Zhang Yueguang’s First Interview: Miaoya Isn’t AI Native
Summary
Zhang Yueguang has ended nearly two years of low-cost exploration and will no longer pursue other directions on the application side, formally committing resources to Doki: first validating PMF with AI PPT, ultimately building “a friend that helps me break through the boundaries of my capabilities.” He accepts that competition on this path is extremely crowded, and admits PPT itself is not worth bankrupting the company over; but if every dollar is lost on an “ability-boundary Agent,” he says, “I can accept that,” because even if someone else ultimately succeeds, the goal will still have been achieved.
His most important correction in understanding AI products was admitting that Miaoya is not AI Native, but “an internet product that made exceptionally good use of AI capabilities.” Miaoya traded away user freedom for stable results by constraining the number of photos, templates, prompts and generation pipeline; a truly AI Native product faces open inputs and open outputs, requiring product design to shift “from being process-oriented to being context-oriented,” while organizations must first explore model capabilities jointly before returning to conventional specialization.
Zhang Yueguang believes the cost advantage only gets released suddenly after an AI product crosses the threshold of being genuinely usable, which is why “90 points equals zero points,” and being first matters less than being first across that line. Miaoya split “real, recognizable and beautiful” across 3 serial models, reaching daily revenue in the million-yuan range and nearly 20,000 inference cards online in just over 3 months; Doki’s one-way-door standard is likewise that users no longer need to rework the output in PowerPoint and others cannot tell it was made by AI, rather than producing a polished but undeliverable demo.
Doki and Manus represent 2 different views of Agent value: the former emphasizes high-frequency human-AI collaboration, low latency and capability expansion, while the latter represents end-to-end execution, offline parallelism and labor replacement. Zhang Yueguang believes the automation path has greater commercial upside, yet deliberately chooses a limited “humanism” value proposition—helping people do things they previously could not; he also says the model companies behind Claude, Gemini and GPT are a long-term threat, while application companies’ temporary room comes from task depth, hybrid models and engineering details model companies do not want to handle.
The AI otome game Starlink is another, slower and less crowded bet, attempting to solve 3 chronic problems in AI companionship: a high user barrier, poor monetization and characters that never develop. Its first closed beta opened to only about 2,000 people and filled in a day; the official Xiaohongshu account gained about 10,000 followers that day, and when the service closed 7 days later users gathered to say goodbye to the male lead. Zhang Yueguang therefore believes the basic hypothesis that “AI companionship can be made into a game” is probably correct, while emphasizing that content games cannot simply use AI to cut costs and still require a long replication cycle of at least 1 year.
His underlying framework for allocating startup resources is to look for a one-way door rather than a short-term wow effect: once users adopt a solution, they should not return to the old one, allowing market share to accumulate day by day. This explains why he rejected experiments such as short-form audio platforms and proactive information aggregation—the former never found content native to voice interaction, while the latter had mature substitutes and tried to save users time in an industry monetized through attention, creating a conflict between demand and business model.
The real change in this startup is not only its direction but also its motivation: he is no longer trying to prove himself, but wants to spend 10 years with the same group of people making something worthwhile even if someone else ultimately completes it. He described media reports of “4 rounds of blind bets without a product, totaling nearly RMB3B” as broadly close but unconfirmed in detail; the financing did not lead to aggressive spending, and he gave himself a score of 60 for his performance as CEO. What he is satisfied with is that he “didn’t mess around” and did not materially damage the company’s cash or the team’s energy; the cost was the anxiety of not understanding the direction for a long time.
Deep dive
1. More Than a Decade of Changing Questions Eventually Forced a Long-Term Commitment
Zhang Yueguang joined Kongming Technology around 2012, then worked on Alipay Wireless, ByteDance’s image and growth projects, his first startup, Alibaba and Youku, before joining Miaoya in 2023. He has touched almost everything from funding, red envelopes, photo albums and social networking to short video, long video and AI image generation.
He does not consider himself good at “climbing upward in a big company,” but his progress was hardly slow: when he joined ByteDance in 2016 he was roughly P3; by around age 30 he had reached P9 at Alibaba and managed nearly 300 people at his peak. What truly dissatisfied him was that after more than a decade, “neither the people nor the work had developed a long, durable slope.”
His product preferences were not clear when he graduated. Early on, he simply did whatever his boss asked, taking the work seriously and seeking promotions and raises. It was only after ByteDance and his first startup that he settled on “the joy of creating” and moving real users as his deeper motivation.
2. His First Startup Correctly Read Short Video but Chose the Wrong Entry Point
When he left ByteDance in 2018, Zhang Yueguang judged that short video was not merely a content platform but an upgrade in information media following text and images. Education, e-commerce and other industries would all become short-video- and livestream-driven. He still believes that macro judgment was broadly right.
The mistake was the entry point: the team wanted to “build another Douyin for selling products,” but short-video commerce naturally sits on the extension of existing short-video platforms. In retrospect, the more rational path would have been to build content supply around the platform rather than recreate the platform itself.
After roughly 1 year, the project had to pivot. Zhang Yueguang describes it as a classic first-time founder mistake: seeing a sufficiently large trend but failing to translate it into a product and commercial position with structural advantages.
3. Yuanyin Was the First Time Abstract Users Became Living People
In 2019, he was inspired by the Japanese game Good Morning, My Boy. Its Live2D robot could only recover memories along fixed branches, and he immediately thought, “It would be great if this thing could hold a free conversation.” The team later built Yuanyin, a Live2D product offering avatar customization, clothing changes and voice-chat social interaction.
Yuanyin quickly reached several tens of thousands of DAUs, peaking near 100,000, with users mainly women and middle-school students. It monetized through skins and outfits. The low ARPU of younger users and the complications around minors’ payments were important practical reasons the product was eventually shut down.
