a single opening can attract 1,000 applicants, the value of relationship-based endorsement is returning, with Maimai restarting professional social networking—the constraint is still the inescapable WeChat ecosystem, while the bet is on a generational shift summed up by “Gen Z saying, ‘Why can’t you add me on Maimai?’”
• 林凡’s read on the era: the internet only changed production relations; AI is a productivity revolution, and To B is China’s biggest hope for the next decade. In To C, “it’s all about foundation models” (Perplexity is struggling, and Grok has taken a large share of this kind of business), while AI lets Chinese To B companies move from “providing process value” to delivering outcomes: “If you improve performance by 10%, give me 5% of the upside—that’s easy to negotiate. If you don’t deliver, I don’t pay.”
• The restructuring of recruitment is already quantifiable: AI can handle 70%, even 90%, of a headhunter’s work. Every “Beike of headhunting” failed because it was a production-relations solution that had to wait for a productivity revolution; within five years, standardized roles will be filled “with one click,” payment will come only after the hire starts, recruitment and job-seeking will be “as simple as hailing a ride,” and “human resources in the cloud” will turn labor costs into variable costs.
• The cleanest technical definition is that AI has an extremely high ceiling and an extremely low floor. It won a gold medal at the IOI, ranking seventh globally, yet “you can tell it to fix a small bug until you’re blue in the face and it still won’t fix it”; accuracy of 85% in a single round can fall to 20-30% after 10 rounds. “The essence of AI to B is raising AI’s floor high enough”—high enough to hire AI employees directly.
• The AI talent market is the year’s most important talent-market event. The share of newly posted jobs containing AI keywords rose from 12% in 2024 to nearly 19% in January-October 2025; after the DeepSeek moment, September job volume was up 11x year on year. In Silicon Valley, discussing AI “entirely in Chinese is not theatrical at all”; ethnic Chinese make up more than half of the technical core, and talent flows run “mostly from China recruiting the US.”
• The China-US innovation gap is a multiplication of probabilities. Chinese founders pursue high-ROI projects and are risk-averse; Americans are willing to bet on projects with a 5-10% chance of success: “The US may have 100 people willing to do it; China may have only 10.” China is extremely strong in 80% of fields and moves faster in vertical applications; the other 20%—long-cycle, high-failure-rate innovation—belongs to the US. His judgment of 梁文锋 is unsparing: “That’s a game of luck. It’s hard to expect him to produce a big move on schedule.”
• AI-native organizations make a three-stage jump: everyone becomes a programmer, then everyone becomes a manager, then everyone becomes a CEO. AI handles 20-30% of work in the first stage, 60-70% in the second, and 95% in the third, with one-person companies everywhere; the end state is “born retired.” Maimai has already shifted 20-30% of its workload to AI (employees have been fired since 2023 if they do not use Copilot), while 林凡 himself spends 50-60% of his working time discussing problems with AI, including management’s “brutal truths” and verbatim communication scripts.
• Maimai has been profitable for 2 consecutive years and does not need external funding, but “when there is a truly big opportunity today, you need to make a big investment,” putting fundraising back at the top of the agenda. Its moat over startups is data—“it should matter for at least a year to a year and a half”—and “I have never worried about startups.” He also knows luck matters: Boss Zhipin becoming so large exceeded his expectations; ultimately, it was “desire happening to match social trends” plus luck.

Deep dive

1. Opening correction: Maimai is more than gossip; he only spent one semester on his PhD

• 李翔 opens with the stereotype that big-tech employees come to Maimai for 2 things: checking negative gossip about their own companies and surveying the market for jobs they want to fill. 林凡 adds the overlooked piece: market, sales and business-development teams use Maimai to reach target customers and “talk about potential partnership opportunities”—the most underestimated part of professional social networking.
• Another correction: he spent only 1 semester on his PhD. In 2002, both he and his wife received US PhD offers; after 9/11, her visa was rejected, so he returned to China to marry her and tried unsuccessfully to take her abroad on an F2 visa. His own visa was then subject to a military review “because I was studying artificial intelligence”—“we may have been doing very, very peripheral analytical work for military projects, but anyway, these things do have an impact.” So he stayed and went to work with 王小川. AI had already become a sensitive field in the US at the time.

