A Conversation with Dai Yusen: From Jumei to ZhenFund in the AI Era
Summary
Dai Yusen reduces the essence of angel investing to “backing people” and “patience,” because everything around an early-stage company can change, while a person’s ability to learn and desire to create may be what carries them across ten years. His first investment was VR player Skybox; after the sector cooled, founder 罗子雄 pivoted into games and spent four years polishing Party Animals. That completely unforeseeable path taught Dai what it means to “see because you believe.”
More important than chasing the next hot sector is finding the person without whom the thing would not exist. “Test-taking founders” often buy a ticket after a successful template has emerged, then face overcompetition from large companies, peers and capital. Dai therefore looks for founders who can articulate a vision, have shown a history of creating independently, and can withstand an extreme stress test of “why start a company?”
ChatGPT confirmed for Dai on November 30, 2022 that AI had crossed the chasm from research into an enabling technology for almost every form of knowledge work. He used it until 4 a.m. that night, warned his team the next day that it was significant, organized a discussion three days later, and invested heavily in projects including 光年之外 and Kimi in early 2023. His allocation analogy: “If you went back to 2010, three years after the iPhone launched, it would not be excessive to put 80% of your money and energy into mobile internet.”
He also believes AI will create “the biggest bubble in human history,” but that a massive bubble may be exactly what pays for real infrastructure and innovation. Scaling law will push the race for compute, infrastructure and talent to extremes, while most projects in the hype cycle will still fail. Early investors cannot avoid bubbles; they must look for “the beer beneath the foam” and distinguish good bubbles that leave behind infrastructure and innovative capacity from bad bubbles that merely inflate the price of existing assets.
Models and applications are not mutually exclusive; an application’s value depends on how much processing lies between model output and the deliverable the user actually wants. Dai compares model output to sashimi and a finished application to boiled fish: if the output is already the product, the wrapper will be thin; if the application must use multiple tools and steps to deliver a PDF, file, video or completed task, it accumulates context, environment, interface, data, trust and habits. Manus disclosed annualized revenue of roughly $90M within months of launch, a factual rebuttal to “applications will inevitably be swallowed by models,” though Dai still acknowledges that time will decide.
Heavy spending by a frontier-model company is not automatically a reason to pass; the key question is whether the leading model can control a massive general-purpose entry point. On the program’s figures, ChatGPT had roughly 400M DAU and close to 1B MAU; OpenAI was valued at about $500B, Anthropic at about $175B, and Kimi’s last round at about $3B. Dai acknowledges that model investments are unlikely to produce 10,000x returns and face heavy dilution, but says that if he could do it again, he would still invest—and “should have put more money into Kimi.”
China’s venture market is recovering from a trough, but a return of heat requires higher standards, not another rush into consensus trades. More than half the founders in ZhenFund’s latest fund are under 25; Dai sees investment activity in AI, robotics and hardware approaching 2020–2021 levels and points to large valuation gaps between Chinese companies and overseas peers. His portfolio stance is “long China,” angel-only, believe in young people during winter, and believe in the cycle while maintaining high standards in summer.
Deep dive
1. Leaving entrepreneurship was not retreating; it was refusing to keep acting out someone else’s script
Dai Yusen sees both dropping out of Stanford and returning to China to start a company in 2009, and leaving Jumei to join ZhenFund in 2017, as the same kind of choice: doing what he most wanted to do at the time rather than what society thought someone with a Tsinghua or Stanford background “should” do.
He started a company at 22, rang the IPO bell at 27 and left at 30. The outside world expected the former No. 2 executive to become CEO again and make up for the company’s market cap falling from tens of billions of dollars to several billion, but he had no interest in power, status or being “No. 1” for its own sake.
What consistently drew him was zero to one: small, exceptional teams, creating something new and learning fast. Managing nearly 1,000 people and handling operational work after the IPO showed him that even another company would eventually make him unhappy once it grew large. “I’ve already seen this movie called entrepreneurship. What’s the next movie?”
2. Moving from deep entrepreneurship to broad investing was an honest reckoning with curiosity
Entrepreneurship requires taking one thing deep, mastering it and pushing it to the limit. After eight years, Dai wanted breadth instead—even if he only understood each field “a little”—so he could keep engaging with different problems in robotics, blockchain, consumer, games and AI.
He initially did not know whether investing would become a long-term career and even considered being only a venture partner before briefly observing the market and starting another company. He later realized angel investments often take five or even ten years to reach an outcome, and a loose part-time commitment could not carry that responsibility. He therefore asked to become a partner.
Investing also extended another motivation: helping more young people like his younger self get the chance to start companies. To him, this was not a retreat from “doing things” to “watching things,” but a conversion of the satisfaction of creating into the work of finding and accompanying people.
3. His dislike of large companies emerged through a sequence of personal experiments
Dai’s first internship was in the finance department at General Electric, where he quickly concluded that “anything that wasn’t internet-related was boring.” He then joined Baidu in 2007 and, because Google at the time was like OpenAI today, moved on to Google China to work in user-experience design.
At Google, even a single interface icon required layers of approval before reaching Marissa Mayer, who then oversaw user experience and design. He began asking himself, “How long would I have to work before I reached a level where I could really make the call myself?” The organizational chain itself gave him a headache.
After meeting entrepreneurs including Wang Xing and Wang Huiwen, he reached a more specific answer: he liked the internet but not large companies; he liked work he could truly control and create himself. He describes his path as repeatedly eliminating things he disliked: “If I don’t like it, I won’t do it.”