Unlike the red-envelope project at ByteDance, which added 5M users in a day but left him looking only at the numbers, Yuanyin users discussed the characters and product every day in QQ groups. On the night the service closed, users held a collective “vigil” until midnight, then wrote long memorial posts in app stores. That feedback convinced him he wanted to build “a product that genuinely moves users.”
Starlink was not a sudden startup idea but a continuation of that attachment. The free-conversation Live2D character he imagined in 2019 only became feasible once model capabilities matured in this cycle.
4. His First Startup Failed First on Motivation
Zhang Yueguang stopped his first startup partly because he was exhausted and partly because, after more than 2 years, he suddenly no longer knew “why I was starting a company.” He later admitted that the most honest original motive was not a mission but “I think I’m really good, and I want to prove myself.”
The result quickly “slapped you in the face directly: you’re not that good.” Once the goal of proving himself was rejected, he had no other thing he genuinely wanted to bring into existence. The confusion therefore came not from the failure of a particular product, but from the failure of the original motivation itself.
To fulfill his responsibilities to shareholders and the team, the company was ultimately taken back by Alibaba, and he returned there as well. Zhang Yueguang sees that as handling his responsibilities, not as proof that the original startup logic remained valid.
5. He Chased AI Early but Voluntarily Left Pure Technology Competition
When the visual-AI startup wave began in 2016, Zhang Yueguang wanted to find a company working on consumer-facing CV, so he joined ByteDance’s graphics and imaging group and worked on the Google Photos-like Time Album and image effects. The group was later absorbed into AI Lab, after which he shifted to projects such as red-envelope growth and the mini-program platform.
His experience in Tsinghua’s computer science department instead led him to give up coding early. In one text-clustering assignment, his K-means code crashed after running for a day, while his roommate finished the job in about 20 minutes. That was when he concluded, “My talent isn’t here.”
He later described himself as “half rational, half emotional.” He is better at judging the paradigms beneath visible phenomena and is willing to examine subtle product differences, rather than competing head-on with top students on pure rationality or engineering talent.
6. One Night by the Kamo River Turned 10 Years into a Non-Negotiable Principle
During the 2023 National Day holiday, Zhang Yueguang sat beside Kyoto’s Kamo River from 2 a.m. to 5 a.m. with only a can of beer, reconsidering the coming period from age 35 to 45. What dissatisfied him was not his position or income, but that he had never spent a long period doing the same thing with the same group of people.
The principle he set that night was: “From 35 to 45, do the same thing with the same group of people.” Businesses, colleagues and organizational arrangements in large companies are subject to forces beyond one’s control, and embracing change is a basic capability. If he wanted to control the object of his next 10 years, he could only organize it himself.
Once the principle was established, a P10 promotion, a raise or a bonus could no longer retain him, because none of those variables satisfied the 10-year requirement. The power of a principle, in his formulation, lies in deleting many seemingly negotiable options in advance.
7. Big-Tech Promotion Has External Battles and Inevitable Internal Battles
Zhang Yueguang divides the big-company game into 2 types. One is defeating external enemies through clear campaigns and business results; the other is defeating people around you, where remaining in the top 3 on the team can also produce continuous promotion.
His view is that internal competition is not a matter of individual character but the “gravity” created by the performance system. The external game is harder, less certain and dependent on periods of incremental growth. Without incremental growth and new factors, people in an organization naturally turn inward and compete with one another.
From 2012 to 2016, it felt like “doing push-ups in an elevator going up,” with many businesses able to generate external results. After Douyin, the internet experienced a relatively long vacuum, and AI once again provided an incremental battlefield. What he truly liked was not promotion itself, but external creation.
8. Big Companies Can Innovate, but Innovation Must Justify the Organizational Cost
After returning to Alibaba, he insisted on initiating a new product almost every year. Miaoya was simply the only one to break out at scale. Zhang Yueguang rejects the claim that big companies cannot create, provided a new cycle opens a sufficiently high ceiling.
Yuanyin peaked near 100,000 DAUs but was not worth allocating long-term big-company resources to because of its young users, low ARPU and unhealthy commercial boundaries. AI raised the potential ceiling again, creating room for both internal big-company innovation and startups.
His core reason for leaving Alibaba was therefore not that innovation was impossible there, but that the organization could not guarantee “the same group of people doing the same thing for 10 years.” Creative space and life principles were separate questions.
9. A Good Startup Motivation Must First Have a High Expected Payoff
The first reasonable motivation Zhang Yueguang offers is, “I just want to be the boss; I don’t want to work for someone else.” It sounds ungrand, but it is achieved on the first day the company is founded, so its subjective expected payoff is already 100%.
The second is, “I really want to make something happen.” If you succeed yourself, that is success; even if you fail and someone else eventually achieves it, you have not lost, because you participated in a proposition larger than your personal win or loss.
Getting rich, becoming famous and proving yourself are not morally inferior in his view, but they are unattractive after multiplying probability by payoff. Staying at a big company may make it easier to earn money, and becoming an executive can bring respect; putting those goals into a startup adds unnecessary exposure to low odds.
10. Youku Confirmed That Long Video Is More Like an Investment Business Than an Internet Platform
Zhang Yueguang believes the decisive variable in long video is content: a well-chosen film performs well, while a bad investment is difficult to rescue through product or technology. In the extreme case, a hit movie brings traffic to a cinema; the cinema did not “distribute” the hit movie into existence.
Adding a recommendation algorithm to the long-video homepage could therefore even be negative. Users often open the app for a specific piece of content. The business is more like using flour to make bread: early losses came from excessive content costs and insufficient payment, and the business later approached breakeven as content prices fell and payment habits improved.
