103. Lovart Founder 陈冕 Looks Back on Two Years of Building an App: This Moment Feels So Damn Good!! Hahahahahaha

Summary

• The core investment thesis: general-purpose models will not subsume verticals. 陈冕’s bet is that vertical data, interfaces, and use cases differ too widely; general-purpose agent entry points will converge around a few winner-take-all players, while vertical agents will proliferate, both called by general-purpose systems and accessed directly by users with their own mindshare. The implicit assumption is that “AGI will not arrive within five years—if AGI arrives tomorrow, quit; we’ll all be working for AI.”
• The full numerical chain from despair to reversal unfolded within a year. In 2023, the first round raised $4M at an $18M valuation (“we priced it too low; we should have asked for $50M”); a subsidy war burned through roughly $3M; the product was taken offline in September for lacking a large-model filing; cash at one point fell to RMB4K, and “nobody wanted it even at a $30M valuation.” In 2024, 2–3 months after the license returned, the company closed three rounds totaling over $20M. By early 2025, ARR was close to, and even above, $10M; the current valuation is “probably a few hundred million dollars.”
• Claude 3.5 unlocked Lovart’s product. The original plan was to have designers build workflow nodes on a canvas, but “designers don’t speak logic; designers speak in feelings.” In December 2024, after discovering the model’s stronger Agentic capabilities plus MCP, the team shifted to having AI plan workflows for designers, causing the barrier to entry to collapse; it took only 3–4 months from idea to execution. Manis’s lesson was that “people only remember the first”: leave general-purpose to Manis; “the true first vertical was Cursor—we want to build the Cursor of creation.”
• The viral breakout proved that timing determines customer-acquisition cost. A post published at midnight drew 800K views overnight, Elon Musk liked it, and the waitlist reached 120K. Customer acquisition for the first-generation product cost just RMB3–4 per person; today it is around RMB20. “The core of product-market fit comes from timing, and timing comes from breakthrough innovation”—miss the window and paid acquisition becomes a bottomless pit; “you can never beat the big platforms.”
• His decade in mobile internet became a library of failure cases. Mobike’s 40–50M rides were sustained by unlimited supply and ignoring supply costs; once it had to break even, “it might do only a few million orders a day,” making it “maybe acceptable as a cost center in a payment war, but unacceptable as a business in its own right.” Daily Fresh showed that “internet plus retail is a fake proposition.” The conclusion: “business model determines almost everything”; since then, he has only pursued pure-online, high-gross-margin businesses with scale effects.
• The business model will move from subscriptions to paying by deliverable. In the short term, professional designers will use subscriptions; in the long term, if production relations change and clients use AI directly, “AI will calculate the tokens and electricity it expects to consume, then quote you a price.” To B “absolutely will not happen in the short term” because it is not the organization’s strength. The end state is “1–2 relative winner-take-most players,” while major ecosystems will still retain their own products, as Adobe coexists with Final Cut and 剪映.
• AI has already deconstructed the organization. Product managers “no longer exist”; the core jobs are teaching AI industry knowledge and coding. His advice to founders: build vertical products, not general-purpose ones; “don’t believe in product managers”; prepare before model intelligence improves—and “the most important thing about getting out there is getting out there.”

Deep dive

1. A Decade of Job-Hopping Became a Library of Failed Business Models

• 陈冕 was born in 1992 and, “apart from Alibaba, worked at every one of BAT and TMD”: an internship at Tencent building QQ Zone, 360 Mobile Assistant, Didi, Mobike as head of product, Meituan, Daily Fresh, and ByteDance. Mobike taught him the deepest lesson: the scale created by unlimited supply and ignoring supply costs was an illusion. To break even, the company would have had to raise prices and remove bikes; “this business might do only a few million orders a day,” neither making money nor generating traffic. If the first business does not make money, where does the company get the cash for its second and third growth curves?
• Daily Fresh drove the point home: “internet plus retail is a fake proposition.” In self-operated retail, “money is scraped together one coin at a time,” while the efficiency gains from the internet are extremely limited—“if you remove the algorithms of an internet company, and remove all of us, its gross margin would be fine.” The conclusion: “business model really matters; it determines almost everything.”
• His version of Murphy’s law is worth preserving: “The thing you initially think is abnormal is abnormal; the problem you think will happen will definitely happen.” During the highs, “everyone mainly watches the mood”; once inside, he found that the business was not growing and he had no room to develop personally.