4. Jumei began not with a grand plan, but a survival pivot after cash fell to RMB300,000
After the team raised $180,000 from Xu Laoshi, it initially built an advertising system for social games. China’s social-game monetization had not yet taken shape, the project quickly failed, and the company was left with roughly RMB300,000—enough to last only a few months.
They tried selling products and found cosmetics relatively easy to move. Because they admired Wang Xing, they borrowed Meituan’s model of selling one local-life group deal a day, reversed the name to “Tuanmei,” and changed the business to selling one cosmetic product a day. “Tuan” was later replaced by “Ju.”
Dai uses the episode to remind investors that early corporate paths are often “a matter of chance, with no plan behind them.” The business Jumei ultimately became was completely different from the social-game advertising system described in its original fundraising pitch.
5. Entrepreneurship brings empathy and product judgment, but also the most dangerous “mature perspective”
Having founded a company, Dai understands resource scarcity, difficult choices, the joy of growth and the pain of downturns. He can also discuss operating details more naturally with product-oriented founders. That feeling that “we have all been in the trenches” is difficult to replace with analysis on paper.
But when he began investing, he carried the perspective of someone who had built a public company and forgot that he himself had been just as raw at 22. He could easily see a team’s weaknesses while overlooking the possibility that people grow, forming the static judgment that “the team is not strong enough, so it cannot succeed.”
The deeper trap was projecting himself into a company: “If I were doing this, what would I do?” A founder’s method might differ from his and produce choices he could never have imagined. He later required himself to forget specific experience while retaining the entrepreneurial spirit, so knowledge would not become a constraint.
6. Pinduoduo showed him that his own success formula could become刻舟求剑
Jumei benefited from traffic dividends and social-media marketing, and also experienced the rise and fall that followed their disappearance. When Pinduoduo grew rapidly through official accounts, mini-programs, group buying and “cutting the price,” Dai inferred that the end of WeChat’s traffic dividend might cause the company to stall as well.
Subsequent events forced him to update his view. After the WeChat dividend, the Pinduoduo team launched the RMB10B subsidy program and then took Temu overseas, demonstrating continuous evolution. “Things we couldn’t do, they can do better.” Having stepped into a pit himself did not mean others would inevitably step into it too.
Li Xiang added that the more familiar someone is with an industry, the easier it is to see only a new project’s flaws. Dai agreed, calling it a blind spot under the lamp, and cited ZhenFund’s relatively limited investment in online education: understanding education too well can also make every new opportunity look immature.
7. The deeper the experience, the greater the risk of being trapped by an old paradigm
Dai extends this curse of knowledge to AI. After AlexNet made neural networks mainstream in 2012, some highly experienced researchers still rejected the large-language-model path. They may eventually be proven right, but their current results cannot simply be ignored.
Li Xiang distinguishes between knowledge becoming outdated and being constrained by knowledge. Dai believes the two are essentially related: old training creates confidence in what is correct, so the more a new paradigm conflicts with existing theory, the more experienced people tend to reject it.
This is especially dangerous in angel investing, where every investment should theoretically be in something new. The more experience one accumulates, the easier it is to fall into the innovator’s dilemma. Large companies are constrained by existing users, organizations and success formulas; the space for startups comes from precisely those constraints.
8. His investing education passed through three stages: invest in everything, invest in nothing, then raise the bar while daring to believe
In the first stage, the novelty made Dai feel that robotics, blockchain, O2O, the sharing economy and AR/VR all had opportunities. Jumei’s smooth path to an IPO in four years reinforced his optimism, and he invested in more than 20 companies in his first year. Many later failed.
Operating a company trains people to quantify goals through revenue or project count, but investing is not about investing more; it is about whether one can back enough companies that become large. Treating annual deal count as a KPI was an early misuse of his operating mindset.
In the second stage, his entrepreneurial experience let him spot problems in every early-stage team almost immediately, so he “didn’t want to invest in anything.” That period lasted roughly one or two years. The third stage was learning to raise standards while remaining tolerant of genuinely innovative, young and immature people.
2022 was particularly confusing. The mobile-internet To C world he knew appeared to be over, while the market shifted toward battery materials, energy storage and new energy—unfamiliar hard-tech fields. ChatGPT brought him back into a product-innovation field he both believed in and understood, with excitement far exceeding 2017–2022.
9. Angel investing gives feedback too slowly; the projects running fastest in the short term may be the most dangerous
Angel decisions are unlike stocks: there is no price feedback the next second. The project a fund considers best two or three years after investing is often not the one that ultimately generates the largest return. The real winner may still be hiding in a quiet, unremarkable portfolio position.
Dai uses Xiaohongshu as an example. ZhenFund strongly believed in founder Mao Wenchao, but the market long viewed Xiaohongshu as merely “small and beautiful” and did not immediately price it as a major platform. Slow growth does not mean value is not accumulating.
Conversely, a company raising round after round within one or two years often reflects sector heat, not just business improvement. Hot money creates competition, large teams, high marketing spend and extravagant habits. When the cycle turns cold, the fundraising feedback mistaken for success becomes a burden.
10. His first investment went from Skybox to Party Animals, proving that the business changes while the person evolves
In May 2017, before formally joining ZhenFund, Dai participated online in an investment in 罗子雄. The latter had led the design of Hammer Technology’s phones and was building VR player Skybox on Google Daydream, where its polished product earned a prominent recommendation at Google I/O.
