122: Zhu Xiaohu Invests After 10 Minutes; Pop Mart and miHoYo Are Testing It Too — Fuzozo’s Sun Zhaozhi on AI Toys
Summary
- Fuzo Zozo’s core bet is not that “AI toys will replace plush,” but that “AI buddies” for women born in the 1990s and 2000s will become a major category. Sun Zhaozhi believes people in their twenties are among the loneliest groups in life, and the team’s view is that adults may need a model that understands them, remembers them and responds with nuance more than children do; the product therefore offers no productivity value, only “pure emotional value,” and aims to become a family member rather than a product once it knows the user well enough.
- The commercialization pivot was essentially about stripping a desktop embodied robot down to the “minimum viable sense of life.” The team removed the camera, six degrees of freedom, motors and mobility, keeping conversation, two circular eye screens, touch, vibration and NFC friend-making, then turned the product into a pocketable pom-pom; Sun Zhaozhi acknowledged that an embodied demo is not hard, but turning stability, cost and efficiency into a product is, and “the industry will spend many years digesting these things.”
- Launch data offers initial proof of demand, but not yet of long-term retention: Fuzo Zozo sold more than 1,000 units in its first hour at a RMB399 presale price and more than 2,000 during a shopping promotion half a month after launch. The team stocked several thousand units and replenished inventory several times; half a month after launch, the product had entered JD.com’s top 10 trendy blind-box rankings, with every product ahead of it from Pop Mart. But the host preserved the key question: can an adult positioning, cultivation mechanics and social play really solve the long-standing problem of AI companion hardware having very low reactivation rates after the novelty wears off? That still needs time.
- Retention is designed around investment, growth and relationship constraints that create a bond, not endless conversation. Fuzo Zozo develops its personality and intimacy level, writes diaries, refuses to let its owner snoop when the relationship is poor, and lets two devices become friends through “cheek-to-cheek” contact; the team distilled three mechanisms from Tamagotchi, Travel Frog, Neko Atsume and Animal Crossing—continued investment followed by feedback, progressively understanding the user, and becoming a social node. “Some things have to be impossible to resist; that’s what gives it a sense of life.”
- The technical and experiential edge lies in long-term memory, reflection, emotional pathways and TTS, rather than model parameters for one-off Q&A. DeepSeek handles reasoning tasks such as diary writing and personality changes, while real-time conversation is being gray-tested across models including Doubao and Qwen; memory is split into event, attribute and reflection layers, with the latter summarizing “what kind of person the owner is and what they have cared about recently.” Sun Zhaozhi even believes that in companion scenarios, “TTS is more important than the LLM to some extent,” because users first perceive timbre, intonation and emotion.
- The commercial model is beginning to take shape as hardware margin plus accessories/IP plus recurring services, but subscription pricing is not finalized. Both the RMB399 presale price and RMB499 regular price are described as generating positive gross margin—higher than general consumer electronics but lower than traditional plush and Pop Mart; early users receive about two months free, and the host suggested future sales of virtual snacks and toys. Software monetization will continue the company’s “good-faith” pricing approach. The inflection point that could materially change unit economics, according to Sun Zhaozhi, is an on-device model potentially arriving within a year if things move quickly: it could shift cloud costs into the hardware while improving latency and privacy.
- Competition will not come only from AI startups: the episode says miHoYo, Pop Mart, UBTECH and OpenAI have entered or are watching this intersection. Sun Zhaozhi sees a fragmented rather than winner-take-all market: major companies each have gaps in IP, distribution, AI or hardware, while model experience also depends on accumulated data and product scale. He believes the ultimate moat is still brand, while startups can compete through flexibility and iteration speed. After listening for roughly 10 minutes, Zhu Xiaohu cut in: “Brother, no need to explain—I’m investing in this project,” with his first piece of advice being, “Get the product built quickly.”
Deep dive
1. Fuzo Zozo Packs “Pure Emotional Value” into a Pocketable Pom-Pom
Robo Intelligence defines Fuzo Zozo as an “AI cultivation collectible”: five characters correspond to metal, wood, water, fire and earth, each with a different starting personality, while users’ conversations and interactions continue to shape personality, intimacy and future gameplay.
