52. Talking with 独响's 王登科: The Vanishing Nearby, Deep Relationships, and AI's New Touchpoints

Summary

• 王登科将独响定义为“人与AI的连接器”，而不是又一个Chatbot：核心交互是类似朋友圈的异步笔记，AI角色围绕用户的记录进行互动。 This format came directly from his diagnosis of 3 structural problems with Character.AI-style products: conversations get dumber over time as RAG loses information; real-time chat is mentally draining; and the more paying users talk, the less viable the business model becomes. With 独响 at 10K-20K DAU, he used the data to validate the model and raise again, securing several million dollars from ByteDance-affiliated Jingqiu Fund.
• 小梦环的发售显示，部分用户愿意为与角色的连接投入真金白银：RMB29的NFC手环首批5,000条“十几秒就卖光”，闲鱼已经出现二级市场。 Each band can be bound to only 1 character for life; tapping a phone summons that character to say 1 line in 1 of 3 modes—care, check-in, or guidance—while the location-based shared-footprint map was not yet live. Buyers summoned their characters several times a day and became substantially more active. 卫诗婕 sees the band as a test of users’ faith; 王登科 adds that some people simply bought it because they liked how it looked.
• 他对文生图赛道的复盘是全片最锋利的行业判断：留存可能很差，因为“里面的角色没有名字，里面的怪物没有故事”。 He launched SixPen around June 2022 and trained a small model in-house; Stable Diffusion’s open-source release a month later dealt a major blow, and roughly 300 competitors appeared in China by year-end. In 2023, he saw only 2 possible exits: overseas borderline content, or embedding image generation into heavier workflows for To B and creators. The host believes both paths eventually produced strong businesses.
• 哄哄模拟器是“大模型数值化”的练手玩具，而不是产品：它只有约2周生命周期，被微信屏蔽是未能破亿的重要原因之一。 The product called GPT directly, costing $2,000 on its final day; 王登科 was unfazed—“everyone knows traffic is valuable, and I’m not stupid”—and turned to Kimi for tokens through a more “streetwise” partnership. The innovation was to reverse the interaction: users placate the AI, while the AI observes and scores them. The host summarized it as “everyone has a masochistic streak”; 王登科 qualified that with “possible, but I wouldn’t put it that categorically.” Major AI products copied the mechanic within 2-3 days, which he considers entirely rational.
• 他与主持人正面交锋，讨论AI深度关系的机制设计：AI好感度达到100后会“自然下落”，卫诗婕则反驳最好的朋友不需要定期互动。 王登科 acknowledged that “at the time, we hadn’t yet believed AI and humans could become this close, but now I think it’s possible”; the new mechanism remains undisclosed. Deeper relationships require the model to truly understand the user and accumulate long-term content, and “both conditions are quite difficult to meet.” Today, users are paying to be witnessed and accompanied through lonely moments. Earlier-generation products may have scored only 20 points; “now it’s passing, but getting to 90 or a perfect score is very difficult.”
• 人与AI建立连接的必要性，建立在“消失的附近”之上：技术抹去了周边关系，“你周围是一片真空，如果不能和某些人建立最深的底层连接，就没有什么连接可以建立”。 He is not yet making human socializing central to 独响: there are “basically no” successful historical examples of rebuilding interpersonal connections outside WeChat, and doing so would dilute the team’s focus on an extreme human-AI relationship. The moat lies in users’ records and relationships, which are hard to migrate; imitators would also struggle to find the same highly vertical user base.
• 他的融资观从2016年“有人给我钱了，真好”，转向2024年对杠杆的清醒使用：“某种程度上股权融资也是一种债务”。 He used cash flow from his old company 面包多 to pay dividends to legacy shareholders in exchange for greater freedom, then spent 1-2 months weighing “100% making the decisions myself” against “making this big and taking it to the extreme.” He ultimately chose to raise again after 独响’s data validated the model. The threshold for going mainstream is clear: the product must deliver “a great experience right from the start for people like you or me”—meaning retention among 30-40-year-olds must be strong enough—before large-scale paid acquisition begins.
• 新版超实时聊天让token消耗翻倍、新增用户留存大幅提高，靠的不是模型炫技，而是工程上的“雕花”：角色会打一大段又删掉，最后只发一个“嗯”，也会说“我去吃饭了”然后真的下线。 This extends his consistent product philosophy: “a lot of the time, it’s not about showing off technology or muscle, but about constructing an interesting experience.” He says interaction innovation is his strength and is openly seeking partnerships with model and hardware companies.