By 2023, he believed that the products, technology and design under his responsibility could no longer create material change. At a biweekly meeting, he announced that he would stop attending Youku’s routine meetings and let the team make its own decisions, shifting his time toward AI and image innovation.
The person in charge at the time, Fan Luyuan, continued to emphasize “content is king” while giving product and technology teams room to operate. Zhang Yueguang admits that when nobody had produced material results, performance evaluations were more like process rewards than precise allocations of real business increment.
11. Miaoya Started from 3 Common-Sense Judgments, Not the Large-Model Hype
ChatGPT was already extremely hot in 2023, but after inventorying the available technology, the internal interest group concluded that Diffusion-based Model technology and its surrounding ecosystem were relatively mature, so it chose images.
The team had no cards, money or people, and could not explain why it should rebuild Midjourney, so it needed a sufficiently vertical entry point. Drawing on his past experience with photo albums, Zhang Yueguang noted that more than roughly 70% of users’ photos were portraits, with screenshots a distant second. The team therefore focused on realistic portraiture.
The Alibaba team’s first instinct was Taobao outfit changes, model images and e-commerce visuals. He judged that the e-commerce chain was too long: if image generation was not the core link, the team would have no pricing power. Portrait photography was ultimately chosen because it could target consumers, be priced end to end, and replace an expensive, slow, asset-heavy traditional solution.
12. “Real, Recognizable and Beautiful” Meant 3 Serial Models, Not a Slogan
Miaoya’s 3 requirements were sequential: “real,” meaning users could not tell it was AI-generated; “recognizable,” meaning users had to confirm that it was about them; and “beautiful,” meaning it had to look slightly better than the real person, because nobody wants to pay to see an unretouched version of themselves.
At the time, a single model could hardly meet all 3 requirements simultaneously. The team split the task: one model learned from large numbers of ordinary people’s photos, preserving real-world texture such as blur and ordinary lighting; another used roughly 20 photos uploaded by the user to preserve identity; a third handled templates and aesthetics.
The 3 models were serial controls, not a classic MoE. The separation also made the template ecosystem scalable: each template could be treated as an independent small model without repeatedly modifying the underlying identity and realism capabilities.
Overall development took about 3 months, with most of the time spent tuning the output. Engineering and launch were fast. Zhang Yueguang attributed the result to “subtle differences,” not functional originality—B612, Meitu and others already had similar features and workflows, but none had crossed the threshold of being real, recognizable and beautiful at the same time.
13. A One-Way Door Means Users Do Not Return to the Old Solution
Zhang Yueguang prefers one-way-door products: once users experience them, they know the need will be solved this way from then on. Cursor and Claude Code replacing traditional IDEs are his most direct examples.
Innovation usually does not create a need from nothing; it gives an existing need a new solution with comprehensive advantages. Even if the current result is not yet clearly superior, it is worth sustained investment if its path toward comprehensive superiority is obvious.
The difference between a one-way door and a “wow” product is that after the novelty fades, users of the latter return to the old workflow and must be repeatedly reactivated. A one-way door, even with a tiny initial share, will “rise day by day,” allowing a founder to build patiently.
14. Miaoya Needed AI but Was Still Not AI Native
It took Zhang Yueguang 2 years to accept a counterintuitive conclusion: Miaoya could not have existed without this generation of AI, but that was only a necessary condition for being AI Native, not a sufficient one. A more accurate definition was “an internet product that made exceptionally good use of the AI technology available at the time.”
Miaoya required a fixed number of photos to train a fixed identity model. Users then chose from fixed templates, with the underlying prompts, models and engineering pipeline all preconfigured. It compressed user freedom in exchange for more stable results under the technology available at the time.
A truly AI Native product has more open inputs and outputs. Users can give instructions in language that are difficult to enumerate in advance, while Agents may call tools and produce multiple types of results. The internet product manager used to play God; now no one can draw every possible path in advance.
15. Product Design’s Core Is Moving from Process to Context
The basic paradigm of the internet’s first 20 years was process-oriented: product managers defined decision trees, designers turned flows into interfaces, and engineers implemented them like construction workers. WeChat may resemble a freely navigable Disneyland, but underneath it still offers the “limited freedom” predesigned by product managers.
With Agents, product managers cannot first assume what users will enter or what models will necessarily output. The first priority should be decomposition: how many interactions will occur between user and model, what information the model needs to complete the task, and which pieces of context determine whether the result meets expectations.
The key action is not abstractly to “guide users,” but to find ways to get users to provide the necessary information. Only by understanding what the model lacks can the team design the process backward and progressively obtain the required context.
A chat box, GUI and generative GUI are merely surface choices. What truly determines the result is how invisible prompts, knowledge, tools, model calls and context work together. The conclusion is therefore: “Product design must move from being process-oriented to being context-oriented.”
16. AI Native Teams Must Explore Together Before Returning to Specialization
When model capabilities have not yet been mapped, the boundaries between product, design and engineering become blurred. Designers must judge what is beautiful, product managers must contribute taste, and engineers must participate in defining containers, layouts and model behavior. No single function can independently provide the answer.
Zhang Yueguang proposes 2-stage collaboration. The first is a relatively messy hybrid team jointly exploring what the model can do and what context it needs. The second begins once the answer is clear: product designs the process for obtaining context, while design and engineering move into formal delivery.
He rejects the idea that startup opportunities can be reduced to “big companies turn too slowly.” Anyone who has worked in a big company knows that is fantasy. The real opportunity comes from organizational species differences: linear sprint management seeks clarity, while model exploration is inherently impossible to fully regularize.