2. “Dianping for the workplace”: the real demand behind the gossip image

• 林凡 admits that the gossip image is “a normal reaction from everyone,” but says it misses the point by a wide margin. During the decade of mobile internet, jobs were easy to find and candidates held 3 or 4 offers to evaluate: “This ByteDance role—does the salary make sense? Is there a trap in the actual work?” Employees would sit on the platform every day and answer: “Don’t come today, this department is about to be laid off,” or “The year-end bonus is exceptionally large—12 months, 24 months.”
• Job seekers gave Maimai the name “Dianping for the workplace”—“a very good platform for finding a good job without stepping into a trap.” Those who treat it as a break-room gossip feed are mostly bystanders who are not there to make a job-search decision.

3. Social networking runs on relationships; communities run on content

• 林凡’s distinction is clear: professional social networking creates career, partnership and learning opportunities through building connections and exchanging information based on those connections. A professional community is built around content—UGC lets high-quality posts reach more people and generate discussion, creating the feeling that “these people understand me, and I understand them,” along with a sense of belonging.
• The old internet question of how communities make money is somewhat easier for Maimai: “We have a very good social foundation.” Every user enters a real company and real identity, which gives recruitment and membership businesses a solid base.

4. Before Maimai: why a people-search engine did not work

• The first product of the 2012 startup was not Maimai but a people-search engine built by scraping data from Weibo, Tencent and other platforms. They later also built a dating product based on relationship networks, then called “Mimi.” After 6 months, they found the fundamental problem: “Finding that person is actually only the beginning.” You still had to establish a connection and communicate to achieve the goal, while the other person “had not put their information on this search engine,” creating a huge trust barrier from the first question: “Where did you find me?”
• The lesson became the foundation of Maimai: “You have to own your own social network.” Users had to know that “I left my information on this platform” before they would accept a link from a stranger. Maimai launched in 2013; in between, out of respect for the founder of his former employer Dajie, he deliberately avoided the LinkedIn direction at first.

5. Why China never produced a LinkedIn: supply and demand are the main factor, WeChat the secondary factor

• The core mechanism is that supply and demand determine the matching model. When jobs are easy to find, submitting a résumé and chatting directly are the most efficient approaches: “Why would a boss spend time talking to a candidate? I only do it when candidates are hard to hire.” That is why Boss Zhipin took off so quickly. Overseas, the norm is that jobs are hard to find and require relationship-based endorsement: “How do you judge the quality of 1,000 applicants? In the end, you still ask friends whether they can recommend someone.” That explains why the LinkedIn model “works in every country outside China.”
• The secondary factor is the ecosystem: “The core barrier you can’t get around is WeChat.” In one real case, a user found a dozen or 20 high-school classmates on Maimai and created a group. “As they chatted, someone said, ‘Why don’t we add each other on WeChat?’ The entire group moved to WeChat, and people came to Maimai less.” Overseas users care more about privacy: they do not add work acquaintances on Facebook and only chat on LinkedIn, so there is no comparable siphoning effect.
• 林凡 is placing his hope in a generational shift. Many people born after 2000 dislike having their bosses add them on WeChat: “Don’t add me on WeChat. Why can’t you add me on Maimai? We actually hope that day will come.”