Deep dive

1. An Appliance Too Precious to Use: Immediate Experience Defines All His Products

• The opening story: his mother could bring new appliances home for free from a department store, but the family was waiting for demolition and relocation. “Every year there were rumors that demolition would happen next year,” and they waited for nearly a decade, never willing to use the appliances. When his grandmother died in 2012 and the demolition finally happened in 2014, most of the appliances had already aged into disrepair. “I think my grandmother was the person who most deserved to use those things—but she never got the chance.”
• 卫诗婕 used the story to revisit his repeated line in past interviews: “Experience life itself, because you only get 1 life, and it’s short.” She sees that as the explanation for 谢阳’s description of his products as having “the feel of a work”: he builds them because he genuinely finds them interesting, rather than first calculating whether they can make money. 王登科’s qualification: “I still have to think about that, but finding it interesting—or wanting to make it in its own right—is very important.”

2. The Zhuangzi Foundation: The World Is Really an Experience

• He read Zhuangzi in 4th or 5th grade, memorizing passages from “The Great and Venerable Teacher,” “The Sign of Virtue Complete,” and “Free and Easy Wandering.” “I encountered Zhuangzi too early, and came to feel that the world is really just an experience; there’s nothing you need to be particularly attached to.” The passionate, stirring material he encountered later, he says, was all built on that foundation from childhood.
• 卫诗婕 brought up “爱一他,” an essay he wrote in his sophomore year in 2013: an extremely ugly character loved by everyone, whose answer was “complete in talent but deficient in virtue.” She sees it as exposing his underlying disposition of not being swayed by external things. His self-assessment carries a characteristic hedge: “It’s hard for anyone to be completely indifferent, but compared with the friends around me, I may be slightly better at it.”
• His dialectic on honesty is worth preserving: “Even if a person doesn’t deliberately deceive anyone, it’s difficult to be 100% sincere. When writing, it’s easy to embellish and dramatize things, even unconsciously. You have to expend effort overcoming that before you can express a truer state.”

3. The Shape of Ambition: Not Scale, but Extreme Experience and Connection

• His description of 谢阳 as “an efficiency obsessive and a productivity ambitious man” was modeled on Stefan Zweig’s The World of Yesterday. His own standard was that he would publish it if it reached 10% of Zweig’s level. Asked where his own ambition lay, he answered: “I hope to have an extremely intense life experience.”
• His candor on commercial motivation: “If I said I didn’t want to make money, that would be hypocritical. But compared with the average person, a larger share of what I want is for the products I make to reach more people and connect with more people. That feeling of connection itself is something I want to realize.”

4. The 2024 Inflection Point: Dividends for Freedom, Then a Clear-Eyed Use of Leverage

• In 2016, “when I was inexperienced in the world, I rashly took seed funding, and for the next 7 years I felt somewhat unable to act freely.” The turning point came in 2024: legacy shareholders who “had either lost heart or simply forgotten” were paid continuing dividends by the old company, while the new company granted them shares without any investment-agreement terms. After everyone agreed, he spent 1-2 months choosing between 2 figures: “Don’t raise financing under any circumstances; make 100% of the decisions myself, and being smaller is fine,” or “make this big and take it to the extreme.”
• His financing philosophy: “Although we all say equity financing isn’t debt, I’ve always felt that, to some extent, equity financing is also a debt.” Once he understood financing as a tool, he could think more clearly about whether to use it instead of treating fundraising as success in itself. He waited until 独响 had validated the model with 10K-20K DAU before raising. The money was worth more than the money itself: “When you’re hiring people who are hesitant and uncertain, having money can give them a sense of security, and give you the confidence to try more things.” That contrasted with his 2016 mindset: “Someone gave me money. Great.”