17. Miaoya Was Really Competing with Haima Body, Not Meitu
Zhang Yueguang has always understood Miaoya as a portrait-photography business rather than a general image tool: “I was never sitting at the same table as Meitu. I was sitting at Haima Body’s table.” Meitu solves high-frequency retouching, while Miaoya completes a one-time portrait session; the two do not naturally generalize to one another in users’ minds.
At launch, he personally wrote the user letter, named the product and reviewed every line of copy, deliberately avoiding any mention of AI because the goal was to build a new consumer brand. Its collaborations with Ctrip on travel photography and with Tongdao Dad on astrology followed the logic of brand recognition rather than tool expansion.
This also determined the business ceiling: portrait photography is low-frequency, high-margin and operations-heavy, and cannot become a high-frequency, low-ARPU product through technology alone. Even if he continued, he would call Miaoya a “fine business,” not the main route of the AI era.
18. From RMB9.9 to RMB99, Then Offline, Was the Full Portrait-Photography Playbook
The first thing Zhang Yueguang wanted to do before leaving was raise the price from RMB9.9 to RMB99. The answer was not a simple price increase but adding a human in the loop to customize templates and styles. As long as users knew someone was providing a service, RMB99 remained reasonable against RMB699 offline.
The second step was to enter shopping malls through pop-up stores and intercept users already preparing to visit portrait studios. Whatever style offline brands were promoting, Miaoya could quickly launch a similar one with lower labor, venue and upfront costs, compressing their cost structure.
Because portrait photography is low-frequency, the final contest comes down to who users think of first when they think about portrait photography. Brand awareness may accommodate only a few winners. He would not choose this direction for an independent startup, however, because winning would require a heavy business and the margins would not be particularly attractive.
19. Miaoya’s Breakout Was Huge, at the Cost of 3 Straight Months of Firefighting
The product went viral in only 1 or 2 days, reaching daily revenue in the million-yuan range. The group mobilized computing resources nationwide, with online inference using nearly 20,000 cards at peak. Zhang Yueguang and the team’s PM watched daily revenue break RMB1M for the first time at a bar in Liangmaqiao and drank themselves senseless.
He later disclosed that Miaoya acquired several million paying users in roughly 3 months, possibly already exceeding Haima Body’s annual customer volume. That also showed why the breakout was unsustainable: the number of people in China willing to pay for portrait photography is limited, and a large share of existing demand had been covered in a short period.
Almost all post-launch work consisted of adding cards, fixing bugs and filling engineering gaps. One backend engineer worked for roughly 72 consecutive hours and refused to go home, saying, “I may never encounter something like this a second time in my life. I don’t want to leave myself any regrets.”
The mini-program was initially a result of having a small team, but iOS virtual-payment restrictions forced the team to build a client in 3 days. It also faced suspected organized smear campaigns, including fabricated claims that users’ photos were used for pornography and complaints to consumer regulators. For nearly 3 months, there was barely any room to think.
20. He Left 3 Months After the Breakout Because of Principle, Not Product Possessiveness
Miaoya became popular around July. Zhang Yueguang raised his resignation after the October holiday and formally finished at Alibaba on November 11. The separation was handled “normally”: the other side only required him not to build Miaoya again, and he never considered creating another Miaoya after leaving.
He rejects treating a product as one’s personal child. Miaoya was a team product, and he did not complete many of the key tasks; his colleagues were capable enough to continue advancing it. If the product later struggled, his staying might not have saved it; if it succeeded, that would obviously be better.
Miaoya was the product that best proved himself in his career, but he does not consider it particularly successful in itself, viewing it only as a “fine business,” not the main route of the AI era. ByteDance’s huge data numbers came more from the platform, while Yuanyin was closer to a personal zero-to-one effort but lacked sufficient results. Miaoya gave him both influence and a moment of happiness he would remember for life.
21. The New Company Started with “Creating an AI Population,” but the Roadmap Was Unclear
After leaving Alibaba, Zhang Yueguang did not immediately raise money and start building. He spent several months deciding whether he really wanted to start a company and also considered joining someone else’s team. He studied Silicon Valley and several large-model companies, but their To B orientation and technical DNA did not fit his preference for building consumer products.
The initial vision was to “create AI friends.” AI should not merely be a service or tool but should become an individual. In a future virtual world, AI people could outnumber humans, acquire, think about and publish information, and connect people with other information.
He now believes ChatGPT may have become that individual to some extent, but at the time of founding he had only a grand hypothesis and no clear roadmap. What he was certain of was that AI would eventually enter the human social network rather than remain a private, isolated assistant for each person.
22. A Startup Does Not Need a Perfect Mission First, but It Must Set 3 Basic Variables
A friend with a Hupan background told him that formulating a long-term mission such as “make it easier to do business everywhere” and having a clear roadmap on day one is extremely rare. Many successful companies gradually define their mission through execution.
The 3 more necessary elements are industry, team culture and business model. A company cannot sell milk tea today and switch to AI tomorrow. Culture is essentially what kind of person the founder really is. If the business model has no answer at all, the risk is extremely high.
When Zhang Yueguang started this company, he had already decided to work on AI for the long term and knew he wanted people with aligned values. The business model was unclear, but the industry already had some evidence of payment, so he was willing to explore while building rather than wait for a fictional complete answer.
“The same group of people for 10 years” is a principle, not the purpose of the startup. The purpose must still be to bring something into existence; the principle only determines which people and what method will carry the process through.
23. “People Own Will, AI Owns Skill” Rewrites Team Composition
Zhang Yueguang compresses the human-machine relationship into one sentence: “People should own will—intention and determination. AI should own skill.” As specialized skills gradually depreciate, human breadth, diversity and taste become the competitive advantages of both individuals and teams.