6. 12 years in 2 phases: 6 months of reflecting on wishful thinking

• The timeline is clear: from 2013 through the first half of 2020, Maimai used LinkedIn as its social-networking benchmark; from the second half of 2020 to today, it has focused on community—comparing offers and discussing the sense of security people seek after layoffs and pay cuts. “If someone earning RMB30,000 a year says on Weibo that they were laid off, everyone thinks they’re showing off. But in the internet industry, there is a group of people who understand you.” The future is a return to social networking: “One day, Chinese professionals will definitely start caring about their long-term professional reputation.”
• Why did the pivot take until 2020? “Founders always have a kind of desire. Investors often criticize me by saying this is wishful thinking.” User numbers kept rising, and Maimai raised a large amount of money in 2018 and 2019, so the negative signals were ignored: connections were moving to WeChat, and “if I wrote a post on Maimai, my boss would ask whether I wanted to quit.” Only when social activity declined, while the community product—run by just 2 or 3 engineers and 2 operations staff—remained active did he start asking whether he had ignored user demand. It took “roughly 6 months” of reflection to abandon the social-networking route.
• This time, the restraint comes from the team. Subordinates stopped him: “Boss, you need to stay clear-headed. Your main business is based on community. Don’t jump straight into social networking and end up killing the community too.” The standard for judging the move has also changed: “When social networking starts growing like crazy without much effort from you, that may mean it is right”—driven by user demand rather than desire.

7. Desire-driven companies are fine, but you have to stay in the game

• 李翔 cites Nvidia, and 林凡 adds BYD: both are companies driven by a desire that lasted 10 or 20 years. His correction is practical: “There’s nothing wrong with being driven by desire. I just think you also have to stay alive. Making graphics cards for gamers was definitely not Huang’s original intention, but to survive he had to do that well first.” Maimai’s equivalent is its community. 雷军’s survival theory also comes up: with an unbreakable team, money you can never spend and a large enough track, “one day you can climb out.”
• Financially, Maimai continued to raise small amounts every 1 or 2 years after 2018 without publicizing them, but “the past 2 years have been profitable.” “If you can support yourself, you don’t need investors’ money.” The qualification comes in the second half of the sentence: “When there is a truly big opportunity today, you need to make a big investment to capture it.”

8. LinkedIn left China before dawn; selling to Microsoft showed a different kind of desire

• LinkedIn’s exit came in 3 stages: the closure of Chitu, which 林凡 recalls as 2019; the closure of another product; and finally the shutdown of its recruitment business in 2023. 林凡 was “a little sad”—having only 1 player left in the sector “shows that the sector is not very large.” 2 things offered some consolation: “LinkedIn China was actually profitable; its exit was geopolitical rather than market-driven.” He also told former LinkedIn colleagues: “You didn’t hold out until dawn. I’m still in the game, and I’m trying to live until the time when this becomes hugely successful.” They had officially entered China in 2014 and waited 10 years without seeing that moment.
• The Microsoft math was striking. When Microsoft acquired LinkedIn for $23B, LinkedIn had only $3B in revenue and a market value of roughly $10B—almost a 1x premium, prompting everyone to ask how it could sell for so much. Today, it generates more than $20B in annual revenue and roughly $5B in net profit, and “everyone thinks it is worth at least $100B.” 林凡 once asked Reid Hoffman face to face why he sold. The answer was “there are more interesting things to do.” 李翔 called it “a very PR answer.” 林凡 replied: “After selling, that’s the only thing you can say.”

9. The internet changed production relations; AI is a productivity revolution—the era of To B

• 林凡’s foundational view is that the internet essentially changed production relations: connections between people, information exchange and online services. What emerged was “business models plus To C products”—ByteDance, Weibo, Meituan and Didi all fit that pattern. AI, by contrast, is a productivity revolution, with technology becoming the most important lever. The US, which is moving faster, shows the pattern: in To C, only foundation-model companies are doing well—“Perplexity is not doing very well right now,” and Grok has taken a large share of this kind of business. In To B, the opportunities span GitHub Copilot, coding, advertising, legal services and healthcare.
• His analogy is the Industrial Revolution: machines first benefited producers on the B side, then the gains flowed through to consumers. “AI will have a huge impact on the service sector. It will serve the B side through optimizing and replacing labor, and then move from the B side back to the C side.”

