175: A Conversation with Liblib's 陈冕: On Surviving—and Every Moment That Came Close to Death
Summary
Low gross margin does not mean burning cash on subsidies. 陈冕 says the company has generated positive cash flow since May, with performance advertising contributing just 3%-4% of revenue and monthly spend below RMB1M; the host cautions that cash flow includes annual-plan prepayments and cannot by itself prove that growth is healthy. 陈冕’s target for early-stage AI applications is gross margin “above zero and no more than 30%.” The model depends on users not exhausting their full credit balances, controllable renewal rates, and falling token costs; if everyone used their full allowance, the current pricing would “definitely lose money.”
LiblibTV won on speed and attention, not because it was first to find PMF. 陈冕 acknowledges that competitors first validated the node-based video canvas and that LiblibTV’s differentiation was “thinner”; engineering innovations could be copied within a week. But once he realized the market’s revenue had exceeded Liblib’s daily revenue, the team changed course within just over a month and launched an approximately $1M, month-long “saturation attack” during the critical window. The product may have gone from zero to roughly $100,000 in daily revenue very quickly, with a one-day peak of $1M within its first month; using the host’s cited $300M annualized-revenue figure, 陈冕 only confirms that LiblibTV contributed more than half.
The applications era has not arrived; the first priority is to survive. 陈冕 believes profits are still concentrated upstream along the chip-model-application stack. For applications to target 70%-80% gross margins now would amount to voluntarily suppressing demand; the real opportunity comes when intelligence and tokens are cheap enough for the application layer to create incremental value. He estimates that global annual revenue from AI video creation could eventually reach $100B, because it will not only replace traditional software but also absorb parts of the production stack such as cameras and actors.
Dependence on model vendors is a real problem, but applications are not simply arbitrage middlemen. 陈冕 does not shy away from saying that applications are “leverage on top of models,” and concedes that the odds of surviving independently between the giants may be below 10%. His path is to “trade time for space, space for resources, and resources to build the moat.” For now, multimodal interfaces and user mindshare are only a “Shu Road”; ultimately, the company must build a “Yangtze defense” based on user scale, content supply, and a two-sided network, or it will remain subject to the model vendors’ “gravity.”
Being first to find PMF is still a world away from ultimately winning the market. 陈冕 accepts that LiblibTV is less original than Lovart and agrees that first movers deserve recognition. But swipe-based short video, the smartphone form factor, and node-based canvases will all converge, leaving the interface itself with little durable intellectual property. His response is that “innovation has a cost”; after inventing the airplane, one can study rockets, but if one cannot defend and operate the airplane business, there will be no resources for the next round of innovation.
Speed saved the company, but it also created organizational debt. After a battle with a better-funded rival, the company once had just RMB4,000 left, while GPT Image 1 suddenly erased a workflow direction the team had spent a year exploring; extreme anxiety became the engine for product pivots and timing. The cost was that roughly half of the early team left, including 4 technology heads who had reported directly to 陈冕, while Lovart at one point accumulated more than 2,000 unanswered customer-service emails. He admits he was good at removing people who did not fit, but failed to show those people enough respect or help them find a new role.
The ultimate bet is not on a particular interface, but on whether humans will still have value over the next decade. If AI concentrates both production and the feeding of humans, chips, top talent, and capital will absorb everything, leaving almost no room for independent applications. 陈冕 is betting that within 10 years people will still create, express themselves, and resist total centralization. He summarizes the company’s mission as “using AI to unleash imagination and make the inner world a little richer”—not merely replacing human production, but making human taste and nonstandard creativity part of the network effect.
Deep dive
1. Excessive attention put a fast-growing application company on trial
The episode’s backdrop is that the company raised $300M at a $2B valuation during a period when AI applications struggled to raise capital. It was simultaneously accused of lacking originality, buying growth with aggressive pricing, and being destined to blow up. 陈冕 believes the attention far exceeded the company’s current business scale simply because there were so few leading applications to discuss outside the model layer.