Roughly half the team comes from large companies and half from independent developers and “high-taste” startups. Strong big-company employees bring process discipline, mission completion and decisive execution. Independent developers often handle product, design and development themselves, making them better suited to context exploration. People from high-taste companies bring aesthetic, product and marketing judgment.
He deliberately avoided making the entire core team former colleagues, reducing convergence of thought and preventing newcomers from facing a closed circle of familiar people. Values should align, but working styles should include different roles: push-oriented, detail-oriented and reflective.
The company has no strict co-founder titles and barely emphasizes titles at all. Core members include both long-term collaborators and new partners introduced by friends. The team has remained broadly stable over the past 2 years, with most of the few departures choosing to start their own companies.
24. The Founder’s Strength Is Paradigm Judgment; the Risk Is Over-Abstraction
Zhang Yueguang believes his strongest ability is judging underlying forces from visible phenomena: what is driving a change, whether it can produce a paradigm shift, and which differences will persist over the long term. He does not claim the ability is absolutely exceptional, only that it stands out relative to his other capabilities.
Another self-assessment is realism: he is willing to admit, “If I don’t understand it, I don’t understand it.” But paradigm abstraction has a downside. Abstracting black-and-white cows directly into “animals” causes meaningful differences to disappear, leaving only the verbal tic, “Isn’t this just…”
Miaoya’s lesson is precisely that subtle differences under the same functional description can produce completely different outcomes. After making an abstract judgment, one must return to the specific object, specific user and specific usability threshold.
25. Large Financing Buys Judgment Time First, Not the Right to Burn Money
The media reported that the company raised 3 or 4 rounds without a product, totaling nearly RMB3B. Zhang Yueguang did not confirm the exact figure, saying only that the scale was “broadly not far off” from public reports. He also admitted that he still does not know exactly how the valuation was calculated.
He summarizes the reasons investors bet on him as follows: first, they believed in the long-term vision of AI friends and AI entering the social network; second, his past had demonstrated some product ability; third, they trusted that he would not pretend to understand what he did not and would not recklessly act when the path was unclear.
Outsiders joked that the company was operating on deposit interest. He says it was not quite that extreme, but spending was indeed restrained and the runway remains long. Exploration cannot cost nothing, but its “tuition” should be kept to the minimum.
For him, the scarcer resource was not financing but the energy of the team and himself. Money can be raised again; repeatedly building products without conviction can break people. Not accelerating or selling grand promises during the exploration period was therefore an active capital-allocation decision.
26. AI Companionship Has Real Demand, but the Industry Has Not Yet Broken Out
The company launched an AI otome game on its first day. Part of the motivation came from his attachment to Yuanyin and Live2D, and part came from market evidence: at the time, both domestic and overseas AI-companionship segments had reached tens of millions of DAUs at the category level.
Zhang Yueguang emphasizes that, apart from general-purpose chatbots and possibly image and video generation, there were few AI applications capable of reaching tens of millions of DAUs at the category level. Companionship products also had user payment, with some companies approaching breakeven or modest profits.
The limited conclusion is that AI companionship is a real demand with commercial value. The fact that it has not fully exploded cannot be explained as a “fake demand”; the common bottleneck must be found elsewhere.
27. Companionship Products Are Blocked by 3 Structural Problems
The first is an excessively high user barrier. Users facing a chat box must derive pleasure from text and a small number of images, requiring strong imagination and expressive ability. Younger users are more likely to imagine a mahjong tile as a tank, which is why these products concentrate among younger and anime-oriented users—not because those users are naturally the only audience.
The second is a poor business model. Most products effectively resell tokens, making it difficult to generate a premium, so gross margin is poor.
The third is that characters do not grow. Real partners and friends inevitably change as relationships evolve, but an AI character remains the same person after 10 days of conversation, making long-term retention difficult for a single character. Youth skew, weak monetization and poor retention are only the surface symptoms; the underlying problems are high barriers, a weak model and static characters.
28. Gamification Addresses the 3 Problems of Companionship One by One
Games turn fully user-driven UGC conversation into a combination of PGC and user initiative. When users do not want to express themselves, they can still consume the story, art and setting; when they want to interact, they can chat freely. Live2D feedback also provides a more concrete anchor for imagination than pure text.
The gaming industry already has mature payment structures. An individual game may fail, but every genre has a business model validated over many years, allowing games to use pricing and content-sales methods that ordinary AI companionship cannot naturally adopt.
PGC characters can also be continuously updated. The same person can occupy different stages in different users’ lifecycles, while the model, story and relationship state can evolve, making “the companion grows” a producible content proposition rather than something expected to emerge from the model on its own.
Zhang Yueguang launched application exploration at the same time because he doubted whether a friend could provide only emotional value. Real friends also help you do things; conversely, an object that reliably provides practical help over the long term will likely develop an emotional relationship naturally.
29. The Core of an Otome IP Is Not Chat, but Reweighting Interaction
Zhang Yueguang breaks any IP into 4 elements: values and attitude form the core, while visual packaging, creative expression and user interaction form the outer layer. Hermes, celebrities and streamers belong to different industries but still require these 4 elements to support one another.
Creative expression and user interaction operate like a seesaw. The more expensive and elevated the IP, the lower the frequency and higher the quality of its work, and the more it must limit interaction. A top celebrity cannot livestream every day, and a luxury brand will not constantly operate inside e-commerce channels, or scarcity and persona will be consumed.
On the other end are interaction-first IPs such as streamers and AKB48, whose value comes from being “visible, tangible and available every day.” AI’s variable is that it makes it possible for otome games to systematically create this type of interactive character.
Copying a top celebrity or established otome male lead into an AI that can chat freely is therefore not simple value addition; it may damage the original IP. Interaction-first and prestige-content-first are different products, and traditional otome companies would need to create separate characters and product lines if they enter the market.