10. Every “Beike of headhunting” failed; that was production relations, while AI is the solution

• Why did every company branded as the “Beike of headhunting” fail? Beike worked because buying and selling a home is basically a one-off transaction—“the probability of working together again is basically zero”—so the platform captures substantial value. Headhunters maintain long-term relationships with both companies and candidates. “Even in the worst case, I can survive on my own; the headhunter can simply be a headhunter.” The platform’s inherent value is therefore limited. The conclusion: “Restructuring production relations is basically unrealistic. You have to wait for the opportunity created by a productivity revolution.”
• The first opportunity is that AI can do “70%, even 90%, of a headhunter’s work,” leaving people to do the 10% they are best at, such as helping a candidate negotiate a better salary—“AI may not be very good at that yet.” The second is “human resources in the cloud”: hiring a salesperson or programmer becomes buying a cloud service, so “labor costs also become variable costs.” 李翔 jokes that the last company to turn a large fixed cost into a variable cost was WeWork. “Exactly—but it had its own problems, so let’s not get into that.”

11. Why incumbents cannot seize the opportunity: does the founder believe technology changes everything?

• 林凡’s analogy is ByteDance. Its business should, in theory, have been built by Baidu or Tencent: “Weren’t ByteDance’s recommendation engineers all poached from Baidu?” The difference was the founder’s belief system. In 2010, 林凡 spoke with 张一鸣 at Sogou. 张一鸣 said personalized recommendation would be revolutionary; 林凡 replied that he had worked on it for 2 years and had only improved click-through rates by 20-30%, so he did not really believe in it. “He said, ‘You don’t understand.’ I didn’t try to persuade him, and he went off and did it himself.” The parallel in recruitment is that founders who succeeded in the past “would not think technology was the only variable in the equation”; they would also rely on customer relationships, marketing and promotion.
• He acknowledges that this cycle is different. After several rounds of technological change, “everyone is highly alert to technology. They started buying a pile of GPUs from day 1.” Incumbents may be more alert than outsiders assume.

12. Gartner’s curve explains itself: believers keep using the technology until one day it works

• 林凡 studied the mechanism behind the curve—pronounced “Gaiteler” in the original audio, likely Gartner—where a new technology rises, falls and rises again. People initially experiment with it, then leave after discovering that “the capability is still quite far off.” The second rise comes because a small group of extremely committed believers keep exploring it at high frequency: “someday, one day, the technology will work,” and then word of mouth spreads. Recruitment is the same: “Many people say technology cannot handle these interview stages, but we are technological extremists. If this approach cannot handle it, we try another. That process is actually our core competitive advantage. Once it works, there will be strong economies of scale and Matthew effects.”
• Across the market, the deepest applications are résumé screening and the first round of communication. AI interviews “are effective only for blue-collar workers,” along with initial screening for fresh graduates; they make only fast behavioral judgments, not complex knowledge judgments. Every other stage still faces major technical challenges.
• The challenge can be quantified. AI accuracy in a single round of conversation is roughly 85%, which is good enough for chatting. “But if you use it to do real work, after 10 rounds of conversation accuracy may fall to 20-30%. You basically conclude it is unusable.” The core target is to maintain 99% accuracy after 10 rounds.

13. AI has an extremely high ceiling and an extremely low floor; To B means raising the floor

• 林凡’s most stripped-down abstraction of the session is that in human society, upper and lower bounds on ability move together. “It is hard to imagine an IMO gold medalist not being able to calculate what 2.2 plus 2.3 equals,” which is why companies like hiring students from Peking University and Tsinghua. Yet AI competed in the IOI and won a gold medal ranking seventh globally, while “sometimes you can tell it to fix a small bug until you’re blue in the face and it still won’t fix it.” Its ceiling is extremely high and its floor extremely low; “in a normal workplace, when you want to hire an employee, AI simply cannot do the job.”
• The definition follows: “The essence of AI to B is whether you can raise AI’s floor high enough”—high enough that you can hire AI employees directly.

14. Maimai’s AI recruitment product: 5,000 résumés reviewed in 3 minutes

• After more than 6 months of refinement, the product went live nearly 2 months ago to address HR’s 2 biggest pain points: slow screening and exhausting communication. Recruiters can state requirements in ordinary language—“I don’t want people who changed jobs twice in 3 years unless each move was a promotion,” or “No To C experience; I only want To B experience”—and the system reads 5,000 résumés in 3 minutes and produces an initial ranking. It can also expand HR’s frame of reference. A client asking for a community-operations hire with 1 year of experience might lead a human to consider only Xiaohongshu, Zhihu and Bilibili; AI surfaced community-operations candidates from NetEase Games and Huawei’s internal freshman forum—“you would never have thought of that before.”
• On the communication side, the product provides a 24/7 online assistant. In one real case, a company hiring backend engineers had AI screen 1,600 candidates, initiate 1,200 conversations and confirm interest from 120 people. HR’s feedback: “The productivity improvement is too obvious. Once you get used to it, it is very hard to go back to manually screening résumés and manually chatting.” A larger, industry-changing innovation is “not convenient to discuss for now.”