2. ByteDance Taught Him Not How to Avoid Mistakes, but How to Get Things Right

• Earlier in his career, he accumulated the ability to “know how to avoid mistakes, but not how to do things correctly.” ByteDance was the only place where he saw a correct decision-making process: first use common sense to judge the end state—how many players will ultimately survive, whether the market will be oligopolistic or fragmented, and whether it will be supply- or demand-driven. But “seeing the end state mostly helps you avoid dead ends; it does not solve how to do every single thing in front of you.”
• The other half was extreme pragmatism. Senior executives such as 张一鸣 “would discuss extremely specific, microscopic issues with you”; the quality of every small decision affected the result. After taking 瓜瓜龙 from zero to one and making it a hit, 陈冕 became ByteDance’s youngest-ever promotee at 28 (4-1), then moved to 剪映 and CapCut’s global commercialization after the education crackdown.
• His honest ten-year retrospective: “Without AI, my biggest regret would be not having started a company earlier…but with AI, I don’t regret my past choices.” Seeing more business models succeed and fail “gave me a better understanding of my strengths and weaknesses”: he is better suited to online creative businesses that require imagination and empathy.

3. ChatGPT Was His Redemption: “Hope Is the Antidote to All Pain”

• He saw ChatGPT, Stable Diffusion, and Midjourney at the end of 2022 and decided to start a company almost immediately, “without any hesitation.” That night he posted to WeChat Moments: “Hope is the antidote to all pain, and the meaning of all pain.” His decade of job-hopping while employers found fault with him “suddenly had meaning in that moment.”
• His assessment of the wave: AI is not mobile internet, which was merely an extension of the scenarios enabled by computers and the internet. It is on the level of “the invention of the computer and the information revolution”: intelligence is to AI what information was to the internet, and “we should be at 1980 now.” Its changes cannot be extrapolated from the mobile-internet playbook.

4. The Initial Positioning: Avoid the Model Mainline and Cut into Multimodal Creation on the 1980s PC Playbook

• He started with self-awareness: “Building a foundation model requires a top scientist; I never considered it.” In 2023, he was early enough to build an application but already late to build a model. His PC-history analogy: Office and Adobe came first as productivity tools, search followed, and social and general entertainment did not arrive until after the 1990s. ChatGPT is essentially Office plus search and “very close to the core capabilities of large models,” so he cut sideways into multimodal creative verticals.
• With neither a model nor traffic, the capability stack had to come from differentiated data and a differentiated interface. The first-generation product concentrated “human creativity”—templates and assets, the parts AI could not replace. The second-generation Lovart’s canvas and chat box “look more like the interface you use to communicate with a designer, rather than the interface two ordinary people use to communicate.”
• His ontological framework for AI: “AI is the new camera and the new video camera, and also the new canvas.” It simultaneously overturns the means of creation and the production relations of creation: diffusion models replace cameras, while AI replaces some designers. “Both will change; that is a much bigger change.”

5. The 2023 Fight for Survival: Subsidies, Delisting, and Just RMB4K in Cash

• The first $4M round went smoothly, but by June 2023 there were already “10 people copying us,” and the subsidy war began: subsidies for creators, subsidies for users, and free GPUs. Around $3M was burned. In September, the product was delisted for failing to complete the large-model filing: “The good news is that all our competitors were delisted too; the bad news is that we had no money left.” By June, the product had already surpassed 1M MAU and was No. 1 in China for image generation.
• The team was cut immediately from 50–60 people to 30, and half of the 20-person founding core team left. Those who stayed earned RMB10K a month; some took no salary. Monthly labor costs were pushed down to RMB100K-plus. When cash reached just RMB4K in March 2024, a rescue strategic investment arrived; the license came through in April. His confidence came from “never feeling the company would fail—I knew why we had reached that point, and what we had done right and wrong.” The worst-case plan was to personally fund two months of servers while waiting for the license.
• Should they have fought the subsidy war? “Of course we had to. How else could we win? Without it, we would have been eliminated.” When 小珺 pressed him on Kimi’s paid traffic, he reversed the question: “The point is whether Kimi would have been able to build a better model without buying traffic.” Once Doubao entered the main battlefield, not buying traffic was even less viable; when a team cannot build SOTA, “users are still your moat.”

6. Fundraising Frustration and the Lesson: Timing Is Everything

• From September 2023 through January of the following year, he met 3–4 groups of investors every day and no one invested: “We had such good data, so many users, and such a strong market position, and nobody wanted us even at a $30M valuation…Investors kept asking question after question, then said they would not invest because we were offline. Isn’t my time time? Don’t I get tired of talking?”
• He took away two lessons. First, on the first startup, he had “zero understanding” of cash-flow management. Second, he had no independent judgment on fundraising cadence: “If you don’t raise when you should, or look for money when you shouldn’t, nobody can help. You have to seize a good window, extend the gains, and prepare deeply enough; otherwise you cannot withstand the difficulties of bad times.”
• The reversal came in 2024. After the license and revenue returned, the company closed 3 rounds totaling over $20M in roughly 2–3 months. He deliberately kept them separate so the money would arrive in batches as quickly as possible—the direct lesson of the prior year. By early 2025, ARR was close to, and even above, $10M; the current valuation is “probably a few hundred million dollars,” perhaps a Series B, and the company is “definitely not profitable.”