30. Starlink’s Single Male Lead, 2D Style and Live2D All Follow the Same Positioning
Interaction-oriented and celebrity-oriented IPs target the same broad group of otome users but appeal in different ways. Some users like both, while others prefer high availability. Zhang Yueguang judges that interactive companionship leans more toward anime-oriented users, while real-world users brought into the market by 3D otome games are more likely to look up to “big stars.”
The team therefore chose a single male lead, a 2D visual style and Live2D, while making the character more service-oriented rather than an aloof domineering CEO. The choice reflected both his attachment to Yuanyin and upfront judgments about interaction frequency, user profile and content cost.
Zhang Yueguang says he chose a female-oriented product from day one and once advised another founder not to build for men because his judgment was that “men do not need companionship.” This is a strong view about the current product opportunity, not a universal conclusion about every future game category.
31. AI Games Cannot Skip the Content Industry with “Cost Reduction and Efficiency”
The Starlink team has about 20 people, mostly game professionals rather than a large number of AI researchers. Model capability does not require simply adding headcount, but content craft—planning, art, copywriting and operations—remains a human labor chain that cannot be erased with a slogan about AI-driven cost reduction.
Otome and anime-oriented users are especially sensitive to AI art and plagiarism. They protect artists and place a high value on originality; even a new character resembling an existing one can be unacceptable. Using AI in the interaction core does not mean every visual asset can be generated by AI.
Zhang Yueguang participates in underlying decisions about the single male lead, 2D style, character direction and AI gameplay. He does not participate in deciding whether the man is handsome, whether the story is moving or whether the art is accurate, because he is not the target user; the team’s planners and content leads understand those questions better.
His definition of an AI Native game is that the user’s core reason for playing is the flexibility AI provides, rather than adding a chat NPC to a traditional game. Real-time video and expansion into the physical world remain out of reach, while systems like “Stanford Town” may also become homogeneous and boring after roughly 100 rounds.
32. The First Closed Beta Proved Emotional Intensity, Not Scale
Starlink’s first closed beta opened to about 2,000 people and filled in a day. The official Xiaohongshu account reached about 10,000 followers that day. Zhang Yueguang described retention and usage duration during the 7-day test as “extremely high,” but did not disclose exact ratios.
Related user topics often reached millions of views, mostly through organic sharing. Zhang Yueguang believes content products inherently carry attitude and emotion, prompting users to spread them voluntarily. The downside is that female users are equally demanding and will criticize the product most harshly when it gets something wrong.
When the service closed, users gathered to say goodbye to the male lead, and backend conversation volume surged. Large numbers of users then posted chat screenshots on Xiaohongshu. This recreated the “living users” he saw when Yuanyin shut down and reinforced his judgment that the hypothesis of solving AI companionship through gamification was probably right.
According to the plan at the time of the interview, the product was expected to launch in the second half of the year. Slow feedback is itself a defense: a large company can copy a tool feature in a week, but it cannot compress 1 to 1.5 years of content production into that timeframe.
33. A New Platform Can Be Born Only from Pairing a New Medium with New Interaction
During exploration, Zhang Yueguang asked why there were many AI tools but few platforms. His answer was that content platforms do not emerge automatically from “new supply” or “new users.” They require a new information medium and native interaction designed for that medium.
Twitter’s short text corresponds to a feed, Instagram’s large images to a single-column stream, and Xiaohongshu’s image-led, text-supported notes to a two-column stream. Long and short video, or Weibo posts and Xiaohongshu notes, may appear similar, but the content they carry and the interaction they enable are different.
He therefore argues that a Sora-like product cannot become “the new Douyin” simply by using AI video. The content remains short video and the interaction remains vertical swiping; viewing derivative videos horizontally is not fundamentally different from tapping into a collection of content using the same effect on Douyin.
ChatGPT did create a real-time generated conversational medium and conversational interaction, so “it must be a platform.” New platforms may still appear in the AI era, but only if someone redefines the medium and interaction rather than swapping AI-generated content into an old short-video format.
34. ChatGPT’s Potential Commercial Power Is Long-Term Influence over Decisions, Not Ads
In response to doubts about “two-sided network effects,” Zhang Yueguang believes ChatGPT already connects users with the information it retrieves, even if it has not formed a conventional interpersonal network. Search and Q&A themselves constitute an information network.
He then proposed a commercial model explicitly labeled as a hypothesis: short video takes away users’ autonomy over information, while chatbots may gradually weaken their autonomy over decisions. As users repeatedly consult a model about buying a home, buying a car or education, the model could anticipate and influence their eventual choices over long periods.
If that control truly forms, monetization may not involve inserting an ad at the moment of decision, but shaping choices across cycles and taking a higher share from merchants. He admits he “cannot fully guess” how the value chain would be divided, but believes control over decisions itself has enormous value.
If Doubao were combined with ByteDance’s information-distribution capabilities, it could theoretically create a loop that “both decides what you see and influences how you think.” This is Zhang Yueguang’s risk scenario, not a factual description of the current state.
35. The Short-Podcast Experiment Found No Native Supply, and the Problem Was Not Just Product Form
The podcast project did not start from a love of podcasts but from a mobile-internet intuition: text and video had both become shorter, so why did audio still often last 1 or 2 hours? He suspected that the traditional GUI was poorly suited to audio when both hands are occupied, and that LLM voice capabilities might create a voice user interface.
The team first put open RSS podcasts into the product simply to gain experience quickly. Zhang Yueguang knew this was not native, just as using Douyin to distribute Youku’s long videos would not be native. What they needed was short-audio supply suitable for “browsing with your mouth.”