15. Soul and Xiaohongshu: 2 strong examples; a workplace Xiaohongshu is hard to build

• 李翔 cites 2 good examples. Against Momo and Tantan’s face-driven dating model, Soul successfully pursued “stranger interests plus emotional value”—“many people have thought about this route, but Soul seems to be the only one that truly made it work.” Xiaohongshu rose despite Douyin and Kuaishou’s dominance. Video directly stimulates the senses and is easy to become addicted to; a few thousand or 10,000 good posts can serve everyone. Text and images are less addictive and require precise supply at the scale of 100,000 posts—1 or 2 orders of magnitude more—to retain users. “It is not that Xiaohongshu’s algorithm is stronger than Douyin and Kuaishou’s. The nature of text and images itself demands a higher threshold.”
• Maimai tried the workplace-Xiaohongshu direction and abandoned it for a simple reason. The good things in life can be learned and replicated—you go somewhere, I can go there too; you cook something, I can learn to cook it. But “the good things at work are both scarce and impossible to replicate.” If ByteDance pays 12 months of year-end bonus, can your company write one too? Share it on Maimai and “you’ll get a pile of abuse below.” The conclusion: “This is not something production relations can solve. It requires productivity”—a powerful agent that takes the annoying work off your hands, such as screening résumés.

16. China’s 10-year To B comeback: from process value to outcome delivery

• Chinese To B software was weak across the board over the past decade. “Many investors became disheartened and decided To B was a false demand in China.” AI happens to favor To B, but “everyone is still racking their brains to start To C companies; the mindset remains stuck in the mobile-internet era.” His blunt advice to founders: “If you build a horoscope AI, how much more can it provide than horoscopes did before?”
• The root of the failure was that “you could only provide process value, not outcome delivery.” A data-analysis SaaS product cannot explain whether it is worth $10K, $100K or $1M, leaving both buyer and seller in an awkward position. AI makes outcome-based pricing viable: sell an improvement in a specific metric; “if you improve performance by 10%, give me 5% or 10% of the upside. That’s easy to negotiate—if you don’t deliver, I don’t pay.” SaaS companies previously could not afford to go heavy; in the future, AI will carry most of the service burden. “Looking ahead 10 years, this may be China’s biggest hope in To B. Someday, if 1 or 2 companies succeed in China, everyone will realize it. We very much hope to be 1 of those first 2.”

17. AI will restructure social networking: you do not need to publish your whole background to a community; just tell AI

• The LinkedIn model requires users to publish their entire profile in exchange for the chance to be discovered, but the data says otherwise: only 30-40% of LinkedIn users have a complete profile, while Maimai is in the low 20% range and another 10 percentage points lower. “Most people have serious reservations about telling others everything about their background.” That is fatal to matching because “I don’t know what is suitable for you.”
• The breakthrough is captured in one sentence: “You have very few psychological reservations about telling these things to AI.” In the future, you explain your profile to AI, and AI handles matching in the background without exposing it to the community. Once a match is made, you talk to the other person; the middle stage is replaced by AI.

