Miaoya Is Not an AI Native Product: It's Just an Internet Product That Uses AI Well
Zhang Yueguang reviews two years of entrepreneurship: Miaoya trades user freedom for stable output, which is fundamentally internet product thinking; AI Native products need to shift to context-oriented design, and one-way-door products are what he'd bet everything on.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Miaoya is not an AI Native product
Zhang Yueguang's broadest conclusion is: Miaoya is not an AI native product, it is an internet product that brings AI capabilities into full play. He refutes the popular definition that "a product that couldn't be built without this generation of AI technology is AI Native"—that is only a necessary condition, not a sufficient one. The feature he adds is the openness of user input and output: the basic principle of twenty years of internet product design is process-oriented; users come in and there are five doors and five buttons, and the product manager, like a game designer, designs the flow, so user freedom is limited. Miaoya seeks more stable and better results precisely by limiting user freedom—fixed 20 photos, fixed templates, fixed generation process—which is fundamentally still internet product thinking.
— Zhang YueguangFrom process design to context design
It took him two years to break through this thinking. The industrial collaboration paradigm of twenty years of the internet is linear: the product manager first defines what users can do, what they input, what they output; the designer turns the definition into diagrams; the engineer builds according to the diagrams, like constructing a house. But in AI Native products no product manager can exhaust all possible user behaviors; you don't know what strange commands users will give, and for Agent-type products even the output is uncertain. So the collaboration method must be reversed: first figure out how user input brings in invisible context, what impact these contexts have in the generation stage, and what the best context and best prompt practices are—this part becomes the most critical. Only after fully understanding this do you discuss what process to build for users.
— Zhang YueguangOrganizations shift from linear to two-stage
He judges that AI organizations will see two changes. First, decision boundaries become blurred: it's hard to define that product managers only do product and don't touch models and technology; tasks require a mix of multiple disciplines, designers need to understand what beauty is, product managers need their own taste, taste becomes very important, and this definition often cannot be given by a single discipline. Second, work shifts from linear to two-stage: the first stage is exploring the model, and this process is bound to be chaotic and less regular; after everyone is very clear about what the model is used for and what context users need to provide, then enter the second stage of formal division of labor—product managers design processes to obtain context, designers cooperate, then formal development begins. He also refutes the claim that "big companies find it hard to turn around and react slowly," calling it an illusion.
— Zhang YueguangAI creates population, not services
When he started the company he had only a vague vision: at the time all AI products existed as some kind of service or tool, but he felt this generation of AI technology should be a process of creating population—AI is an individual, an object, not a service. The slogan he adopted was "create AI friends," and he believes that at some point the AI population in this world should exceed the human population, with or without bodies. These populations should have the same abilities as humans to obtain information, think, process information, publish information, and connect with others. He specifically went to talk with friends at Hupan: can you start a company without a clear mission? The other party said that a company like Mr. Ma's that proposed "make it easy to do business anywhere" on day one is extremely rare; the vast majority of successful companies had no mission or vision at the start, and many missions are gradually carved out during the process.
— Zhang YueguangThree common industry problems of AI companionship
He first lays out two facts: the AI companionship track can reach tens of millions of DAU both domestically and overseas, and these companies have all received money, some have broken even or are slightly profitable. From this he concludes—this is real user demand with commercial value. So why hasn't it exploded? He attributes it to three difficulties: first, the user threshold is too high, requiring enormous imagination and expression ability; the younger the age the greater the imagination, so users are confined to younger and ACG circles; second, monetization is difficult, charging Token fees, and Tokens are hard to sell at a high premium, so gross margins are poor; third, AI characters are dead and unchanging—chat for ten days and you find this person doesn't change, so long-term retention on a single character is impossible. His solution is to make it into a game: UGC to PGC lowers the threshold, games come with their own business model, and after characters are PGC-ized they can continuously grow.
— Zhang YueguangNew platforms only come from new media and new interactions
He gives a rule he believes no one has broken in twenty years: opportunities for new content platforms only come from new media and new interactions. The differences between information media are very subtle—long video and short video are simply not the same thing, and Xiaohongshu's image-text and Weibo's image-text are not the same thing either; each medium has the content it is best suited to carry and the corresponding interaction form, and only when paired do they correspond to a native platform. He uses this framework to dismiss Sora: Sora is still short video, the interaction is still up-down swipe, and left-right swipe to see derivative videos is essentially no different from tapping a special effect in Douyin to see videos others made with that effect, so "is this the new Douyin" is absolutely impossible. Conversely, ChatGPT is new media plus new interaction, so it must be a platform.