His reaction to nicknames such as “Mian Emperor” was not to enjoy the halo, but to say, “I don’t deserve that much commentary right now.” His recent sleep problems are tied both to the business and to the controversy; what worries him are his own blind spots, including overly direct speech, management style, and the consequences that spill out from both.
陈冕 accepts that aggression and anxiety have become part of the company’s image, but rejects every story inferred from them. He repeatedly separates facts from emotion: “Being misunderstood is the fate of anyone who expresses themselves,” but the person speaking must still bear the backlash that expression creates.
2. Cash flow is positive, but annual prepayments require a closer look
陈冕 says the company turned cash-flow positive in May, with cash on the balance sheet continuing to rise in May and June. That is why he was “very confused” to see claims that the company was about to blow up. This is a cash-flow measure, not a disclosure of accounting profit.
Performance advertising contributes only 3%-4% of revenue, with monthly spending of about RMB1M, possibly less. The company is not refusing to spend; mainstream video platforms may restrict AI tools, leaving it in a position where it “can’t spend even if it wants to.”
The host pressed him on the fact that positive cash flow includes annual-plan prepayments, which naturally make the near-term picture look better. 陈冕 agreed that cash flow includes prepayments, but said the health of growth should still be judged by gross margin and retention, not by annual-plan collections alone.
3. The “39% price” applied to the annual plan, not the Seedance API
Lovart did run an aggressive subsidy campaign for roughly 2 months. A larger overseas platform had first subsidized image generation, and the company feared losing the customer-acquisition window if it did not follow. It moved in while the first-generation Nano Banana was relatively cheap; after global compute tightened and subsequent model costs rose, that subsidy strategy could now be “punched through immediately.”
The “39%” price advertised when LiblibTV launched was the annual-plan price after converting the monthly fee, not 39% of the original Seedance API price. The host pointed out that the copy could easily mislead founders; 陈冕 shot back, “My ad is for my users, not for founders who understand APIs.”
陈冕 says Seedance likely offers no discount anywhere globally and maintains tight price controls. The company therefore was not arbitraging a special upstream deal. The aggressive assumption was about consumption, renewal, and early-stage margins—not access to a price unavailable to others.
4. The underlying AI subscription model is still a “gym model,” with token costs layered on top
Pricing does not simply apply a discount to the API. The company starts with the cost of generating one image or video, maps that into credits, builds price tiers and credit packages, and then layers in unused capacity, renewal rates, and full LTV. 陈冕 calls it an average-consumption model, not a model that assumes every user exhausts the full annual quota.
The host’s challenge was blunt: if users consume their entire original allowance, does the company make money or lose money? 陈冕 answered, “Then it definitely loses money,” and confirmed that when a user buys 1M credits but uses only 200,000, the cost saved on the remaining 800,000 becomes a source of profit.
He rejects the binary choice between “no gross margin” and “no retention.” A user opening ChatGPT or 豆包 every month does not mean the user must exhaust the monthly limit; retention and full consumption are two different things. These tools are not daily-active products: a stable user may be active only for several months of the year and only a few days in any given month.
The unresolved issue is the business model itself. Pure usage pricing can hurt renewals and loyalty, while pure subscriptions are exposed to token-consumption volatility. “Membership fees for renewal, usage billed separately” may be more rational, but would leave the company at a disadvantage while competitors are all selling low-priced annual plans.
5. Annual plans lock in LTV, but can hide weak usage behind attractive cash flow
陈冕 estimated on the show that annual plans account for roughly 20%-30% of orders. He acknowledged that a higher share is generally good for the company because it locks in longer LTV, but said the ideal outcome is not inflated bookings—it is that users continue using the product after purchase.
The host cited Uber’s CEO warning against annual orders: cash flow can look excellent at one point in time, only for a later review to show that users never used the product. The result can be ugly. 陈冕 said the warning should be framed as a warning against “non-use,” not as an argument against annual plans themselves.