18. “One-click placement” within 5 years: recruitment as simple as hailing a ride

• The end state has 2 defining features. Companies born in the AI era will be technical companies—mobile internet still allowed business-model-driven companies to grow—and their interfaces will be extremely simple while the technology underneath is extremely complex, like a search engine or ChatGPT. A candidate says, “I want a company that pays well, has light work and is close to home”; HR says, “I need 5 senior frontend engineers in Beijing who have built large-model applications.” The 2 AIs match them directly. If the candidate asks for light work, “I’ll reply: impossible. There is no job that is light, pays well and is close to home. Would you consider something else?”
• The timeline and business model are clear: “Within the next 5 years, standardized roles should basically be filled with one click.” Payment comes only after the person starts: “If they join, you charge; if they don’t, you don’t.” Recruitment and job-seeking become “as simple as hailing a ride,” and “traditional recruitment websites will have no role.” AI will also record why you left after 3 months and prevent you from applying to similar companies again, becoming a career adviser. “AI will tell you some things: this is your problem; you need to solve it.”
• Asked whether data trapped in old apps creates a Doubao-phone-style challenge, 林凡 says it is not that important. Your job preferences and workplace experience “were not published on any platform”; you can simply tell AI. The fundamental conflict around Doubao is “not information first, but more about services”—the services on those platforms are not exposed through APIs, allowing them to be bypassed. Just as HR tells a headhunter the real requirements that do not appear in a job description, “in the future you can tell AI.”

19. Internal experiment: no Copilot, no job; everyone writes code

• Maimai built an “AI Island” integrating ChatGPT, Gemini, DeepSeek and other models into one platform. It did this as early as 2023, solving employees’ access problems and other frictions. Starting in 2023, engineers were required to use GitHub Copilot: “If you didn’t use Copilot enough, you were fired.” The company paid $20 per account, and after 1 month everyone said, “It’s really good.” The marketing team used to have 2 or 3 people write 1 PPT; now 1 person writes 1 or 2. Efficiency is higher, but so is the pressure.
• Across the company, roughly 20-30% of the workload has already been handed to AI. “For a 10-year-old company, I can proudly say we are definitely top-tier,” although AI-native startups are already at 60-70%. Maimai is now pushing the idea that “everyone is a programmer”: everyone writes code, with game competitions and AI hackathons. “Not everyone accepts this idea yet, but I keep pushing it.”

20. Half the CEO’s work is talking to AI; even the “brutal truths” come from it

• 林凡 himself spends 50-60% of his work discussing problems with AI. After 3 rounds of iteration on a PPT framework, he hands it to the team to complete. He writes down management problems and asks, “What should I do?” HR rates AI’s management advice as 90% reliable. “As someone who can only come up with answers that are 30% reliable, the other 60% is my value-add.” Brainstorming and aligning with senior executives takes 10 or 20 minutes; AI understands him after 2 or 3 questions and produces 12 possibilities. “I know 8 of them are definitely unreliable, and 1 is brilliant.” That judgment is a form of reinforcement learning: “you know what is right.” So “I’m the one who decides whether it is right; AI does not have to take the blame.” The side effect is identity anxiety: “When I watch it work so hard and then go do leisure activities, I feel uneasy. Am I being a sufficiently responsible CEO? But my work throughput is much higher than before.”
• The use that affected him most was asking GPT, after a long conversation, to tell him “the truths that look correct but I may not be aware of, and that are especially close to brutal.” It returned 7; “at least 5 were right,” including: “You have very high standards for people, but you do not replace unqualified people quickly enough,” creating a double illusion for everyone and causing motivation to decline. He even asks AI to draft verbatim management scripts: “I cannot say this part out loud. It produces another version, and the next time it is more accurate.” 李翔 jokes that if the company goes under, it can absolutely blame AI.
• Should he talk to 王兴 and 王慧文 or to AI? “In their areas of expertise, they are definitely more reliable. Their ceiling is already the ceiling of AI’s ceiling. But management has many styles and schools of thought, and the methods they know may not suit me. AI has talked with me for so long that it knows where I am conflicted and where I am in pain.”