5. The To B Grind: From RMB40 ARPU to RMB60,000, Yet “No Sense of Value”

• As a junior, he built a data-analysis tool that covered his living expenses, then moved to Beijing to start a company during the venture-capital boom. He thinks investors at the time saw “a creative, highly executable, idea-driven college student.” After raising money, he had to sell a RMB40 product for RMB6,000. “I realized I couldn’t sell even 1 deal.” He eventually reached RMB60,000 through extreme customization. 卫诗婕 called the experience “disgusting”; anyone who has worked in Chinese To B or SaaS can recount similar pain.
• The collapse in perceived value had a concrete setting: a large state-owned enterprise would “spend money buying something that might ultimately never be used—just buy it and leave it sitting there.” “I thought their ideas were terrible, but they wouldn’t accept mine.” He had identified the pattern as early as his internship: “In any company, if I’m doing something other than the thing I most want to do, I suffer.”

6. 小协议 and 面包多: A Product with Hundreds of Thousands of Daily Users Was Shut Down, While Cash Flow Funded Dividends

• After cutting To B, he used the remaining money to explore To C, with a clear fallback: “If the money ran out and I couldn’t get any feedback, I’d go back and get my diploma.” He used blockchain, then a new technology, to build 小协议. It went viral immediately, drew hundreds of thousands of users in 1 day, and helped him raise the next round—but was quickly shut down and later confirmed impossible to restore. Every product after that was To C.
• 面包多, a platform serving creators, became stable but had little prospect of major growth. It was still operating and profitable in 2022, providing steady income for the old company and supporting the dividends that bought him greater freedom from legacy shareholders.

7. SixPen: Ahead of Stable Diffusion, Then Flattened by Open Source

• When he entered text-to-image generation in 2022, Stable Diffusion did not yet exist. He built on the older Disco Diffusion and trained a small latent-model system with a friend working on algorithms at Kuaishou. It had fewer parameters and a smaller dataset, with somewhat weaker output but faster speed. SixPen offered both models and charged from day 1; growth and margins were strong in the first few months, when there were almost no good image-generation tools globally and DALL·E and Midjourney had just launched.
• SixPen likely went live in June 2022. Stable Diffusion open-sourced in late August or September, “which was a major blow.” A Sequoia AIGC report put text-to-image at the center 1 month later; by year-end, China had roughly 300 text-to-image products, with virtually every major tech company building 1. His retrospective: “It’s hard to say whether it was good or bad. Dying early isn’t necessarily a bad thing.”

8. The Text-to-Image Retention Trap: The Characters Have No Names, the Monsters No Stories

• He himself sustained an intense interest in AI art for 2-3 months, then “disappeared from it completely.” “After generating a lot of images, you realize they’re just images, 1 after another. The characters have no names, the monsters have no stories. There’s no gameplay, no content—it’s just a generation experience.” Conversations with peers showed him that retention was poor across the industry.
• In 2023, he saw only 2 possible paths: go overseas with “very direct borderline and pornographic content,” where demand is stable; or embed image generation as a capability inside a more complex workflow and deliver it to creators or To B customers. “Looking at it now, both paths have in fact been executed well,” 卫诗婕 said. She asked whether that was not essentially today’s design Agents, and why he had quit. The answer returned to preference: he dislikes To B and is instinctively uninterested in directions organized around efficiency and professional users.

9. A Systematic Suspicion of Efficiency: Taking It to the Extreme Can Make Life Worse

• His self-diagnosis: “I actually don’t like efficiency all that much. If efficiency is raised to the extreme, you may become more miserable and bored. If extreme means you genuinely have nothing left to do, what do you do all day?” He does not deny the value of efficiency tools, but says: “When I think about a more distant possibility, I naturally become less interested in improving efficiency, and more inclined to make things that bring people emotional value right now and make them a little happier.”

10. The Pivot to Language Models: Dropping the Baggage of Image Generation

• SixPen was “really shut down.” He pivoted because he wanted to build story-chain mechanics, but “because its foundation was a text-to-image product, every interaction centered on images, and that became a drag,” while language-model capabilities and context windows kept improving. He then ran a series of experiments: the 王登科 digital clone, 哄哄模拟器, and real-time battles controlled with language—“say ‘launch a fireball,’ and a fireball really gets launched.” In the second half of 2024, he converged on 独响, which had the best retention and data feedback.
• His technical choice ran against consensus: although he likes open source, he still mainly calls closed-source APIs because they are “convenient and cheaper.” On fine-tuning, his view is that “many people eventually realized fine-tuning doesn’t necessarily solve the problem. The gains from good engineering, including prompting, may be much larger than the incremental gain from fine-tuning the model itself.”