The test left by the exchange is straightforward: prepayments solve a timing issue, while retention, renewal, and consumption rates determine whether the model works over the long term. 陈冕 did not disclose a specific gross margin, but confirmed that LiblibTV was not selling at a loss and that Lovart briefly operated at negative gross margin.
6. Gross margin is capped at 30% because applications currently need users more than profits
陈冕’s target range for early-stage AI applications is gross margin above zero and no more than 30%. His comparison is that storage and chips can earn 80%-90% gross margins, while SOTA models can earn 70%-80%; for an application launched less than a year ago to demand the same 70%-80% would be “not reasonable.”
The logic is that applications do not control the underlying model, so their most valuable asset is the user. Raising prices to protect gross margin and suppressing demand only makes the application layer thinner and leaves even less residual value. If consumption keeps rising, prices can go up, while model and token costs should fall over time.
“This is not the applications era yet” is his core view. The main narrative remains models, or even infrastructure. Applications need to stay alive until intelligence becomes cheap and readily available. Low margin is a phase-specific choice—not the end state and not a self-destructive subsidy program.
7. The RMB4,000 balance was the endpoint of the first failed cash-management exercise
After nearly a year of negative gross margins under Liblib’s freemium model, the company really did have only about RMB4,000 in its account before a payment arrived. The real panic began when cash fell to several million yuan; by the time it reached RMB4,000, the next financing round had already been secured.
At the time, the company was a roughly $4M seed-stage business fighting an equally aggressive competitor that had raised $70M-$80M in its latest round. It burned roughly $3M over several months. 陈冕 barely watched cash and believed that if the business worked, financing would inevitably follow.
More damagingly, when the business was performing best and fundraising was easiest, he rejected every approach—even declining to respond on investors’ WeChat messages—because “I was going to war; I had to win.” Once the battle ended and cash had fallen to RMB6M, he started fundraising intensively, only to find that investors were less willing to back a company close to collapse. He calls it the first major mistake of his entrepreneurial career.
8. Strategy can ignore competitors; an existential battle cannot
The first-generation Liblib was a marketplace and community. Once design-related supply had a clear quality lead, supply and traffic were reinforcing each other, and the company had grown to more than twice the rival’s scale, 陈冕 judged that “the game is over.”
His correction to “don’t drive while watching competitors” is: vision should not look at competitors; battles must. Liblib had targeted design and creative production from day one rather than anime, gaming, or general entertainment. But a better-funded rival capable of killing the company in the present could not be dismissed as irrelevant.
Liblib subsidized only design creators, while the rival also poached general-entertainment creators. 陈冕 does not regard that as a failure: “Those creators were people I didn’t want in the first place.” Competition should serve the existing strategy, not rewrite it.
9. High-value production matters more than total users—Liblib’s earliest nonconsensus view
陈冕’s time at 剪映 taught him that Adobe was the world’s most profitable creative tool of the previous era, even though Adobe did not have the largest user base. The lesson was that productivity tools are not a game of “whoever has more users wins,” but of who controls the highest-value production.
The business model of a tool is fundamentally a take rate on production value. That is even clearer in AI: coding and B2B commercialized first because the economic value generated after consuming tokens is higher. If revenue depends on high-value tokens, the company should not begin by chasing low-value, purely entertainment use cases.
He calls this a “simple nonconsensus”: after hearing it, people think it was obvious all along. But when consumer internet success was the industry’s strongest reference point, it was easy to transfer the feel of scale directly into AI creation.
10. A decade of watching from the sidelines left no standard answer, only a constantly revised reward model
陈冕 admits he has repeatedly been wrong. He once thought 360 Mobile Assistant was the gateway, believed bike sharing had enormous potential, and thought retail could improve people’s lives. He later concluded in each case that “the answer is not,” while retail proved too dependent on offline operations and supply chains for his capability set.
While observing Douyin at ByteDance, he was less interested in who first built short video than in why the eventual winner could change even after products such as 美拍 already existed. After founding a company, he found that macro factors are not fixed parameters; countless small decisions, individual efforts, and value judgments can alter the outcome.