7. Lovart’s Birth: Claude 3.5 Unlocked “AI Planning Workflows for Designers”

• The limitations of the first-generation product were clear: only the most intensely competitive designers—those competing to learn AI—would use it. The barrier was too high to push penetration toward the majority. The second generation had to “massively increase productivity while keeping operation extremely simple”; the end state could even be clients using it directly. That meant a change in production relations themselves.
• The key correction was conceptual. He initially wanted to build workflow nodes on a canvas, but “designers don’t speak logic; designers speak in feelings,” and the project stalled. Around December 2024, he found that Claude 3.5’s Agentic capability had strengthened, alongside MCP and other tools: “AI could help designers plan workflows, and the barrier collapsed.” The form had already been imagined: “The canvas is the table, the person is the chat box, and with a toolbox and editor, the form emerged.” It took roughly 3–4 months from idea to execution, including taking only 2 days off over the May holiday and working consecutive all-nighters.
• Manis provided the urgency: “Build the Agent now, and move fast—the first and second may not differ much in experience or technology, but people only remember the first.” Let someone else take the general-purpose first; “I’m happy to be the first in a vertical.” And “the true first vertical made money quietly long ago: Cursor—we want to build the Cursor of creation.”

8. The Economics of Going Viral Overnight: Timing Determines Customer-Acquisition Cost

• He posted at midnight Beijing time. By 3 a.m. it had 100K views; “I slept, woke up, and it was at 800K.” Elon Musk liked it, and the Chinese tech circle exploded the next day. The waitlist still has 120K people; nearly 200K have requested codes, but only tens of thousands have received them. One important reason for limiting invites is that commercialization features were not ready: “If we release more, costs will explode.” Retention is significantly better than on the already mature first-generation product.
• The acquisition ledger tells the story: RMB3–4 per user for the first-generation product, around RMB20 today. “Product and performance in one—the core of the product comes from timing, and timing comes from breakthrough innovation and the moment you enter.” Without that, paid acquisition becomes prohibitively expensive and “you can never beat the big platforms.” Subsidies have a clear boundary: they work only in China; “subsidies in North America are useless.”
• Poverty produced its own anecdote. The team registered only lovart.ai and lovart.art; once the product took off, every other suffix had been taken, and “Google is full of pirated copies.” “Buying 10 domains would cost RMB10K. We were so scarred by being poor that we saved the money.”

9. The End State of the Ecosystem: General-Purpose Entrances Converge, Verticals Flourish, and Deliverables Get the Bill

• In the pre-AGI era, general-purpose agents are the entry point, like search and app stores. “General-purpose will inevitably be winner-take-all—if you have a distinct mindshare, you are no longer general-purpose,” so the fight will converge around a few players. Verticals will multiply, both called by general-purpose systems and visited directly by deep users. The end state in this category is “1–2 relative winner-take-most players,” while major ecosystems retain their own products: Adobe is the biggest, but Apple’s ecosystem still uses Final Cut, and Douyin creators still use 剪映.
• The business-model split depends on who the end user is. Professional workers produce continuously and fit subscriptions; clients “buy the result of one project and leave,” so they should pay according to the type and value of the deliverable: “AI calculates the tokens and electricity it will consume, then quotes you a price.”
• His closing bet, and the assumption nested inside it: “There will not be one general-purpose model that eats all verticals,” because vertical data, interfaces, and use cases are different. Behind it sits the belief that “AGI will not arrive within five years—if it arrives tomorrow, quit.” He is “actually quite uneasy” about AI comprehensively surpassing humans: “That would be something like a new species surpassing us. Would humans still be the leaders of this planet?”

10. Organizational Deconstruction and a Fighter CEO: “This Moment Feels So Damn Good”

• The organization has been rebuilt for the AI era. Product managers “no longer exist; they have been deconstructed.” The core jobs are teaching AI industry knowledge and coding; front-end, back-end, and testing have been dramatically simplified by AI-assisted coding. The team is close to 100 people, mostly operations staff from the first-generation product. Lovart itself has very few people and is mainly based in San Francisco, with a night-owl rhythm that has them “leaving the office at 3 a.m.”
• He described the best moment of entrepreneurship twice: “The future you predicted actually happened, and you actually made it happen…The longer you grope around in the dark, the brighter the light looks.” Neither generation was marketed; both relied on judgment. The first reached PMF, and the second at least showed a PMF trajectory. “You can’t buy this for any amount of money.” ByteDance’s title “means nothing to me”; no conditions could make him go back. Anxiety does not need to be solved: “Use a high level of anxiety to stay sensitive.”
• The biggest disappointment was also people: after meeting so many institutions, no one came to the rescue, and half the founding team left. “But if I were in their place, I might have made the same choice; that makes the people who stayed extraordinarily precious.” His self-image is 杨过 from Jin Yong: “Make a huge scene, then leave quietly.” His parting advice to founders: build verticals, teach AI industry knowledge, embrace AI completely, and prepare before model intelligence improves—“the most important thing about getting out there is getting out there.”