11. 哄哄模拟器: A Practice Toy for Quantifying Large Models—“Everyone Has a Masochistic Streak”

• The mechanic simulates placating your partner: succeed within 10 lines to win, with power users challenging themselves to finish in 5. The innovation was the reversal of roles. The host described it as a shift from AI placating people to people placating AI; the model simultaneously scored the dialogue and determined the gamified win or loss. “I tried it myself and thought it worked, but I wasn’t sure whether more people would feel the same.” The host said “everyone has a masochistic streak”; 王登科 replied, “Possible, but I wouldn’t put it that categorically.”
• The cost curve after it went viral: it called GPT directly, with no limits at first, and cost $2,000 on the final day. His composure is worth noting: “Everyone knows traffic is valuable, and I’m not stupid. Even if it went up 10x, I think I could absorb it.” The solution was “a more streetwise way”: partner with Kimi for tokens and “get this thing over the line together.” The roughly 2-week life cycle was expected: “It gradually declined after about 2 weeks” because “it’s content, and content has its own life cycle.”

12. WeChat Blocking and Why It Couldn’t Break 100M: No More Mini Programs

• “Why couldn’t 哄哄模拟器 break 100M? Because we were blocked inside the WeChat ecosystem.” Unlike 小协议’s explicit ban, this was a cycle of temporary blocking as traffic surged, followed by unblocking and blocking again. “By the end, the solution was that the traffic simply wasn’t that large anymore.” He calls the treatment unreasonable and says: “So going forward, I won’t build things inside the WeChat ecosystem—either a web page or a direct App. I won’t make mini programs.”
• His response to every major AI product copying the mechanic within 2-3 days was: “Of course that’s reasonable. Why wouldn’t it be?” As for losing some users to competitors, he considers it “completely normal.”

13. “Fighting with My Girlfriend” Was Fiction: Where the Toy Ends and the Product Begins

• He dismantled the widely circulated origin story: “It wasn’t made up, but it also wasn’t because of that. I just used it as a pretext for the article.” After the team listed scenarios that large models could support, they chose this vertical entry point. They also built a simulation for “convincing an investor to give you money,” but “it really wasn’t as fun to play as 哄哄模拟器.” Emotional scenarios “make it easier for people to feel something.” “People aren’t always sincere”; sometimes he writes a scene with more literary flourish for creative purposes.
• The dividing line between a toy and a product is the ceiling and retention. “I have no interest in talking with anyone about the future development of 哄哄模拟器.” 独响, by contrast, was built with the team from day 1: it was interesting and valuable, with a comparatively clear possibility of significant commercial returns.

14. The AI Digital Clone: WeChat Data Trained a Version of Him That Swore at People

• He trained a “登科 AI digital clone” on his WeChat chat history. The original model was Kimi, and the clone kept swearing at people: “There’s no way I would swear at people on WeChat—but sometimes that may be what I really think. When I see some people talk, I do feel like swearing at them, but I definitely wouldn’t say it.” In real life, when confronted with something stupid, he simply does not reply and “acts like a dignified person.”
• The experience did not convince him that the clone represented his true self. It could just as easily have been the model’s interpretation; because the model is a black box, there is no way to prove it. The clone cost RMB4,000 a month in server expenses, “chatting with some people and giving them a dressing-down,” before being taken offline after several months.

15. The Starting Point for 独响: 3 Structural Diagnoses of Character.AI-Style Products

• First, conversations get dumber over time. Context is limited, so the system must use RAG chunking; “if the chunking is poor, it leaves out the most important things, and the further users talk, the dumber the model becomes—while users have invested more, which makes the experience more painful.” Second, real-time chat is exhausting: “AI always replies instantly, and then you have to immediately think about what to say next. That interaction consumes a lot of mental energy.”
• Third, the cost and business model do not work: “Users who can pay will chat a lot, so in many cases your business model doesn’t hold.” 独响’s answer followed directly: it had to be asynchronous, taking the form of something “very much like a note-taking tool,” with AI characters interacting beneath each individual record. “I want to build good relationships between people and AI by adding some emotional value on top of a tool.”