After the VR boom faded, the company fell into difficulty. In late 2018 or 2019, 罗子雄 proposed making a family-friendly game. Dai initially thought it was merely a side project to keep the company alive, but the team polished it for four years before releasing Party Animals in 2023.
By Dai’s account on the program, Party Animals sold several million copies, making it the Chinese game with the second-highest Steam sales after Black Myth: Wukong. Its wishlist ranking once reached No. 2, behind only Elden Ring.
The game was nowhere in the original investment thesis, but design taste, product capability and the resilience to search for a new direction remained constant. It echoed Jumei’s shift from game advertising to cosmetics: “Even if you understand the person, what they do may still change.”
11. The person worth backing is the leader, not someone waiting for investors to set the question
After reviewing its historical investments, ZhenFund strengthened one standard: back the leader “without whom this thing would not exist in the world.” Founders who wait until a trend has emerged and ask, “What’s hot recently? I’ll build it for you,” are what the firm calls “test-taking founders.”
Once labels such as agents, AI companions and humanoid robots become standard answers, large numbers of teams enter simultaneously. But a trend may ultimately support only one or two large companies—or none at all. The metaverse and AR/VR’s failure to cross the chasm for years are cautionary examples.
Entrepreneurs also underestimate the intensity of competition inside a hot sector. Large companies, strong founders and capital have all seen the same opportunity. A profitable business becomes unprofitable through crowding, and a market that was once large enough gets divided among multiple companies.
12. An opportunity dismissed by the giants may be the one that leaves a window for startups
Dai believes it is difficult today to expect large companies not to see an opportunity at all. Strategy departments and bosses are sufficiently diligent. If a giant immediately thinks a market looks attractive, a startup will probably enter a direct contest with an overwhelming resource disadvantage.
The intense competition between Kimi and ByteDance and other large companies shows that a correct direction does not necessarily give a startup good odds. By contrast, Manus was initially dismissed by many researchers as a wrapper with no technical substance. That contempt may have created room to explore.
Li Xiang compared this to the mobile-internet era’s “stupid window”: when everyone thinks an idea is stupid, an entrepreneur can grow quickly. Airbnb, which let strangers stay in one another’s homes, was first considered crazy and only later revealed its industry-disrupting potential.
13. Entrepreneurship turns small cracks in a team into structural failure
Dai compares entrepreneurship to Brad Pitt’s F1 film: a race car operates under extreme conditions, so any tiny crack can rapidly become a break. Starting a company likewise acts as an extreme stress test of team composition, founding motives and motivation.
A prestigious university or large-company résumé cannot prove that someone is ready to persist in an environment of extreme resource scarcity, highly uncertain outcomes, intense competition and simultaneous declines in income and social status.
That is why one of ZhenFund’s most common recent questions is, “Why do you want to start a company?” After mass entrepreneurship and the proliferation of “30 Under 30” lists, being a CEO briefly became a life checkpoint and social currency. But entrepreneurship is “extremely hard,” and status-seeking cannot carry someone through a trough.
14. Entrepreneurial motivation cannot be judged by polished answers; return to the person’s prior actions
The best entrepreneurs may not succeed on their first attempt, but they often discover early that employment or academic research is not the life that suits them. Creating, tinkering and implementing personal ideas did not suddenly begin when a starting gun went off.
Dai compares people to models continuously aligning with a reward function. After 20 years at a large company, an individual’s reward mechanism is usually deeply aligned with the organization; moving into entrepreneurship then requires overcoming existing identity, certainty and behavioral inertia.
ZhenFund tracks what candidates did at key career forks: whether they pursued their own interests rather than social approval, whether they created proactively, and whether they had articulated a vision of the future and gathered people around it. Communication skills are increasingly trainable; real actions contain more information than expression itself.
15. Everyone says they want to be contrarian, but collective action keeps creating consensus trades
Li Xiang points to a venture-capital paradox: almost no GP will admit to wanting consensus investments, yet the industry’s collective behavior continually creates crowded sectors. Dai believes this begins with the security of short-term fundraising: someone quickly takes the next round, the project’s paper valuation rises, and it feels like receiving a high score on a capital-markets exam.
Hot sectors usually have a rational basis. There may be a technology inflection point, or one company may already have shown that the path works. Once smart people understand those rational factors, it becomes easy to package following the crowd as rigorous reasoning while overlooking how competition and valuation have changed the return profile.
More commonly, people “buy a ticket after the train leaves”: after missing Douyin, they build a vertical Douyin; after missing Uber and Didi, they search for a “Didi” in every industry. Besides shared charging banks and shared bikes, there were shared massage chairs, gyms and even toilet paper—the mechanical extrapolation of one successful logic.
16. ZhenFund does not divide sectors into fiefdoms; it builds agency into the organization
ZhenFund does not set three sectors every year and assign partners to cover them. Dai believes making AI one partner’s exclusive territory would kill everyone else’s curiosity and innovation. Most of the team are generalists anyway.
From Xu Laoshi to analysts, everyone must find projects, register them, write up a share and push them into the meeting. Dai calls the traditional model—young people source deals while partners only judge—a form of “partner exploitation of children.”
Every Monday from roughly 10 a.m. to 2 p.m., the team spends four hours sharing what it learned the previous week. The rotation runs alphabetically from A to Z, then reverses from Z to A the following week. Founding partners and analysts sit on the same table and carry the same learning obligations.
Returns are also shared by the team rather than allocated according to whose project it is. This reduces destructive competition for sectors and deals, allowing Manus, invested in by Liu Yuan, to receive recruiting, operating and overseas-event support from Dai, Anna and others at the same time.
