AI Applications Haven't Found Their Scenarios, and DeepSeek Didn't Make the Startup Ecosystem Boom Either
Crystal Stream Capital's Wang Mengqiu (王梦秋) on why the capital winter isn't over, how DeepSeek flattened the information gap without creating new scenarios, why the general-purpose Agent may be a false premise, and why the real need is vertical workflows.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The capital winter hasn't actually ended
Wang Mengqiu looks at it from two angles. On money: more capital is willing to write checks this year than in the past two years, but the increment comes mainly from institutions that were previously sitting on the sidelines, and it isn't large relative to the volume of state capital. On investable projects: the primary market is no more prosperous than it was over the past two years — still AI, hard tech, low-altitude economy, those few sectors. Consumption remains very inactive; nobody is willing to burn money on consumer, and what interests people more is companies that are already profitable and have reached a certain scale. So if you called it a winter two years ago, it hasn't fully emerged from that yet.
— Wang MengqiuDeepSeek didn't make the application ecosystem boom
Wang Mengqiu argues the core problem for AI application startups isn't whether the underlying large models are cheap enough — it's that applications haven't found a genuinely exciting scenario, and that isn't something DeepSeek can solve. DeepSeek going open source amounts to information equalisation: it flattens the possible differences between big and small companies in usage. But everyone's biggest problem was compute anyway — doesn't DeepSeek need good compute too? So in essence it hasn't made the application ecosystem boom. The reason no particularly good applications have appeared in China over the past two years is that nobody has found a good application scenario yet.
— Wang MengqiuAI hasn't created new scenarios or new demand
Wang Mengqiu says everyone feels this generation of AI is a technological revolution, but what it can currently satisfy and deliver has not created new scenarios or new demand. Take Agents: the industry started working on them in 2023, beginning with the auto GPT concept, and both China and the US have been exploring for two years, yet they've never taken off. Even for the most vertical case — travel-planning Agents, which plenty of companies in both countries are building, trying to automatically plan itineraries and complete hotel and flight bookings — in actual use the difference from what a human produces is still quite large. People keep revising their needs along the way: they first want to stay somewhere, search and find it's too expensive, then switch location or even switch destination. So in satisfying existing demand, AI in many domains hasn't achieved a leap in experience, let alone created new demand.
— Wang MengqiuDemand analysis matters more than AI technology
Wang Mengqiu thinks the most important thing isn't AI technology itself but demand analysis: which needs are currently poorly served and could achieve a leap in experience through this generation of AI. It can't just be a minor improvement — a minor improvement isn't enough to support a truly independent business. It has to be a leap in experience: saving me a lot of time, giving me results that surprise me. Just as AI talking bots were surprising at first, and DeepSeek too — never mind performance or compute savings, the answers it gives are still far from accurate and its hallucinations are very serious, but the way it talks to you makes you feel it's more human than previous bots, and Chinese people especially are more willing to talk to it.
— Wang MengqiuUsing DeepSeek as a fortune teller
Wang Mengqiu says she uses DeepSeek as a fortune teller — when she needs emotional value or is unsure about something, she goes and asks it. You know it's talking nonsense, but it makes you feel very comfortable. She speculates this is because reinforcement learning was used in the post-training stage: the more feedback you give it, the more it learns. For example, if you read its answers carefully or spend a long time reading them, it may take that as good feedback; or you just say outright, "you're right" or "you're talking nonsense." She thinks some Yuanbao instances also have a lot of emotional value, but use them too much and you feel it's too accommodating, too aware of what you want — a bit of a simp.
— Wang MengqiuWeChat Yuanbao is expected to be a person
Wang Mengqiu uses WeChat Yuanbao to do background checks on people, asking whether it knows a certain person at a certain company and how it evaluates them. Her original intent was for it to dig up information about that person scattered across forums and Xiaohongshu. She admits the demand is unreasonable, but for some reason she feels it's a person, so she can have that demand of it. In the end it found nothing, and she thinks that's explainable: a lot of forums don't let AI bots scrape, and a bot that follows protocol shouldn't scrape either; and whether scraped content is used for training also raises authenticity questions. But from the user-contact level she feels it falls far below expectations, and from that she wonders: when Yuanbao presents itself as a WeChat friend, the things users want it to satisfy should be different from what Yuanbao as an app should do.