They tried minute-long news, stories and novels, as well as cutting long dramas into short segments. But even when a drama is cut into short pieces, it remains continuous long-form content; users do not need frequent voice interaction, and the value of a VUI is not fully realized.
The deeper hypothesis was that audio often accompanies visual tasks such as driving and housework, leaving limited attention for listening. Making content shorter, faster and denser could therefore become uncomfortable. After roughly 3 to 4 months, the team stopped exploring deeply, though he still treats whether a voice-audio platform can exist as an open question.
36. Proactive Information Feeds Lack Both a One-Way Door and a Compatible Attention Business
Another experiment, never publicly launched, used a large model to proactively push subscribed topics. The system first collected content, broke it into atomic information and then recombined it according to user interests, attempting to move mobile internet’s “recommendation replacing search” into a chatbot.
Internal testing found it useful, but the team discovered that genuinely important information had already been curated by humans. AI topics overlapped heavily with leading technology newsletters, and business news overlapped with established morning briefings. Information that nobody had handled was usually too niche to create mainstream value.
The more fatal issue was the business model. Information distribution depends on competing for attention: the longer users stay, the greater the commercial opportunity. An aggregator that saves users time would have to make money in the same market by reducing attention, creating a direct conflict between demand and monetization.
Zhang Yueguang is not rejecting every product that saves time; he is rejecting saving time in information distribution. Productivity tools can certainly save time, but the revenue mechanism of attention platforms naturally tends toward killing time.
37. Short Video May Already Be Near the Maximum Stimulation the Brain Can Sustain
Zhang Yueguang is not optimistic about the next generation of kill-time products continuing to raise stimulation. Recommendation systems do have algorithmic buckets designed to increase time spent, but people working on short video know these tactics may improve short-term duration and certain metrics while reducing long-term retention.
His analogy is that short video is already a powerful stimulus. Raising the intensity further may not produce a better business and could even have negative effects. Pushing daily consumption from just over 2 hours to 3 or 5 hours does not necessarily create a healthier or larger business.
Xiaohongshu returns some information autonomy to users: they browse a two-column feed first, then decide whether to enter a post, and can leave when they are tired of browsing. It is unclear whether marijuana or heroin is the better business, but physiologically, ethically and over the long-term data, stronger stimulation is not necessarily better.
38. The Goal of Exploration Was Not to Build Another Miaoya, but to Preserve the Ability to Bet
Over more than 1 year, Zhang Yueguang tried 3 or 4 categories of small products, including podcasts, information and images. Before each project, the team explicitly defined it as an experiment rather than packaging it as a high-confidence direction. Most of the team worked normal hours; overtime was not used to manufacture false certainty.
He also told investors in advance: “Don’t expect me to make another Miaoya. The probability is very low.” Before a fundamental value insight emerged, forcing an all-in would only magnify the loss of resources and energy.
The stopping condition was not necessarily failed data, but a flaw in the core hypothesis combined with the absence of a sufficiently good next step. He would change pace only after finding a one-way door worth betting the company’s resources on.
39. Slow Feedback in Content Is a Defense against Big Companies
Games are fundamentally a content industry. Films and content-heavy games often take 2 years to reveal whether they are good or bad. Big companies need layer-by-layer evaluation and clear feedback; that system suits fast tool competition but is inherently poorly suited to long-term, ambiguous content production.
Zhang Yueguang is not afraid of big companies copying Starlink after seeing its closed beta: “If you want to do it, see you in 1.5 years.” Competitors can make something bigger and more expensive, but they cannot compress a complete content pipeline into a week.
The first game still needs controlled investment: validate the model with small amounts of content, polish the team and pipeline, and increase costs gradually. The growth path of a game company is to make a small work first and then iterate through a series toward larger content, rather than betting RMB1B on an IP on the first try.
This is also why the company works on both games and Agents. The game table is relatively uncrowded but does not sit on the steepest slope of AI capability improvement; Agents are extremely crowded but can directly absorb growth in model capabilities.
40. Once Platform Capabilities Are Abundant, Good Content Becomes the New Traffic Dividend
Zhang Yueguang distinguishes himself from other AI-companionship founders this way: “I make content; everyone else makes internet products.” Most companionship products let users or creators design characters through UGC, with the platform providing distribution. Starlink instead defines the characters, stories and experience directly.
A platform may of course be a better business, but the previous generation’s platform capabilities are already controlled by giants. As long as a product still depends on the “ByteDance three-piece set” of recommendation algorithms, user growth and monetization loops, a new company is fighting the old champion with the weapons it knows best.
Red Fruit short dramas and Tomato Novels were ultimately strengthened by ByteDance not because it started first, but because the factors of production had not changed. By contrast, content companies such as Paper Games, Game Science and Pop Mart depend more heavily on specific works and the depth of their IP.
Today everyone has social media. Good content does not need to wait for a new device to become widespread or necessarily buy massive traffic; truly strong content will be shared by users. Platform capacity may be excessive while good content remains scarce, which may explain structurally why content companies have continued to emerge in recent years.
41. Agents Have 2 Clear Sources of Value: Replacing Labor and Expanding Capability
By early 2026, Zhang Yueguang believed AI had to answer the question, “How do ordinary people actually use it?” Most people still treat products such as Doubao as search or encyclopedia Q&A and have not felt a material change in their work or capabilities.
The first category of Agent value is replacing labor end to end, saving time and potentially causing layoffs and cost reductions. General-purpose Agents, vertical Agents and products such as Harvey are all delivering on this direction.
The second is enabling users to do things they previously could not. People who cannot code can use coding Agents such as Lovable and Cursor to build products; people who cannot draw or make videos can create work that was previously impossible for them because of missing skills.
Tool use amplifies this capability further. Models can call external software, engineering services and other models rather than relying only on their internal knowledge. Whether through efficiency substitution or capability expansion, Zhang Yueguang believes good companies will emerge from Agents.