17. Cross-sector investing is not constant market timing; it is familiar people heading in different directions
Dai invested in Wen Hsiang not because he had a firm view on fragrance, but because the founder had been his colleague at Jumei. She was still choosing between tea and fragrance, and ZhenFund also took part in discussing the direction.
His investment in Orienspace likewise came from long-term confidence in 姚颂’s capabilities. 姚颂 had previously founded DeePhi Tech and sold it to Xilinx. Dai is a space enthusiast, but the investment logic was first “I want to invest in 姚颂,” not “I have decided to allocate to rockets.”
Dai studies AI deeply but still does not try to divide it into fixed sub-sectors and scan each one. The more fundamental question is: “What are the best people of this era like, where are they, and what are they doing?”
18. A four-part founder framework makes judgment discussable, but cannot become another cage of knowledge
As someone with an engineering background and a T-shaped way of thinking, Dai has tried since joining ZhenFund to distill the experience in senior investors’ heads into a system that can be learned and passed on. ZhenFund’s generational transition after Xu Laoshi’s retirement also required a method that did not depend on one person’s intuition.
The team categorizes founders into four archetypes: “young genius, veteran driver, scientist and operator,” with particular attention to the first two. Judgment cannot stop at “I like this person”; the investor must explain which type the founder represents and what objective behavioral evidence supports that view.
But angel investing is precisely about finding outliers. Any pattern can only improve probabilities, while the framework itself may exclude the most important exceptions. So the team builds a common language while repeatedly reminding itself that frameworks become outdated and always have exceptions.
19. Missing Pop Mart showed that even a complete framework has limits
ZhenFund met Wang Ning several times and visited Pop Mart stores but did not invest. The company may still have been operating a shelf-rental model, making the platform difficult to foresee from the business itself. In retrospect, the question is whether they could have seen Wang Ning’s still-emerging ambition earlier.
Dai admits that this kind of judgment is difficult to reconstruct using today’s understanding of the founder. Even a redesigned framework might not capture every type. Angel investing does not require backing every good company; executing well on the projects inside one’s own capabilities and culture matters just as much.
A miss can still expand the boundary of understanding. Consumer-brand founders have a typical profile different from technology founders. A fund must decide which exceptions are worth updating the framework for and which reflect blind spots where it cannot build an edge.
20. Consumer and hard-tech founders are identifiable in completely different ways; people-business fit matters more than a unified education standard
Consumer products have a “universal” quality: anyone can open a tea shop, and the number of competitors is enormous. ZhenFund cannot easily judge whose tea tastes better before a founder has opened the first store. When Heytea entered its field of view, it already had queues; when Chagee submitted its first BP, it already had roughly 100 stores.
Candidates training large models or building robots, rockets or chips may number only a few dozen or fewer. Technical barriers make the identification set much narrower. Yang Zhilin and Wang Huiwen are standard examples of ZhenFund’s “young genius” and “veteran driver” archetypes, but the model-building case also demands deep technical understanding.
Dai therefore revised the simplistic statement that “education does not matter.” A consumer brand does not necessarily require an elite educational background, while model training requires matching technical ability. The question is whether the person can do the job, not whether one résumé standard can cover every industry.
21. AI differs from previous booms because breadth and maturity crossed the threshold together
Dai places AI alongside the internet, mobile internet and semiconductor revolutions because it is not a tool for one sector. It can augment or replace knowledge work across law, finance, research, programming, images, video, voice and music.
The second axis is maturity, as described in Crossing the Chasm. Early markets can become excessively excited about a technology while mainstream markets have yet to accept it, leaving the technology stranded in the middle. VR remained there for years; Quest’s tens of millions of units looked more like a “BlackBerry moment” than an iPhone moment.
ChatGPT’s significance was turning the model into a general-purpose product that anyone could use by talking. It had broad use cases, dramatically improved memory and answer quality, and kept getting better and cheaper. Large language models were no longer merely AlphaGo-style research demonstrations; they were entering the mainstream market.
Cursor and Claude Code have pushed AI programming across a similar threshold. Dai’s aggressive view is that programmers who do not use AI may be 10x less productive than those who do. Once “not using it no longer really works,” diffusion is no longer just a hype narrative.
22. Manus’s email entry point shows how an agent can embed itself in real workflows
Dai’s representative scenario is this: after receiving an email, the user does not need to copy the content, switch to GPT or Doubao, then move the result back. They simply forward it to Manus’s dedicated address, allowing Manus to read the attachments, complete a research report and reply.
This is more like a boss forwarding a task directly to an assistant than personally operating a tool: “You really spend the time of one tool.” If other people can also communicate first with a user’s AI bot, an agent may take over more information filtering and execution.
Li Xiang questioned whether such an entry point would be easy to copy. Dai admitted that “once you have pierced through it, none of it is that difficult,” but early entrants accumulate data, trust, brand and user habits. Once forwarding a task becomes the default action, later entrants may copy the function without copying the relationship.
23. The bigger AI gets, the bigger the bubble; early investors must learn to coexist with it
In a 2023 presentation, Dai predicted: “AI will be the biggest bubble in human history. Compared with the AI bubble, the internet bubble may not even count as much.” Science-fiction narratives can generate hundreds of use cases, while companies, individuals and governments all have incentives to win the race.
Scaling law reinforces the capital instinct that bigger and more is always better, pushing compute, infrastructure and talent prices to extremes. He cited Oracle’s disclosed future orders, which at the time far exceeded analysts’ expectations, as evidence that the infrastructure expansion was already visible in financial data.