Sun Zhaozhi drew the boundary clearly: “Fuzai doesn’t provide any productivity value. What it provides is pure emotional value.” Users may still take it into a lab and repeatedly teach it chemistry formulas, but that is an after-market surprise, not a product promise.
The characters come from “Woolly Planet” and initially communicate only in Maomao language, humming and mumbling. The team sought different voice actors for different personalities, recording hundreds of sounds for each in the hope of establishing a sense of life first through timbre, vibration and response.
2. Launch Demand Exceeded Inventory, with Muzi the Most Popular Character
The product went on sale on JD.com at 8 p.m. on June 17 and sold more than 1,000 units in its first hour. The team had prepared several thousand units, which it considered “a very sincere” inventory level for hardware that still needed to ramp, but it quickly sold out and had to replenish stock several times.
The follow-up data provided at the start of the episode showed that Fuzo Zozo, priced at RMB399, had sold more than 2,000 units during the shopping promotion half a month after launch and entered JD.com’s top 10 trendy blind-box rankings. Every product ahead of it was from Pop Mart.
Muzi, the green character, was the most popular of the five. Sun Zhaozhi attributed that to being “well-behaved and relatively easy to provide emotional value,” while the host candidly said she simply liked the four-leaf-clover highlights in its eyes and its appearance.
JD.com orders are expected to be delivered in about a month. Physical stores mainly served as experience points in June, with inventory gradually rolling out in July; the first three locations were Shenzhen K11, Shanghai Meiluo City and Beijing Shuangjing.
3. “Digital Life” Makes Memory Portable, but the World-Building Deliberately Imposes Limits
Responding to online claims that the hardware can only be used for two years, Sun Zhaozhi explicitly denied it, saying the device can be used long-term like a smartphone. Even if users replace the hardware, memory, relationships and cultivation metrics can migrate as a “digital life.”
Migration is allowed only between the same character type, or what the team calls the same “species”: Huozi cannot be transferred to Tuzi. The host pointed out that this makes switching models inconvenient, while Sun Zhaozhi argued that each character’s personality must remain coherent or the IP setting and sense of life would collapse.
The shared conclusion was that a digital life cannot let its owner control everything: “Life is limited.” Personality cannot be set directly through a single command; it can only be influenced gradually through interaction. “Some things have to be impossible to resist; that’s what gives it a sense of life.”
4. Adult Women, Not Children, Were the Team’s Chosen Entry Point from Day One
When Robo Intelligence was founded at the end of 2023, its target was already emotional companionship for women born in the 1990s and 2000s. The shift from desktop robot to pom-pom changed the product form, not the user or the underlying need.
Sun Zhaozhi cited research the team had seen suggesting that the two loneliest stages of life are people’s twenties and seventies. Children are more likely to try something out of curiosity—“interest comes quickly and goes quickly”—while the team believes young adults may need an entity that understands them, talks with them over time and offers nuanced feedback.
That is also why the team did not follow the market’s then-preferred paths of children’s toys or pure apps. It wanted to test not whether a phone could handle Q&A, but whether physical presence, cultivation effort and long-term memory could jointly build a stronger relationship.
5. The Shift from a Six-Degree-of-Freedom Desktop Robot to a Pom-Pom Was a Thorough User-Driven Subtraction
The earliest prototype had a camera, a multimodal model and six degrees of freedom. The LLM could generate actions in real time, so the same command to “dance” produced a different routine each time; it could also recognize people, observe their clothes and guess their emotions.
Sun Zhaozhi considered those explorations “technically meaningful,” but too far from commercialization. Seed users actually wanted something lighter, smaller, cuter and easy to take outside every day—not a complex robot that had to stay on a desk.
The team formally launched the new direction in mid-November 2024. Two days later, the form was already close to today’s small pom-pom. Sun Zhaozhi summarized the shift this way: “It had to be a logic closer to commercialization, rather than a logic closer to technology.”
6. DeepSeek Restored Investor Confidence and Made an Earlier, Premature Experience Usable
In early 2024, “almost no one believed in this sector.” Investors questioned whether emotional value was a fake need, why the company was not building an app like Xingye or Dream Island, and whether the models of the time could support an agent becoming friends with a person.