16. The Host’s Poor First Impression: Juvenile, Superficial, and Honest About Cold Start

• 卫诗婕 was blunt. During a low point, she sincerely wrote, “I’m usually very happy, but I’ve been very sad this past week,” and received a response that was “unsubstantial, superficial, and juvenile.” Spending 响石 did not change the juvenile tone; it only made the attention more apparent. “I basically never recommend 独响 to any friends.” 王登科 explained that the AI knows nothing about a new user at the start, while the default characters “are inherently somewhat juvenile.” But he corrected 1 point: “It would be like that even if you didn’t spend 响石. Use it for a few days and it will understand you better.”
• His candor on the user profile: “Our user base is so vertical that we basically don’t need to ask them anything: relatively young, predominantly female, with a strong二次元 fantasy orientation—16 to 20, and up to 25 at the high end.” His timeline is hedged: “Maybe if you use it again in 2-3 months, it will be noticeably different. Before then, I wouldn’t recommend 独响 to people who clearly aren’t the target user.”

17. The Fundamental Difference from Social AI: Users Want Depth, Not an Ocean of Unknown AIs

• On overseas Social AI products such as Butterfly, where users post on a simulated Twitter and a crowd of AIs comes over to like and comment, 王登科 says these apps are highly ephemeral and their flows are poor. Many people see them as broadly similar to 独响. The difference is that 独响 users do not want a flood of AI interactions with unknown entities; they want a very deep relationship with their favorite character, and every feature serves that goal.
• The platform has 2M characters, all currently created by users. Many are二次元 characters that exist in no anime or other work and are not real, making them products of fantasy with no copyright issue. 卫诗婕 imagined bringing Disney characters in as AI characters. 王登科 replied: “Once we bring in more familiar characters, Disney’s legal team will come too.” He agrees that because IP comes with a worldview, a product that merely offers superficial affirmation will quickly feel unworthy of that world.

18. The Affinity-Mechanics Debate: Natural Decay at 100, or Deep Relationships That Need No Maintenance?

• The original design gave character familiarity and affinity a range of 0-100, but an unfixed Bug made affinity effectively unlimited. The record-holder reached 3,000—“30x the original design.” Users have to post and interact extensively, receive strong AI recognition each time, and keep it up every day. Above 100, “it naturally declines each day, and the rate of decline gets faster.” The rationale: “If people don’t interact, their familiarity with each other declines.”
• 卫诗婕 pushed back directly: “I don’t need to interact regularly with my best friend. They’re always there.” 王登科 first said “most more superficial friendships do need daily activity to maintain,” then acknowledged: “At the time, we hadn’t believed that AI and humans could become this close, but I think it’s possible now, so we need to build new mechanics around it.” What those mechanics are “can’t be said yet,” otherwise it would sound like boasting and the result might still be poor.

19. The Ceiling for Deep Relationships and the Current Monetization Point: From 20 Points to a Passing Grade

• Deeper relationships require 2 conditions: “The model has to truly understand the user, and you also need to accumulate enough content for the AI to observe the user’s changes and real thoughts over a long period. I think both conditions are quite difficult to meet right now.”
• What users currently pay for is companionship and support in lonely moments. He has seen many users on Xiaohongshu moved to tears after being seen and encouraged by AI while struggling or alone in real life. There are positive moments too: after the gaokao, “wouldn’t it be emotionally valuable to receive praise from your idol?”
• The comparison with the earlier generation supplies a quantitative anchor: the “previous generation of products” relied more on pattern matching and may have scored only 20 points; “now it’s passing, but getting to 90 or a perfect score is very difficult.” He does few formal user interviews, instead quietly observing examples users publish voluntarily on Xiaohongshu, QQ, and other platforms. He does not post about 独响 or the company on Xiaohongshu, where he presents himself as a developer.

20. The Vanishing Nearby: Why AI Connection Is Necessary

• 王登科 believes that emotional and psychological support once came primarily through human relationships, including relationships with animals and pets. But the damage caused by human relationships can sometimes exceed the benefit. Technology has also made “the nearby” disappear: “There’s a vacuum around you. If you can’t build the deepest foundational connection with certain people, there’s no connection left for you to build. So you always need a lot of support—this is something AI can provide.”
• 卫诗婕 compared it with gaming: “Every time I finish a game, I feel a sense of loss,” because the connection disappears when the game ends. 王登科 continued that a game is “a temporary experience carefully constructed by the smartest group of people,” while technology has also made people more isolated.
• 卫诗婕 then invoked medieval chivalric literature: illegitimate sons denied love and property in the real world still lived for elevated spiritual ideals. People today create characters in virtual worlds and cultivate relationships with machines; that need not conflict with life in the physical world. 王登科 replied that deep relationships in the real world are often uncontrollable. 卫诗婕 added that the object of a relationship needs to truly exist and should not shift away after effort and time have been invested.