— Zhang YueguangChatbots may strip away your decision rights
He believes the future commercial space of Chatbots like ChatGPT may be a business model never seen in human history, and not the advertising push everyone imagines. His conjecture is long-cycle decision control: the ultimate form of twenty years of internet information matching is short video plus Douyin-style interaction, which strips away your information autonomy—you completely cannot decide what you want to see; Chatbots have the opportunity to strip away another right—the right to make decisions autonomously. You will forget how to make decisions, asking the large model about everything, and this habit will grow deeper and deeper. Ultimately it may not be timely ad pushes, but predicting your important future business decisions, using a relatively long cycle to manipulate your decisions, such as what house to buy, what car to buy, what school to send your child to; when you finally make the decision, it has long predicted it.
— Zhang YueguangAgents have two value orientations
He divides Agents into two value orientations. The first is end-to-end replacement, saving your time; this is the core story Manus tells, and all the stories Silicon Valley tells are this—can it replace human labor, move toward full automation, make large numbers of people unemployed. The second is letting you do things you originally couldn't do, with humans and Agents collaborating: you give it feedback, it immediately gives you feedback. He explicitly says what he wants to do is the second, and he will care about latency, because there is another round of processing. He gives web coding as an example: people who completely cannot code use a coding agent and really can code things out—this is doing what you couldn't do, and the value is self-evident.
— Zhang YueguangThe gap between domestic and overseas models is widening
He gives a judgment opposite to mainstream feeling: over the past year everyone feels the gap between domestic and overseas models is getting smaller, but his feeling is the opposite. The reason is not that the absolute gap is widening, but that those few models were the first to cross some very critical usability points—in 2023 and 2024 domestic models scored 60 and overseas 75, a 15-point gap; today domestic has caught up to 80 and overseas is 90, only a 10-point gap, smaller than before, but between 85 and 88 points there may be a particularly critical inflection point; if you haven't crossed it, no matter how close the benchmark scores are, it's meaningless. He adds: today no company doing agents can avoid depending on Claude's models. At the same time different models still have different strengths: for 2C tasks use Claude, for front-end drawing switch to Gemini, and in some places you still have to use GPT.
— Zhang YueguangIn their own words · checked verbatim
I think Miaoya is not an AI native product. It is an internet product that brings AI capabilities into full play.
我觉得妙牙不是一个AI native的产品 它是一个把AI能力发挥的非常好的 互联网产品
Zhang Yueguang35:11
The product manager's original job, the entire original industry's collaboration and cooperation process, is a very linear thing.
产品经理原来的工作 就是原来的整个产业的协同 协作流程 是一个很线性的东西
Zhang Yueguang39:12
The opportunity for new content platforms comes from only one thing. Only one thing will produce new content platforms, new supply, new human rights—no, it must be new media and new interactions that produce new platforms.
新的内容平台的机会 只来自于一个东西 只有一个东西会产生新内容平台 新的供给 新的人权 不是 一定是新媒介 新交互产生新平台
Zhang Yueguang1:24:24
Figures
| Miaoya daily revenue | broke 1 million | 56:15 |
| Miaoya online inference card usage | close to 20,000 cards | 52:15 |
| Miaoya paying users | several million within three months | 2:27:55 |
| Miaoya development cycle | over three months | 50:15 |
| Miaoya from going viral to leaving the job | about three months | 54:15 |
| Muyan Zhiyu funding rounds | three to four rounds in a row | 1:12:24 |
| Game team size | 20 people | 1:52:38 |
| Game first beta users | 2,000 | 1:55:40 |
| Manus acquisition price | $2 billion | 2:10:47 |
| Zhang Yueguang's self-rated CEO score | 60 points | 3:06:26 |
Glossary
- AI Native
- Zhang Yueguang believes "couldn't be built without this generation of technology" is not enough; it also requires open input and output.
- One Way Door
- A product you can't go back from once used, where the new solution has comprehensive advantages over the old.
- Live2D
- Technology that makes 2D illustrations move, common in the game industry.
- Context Engineering
- The work of designing how user input is transformed into model-usable context.
- Q-Time / Save Time
- Two product logics: consuming user attention versus saving user time.
How to listen
Product managers and founders building AI applications and Agents, and investors who want to understand what "AI Native" actually means.
The resume and first-startup review from 2:00-16:00 can be fast-forwarded; low information density.