AI Application Startup: General Agents Can't Eat Verticals, and Year One Nearly Killed Us
Chen Mian's post-mortem: the first-generation product won the subsidy war but got taken down, with only 4,000 yuan left in the bank; the opportunity for vertical Agents lies in general models being unable to digest industry data and interaction — provided AGI doesn't arrive within five years.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The peak is when the fight is at its fiercest; afterwards it's a mess
Chen Mian went through two fights, at Didi and Mobike, and found that the fighting itself was the high point of the business model, and that after the fighting was over the model itself turned out not to be that good. Mobike looked like a huge business of forty to fifty million rides, but that came from unlimited supply and unlimited supply costs: users weren't charged, the bikes didn't make money, and the losses were a disaster. His conclusion: if supply is limited and supply has a cost, it's impossible to sustain that kind of scale; to break even, users will inevitably get more expensive and bikes will inevitably decrease, and the business might do just a few million rides a day, with a low ticket price and low gross margin — neither profitable nor particularly useful. So you still have to respect the nature of the business and common sense.
— Chen MianBurned two million US dollars in three months, then got taken down
From June to September 2023, Chen Mian went all in on the subsidy war — subsidising creators, subsidising users, poaching people — and burned over two million US dollars in three months. The first-generation product became number one in China in the image generation category, with over one million monthly active users. But in September it was taken down. The good news was that all the competitors were taken down too; the bad news was that the money was gone, with only a few million yuan left in the bank, and the team was cut from fifty or sixty people to thirty. His post-mortem: the strategy was right, the business was right, and they won the competition — the mistake was that as a first-time founder he had no experience managing cash flow and no judgement about funding cadence: he didn't raise when he should have, and when he shouldn't have, there was no point talking to anyone.
— Chen MianWith only 4,000 yuan left in the bank, he paid out of his own pocket to last two more months
After being taken down, fundraising was extremely hard. Chen Mian met three or four groups of investors a day, talking for over an hour each, speaking for five or six hours a day, and after seeing every institution, nobody invested. At his lowest point the team was threatening to leave every day, and investors kept asking piles of questions, finally saying ‘you're not listed, so we won't invest’. He calculated that the company had only 4,000 yuan left in the bank, but by then the team was down to just over ten or twenty people, each taking only 10,000 yuan, some taking nothing, so the monthly payroll cost was just over 100,000 yuan, and revenue was nearly at break even. He said that if that firm didn't invest, he would put in two months of his own money to hold on a bit longer, until the licence came through in April. He never felt the company would fail, because he knew why it had come to this, and he knew what he had done wrong and what he had done right.
— Chen MianCustomer acquisition cost of three or four yuan; paid traffic is just a pipe
Chen Mian says acquiring a user now is enormously expensive, but the first-generation product's customer acquisition cost was very low, only three or four yuan, because at that time Chinese users needed this thing, and they were the best at it. He stresses that the timing window matters: when Kimi first ran ads, there were barely any AI advertisers on Bilibili, so of course it was cheap; later, when Doubao, Yuanbao and others flooded in, it got expensive. Paid traffic is a pipe, a stimulus, not the driving force of customer acquisition — once people know you, the cost of running traffic to them is low; if people don't know you, pushing ads hard is expensive. So the core of combining brand and performance is the product's timing, and timing comes from innovation and the moment of entry.
— Chen MianCanvas plus dialogue box, waiting for models' agentic capabilities to strengthen
Lovart's pre-research had two parts: the canvas was prepared very early, starting in the second half of 2024 when the company came back to life, because they judged that the future intermediate form would be a canvas. But agent capabilities take time. At first they wanted to do workflow, letting users build nodes on the Canvas, only to find that designers don't talk logic, they talk feeling, and building nodes simply didn't work. Then around December they found that Cloud 3.5 worked very well, the model's agentic capabilities had strengthened, and it occurred to them that AI could help designers plan workflows, greatly lowering the barrier and rapidly improving the experience. So the form took shape: the canvas is the table, the dialogue box is the person, the toolbox is the tools, plus an editor. From idea to execution took about three or four months.