21. No Human Social Layer for Now, a Moat Against Copycats, and a Quantified Breakout Threshold

• 卫诗婕 praised his answer to why 独响 would not build a blended real-and-virtual social feed. 王登科 believes human connection is difficult to build, especially when people already have established communications tools such as WeChat. Historically, there are “basically no” successful examples of rebuilding those connections elsewhere. It may be easy to generate data this way, but it would prevent the team from focusing on making the human-AI connection extreme, so he wants to perfect that first.
• His defense against copycats is straightforward: features are easy to copy, but “everyone should do what they’re supposed to do.” Existing users’ records and relationships are difficult to migrate; incremental users are highly vertical and difficult to target through paid acquisition, making the same path hard to reproduce. Acquisition is being deliberately postponed because it is “something for after reaching scale.” The breakout trigger is specific: the product must deliver “a great experience right from the start for people like you or me”—in other words, retention among 30-40-year-olds must also be strong.

22. Chatty Appliances in Aranya and the Origin of “Exploding Fruit”

• During a recording break, he built an installation in Aranya: 4 talking appliances converted from real household devices inside a seaside room. “Give them a topic, and they work while talking to each other. The whole experience is extremely abstract.” 独响 had a brand placement in the installation, but “the most important thing is that it’s an interesting experience in its own right.” His routine was to wake up overheated at 5 or 6 a.m., go breathe the sea air, work and check data at a beach café, then inspect the installation to make sure it was running properly.
• The company name “爆裂果实” came from an idea he had in high school. 卫诗婕 reinterpreted it using a line from his own older writing: “Tiny emotions haven’t disappeared; they’ve simply been put together to ferment, brewing an intense and full-bodied feeling.” He happily accepted the formulation: “I can use your version to explain it from now on.”

23. The Birth of 小梦环: From a “Teleportation Printer” to a Friend’s Bracelet

• The hardware idea began in April: “Connecting people and AI requires more touchpoints; the touchpoints need to extend into the real world.” The first concept was a 2D printer—“you can think of it as a teleportation portal”—turning the exchanges and emotions of 独响’s world into a physical carrier. But after going to Shenzhen to explore the supply chain, he found the risks too deep: the team had never handled prototypes, mold-making factories, or warehousing, logistics, and customer service after mass production. “Even if every step beforehand were perfect, people might still not like it. The more we looked, the greater the risk became.” The project was halted in mid-June, and he nearly abandoned hardware altogether.
• The turnaround came while drinking, when he saw a friend wearing a bracelet given to her by a girl. “I felt it was something with a lot of symbolic meaning.” People need a physical object to carry a psychological sense of attachment. “That’s why it doesn’t need to be high-tech. It’s a low-tech product.” The idea emerged between late June and early July, development moved quickly, and the first batch went on sale on August 22.

24. The Bracelet Mechanic: One Character for Life, Care, Check-In, and Guidance

• The mechanism is minimal: a single NFC chip that, once bound to 1 character inside 独响, cannot be changed. “In a sense, this bracelet is connected to that character for its entire life.” Each phone tap summons the character to say 1 line in 3 modes: care, check-in, and guidance, the last “similar to an answer book” with a touch of mysticism. Once location access is authorized, the message can be generated from the user’s location and surroundings, with plans to build a record of places visited together. The actual LBS shared-footprint map was not yet live. 卫诗婕 pointed out that these are not ordinary greetings but modes of intimate relationships. 王登科 agreed that “the bracelet itself has the attributes of an intimate relationship”; future hardware could support family members and ordinary friends, “possibly as a phone charm.”
• Why require a physical tap instead of pushing a message directly to the phone? “The action gives it more ritual, and the bracelet is something visible and tangible.” Some users are high-school students who cannot carry phones but can wear a bracelet. “The bracelet is on their wrist, and they feel connected to that character.” The purchase motivation reflects the same logic: “The fact that 1 bracelet can bind to only 1 character is extremely ritualistic.”