But the bubble is not a reason to avoid AI. Dai says early investors should “try to find the beer beneath the foam.” Without bubbles, Amazon and Google might not have received funding, vast amounts of fiber might not have been built ahead of demand, and later companies would have had no reusable infrastructure.
24. A good bubble bets that the future will be different; a bad bubble bets that old demand will last forever
Li Xiang introduces the distinction between good and bad bubbles. A good bubble believes AI, robotics and commercial spaceflight will change the future, leaving behind infrastructure, technology and talent after it bursts. A bad bubble assumes housing prices, existing consumption or asset prices will keep rising.
Dai adds that good bubbles are usually paid for by professional risk-takers such as VCs and entrepreneurs, generating positive externalities that advance innovation. In real estate or subprime-style bubbles, ordinary people often bear the final cost, without creating new supply capacity.
AR/VR’s long-term vision may be correct, but the technology matured too early and had too little beer beneath the foam, so the hype mainly expressed itself as a bubble. ChatGPT was different not because its vision was grander, but because ordinary users could already access productivity directly.
25. His greatest humility about the future comes from the fact that no top VC discussed ChatGPT in advance in 2022
Dai recalls visiting Silicon Valley in May 2022 and speaking with several top VCs. Almost no one treated AI as the most important topic. The common subjects were biotech, inflation and similar issues, while autonomous driving was considered “more or less done.”
Six months later, ChatGPT rapidly changed the industry’s attention. If the industry failed to see a major change one year ahead, it was difficult to predict three or five years ahead with confidence. He quotes a Zhang Xiaolong-style reminder to himself: “Everything I say is wrong,” and he must always be ready to be proven wrong.
This is also the weak-side logic of backing people: not knowing what the future will bring, and therefore searching for relatively stable qualities such as character and creative ability. Once a major change crosses the chasm, investors should dare to concentrate resources rather than continue searching for “the next one” merely for formal diversification.
26. Whether to go all in on a wave depends on both its breadth and whether it has begun crossing the chasm
Dai’s thought experiment is that if a fund manager returned to 2010, roughly three years after the iPhone launched, putting 80% of the fund’s money and energy into mobile internet would not be excessive. When a true technology platform appears, the duration and scale of the change are often still underestimated.
But investing in mobile internet before the iPhone might have left only an acquisition outcome like UC Browser. The direction could be right while the timing was several years early. Knowing a trend too early can also cause smart people to miss the real commercialization point because it already feels “unoriginal.”
He therefore uses two criteria to distinguish AI from lookalike trends: whether it can broadly empower industries, and whether products have already delivered clear value to mainstream users. In his view, AI satisfies both; AR/VR still lacks the second.
27. Angel investing has only three jobs—find people, assess people, invest in people—and only one shot
ZhenFund breaks the work into finding excellent people, judging excellent people and putting money in. Brand, friendly terms and founder reputation make the third step relatively easy. The bigger bottleneck is meeting someone early enough.
Because ZhenFund usually seeks the first round, or at most the second, confirming that a company is excellent after the third round means it has already moved outside the firm’s circle of competence. “We only have this one shot.” Early-stage investing therefore requires building relationships early.
Dai contrasts primary and secondary markets. Secondary markets believe because they see, based on financial reports. Angel investing must see because it believes: believing in a founder before the company even exists, when the founder may still turn an unknown idea into reality.
28. Finding people is treated as a product, with different entry points for different stages of life
For students and very young founders, ZhenFund uses its “Post-00s Plan,” a campus program running for nearly ten years, school visits and exchanges with portfolio companies to build relationships. For middle managers with three to five years of experience, it organizes professional discussions around topics such as robotics.
For senior candidates such as unicorn co-founders, the “Ostrich Club” emphasizes confidentiality and peer-level exchange. They do not want the outside world to know prematurely that they are preparing to start a company, and they need to learn about fundraising, legal matters and company-building as they move from business leader to CEO.
Dai admits that the best projects still come disproportionately from trusted referrals, because the referrer puts personal credibility behind the recommendation. Events are not only about meeting an investment target on the day; they build weak-tie networks that can continue producing deal flow.
Manus’s relationship began roughly ten years ago, when investor Liu Yuan met Xiao Hong, then a college senior, at a hackathon and wrote the first check. After one startup, an acquisition and a second investment, the relationship eventually led to Manus.
29. ChatGPT was Dai’s lightning moment, not a gradual industry call
On November 30, 2022, Dai first used ChatGPT until 4 a.m. The experience reminded him of seeing Google for the first time in early 1999: a simple input box could quickly return rich feedback to almost any question, “like being struck by lightning.”
He had followed GPT-3 before, but mostly understood it as a writing model that continued text. The conversational question-and-answer format was the first time he saw next-token prediction produce an ability to solve problems, turning AI from a research capability into a product everyone could use.
The next day he told the team group that “this is extremely significant; it showed me the dawn of a new era,” and organized the first large-model discussion three days later. The more he analyzed it with researchers, the more he believed this was not a superficial product improvement but a turn in underlying capability.
30. The most effective way to feel the future is to pay for and use the frontier products yourself
Dai sees going online early, buying an iPhone early and subscribing to AI products today as the same kind of cognitive investment: “It is important to spend a little time and money feeling the future.” Many products cost only $20 and can let users experience the most advanced productivity available at the time.
When Devin launched and required a $500 prepayment, he tried it immediately and offered the program’s most vivid comparison: “$500 can buy a bottle of Moutai, or one month of Devin. With Moutai, you can only get drunk and see a fake future; with Devin, you can see the real future.”