— Chen MianRacing to be the world's first vertical Agent; being first in mindshare matters
Chen Mian believes the payoff of being ‘the world's first XX Agent’ is a wave of attention, first-mover traffic and a brand impression. Doing PMF in North America, the hardest part is really 0 to 1; after that it's a contest of product strength, but how product strength reaches the public eye matters a great deal. In the 0 to 1 stage, attention is limited and product competition is fierce, so you need some way to step in front of the public. Manus being the first to do a general Agent was meaningful: it proved that the technology had given products the conditions to be realised. The first and the second may not differ much in experience and technology, but people only remember the first. Lovart is the first in the vertical category; the real first in verticals was Cursor, and they want to be the Cursor of design, the Cursor of creation.
— Chen MianProduct managers are useless; you only need people who teach AI industry knowledge
Chen Mian says that in the know-how of AI application startups, product managers really weren't much use; industry data and industry understanding are what matter. In terms of organisational form, the most core people now are those who teach AI industry knowledge and those who can code: if someone who can code can also teach AI industry knowledge, then you only need people who can code; if everyone can code, then you only need people who teach AI industry knowledge. The old product manager role has been deconstructed, and the R&D team has been simplified too — no need for frontend, backend or testing, only R&D people who can use AI coding. The team is now close to 100 people, but the second-generation product has very few people, mainly in San Francisco.
— Chen MianCoexisting with anxiety; no time to be anxious, just do it
Chen Mian says he is a combative person, and when he's excited he's really excited. Asked how he regulates his mindset, he says there's no time to be anxious, you just do it; coexisting with anxiety can't be solved, and he doesn't want to solve it — without that strong anxiety, you wouldn't be that sensitive, wouldn't be able to judge many things. You need a high degree of anxiety to stay sensitive, and thereby stay sharp about business judgement and business plans, and hungry for execution speed. The happiest moment is when the things you imagined would happen in the future actually happen, and you actually pull them off — for example, when Lovart's product form got very good feedback, he scrolled through every post and every piece of user feedback on Xiaohongshu, and concluded that this form should be right.
— Chen MianIn their own words · checked verbatim
The things you initially thought were abnormal — they are abnormal. The problems you think will happen — when something is out of the ordinary, there's something wrong. Yes, the problems you think will happen will definitely happen. Murphy's Law.
你最开始觉得不正常的事情 它就是不正常 你觉得会发生的问题 事出反常有妖 对 你觉得会发生的问题 它就一定会发生 莫非定律
Chen Mian7:18
The good news is that all the competitors were taken down. The bad news is that the money was gone, all burned up, only a few million yuan left.
好消息是 所有的竞争对手都被下架了 坏消息是没钱了 全烧完了 只有几百万人民币了
Chen Mian39:03
We're willing to pay a lot of pain for it, but the longer you grope in the dark, the brighter the light you see.
我们愿意为它付出很多的痛苦 但是你在黑暗中摸索的越久 你看到的这个光就越亮
Chen Mian1:26:03
Innovation in an unfamiliar field is like striking a match against wet wood, again and again, lighting and going out, until one day you catch a crack, set the wood alight, let the fire spread through the whole cave, and set the mountains and plains ablaze.
在陌生领域的创新 就像用火柴在潮湿的木头上 反复的滑动 点燃又熄灭 直到有一天 你抓住了某一个缝隙 把木材点燃 让火势弥漫到整个山洞 燃起了漫山遍野的大火
Chen Mian1:32:35
Figures
| Subsidy war burn | Over two million US dollars burned in three months, about three million US dollars in total | 37:43 |
| Cash remaining in the company's bank account | 4,000 yuan | 44:09 |
| First-generation product customer acquisition cost | Three or four yuan | 45:17 |
| Total funding in 2024 | Over twenty million US dollars | 50:39 |
| ARR in early 2025 | Close to ten million US dollars, over ten million US dollars | 51:39 |
| Lovart waitlist queue | 120,000 people | 1:01:47 |
| Reads after Lovart's post | 800,000 after a night's sleep | 1:03:45 |
| Team size | Close to 100 people | 1:18:58 |
Glossary
- PMF / product-market fit
- A product finding a market that genuinely wants to keep using it; the hardest 0 to 1 stage of early startup life.
- agentic
- A model's ability to plan autonomously, call tools and complete multi-step tasks.
- workflow
- A way of executing tasks by manually breaking them into nodes and connecting them into a process.
- break even
- The threshold at which revenue covers costs and the business stops losing money.
How to listen
Founders, product managers and investors building AI applications or considering starting a company, especially those focused on vertical Agents, going global and early-stage funding cadence.
From 1:14:55 to 1:16:57, the section on North American team integration and the competitive landscape is fairly general and can be skipped.