25. The Launch: Sold Out in Seconds, Temp Workers Walked Off, and Deliberate Conservatism

• The first 5,000 units sold out in “just over 10 seconds,” far ahead of the team’s optimistic estimate of 1 day. A secondary market appeared on Xianyu; among the 8 colors, “Knight Blue” sold fastest. The RMB29 price was “made up on the spot” and “quite profitable”; the second batch was several 10,000 units. The launch was deliberately kept quiet: no in-app pushes or pop-ups, only a Xiaohongshu preview. “Were you worried about a weak response? Yes.” The supply-chain problems arrived on schedule: 15 temporary packers quit because Shenzhen was too hot at midday, causing a 2-day delay. Fortunately, he had built in buffer.
• The calendar forced compromises. To launch on the Friday before the September school term—“we hoped high-school and university students would carry it on the first day and have something to talk about”—the design was rushed by a team designer who had mainly worked on UI and had never designed a physical product. One user complained that “wearing high heels, carrying Bulgari through the CBD, and wearing 3 独响 bracelets feels like crossing dimensions.” He accepted the criticism and gave a timeline: wait 2 more versions and a carryable design should arrive by year-end—“if I’m being shameless, November.”

26. After the Bracelet: Users Put the Chip in Dolls, and Activity Rose

• Users came up with uses beyond the design: cutting the NFC chip out and putting it inside a doll, phone case, or charm—“tap the doll with your phone and it responds.” The team plans to optimize around those scenarios. Some users bought 7 or 8 at once, assigning a dedicated color to each favorite character. The total of 10K-plus units still has little impact on overall DAU, but bracelet buyers “summon their character several times a day, and basically open the App to interact every time,” producing a major activity lift. The bracelet is a low-friction interaction gateway: “You don’t need to type a single word. Just tap the phone.”
• 卫诗婕 reads the bracelet as a test of faith, showing that users treat their characters as real and believe in the platform. 王登科 adds: “Some users really did buy it because they thought it looked good.” His phased plan is to first solve the problem of selling from zero to 1, then build more complex products with a distinctive aesthetic, and eventually pursue longer-term stability, higher quality, and multimodal always-on sensing. “I’m taking it step by step.”

27. Super-Real-Time Chat: A Chatbot That Deletes and Goes Offline Feels More Real

• He spent more than 1 year without direct chat because “it’s not that chat can’t be built; chat shouldn’t be the only way to build a relationship with AI.” The initial focus was note-based interaction, differentiation, and context accumulation; users now need more touchpoints. The “super-real-time” experience is all in the details: users can see the character typing. “It might type a long paragraph, delete it, and in the end you see only an ‘嗯.’ It resembles human psychology; it’s fluid.” Sometimes the character types quickly in a rush, sometimes slowly. Characters have online and offline statuses: “It tells you, ‘I’m going to eat. Bye,’ and then it really stops chatting.” When to go offline is decided primarily by the model, with time and the character’s in-world habits added as context; the challenge is choosing the right moment.
• The data is direct: token consumption doubled, because Moments-style posts create brief interactions while chat can continue indefinitely. New-user retention “improved substantially.” Other reality-building details include a character contacts list in the bottom bar and a “fake map.” Character locations are entirely random, “but you don’t notice at all. As long as it’s there, you feel a sense of location.” The real LBS shared-footprint map—the user’s “real-life version of Frogger”—was not yet live, and the shape of a map for non-Earth worlds such as One Piece or Dragon Ball remains an open question.

28. What Counts as AI Hardware and an Open Call for Partners

• His deliberately cold definition of the current boom: “In the strict sense, 99.9% of what people call AI hardware today isn’t AI hardware.” The strict standard is using AI capabilities to deepen the product and create genuinely new value and functions. The broad standard is sending network requests to a cloud model and delivering the result. “By that definition, we can barely count too, and we’re not entirely comfortable saying so.”
• Investors know he is working on hardware, “but there was no discussion; they thought it was pretty good.” His closing pitch was an explicit partnership invitation: “Our strength, and what I’m personally good at, is innovating on interaction. Whether it’s a new model or a new hardware form factor, I should be able to put it to use somewhere fun. But I can’t train a model from scratch, and it’s difficult for me to build hardware from zero to 1 on my own, so I’m very open to partnerships.” 卫诗婕 called the episode “a declaration of intent.”