His view therefore shifted from “the times make the hero” to “heroes also shape the times.” People cannot change the direction of the tide, but they may delay a wave and alter volatility and the distribution of resources. BAT was not history’s only possible answer; it was the result of a particular competitive process.
11. First movers deserve rewards, but markets do not permanently belong to them
陈冕 explicitly acknowledges that LiblibTV was not the first product in the market to find PMF; a competitor provided the signal first. Criticism that its originality falls short of Lovart’s is “valid to a certain extent.”
His defense is not to deny the similarity, but to argue that the optimal interface converges. Node-based multimodal tools, swipe-based short video, and full-screen smartphones all move toward similar forms. The interface is visible and easy to spread, but it is also the thinnest and easiest layer to copy.
He understands why the industry sympathizes with first movers: “Being the first to find PMF is also a form of innovation,” and should be rewarded. But “being first to discover PMF and ultimately winning the market are a hundred thousand miles apart.” Discovery does not automatically come with a permanent market share.
On accusations of bullying the weak with scale, his answer is that the market LiblibTV entered was larger than the company’s original business, and every team was chasing the same larger prize. The company was not a giant choosing whether or not to enter; once the new category’s revenue exceeded the core business, not pursuing it would have been anti-commercial logic.
12. One failed experience made the company underestimate the video canvas; revenue signals forced a one-month reversal
In 2024, the company had renamed “星流” and converted it into a node-based workflow, but the product performed relatively poorly. That left 陈冕 with the instinct that the market might not be large. When he first saw the new video-canvas team, he was thinking about communication, investment, and supporting innovation—not immediately entering the market.
After the other side did not respond, the company continued researching and found that revenue in the category had exceeded Liblib’s daily revenue—and Liblib was still a leading image platform at the time. Within just over a month, 陈冕 changed the judgment from “we could invest a little” to “this is bigger than the core business; get in and do it ourselves.”
He denies ever making a formal acquisition offer: “I couldn’t even get a conversation with them—how could I make an offer?” The episode was better understood as a correction of priors: the decision was not initially to copy the product; an old product failure had created the wrong prior, which real revenue scale then overturned.
13. Engineering differentiation lasts a week; the real scarcity is catching the launch window
Around LiblibTV’s launch, the team built practical features such as Director’s Desk, along with 1 or 2 early standalone features that generated word of mouth. But 陈冕 acknowledges they were not killer features; once built, the whole industry followed quickly. His view is that engineering innovation has roughly a one-week shelf life, while even a model advantage can disappear within a month.
The product brief was therefore simple: launch quickly, make the experience good, and create at least one shareable point of word of mouth. LiblibTV used the early AI-video buzz to position itself as “the first video canvas that humans and AI could use together”—true, but “not that remarkable” as a lasting advantage.
Once PMF had been validated by someone else, the company did not need to prove demand again. It needed to find a moment when attention was available and amplify a thin differentiator. 陈冕 concedes that LiblibTV’s differentiation was thinner than Lovart’s; the market was simply large enough to magnify several small advantages.
14. “Build attention first” was not a faith in paid traffic, but a fight for breakout-period demand
陈冕’s point is that finding PMF is not the finish line. If incremental demand breaks out at a particular moment and the company fails to capture enough of it, its existing installed base will quickly lose relevance. Application companies must build attention and build ability at the same time.
LiblibTV’s key window lasted about a month, with roughly $1M spent on brand and content. He calls it a “saturation information attack,” but it was not simply buying performance ads. Without a new-model experience or differentiated content, even dense exposure would not generate strong returns.
After the investment, the product may have gone from zero to roughly $100,000 in daily revenue very quickly. 陈冕’s first reaction was, “This market is that big?” Paid users reached several hundred thousand, while roughly 90% of acquisition came from channels other than direct performance marketing; performance ads generated only about 4% of revenue.