In early 2023, he organized the prompts accumulated from daily use into a test set and built an early large-model benchmark, Z-Bench. Model companies and relevant regulators came to exchange views. Although the team lacked the personnel to iterate on it for the long term, the project embodied hands-on use rather than listening only to reports.
31. The best AI products create a sense of magic through technical progress in the right form
Dai observed that when products such as Manus and Nano Banana took off, they did not rely on large marketing budgets. Users spread them organically after seeing the capability for the first time. Technical progress must be converted through product form into an experience that needs no explanation.
Manus lets AI use its own computer environment to complete an end-to-end task, while Nano Banana turns new generative capability directly into an experience. These moments that “look like magic” are important signals for a product investor judging whether a technology has crossed the chasm.
Before ChatGPT, Dai had only broad interest in AlphaGo, autonomous driving and robotics because the commercial product path remained unclear. After ChatGPT, he began “talking to everyone who knows the most about anything worth discussing, and using every frontier product available.”
32. ZhenFund moved quickly in 2023 and converted shared learning into portfolio collaboration
Dai recalls investing in 光年之外 and Kimi in January and February 2023, then in an AI education project in March and April. Butterfly Effect had entered the portfolio in August or September 2022 and quickly pivoted to Monica after ChatGPT appeared.
Genspark received investment around August or September 2023 and later reached a valuation of roughly $1B, according to the program. Dai’s view is that the sharpest teams were generally founded or pivoted quickly in 2023.
Weekly sharing turned individual discoveries into fund-wide information. New products such as ChatGPT and Devin were demonstrated live, and the team even planned to buy Devin and compete over who could build something, rather than leaving new technology only to the partner responsible for the investment.
33. Heavy support for Manus was a special response, not an attempt to turn the fund into an incubator
ZhenFund owns close to 20% of Manus’s parent company and is its largest shareholder. The fund entered an unusually intensive support mode only after the product suddenly became a global hit and the team was severely understaffed.
From the March 6 launch to early April, Anna, whose English was the strongest, replied one by one to Twitter and LinkedIn messages and helped contact KOLs. Dai held multiple Meetups with team members in the United States, while other partners helped with recruiting and traffic issues.
The intervention was built on years of trust and a clear need from the company, not on investors taking control. Dai emphasizes that compared with what founders create, all of the help is merely “a small favor”; ZhenFund would never believe, “Without me, you could not have done it.”
34. In the “wrapper” debate, separate factual questions from questions of principle
Dai responded to doubts about Manus’s paid conversion and renewal ability with disclosed company data: only months after launch, annualized revenue had reached roughly $90M, among the fastest in the industry. Such questions can continue to be tested against operating facts.
Whether “wrappers have value” is a disagreement in principle. Dai rejects the view that if model companies build good models, applications will inevitably disappear. Model capability needs context, environment and interface to become user value, though he also admits he may be wrong.
His attitude is not to defend every criticism, but to return attention to the product: “Users like you and are willing to pay you.” Misunderstandings will gradually disappear. Investment alpha requires differentiated views; controversy cannot automatically turn a non-consensus view into a mistake.
35. An application’s moat depends on how much processing lies between deliverable and model output
Claude Code’s impact on Cursor forced Dai to update his view of applications. If the model directly outputs code and the product’s deliverable is also just code, the application layer adds little value, overlaps heavily with the underlying model and is vulnerable to upward integration.
Manus may deliver a PDF, file, video or complete result produced across multiple tools, requiring several different types of processing along the way. The farther model output is from the user’s desired result, the more an application can create independent value through orchestration, environment, data and workflow.
His analogy is: “If you are making sashimi, the dish is the fish itself.” A fish seller can easily replace the restaurant. Boiled fish requires heat control, sauce and cooking skill; the raw-material supplier is not the same as the finished product.
Li Xiang asked whether this should become a standard for investing in applications. Dai still brings the boundary back to backing people: the mechanism helps explain the business, but cannot replace judgment of the team, because product direction in the early stage of technology will continue to change.
36. The shift from Monica to Manus again proved that strong teams outlast thin-or-thick product labels
When Monica was a browser plug-in, almost every Chinese VC knew the product and Xiao Hong, and most judged it as “a plug-in, a thin wrapper.” ZhenFund still made a large investment because it was backing the team, not that particular product form.
The team later considered an AI browser, abandoned the direction and ultimately launched the general agent Manus ahead of the market. The product, user and entrepreneurial experience accumulated through Monica prepared the next leap, but the outcome could not have been predicted a year earlier.
Dai is wary of founders who refuse to change course in the early stages of technology. When capability boundaries and product forms are still evolving, exploration and pivots are normal. Being prematurely certain that the original plan must be right may be a form of arrogance.
37. Large-model startups are expensive, but may control AI’s most valuable general-purpose entry point
Responding to the claim that large models belong only to large companies, Dai notes that OpenAI and Anthropic themselves are VC-backed startups. China’s DeepSeek, Kimi and MiniMax likewise show that startups can participate, though they are “large startups.”
He believes OpenAI may still have been undervalued at the time of the program. ChatGPT had roughly 400M DAU and close to 1B MAU. Users were moving from entertainment questions such as writing poems toward work, health, shopping, travel and life decisions; trust and commercial value increase with the importance of the questions.
The model market may become highly concentrated; there will not be three ChatGPTs with 1B DAU each. VCs therefore need to invest in the first tier. Heavy spending is not sufficient grounds to object; the key is whether the outcome produced by the investment can be large enough.