— Wang MengqiuThe general-purpose Agent may be a false premise
Wang Mengqiu says Tencent isn't currently committed to building a general-purpose Agent, and she even thinks the general-purpose Agent may be a false premise — whether there's really a split between general and vertical is also uncertain. Crystal Stream has been looking at Agent startups from last year through this year: some do general-purpose, others turn a particular workflow — video editing, say — into an Agent to satisfy it, and the latter is quite a real need. She finds this a bit like the old distinction between large models and the applications above them: some teams want to build an Agent framework so that more application-layer and workflow Agents can be built on top of it, and that layer is like the large model as foundation back then. She has always felt a big question mark hangs over whether the general-purpose Agent holds up — at least there's no general-purpose Agent on the market right now, or none that's done well.
— Wang MengqiuSmall organisations favour young founders
Wang Mengqiu says that perhaps, at this moment, post-00s founders will have a relatively higher probability of success than post-90s founders did in the past, because organisations now tend toward small organisations. The biggest problem with founders who were seen as too young used to be management ability and life experience: building a business takes not only a good product idea and good technical ability, but even more the ability to manage an organisation and judge the whole industry, and that requires life experience. But in the previous mobile network era, the number of people you ended up managing was large and the complexity was there; in this generation's AI application layer, a small organisation may be able to build a big business, so post-00s without that much life experience have a certain probability of making it. But she stresses you don't have to chase post-00s to invest in — dividing founders by age is a very reductive thing to do.
— Wang MengqiuEmbodied intelligence most lacks simulators
Wang Mengqiu says Crystal Stream has invested in embodied intelligence — in Chen Jianyu's (陈建宇) Robot Era (星动纪元) — and will recently try to invest in a company that makes simulators. Because for embodied intelligence, the most damnable problem is that there isn't that much data in the real world: you have to simulate well enough in a virtual environment, then transfer the algorithm to the real world, and also find ways to bridge the sim2real gap. So simulators are the infrastructure of the embodied intelligence industry, but this is a very, very hard thing to do. Crystal Stream is following Robot Era to look for opportunities along the industry chain, just as it invested in new energy vehicles back then and found industry-chain opportunities along the OEMs — but it's still a bit early now, because nobody has entered mass production yet, and supply-chain opportunities haven't shown up yet.
— Wang MengqiuGlasses replacing phones is unnatural
Wang Mengqiu bought Meta's Ray-Ban glasses herself and plays with them when travelling. The reasons she's willing to wear them: first, they look fairly good and fashionable; second, they look no different from ordinary glasses. She doesn't like devices that look distinctive — VR headsets are unsustainable, not wearable day to day, a violation of human nature. The glasses combine sunglasses and earphones, with bone-conduction earphones, and she really dislikes earphones stuck in her ears. But she won't wear them often either, since she isn't short-sighted. She thinks the people making AI glasses now are still very shallow; the core is to figure out what problem it solves and when a person needs to wear the thing. She doesn't believe glasses can replace phones, because it's unnatural — having something growing on your face all the time is uncomfortable, whereas a phone you hold when it's heavy and put down when you're done.
— Wang MengqiuTravel is looking inward
Wang Mengqiu went out roughly every month last year — Saudi Arabia, Germany, Iceland, Japan. She says going to Saudi Arabia felt like a country embracing modernisation very actively, very young, thriving; the local tour guides didn't rely on it for money at all, even paying out of their own pockets for coffee and ice cream. Germany is an already very well done country; going to Germany is seeking inner peace. Iceland she visited during the polar night, with only three or four hours of daylight a day, a very end-of-the-world feeling. She says travelling now is different from in her twenties: first, she has more money than in her twenties; second, she no longer wishes she could hit ten-odd places in a day — it's more about looking inward, the changing external environment projected onto the inner self.
— Wang MengqiuGreatness cannot be planned
Wang Mengqiu read the biographies of Jensen Huang, Sam Altman and Musk, and made an internal PPT titled "Greatness cannot be planned" — when you truly build a company that great, you actually can't foresee or plan it in advance, but it can be learned. She tried to summarise whether the three have anything in common, and in the end found not much — but Jensen Huang and Musk share one human trait: they didn't try to be good people, and in daily life, especially working with their teams, they are not easy people to get along with. Her conclusion: if you want to become a great entrepreneur, you probably shouldn't try to be a good person; there must be very harsh situations and very harsh relationships, and then you have to behave harsh.