His judgment on the competitor is that it may have got the product shape right but was not aggressive enough in pursuing attention, continuing to run a rapidly expanding market at its old operating speed. Share captured during the window, rather than first-launch status, created LiblibTV’s initial first-mover advantage.
15. LiblibTV is now the revenue engine, and the ceiling for video has moved to $100B
The host asked about product mix using the company’s roughly $300M annualized-revenue figure. 陈冕 declined to disclose the breakdown but confirmed that LiblibTV contributes more than half—at least the $150M-plus implied by the host. Lovart’s revenue is also “not small.”
LiblibTV launched in March and broke out rapidly. Its one-day peak reached $1M within the first month, but 陈冕 emphasized that a peak is not an average daily run rate. He now only says it is “growing rapidly,” withholding more detailed figures to avoid revealing the competitive situation.
He estimates that global annual creative revenue generated by video models could ultimately reach $100B. Traditional Adobe is already a tool business measured in the tens of billions of dollars, while AI video could also replace more production costs, including cameras and actors—hence his rough 10x multiplier.
No single company will capture the entire market. Beyond general-purpose tools, there will be a large number of vertical products. He expects pure production tools to settle into a structure of 2 or 3 players in China and 2 or 3 overseas, with the remaining value distributed across different video applications.
16. “ByteDance’s largest agent” is not entirely wrong, but the dilution accusation touches a long-term interest
陈冕 does not avoid the agency characterization: “As an application layer, everything we do is leverage on models; of course we sell models.” The value of an application is the product, users, scenarios, and service it layers on top of the model.
Lovart may eventually stop putting the model name front and center, functioning more like a general interface that selects models for the user. The host believes video is still far from that point and described Seedance as “miles ahead.” 陈冕 then addressed the user’s concern about dilution.
He categorically denies silently routing users who choose Seedance to a cheaper model: “We could never do that,” because it would damage long-term trust, and the company is not operating at negative gross margin. If the product eventually becomes explicitly model-agnostic and automatically routes after the platform’s own evaluations, that would be a different product promise.
The fact that AI video has largely moved away from freemium also improves the cost structure. High-value models opened for free would be farmed to death by abuse networks. Requiring payment upfront removes the cost of free trials and gives pricing more elasticity than early image tools had.
17. The management objective is not free exploration, but sustained delivery from AI-native talent
On specific allegations in reports, 陈冕 says the company has not gone through 7 CTOs; only 1 CTO actually left the company. The historical number of technology heads who reported directly to him and later left for various reasons is 4. He also denies using educational accounts or a diluted Claude 3.2 API.
His starting point is that survival for an AI application company is brutally difficult. So the claim that “management has become unprecedentedly unimportant” needs to be completed as follows: management is relatively less important, but organization and culture still matter. Management exists to keep high-talent-density, AI-native employees who meet the business’s requirements on the platform.
Model companies can recruit the best researchers and give them freedom to investigate. Applications companies must screen people in a chaotic market and coordinate around the one answer that matters at a given stage. They are not well-funded labs searching for the next technical breakthrough; they are fast bridges between changing technology and user demand.
Lovart has found a relatively correct product paradigm. The team’s first task is to make it deeper, more complete, and more durable while staying sensitive to the next technology shift. Free exploration cannot replace reliable delivery, or the product window will close first.
18. 陈冕 is good at removing people who do not fit, but overlooked how they could move forward afterward
He believes one thing the company did well was quickly and honestly judging whether someone fit a role at the current stage, then changing personnel when necessary. Without that results-oriented approach, the company might already have stopped growing.
But “that isn’t management.” What was missing was another reward model: someone removed from a role may not lack ability; they may simply be a temporary mismatch. The company should recognize the effort, show respect, and help find the next position instead of saying, “Whatever happened to that person is no longer our problem.”
Roughly 4 to 5 members of the original founding team remain. More than 10 people have left, about half the group. 陈冕 has not reviewed average tenure and acknowledges that attrition is not low, but says, “A CEO whose goal is very low attrition is not the CEO I aspire to be.”