21. The AI talent market: jobs up 11x in a year; “in the future, everyone will be AI talent”

• The underlying idea behind Maimai’s talent-flow report is that “one day in the future, everyone will be AI talent.” The analogy is 20 years ago, when knowing how to use a computer meant being a senior white-collar worker. Today, no job description says “proficient in Word and Excel.” The data: the share of newly posted jobs containing AI and large-model keywords was roughly 12% in 2024 and had reached nearly 19% in January-October 2025, a 7-percentage-point increase in 1 year. “I estimate that when it reaches 50-60%, people will stop writing it.” After the DeepSeek moment, AI job volume surged month after month starting in February 2025; in September, it was up 11x year on year. “This is the most important thing in the talent market this year.”
• The report has 2 audiences. It helps companies plan organizational and talent moves earlier, “instead of waiting until they can no longer compete for talent,” and helps professionals plan what to learn and where to pivot over the next 3-5 years. Revenue is hard to separate because the industry charges by account and service package, but by job volume roughly 20% of revenue is AI-related. “In reality it should be higher—the willingness to pay ahead of time is much stronger.”

22. The talent battle: every offer was rejected until the CEO took the final interview

• For a period, Maimai’s effort to recruit AI talent was “truly miserable. Every offer we sent out was rejected.” Even on the day someone was supposed to join, they would say, “Sorry, I went to some big tech company.” The company changed the process: 林凡 now takes the final interview for every AI hire. “I’m not going there to interview them. I’m going there to attract them.” He describes the revolutionary possibilities in recruitment, and the joining rate rose to roughly 70-80%. What is the key attraction? “I can’t tell you. When we become 1 of the first 2 truly profitable AI To B companies in China, I’ll share it.”
• The split between big tech and startups is clear. Execution-oriented talent still chooses big tech; when offers from the “six little dragons,” Unitree and similar companies are compared with big-tech offers, many still choose big tech. But “the top 1% actually choose startups”—“they feel startups give them more room to perform.” As for model strength, “it is hard to say that big tech’s models are better.” It is difficult to say who is better or worse among Kimi, Zhipu, Doubao and Qianwen. Users may feel that big-tech products are better because they have more employees and have been finely adapted, not necessarily because the underlying model is stronger.

23. Silicon Valley discusses AI entirely in Chinese; the innovation gap is 100 people betting versus 10

• 傅盛’s video saying that “you do not need to speak English to discuss AI in Silicon Valley” is, in 林凡’s judgment, “not theatrical at all.” Before the pandemic, he used English throughout his US visits; last year, 80-90% of conversations were in Chinese; this year, all of them are. Ethnic Chinese account for more than half of the technical core at both big-tech teams and startups. Conversations with Westerners are more polite and official—see Reid Hoffman’s answer—while conversations with Chinese people are more likely to produce valuable information. Cross-border mobility is limited: both sides are largely self-sufficient in basic talent, with only a small amount of movement at the very top. “There is very little US recruiting China; a lot of it is China recruiting the US.” As for the UK, home of DeepMind: “It is still far behind.”
• Talent density in China and the US is “basically very close.” The difference is style. Chinese people generally pursue high-ROI projects and are risk-averse; Americans say, “I just like doing this, and even if it looks particularly unreliable, I firmly believe it is worth doing.” Peter Thiel’s interview question—“What important truth do few people agree with that you believe is correct?”—encourages people to think differently. For a project with a 5-10% chance of success, “the US may have 100 people willing to do it; China may have only 10.” Multiply those probabilities over time and you get a persistent innovation gap: “OpenAI keeps producing things; DeepSeek appeared once and flashed brightly.” The conclusion: Chinese companies and talent are extremely strong in 80% of fields and move faster in vertical applications; in the 20% requiring long-term investment and carrying high failure rates, the US is more willing to take the risk.
• Asked in the community whether 梁文锋 is biding his time to prepare a big move or has already handed in a blank paper, 林凡 is unsparing: “It has nothing to do with him. That’s a game of luck. It’s hard to expect him to produce big moves on schedule.”