— Wang MengqiuFilm must first answer why go to the cinema
Wang Mengqiu says, why do today's moviegoers still go to the movies? There are too many ways to seek emotional value, and going to the cinema is an extremely high-threshold activity. People used to watch films because there weren't more options; now scrolling short dramas and short videos costs almost nothing and is more comfortable — a couple can sit face to face in a café scrolling short dramas. Going into a cinema used to feel like a social occasion, but really, sitting in the dark watching a film, who's talking to whom? There's nothing social about it. So the film industry first has to answer one question: why should people still go to the cinema to watch a movie. She thinks many people in the industry, especially producers, haven't realised this — they haven't thought about answering why the audience should go, what genre they want to see, what audience to target.
— Wang MengqiuHer Story rides the feminist narrative
Wang Mengqiu thinks Her Story (好东西) succeeded because it rides the current feminist narrative — it is a purely feminist narrative, and different from the previous sister film B for Busy (爱情神话), which was more about doubling the various possibilities of refined middle-class Shanghai life and finally telling a myth of love. She heard Her Story did well at the box office in lower-tier cities, which is a very good breakthrough, because B for Busy was more in first-tier cities, whereas Her Story let women in second- and third-tier cities see that women in first-tier cities can live like that. She likes the film and has seen it twice, but finds some of the lines a bit too slogan-like, and that all the men in it are portrayed as very weak — she has good, strong men around her too.
— Wang MengqiuShe hates March 8th Women's Day
Wang Mengqiu says she has always hated International Women's Day on March 8th, because "I think this thing is really because you're still too weak, so people specially make a holiday for you." So she never celebrates it, hates the holiday especially, and it's even been split from Girls' Day — Tsinghua has a March 7th Goddess Day. She says she has never encountered anything particularly unfair to women in her own life, whether studying, working at Baidu or doing investment — no discrimination, no condescension, no mansplaining. She thinks on the whole it should be equal rights: men and women both seen as independent people, and it's enough that behaviour is reasonable; don't demand things just because someone is a woman, don't make everything a big deal — that, in fact, isn't true equality.
— Wang MengqiuIn their own words · checked verbatim
What it can currently satisfy and deliver has not created new scenarios or new demand. Everyone feels AI is already very strong, but if you look at it statically, what today's large models can deliver still hasn't created new scenarios or new demand.
它现在能够满足和实现的东西 并没有能够创造新的场景和新的需求 就它现在大家都觉得AI已经很强了 但是实际上你静止的去看的话 现在的AI大模型能够实现的东西 仍然没有能够创造新的场景和新的需求
Wang Mengqiu5:05
Some Yuanbao instances have a lot of emotional value too, but I think use them too much and you feel it's too accommodating, too aware of what you want — a bit of a simp, right?
有些元宝也非常的有情绪价值 但是我觉得用多了 就会觉得他太顺着你了 他太知道你想要什么 就有点舔狗是吧
Wang Mengqiu13:08
So I've always felt a big question mark hangs over whether the general-purpose Agent holds up. And at least right now there's no general-purpose Agent on the market, or none that's done well.
所以我一直觉得通用agent 这件事情是不是成立 我其实是打一句大的问号的 对 然后 至少目前市场上也没有什么通用agent 或者做的好的
Wang Mengqiu26:11
The reason I've always hated International Women's Day is this: I think this thing is really because you're still too weak, so people specially make a holiday for you. So I never celebrate it, I especially hate this holiday.
我自己一直很讨厌三八妇女节的原因 就是这个 我觉得这件事情其实是因为你 还太弱 所以人家专门要为你笑一个节日 对 所以我从来不过这个节 我特别讨厌这个节日
Wang Mengqiu1:23:46
Figures
| Number of projects Crystal Stream Capital invested in last year | 2 | 1:01:35 |
| Wang Mengqiu's age at marriage | 23 | 1:26:47 |
| When Wang Mengqiu and her husband started dating | Started dating in sophomore year, dated for four years | 1:26:47 |
| Hours of daylight per day during Iceland's polar night | Three to four hours | 1:00:34 |
| Reading progress on Ten Days Ending (“十日终焉”) | Read past 100-plus chapters, of 900-plus chapters in total | 1:09:39 |
Glossary
- sim2real
- Transferring algorithms trained in a virtual environment to the real world, and bridging the gap between the two.
- general purpose agent
- An AI Agent that can autonomously complete many kinds of tasks across scenarios; Wang Mengqiu doubts whether it holds up.
- reductive
- Reducing something complex to a single dimension, such as dividing founders by age.
How to listen
Investors and founders watching AI application-layer startups and the temperature of the primary market, and anyone who wants Wang Mengqiu's real judgments on Agents, embodied intelligence and wearables.
The books, film and feminism chat after 1:03:55 — low information density, skippable.