42. Manus’s Acquisition Validated the Paradigm, Not the Completion of Mass Adoption
When Manus first launched, Zhang Yueguang immediately found an invitation code and asked it to plan an offsite trip for his team outside Beijing. The system ran for roughly 30 to 40 minutes but misunderstood the trip from the first step, turning it into a Beijing suburban plan. At the time, he found it “completely unusable.”
After understanding the end-to-end Agent paradigm, he came to regard Manus as the strongest representative of “completing a task for you.” It saves users time and takes over tasks they are bad at or do not want to do, even though many people still cannot find a suitable use case today.
He heard that the transaction value was around $2B but explicitly said he did not know the exact figure. Assuming a 20x P/S for comparison, he believes it was not expensive from Meta’s perspective and that Meta acquired a leading asset in the paradigm.
He refuses to answer on behalf of the founders why Manus was sold, offering only the guess that general-purpose Agents may be a superhighway and the competitors are all major players, making it necessary for an independent company to rely on a giant. He explicitly maintains this as a maybe, not an insider explanation.
43. He Chose High-Frequency Collaboration Rather Than Handing the Task Entirely to an Agent
Zhang Yueguang defines the Manus-style route as end to end, offline and parallel. After the user assigns a task, they forget about it; the Agent may deliver in a minute, an hour or a week, as long as the final result is large enough and good enough.
Doki pursues the opposite rhythm: the user gives an intention, AI responds quickly, the user immediately handles the next round, and AI continues executing. He cares deeply about latency because the value lies in continuous iteration rather than the one-time completion rate of the longest possible task.
The first paradigm focuses on tokens per user, offline execution volume and how much labor is replaced. The second makes people feel they have gained a capability they did not previously possess. “It’s not handing the task to an Agent; it’s you and the Agent doing the task together.”
44. PPT Is a Validated Need; the Only Question Is How to Be Clearly Better
Choosing PPT is not novel. Zhang Yueguang even says the team is “almost the last one.” The advantage is that the category already has PMF, so it does not need to answer what or why; it only needs to answer how, and why a late entrant can still create a one-way door.
The first category of AI PPT products uses fixed templates, such as Kimi at the time. The model generates text first and then forces it into an unchangeable layout. The design does not follow the expression; the expression is forced to fit the template.
The second category generates slides through frontend code, offering more visual freedom but potentially making post-download editing harder than starting over. The third, Banana, directly generates images. The result can be attractive, but it may omit complex information, be difficult to edit and “look obviously AI-made.”
For people who need to present formally to a boss, client or conference audience, these solutions easily become “just going through the motions.” Doki’s standard is that the entire job must be completed inside the product, with no second round of editing in PowerPoint or Google Slides.
45. Doki Polished the Smallest Generation Unit First, Then Worked Backward to the Product Flow
The team was initially still influenced by internet-era habits. It first imagined how people would collaborate with an Agent, then built a complete interaction and workflow, only to discover that generation quality was poor and many preset actions simply did not hold at the model layer.
The team then paused interface-first development and focused on polishing the smallest generation unit. Designers defined aesthetics, engineers handled overflow, layout and containers, and product and model specialists injected knowledge. Together they searched for stable output.
Once the generation unit was sufficiently clear, the team knew what context users had to provide before entering the stage and could design the steps for obtaining that information backward. This was a practical example of “context first, process second.”
Doki uses a proprietary Agent framework rather than simply copying a Claude Code Agent with skills. Zhang Yueguang does not treat the framework itself as a selling point: users perceive only the final result, and context, tools and model orchestration must all serve a deliverable outcome.
46. “90 Points Equals Zero Points” Ultimately Became a Company-Wide Bet
Before Miaoya, Zhang Yueguang had already proposed that “90 points equals zero points.” If people can immediately tell something is AI-made and it cannot enter a real workflow, its value has a low ceiling regardless of cost; once it crosses the usability line, the cost advantage can be amplified without limit.
Doki’s internal proof came when he used it to complete the formal presentation deck for the Baidu World Conference, including converting the conference template into a custom template. He did not manually edit it afterward, and nobody realized it had been generated by AI. He even judged the content organization, logic and punchlines better than his own, while the visual output made up for a design weakness he completely lacked.
The model layer remains a real threat. Claude is “in a different league” on tool-call accuracy, Gemini is used more for frontend aesthetics, and GPT performs better on some tasks. The gap in benchmark scores between domestic and overseas models may have narrowed, but overseas models crossed key usability inflection points first, potentially making the actual product gap larger.
Application companies’ room comes from hybrid models, task depth and engineering details model companies do not want to handle, such as converting HTML with high visual fidelity into PPTX. Zhang Yueguang admits this may be a temporary business, but deliberately chooses the limited value of “technology serving people” rather than removing people from the production process.
PPT is only Doki’s first step. Its long-term positioning is an Agent that helps users “create and express,” and it may gradually develop a personality. Starlink may also acquire practical capabilities in the opposite direction; the 2 lines could eventually converge on “AI friends that combine help with relationship.”
The company has now ended application exploration and does only Doki apart from continuing to mass-produce game content. Zhang Yueguang gives himself a score of 60 for his first 2 years as CEO, satisfied that he “didn’t mess around” and did not exhaust the company’s money or the team’s energy. He now tells the team clearly: “If we lose, we lose everything here. We will not leave ourselves a way out.”
He still insists that the LLM is the intelligence core of an Agent and rejects the idea that language cannot reach the endpoint of intelligence. When people speak, they may also be continually predicting the next token, with logic sometimes reinterpreted only after the sentence is complete. “Language is the endpoint of intelligence. Language is the essence.”