Another mistake was careless hiring. Early in a startup, options are limited, and people are sometimes simply “kicked into” roles. When the fit proves wrong, the organizational and emotional cost of replacement is actually higher.
19. Ten “co-founders” exposed distorted promise language, not an equity-design problem
Early on, 陈冕 brought a group of former colleagues and friends into the company. Influenced by ByteDance’s weak-title culture, he believed everyone was equal internally and contribution would determine the outcome. Externally, he casually allowed everyone to call themselves “co-founders,” at one point possibly creating 10 co-founders.
Their equity stakes were far below 陈冕’s; in substance, they were day-one employees. The host pointed out that 陈冕 might not take the title seriously while the other person would interpret it as a promise. 陈冕 accepts that this was his own learning and that casually bringing friends together ultimately damaged the relationships.
The company is now establishing clear standards for “partners,” who must withstand long-term contribution and evaluation. 陈冕 says it was only in the past year that people emerged who could genuinely tell him no; more partners may emerge in the future.
He rejects the label of tyrant but acknowledges that his views are extremely strong. If an objection lacks a solid reason, or the relevant risk has already been considered, then in a brutal competitive environment people should “please just go do it.” That also explains why the early team struggled to create effective checks and balances.
20. The company needs a research partner, but cannot confuse an “innovation vibe” with innovation
What 陈冕 most needs now is a more research-oriented partner. He admits that research will not drive an application company in its early phase; otherwise it should simply build a model. But if the goal is to be a technology company, the end state must still be driven jointly by technology and product.
He only recently understood why he insists on building a technology company: “I enjoy creating more than I enjoy operating.” It is not entirely wrong for outsiders to classify him as a businessman, but commercial execution is only the slice currently visible; it does not capture his full motivation.
His definition of innovation includes vision, imagination, resilience, and dependable execution: “Don’t innovate for the vibe of innovation; innovate for the result of innovation.” Freely generating many ideas is not necessarily creation. Making one direction work is.
“Innovation has a cost” is his final answer to the LiblibTV controversy. After inventing the airplane, one can study rockets, but the airplane company still has to operate the airplane business, defend the result, and generate profits. Without that, there is neither a market nor the resources to support the next round of technical exploration.
21. Lovart’s 2,000-plus backlog emails showed that product innovation cannot replace reliable delivery
At Lovart’s most dangerous point, customer service had accumulated more than 2,000 emails. The issues were complex and the support tooling had not kept up. The entire company ultimately spent about 4 hours, with each employee handling several messages, to clear the backlog. That was how close the operation had come to the edge.
The host’s explanation was a mismatch in talent expectations. Some “top-tier horses” may have joined for the technology and innovation vision, only to find that the priority was stability, commercialization, and delivery. 陈冕 says many were not hired by him directly but by the technology lead who built the team; he had delegated heavily for years without checking whether expectations around team-building matched the company.
In August or September last year, 陈冕 had been in the US pushing international expansion when he realized “the backyard was on fire” and returned to China. He first aligned the business plan, then addressed conflicts between product and engineering, and subsequently found problems in delivery quality, capabilities, and the leaders reporting directly to him.
Lovart’s experience was poor for a period, and the team became unstable. 陈冕’s reflection is that because he “didn’t understand technology,” he treated technical details with excessive deference and failed to check whether the macro-level technical judgment was actually aligned. Delegation had turned into loss of focus.
22. GPT Image 1 erased the old path overnight and forced Lovart toward agents
In January last year, after testing GPT Image 1, 陈冕 concluded that the model had internalized controls that previously required a node-based workflow. The intermediate state on which “星流” had depended after nearly a year of exploration was suddenly flattened. He recalls being hit hard enough to cry.
When an investor called to ask for his view, he initially wanted to put a positive spin on it, but ended up saying, “We’re screwed. Give me some time to think about what to do.” The investor reassured him instead. On the flight from Shanghai to Beijing, unable to get online, he reworked the problem; after landing, he called back and said, “I have a way.”