24. Organization: learn the method, not the appearance; rotation works only when every move succeeds

• Maimai studied the management methods of Alibaba, ByteDance, Huawei, Google and Amazon, but “we genuinely do not have the complete conditions to replicate any big-tech system.” The rule is “learn the method, not the appearance”: break down the underlying principles and match them to the company’s stage, business and people. The operating differences are real. A big tech company can spend 2 weeks repeatedly discussing quarterly OKR alignment; “for a small company like ours, spending 1 day already feels painful.” Company-wide OKRs do not work everywhere—“the sales organization is completely unsuitable for OKRs; it can only use KPIs.”
• After returning from Hupan, Maimai copied Alibaba’s executive rotation system and abandoned it after 1 or 2 years. The lesson became systematic: rotation works only during periods of rapid business growth—“everything is succeeding, so you sacrifice some growth in exchange for the person’s broader experience and perspective.” In a red-ocean market or a period of slow growth, “frequent executive rotation is extremely dangerous.” Changing commanders in the middle of a battle can kill the business. Alibaba’s reduced use of rotation after growth stalled was cited as evidence, along with the return of someone whose name sounded like “吴钊” in the original audio to run DingTalk: “I thought, ‘If I replace this person, I can do the job just as well.’ In reality, that is not how it works.”

25. The 3-stage jump for AI-native organizations: programmer → manager → CEO

• The definition is “an organizational form in which humans and AI collaborate best and combine their capabilities to the fullest.” The endpoint is “humans assisting AI to work, rather than AI assisting humans to work.” The progression has 3 stages. First, everyone becomes a programmer: AI handles 20-30% of each role’s tasks, while humans break down the work and give instructions; “1 person becomes 1.5 people,” and the coordination link from product manager to programmer is skipped. Second, everyone becomes a manager: agents are orchestrated into complex workflows, AI handles 60-70%, and “where you previously needed 3-5 people for a recruitment director’s work, 1 workflow now solves it,” causing functions to consolidate. Third, everyone becomes a CEO: AI completes 95% of the work and acts as the coordination center; “you only need to understand the business model and user demand,” and one-person companies proliferate. The end state, 20 years from now, is that “AI may do all the work. You are born retired, and whether you want to work depends on your interests.” 李翔 pushes back that people imagined this years ago and still work longer hours. “This is the darkness before dawn.”
• The defense of AI-centered organizations uses industrialization as a mirror. The point of designing an organization around a machine was not that the machine was large and heavy; “it was that the machine could do the work of 1,000 people.” AI can do the work of 1,000 knowledge workers, so processes must be rebuilt around it. The direction is different: steam engines concentrated people in factories, while AI is decentralized—“you can collaborate with AI in the cloud from home, and work from home again.”

26. The only anxiety is not moving fast enough; the data advantage should last at least 1-1.5 years

• 林凡 calls himself a technology optimist and says he has only 1 anxiety: “Our generation is standing at a point where opportunity is being redistributed. This may be bigger than the internet and mobile internet. In a window of opportunity like this, how can we move a little faster?” The CEO’s priorities have changed accordingly: “At this stage, finding money is the No. 1 priority.” The prerequisite is that the new business form, technology architecture and business model have already been figured out; he has been thinking and researching them from 2023 through 2025, and “by now they are extremely clear.”
• Building an AI company on Maimai’s foundation is easier than starting over, and his answer is “very clear”: AI needs data, and Maimai’s existing data “should matter for at least a year to a year and a half.” “I have never worried about startups; they do not have this data.” Could a high valuation make fundraising difficult? “There are definitely technical ways to solve that.” The real questions are how big people think the opportunity is and whether you are the person who can make it work.
• His closing self-reflection is that Boss Zhipin’s scale exceeded his expectations. “In 2017 and 2018, we were larger than Boss Zhipin.” Its only difference was adding a faster instant-messaging channel, yet the market outcome far exceeded expectations. He makes a similar comparison with Pinduoduo, which briefly caught up with Alibaba in market value during a period of consumption downgrading despite Taobao holding 95% of China’s e-commerce market: “Without this new AI opportunity, Alibaba’s market value might not have recovered.” The final formula is: “What really matters is social demand and trends. Every founder tries to adapt to them, and whether you have enough luck for your desire to happen to match social trends—Huang is very lucky.” 李翔 adds: “The main thing is that selling graphics cards for games was also very profitable.” “Yes, that is fairly important.”

Verification Notes

• The original audio’s “Gaiteler curve” may refer to Gartner, but this cannot be confirmed from the captions.
• The name “吴钊” and whether he was the person in charge of DingTalk cannot be confirmed from the captions.