These projects are unlikely to replicate the program’s claimed $500 valuation investment in Xiaohongshu and its later rise to roughly a $50B market cap. But the absolute dollar return may still be enormous, making them suitable for large funds and less suitable for institutions pursuing only 1,000x opportunities.
38. The Kimi odds debate is a strategic split between high multiples and high absolute value
If he returned to 2023, Dai says clearly that he would still invest in model companies and “should have put more money into Kimi.” Describing model investing as a painful mistake after DeepSeek appeared, he believes, confuses failure to lead the consumer entry point with failure to create value.
He calls Kimi the best open-source coding model currently available. Facing powerful competitors such as ByteDance and DeepSeek, Kimi still has an excellent team, resources and international open-source standing; it does not need to become the sole winner to prove its significance.
On the program’s figures, OpenAI was valued at roughly $500B, Anthropic at roughly $175B and Kimi’s last round at roughly $3B—less than 1% of OpenAI. Mistral had already stopped training models, and Dai went so far as to call Kimi “the only investable company in the world” for the coding-agent direction.
He also acknowledges that “where your butt sits determines what your head thinks”: as a Kimi investor, he inevitably has a position. A fund such as Benchmark that seeks 1,000x returns is entirely rational not to invest in models. The debate is not whether models deserve to exist, but whether the return profile fits the fund’s size and strategy.
39. AI’s next phase will move from training to inference, and from model capability to application delivery
Dai believes the market still underestimates AI’s long-term impact. Model capability is already very strong; what is missing is product form that makes the capability visible. After huge sums are spent on training, industry attention will inevitably shift toward inference, applications and concrete user scenarios.
Manus shows what current models can do on agent tasks, while programming has already undergone a structural shift. Junior programmers face the most direct impact; experienced programmers may become more valuable because they are better at defining tasks, checking results and directing AI.
Images, video, music, mathematics and programming may each produce a “Li Shishi moment”—AI not merely reaching average human performance, but exceeding top humans. Much work will then shift from doing the task personally to deciding what to do, evaluating multiple outputs and taking final responsibility.
The fundamental product leap in Dai’s view is from an answering machine to a task executor: “When AI can complete most of what we need to do, what is our meaning as a person?” He admits the question sounds philosophical, but programmers who use AI heavily have already experienced it directly.
40. Everyone will become an AI boss, but not everyone has learned to be a boss yet
In the past, only a small minority truly managed other people. Agents require ordinary people to break down goals, allocate resources, inspect quality and provide continuous feedback. A large capability gap may open between those who learn to direct AI early and those who learn later.
In writing, research and programming, people who still use no AI are already materially less efficient. The secondary market is especially aggressive because information-processing advantages can translate directly into returns: whoever leads can make money.
Dai cites William Gibson: “The future is already here—it’s just not evenly distributed.” He discusses AI through podcasts partly to help distribute this productivity more evenly, not merely to create a narrative for his portfolio.
41. Angel investing may be replaced relatively late, but AI has already restructured desk research
Li Xiang suggests that angel investing depends on face-to-face trust and long-term relationships and may be among the last investment functions replaced by AI. Dai adds that the “finding people” step is already using substantial amounts of AI.
Dai mentions that ZhenFund has many AI trackers: they automatically follow new high-star GitHub projects, highly cited papers and new researchers, after which AI searches, summarizes and reports on them. Humans handle only the people worth contacting further.
Search and knowledge organization that once required well-educated analysts to perform every day have already been heavily automated. AI does not need to replace a complete investor in one step; it can first compress a large amount of junior desk work.
42. New technology first widens gaps, then raises overall living standards through spillovers
Li Xiang asks whether AI will make people work harder until they have nowhere to go, or create more opportunities. Dai believes both always happen in parallel. Technologies such as the steam engine first gave their owners huge advantages, after which productivity and products spread to everyone.
Li Xiang adds that technologies can also be divided into those that promote equality and those that expand asymmetry. Early firearms lowered the military barrier, while a full medieval knight’s armor was expensive and monopolized. The ownership structure of technology affects how gains are distributed.
Dai responds with radar, GPS and the internet: many technologies initially served the military but later benefited ordinary people. Whether AI will follow the same path remains unsettled, but “widen first, then spill over” is the historical framework he finds more persuasive.
43. The new generation of entrepreneurs is stronger, earlier and more global, but has less tolerance for mistakes
Dai believes today’s entrepreneurs are generally stronger than his generation in education, entrepreneurial knowledge and international perspective. In the past, being the first person to read TechCrunch could create a time gap for importing an overseas model into China. Today, innovations are translated and discussed the same day.
AI products naturally serve global users. Manus could provide similar programming and office capabilities in different languages from its first day, while young teams start companies earlier and see entrepreneurship as a normal life option rather than a rare gamble.
But competition during the Jumei era was more like “amateurs fighting amateurs”: everyone lacked experience, and if you made five mistakes while I made four, I might win. Today’s founders face established giants, serial entrepreneurs and global peers simultaneously. A few key mistakes can make a company disappear.
More than half the founders in ZhenFund’s latest fund are under 25. Dai does not require a first-time founder to achieve immediate greatness; he values starting early and continuing to experiment until entrepreneurship becomes a way of life.
44. Education is only a proxy for learning ability; what matters is the slope of growth
Because of New Oriental’s network and the influence of its founding partners, ZhenFund naturally encountered many students from overseas elite universities and acquired an external label of “backing elite-school returnees.” Today the team no longer deliberately tracks the split between returnees and domestic graduates or between elite and non-elite schools.