The new answer was context plus an agentic flow: from manually editing generated assets in Photoshop, to building workflows for AI to execute, to giving the user only the context and letting an agent handle the rest. Professional creation still needs an interface different from ChatGPT, giving Lovart a reason to exist independently.
He then “leapt for joy” and went to the head of marketing, demanding that all resources be committed and that the product launch within 1 to 3 months. He acknowledges that he may have been close to “going crazy” at that moment because he believed there would be no opportunity if they were late.
23. Manus made agentic the consensus, while the extreme rush exhausted the first-launch team
After Manus launched, 陈冕 concluded that agentic systems would quickly become an industry consensus and that Manus had already captured the startup sector’s tech brand. That further compressed Lovart’s window. Even though the beta was not yet performing well, the team still had to launch first.
In hindsight, he admits he may have rushed too much. If Lovart had been 1 or even 2 months later, it might still have been the fastest. But there was no way to know competitors’ real progress at the time: “I could only be maximally fast.”
After the high-pressure launch generated strong positive feedback, the company added bonuses and options. Many people nevertheless left later because of exhaustion and organizational problems. Speed produced no one-way victory; first-mover advantage and personnel attrition happened simultaneously.
陈冕 told the team that 100 similar products would appear within a month, so first-month growth would determine the initial advantage. As soon as the product launched, he went straight to the US to advance specific projects and left no recovery period.
24. Vibe coding makes ideas cheaper and judgment the startup’s biggest lever
The technology team had been “doing everything,” including many good ideas. But 陈冕’s conclusion is simple: idea is cheap. Vibe coding makes implementation easier and easier; among a vast universe of ideas, finding the one that is right now is where human value lies.
A creative person can generate many ideas every day. Determining which idea is the best choice for which moment is what judgment means. 陈冕 sees that as the most important lever for a startup—and one reason that greater individual capability does not automatically benefit individual entrepreneurship.
He also admits the company once allowed short- and medium-term goals to drift away from the long term. Employees did not know what the founder truly wanted, or felt only the pressure of data growth. He eventually understood that expressing oneself does not mean others will naturally understand; demanding that others understand you is “a form of laziness and idealism.”
A better hiring process must explain the vision, understand the other person’s motivation, and then judge whether the working style is a fit. Being misunderstood may be fate, but alignment remains a management responsibility.
25. Three product lines proved the ability to capture opportunity—and stretched the organization to its limit
陈冕 believes that if Liblib, Lovart, and LiblibTV had not been built by the same company, the market might instead have produced 3 companies, or some of the opportunities might ultimately have gone to big tech. An individual cannot change the direction of the wave, but can change who captures the demand.
Liblib and Lovart are still growing, but investment in Liblib has been cut substantially. Lovart’s overseas penetration remains low and its runway large, yet domestic product and research problems once interrupted expansion. Once the organizational issues eased in January this year, resources were rapidly pulled into LiblibTV.
When legacy product teams asked whether their products still mattered, the stage-specific answer was simply do the new thing first. 陈冕 has no stable roadmap: “You don’t know when a product is going to arrive, so you can only do the right thing; once it arrives, do the right thing again.”
His split from general-purpose agent companies such as Manus is a choice of vertical, multimodal, and independent survival over a general-purpose strategy. 陈冕 denies ever making a formal acquisition offer; even hypothetically, if someone offered $10B, he says he probably would not sell now.
26. Fighting giants means unifying Jiangdong before competing for the Central Plains
陈冕 admits he has no firm confidence that the company can survive competition with the giants. His initial estimate of success was 10%-30%; after further questioning, he said it might be below 10%. But he does not see entrepreneurship as a wager with fixed odds; it is a search through a dark wilderness for a small path that actually exists.
His Three Kingdoms analogy is that the giants must first “compete for the Central Plains,” where general-purpose models and core battlefields have the highest priority. Startups can use the time gap to “unify Jiangdong.” The giants know Jiangdong is being consolidated and may send troops there; whether the startup can hold depends on the startup itself.
The key sequence is: trade time for space, space for resources, and resources to build the moat. Any moat built overnight is not a real moat. The real defense can only form gradually while the giants have not yet made the vertical market a top priority.
He retains the coldest fallback: even if the company cannot remain independent forever, a sufficiently strong “Yangtze defense” could at least raise the cost of resistance and improve negotiating leverage. Survival does not mean a small company barely staying alive; it means eventually controlling a base large enough to matter.
27. The multimodal interface is only the “Shu Road”; the user network may become the Yangtze
陈冕 sees a “hard road through Shu” between general-purpose text products and multimodal professional creation. Put a complex creative interface into general-purpose chat and the product loses its purity, just as people do not open an iPad every time they chat but still need a computer and canvas for professional expression.
Yet that interface difference is not a thick moat; it is only a temporary geographic advantage. A startup must turn “Bashu” into “Jiangdong” and build its own Yangtze. Until the moat truly works, he is unwilling to claim that he already knows where the Yangtze is.
Ultimately, the company must use user scale to escape the model vendors’ “gravity,” then go further and make users create irreplaceable value for the platform. Only when content, supply, collaboration, or transactions form a two-sided network will model upgrades become less capable of erasing the platform’s value.
Liblib itself is not a two-sided network, and LiblibTV has not yet clearly formed a loop in which users create value that keeps other users on the platform. The current product form may therefore not be the end state; it is only one step on the path toward network effects.
28. The deepest commercial bet is that humans will still choose to create rather than simply be fed by AI
If people no longer produce or create value, and AI centralizes all production, value returns to chips, top talent, and enormous pools of capital, leaving almost no room for new independent startups. 陈冕 is really betting that humans will still have value over the next 10 years, not that people can never be changed by ASI.
He distinguishes between 2 AI paths. One is “creating wealth for humans,” meaning replacing human production. The other is “creating wealth through humans,” meaning enabling human creativity. The latter depends on taste, nonstandard expression, and multimodal content—precisely where application companies may build network effects.
Most people do not want a few giants to consume everything, which itself supports the alternative path. 陈冕 therefore describes the company’s mission as “empower imagination, enrich mind”: helping people with imagination but limited production skills turn their ideas into content.
His ambition remains to build a $100B company, but “winning” is not simply beating others; it is making the world a little richer. He says that if he ever becomes convinced people have no meaning and product managers are no longer needed to mediate between people and technology, “I will sell the company, because I should not be the CEO.”
29. The supercar being repaired while moving explains the speed, heroism, and every goodbye left unspoken
In high school, 陈冕 skipped self-study through computer competitions and gaming. When he failed to secure an early admission to Fudan, only 6 months remained before the gaokao. On a rainy night outside the computer lab, he lay down and cried, then “hacked the reward” by drilling problems repeatedly and eventually entered Southeast University. Games, theater, and debate shaped his instincts for product, content, and expression.
He identifies most with 杨过 in Jin Yong’s fiction: a rebellious youth who becomes the Condor Hero devoted to “the country and its people,” then “stages a great spectacle and quietly leaves.” The host pointed out the strong individual-hero narrative and weak emphasis on long-term organization. 陈冕 accepts that a founder’s strengths carry the company from zero to one, but from one to 10 requires fixing the organizational weaknesses.
Anxiety and speed may have been the strengths selected by the era, but they also exacted a physical price: insomnia, a racing heart, daily herbal medicine for a period, and panic whenever his phone vibrated. He offered no psychological counterweight; the practical relief came when “the company got better, revenue rose, and the money came in.”
His metaphor for 3 years of entrepreneurship is: “You start driving a super-fast car, and repair it while driving.” The people in the car and the scenery outside keep changing, but there is no time to say hello or goodbye; all he can do is grip the wheel. A glance backward may be read as hatred, but the real emotion is reluctance to let go.