Dai reduces the core of education to learning ability. A strong degree increases the probability that someone learns well, but does not guarantee continued willingness to learn. People who did not attend top universities may still demonstrate extremely fast learning through entrepreneurship and product work.
Robotics competitions expose metric bias particularly well. Hands-on projects such as RoboMaster consume enormous amounts of time, and students optimizing exams and papers may conclude that they are “not worth it.” Those willing to build things may not have the highest grades but may be better suited to innovation.
ZhenFund once searched for “people who did not fit in at elite schools”: getting in proved baseline ability, but they did not simply follow established rules. Today this has expanded into observing a founder’s growth slope rather than treating an educational credential as ability itself.
45. China’s global advantages lie in applications, hardware, supply chains and the courage to build To C
Dai believes Chinese teams have demonstrated world-class competitiveness in mobile applications, cross-border business, robotics, home hardware and supply-chain products. During the mobile-internet era, perhaps half of the top 10 downloads in the United States were Chinese products, showing that globalization did not begin with AI.
The new generation also understands Silicon Valley products, methods and aesthetics at a comparable level. Many overseas entrepreneurs, by contrast, study Chinese innovation less. Products such as Manus and Genspark no longer look recognizably “Chinese” in interface or experience.
Hard work, pragmatism, speed of execution and careful spending remain common advantages. More unusually, Chinese teams are willing to build general agents and consumer-facing AI products, while the U.S. YC ecosystem more strongly encourages vertical and enterprise services, even producing startups that build bookkeeping, reimbursement and attendance tools for one another.
Li Xiang suggests this may also reflect the difficulty of doing To B business in China. Dai agrees that the environment shapes sensitivity. The United States went years without a major new To C product, while China experienced sustained competition in consumer applications and therefore feels more willing to enter this “dangerous zone.”
46. Chinese companies’ overseas weakness is not only distribution; it is also vision and cultural expression
Entering the U.S. enterprise-services market brings concrete challenges in customer acquisition, BD, renewals and local organizational development. Teams cannot simply copy domestic products and cost advantages unchanged.
Dai believes the global market also demands a stronger narrative: users, employees and partners must understand why the company exists. OpenAI, Anthropic and Cursor have each developed sharply distinctive cultures, allowing them to attract first-rate talent that identifies with their direction.
Chinese companies historically earned trust mainly through good products and service. To recruit top global talent, money will not be the only variable; vision and values will become increasingly important. Li Xiang compares this difference to the United States’ strength in soft power and China’s emphasis on manufacturing strength.
47. China’s primary market is recovering, but recovery can recreate the risk of chasing hot sectors
Dai’s sense is that entrepreneurship and investing were in a relative trough over the past two years, while the economy and innovation activity both needed to recover. At the time of the program, robotics, hardware and AI had clearly heated up, with activity approaching 2020–2021 levels in some areas.
Global attention to Chinese innovation from DeepSeek and Manus weakened the single narrative that China lacked compute and research talent and therefore could not participate in AI. Overseas investors were returning to study models, robotics and applications, while domestic founders and VCs were also becoming more active.
He expects the next year may be hotter, but warns that winter and summer require different disciplines. In downturns, believe in young people and Chinese entrepreneurship; in a boom, believe in the cycle while raising standards, rather than mistaking a recovering environment for proof that every project deserves investment.
ZhenFund summarizes its mission as angel-only and long China. When peers abandon the market in a trough and rush back after the heat returns, the companies already in the portfolio are the result of continuous investment rather than market timing.
48. Overseas LPs care most about three things: are there still entrepreneurs, are there still big fish, and are valuations cheap enough?
The first question is whether young Chinese people still want to start companies. Dai’s answer is yes. The mobile internet’s mature-stage competition once made young people feel they could not beat large companies, while AI and robotics have reopened a technology-transition window.
The second question is whether China can still produce large companies capable of supporting power-law returns. VCs can tolerate 80% of projects failing, but cannot operate without “big fish.” Dai cites Xiaohongshu’s potential for tens of billions of dollars in profits and the growth of companies across robotics and the AI supply chain as evidence that the scale remains sufficient.
The third question is systematic valuation discount. On the program’s figures, ByteDance’s revenue had exceeded Meta’s for two consecutive quarters, while its secondary valuation was in the low-$300Bs versus Meta at roughly $2T. Kimi was valued at about $3B versus OpenAI at about $500B. Whether those gaps are rational is drawing smart capital back for reassessment.
Dai does not equate undervaluation with inevitable upside, but believes capital will search for markets that still have innovators, can still produce large companies and are visibly cheap. Over roughly the past six months, overseas LP visits to China and related interest had increased significantly.
49. After eight years of investing, his best advice remains the simplest: back people and be patient
When Dai first joined ZhenFund, he wondered, “Is it because you don’t understand the business that you invest in people?” He believed his entrepreneurial experience might let him judge the “thing” more accurately. Years of pivots and failures have made him a committed believer in backing people.
His second piece of advice is patience. A company can schedule a three-month test and a six-month launch, but an angel investor cannot schedule when the next unicorn will appear. Someone met tomorrow may not reveal their value for five years, with almost no reliable immediate feedback in between.
Angel capital often needs to wait ten years, forcing investors to ask what matters most over that span. Sectors, products and markets all change; what remains relatively durable is a person’s ability to learn, ambition, resilience and continued creativity. “We have to conclude that people are the most important thing.”