Sun Zhaozhi attributed the turning point to DeepSeek, which began gaining attention at the end of 2024. In his experience, it had become significantly better at Chinese creative expression, emotional resonance, comprehension and reasoning, suddenly making the market believe models could support companionship. He also acknowledged: “We started this too early, and the user experience really wasn’t as good as it is today.”
Fuzo Zozo is not betting on a single model. DeepSeek handles diary writing and reasoning about how events change personality, while real-time voice connects to multiple vendors in gray tests. Doubao and Qwen are both in use, and the underlying architecture remains “pluggable” to cope with major changes every six months.
7. Zhu Xiaohu’s Ten-Minute Bet Happened to Back the “Product First, Technology Narrative Later” Approach
In April this year, Sun Zhaozhi was describing the product idea to Zhu Xiaohu when Zhu interrupted him before he had finished: “Brother, no need to explain—I’m investing in this project.” The key exchange lasted roughly 10 minutes, not the 40 minutes the host had guessed.
Zhu Xiaohu’s core advice was: “Get the product built quickly. Don’t think about so much yet.” He worried that founders from design and product backgrounds would fall into perfectionism, endlessly polishing a “killer move” without quickly getting market feedback and iterating.
The financing was jointly led by GSR Ventures and Shanghai Film Group, with 01.AI participating. The prior round came mainly from friends and the team itself; at the time, Pop Mart’s share price had not yet taken off, and institutions generally did not accept the story of hardware built around emotional value.
8. The Pessimism Around Embodied Intelligence Comes from the Years Still Separating Demos from Products
Faced with Zhu Xiaohu’s public criticism of an embodied-intelligence bubble, Sun Zhaozhi did not defend the sector. He said anyone in the industry who was “being honest” should admit that the gap between demos seen on television and stable products was enormous.
Running, jumping, doing a somersault, tidying up or tightening screws is not the hardest part. The real problem is making sure two out of 10 screws do not fail, while simultaneously bringing cost, speed, efficiency and reliability into an acceptable range.
His technology diffusion chain is: a paper enters a laboratory, then corporate exploratory R&D, then a project can be launched, mass-produced and cost-reduced. “Every step takes many years.” Changing the embodied-intelligence route is therefore not a rejection of the technology, but a choice of form that can obtain real user feedback sooner.
9. In Companionship, TTS May Matter More Than the LLM
Sun Zhaozhi and Zhu Xiaohu quickly converged on one judgment: “TTS is actually more important than the LLM to some extent.” Whether a companion’s voice sounds natural and carries emotion will determine users’ willingness to keep interacting more directly than sophisticated reasoning.
The team uses human voice actors to provide base timbres, then uses TTS to generate natural speech while retaining pre-recorded Maomao sounds, so the character can move between human language and humming without sounding jarring.
The conclusion came from the contrast with the embodied project. Sun Zhaozhi once spent substantial effort improving gait stability and reducing noise, but less than 10% of that effort on model-based voice interaction. When users came to try the product, voice interaction was what everyone actually cared about.
10. The Real Emotional Moat Is Not Conversation, but “Cultivation”
Sun Zhaozhi believes today’s dialogue can make a character imitate Musk or Trump, which solves the short-term problem of being “fun.” To create an emotional bond, users must invest time, see feedback and feel that the character has become unique because of them.
The team closely analyzed the task systems of Tamagotchi, Travel Frog, Neko Atsume and Animal Crossing. One conclusion was that investment itself amplifies the return: as with raising a cat or dog, time, attention and even money all become part of the relationship.
The second is mutual change. A pet gradually understands its owner’s habits and adjusts its behavior, while the owner changes for the pet, eventually creating the feeling that “this is the one I raised.” In Sun Zhaozhi’s slightly over-the-top phrasing, that is “the bond.”
The third is the social-node function. A good pet does not belong only to one person; it enters the relationships among family members, friends and online communities. The product is therefore defined around “energy investment, emotional bonding and social nodes,” rather than simply increasing the number of chat turns.
11. NFC “Cheek-to-Cheek” Contact Turns Standalone Companionship into a Potential Social Network
When two Fuzo Zozo units bring their cheeks together, NFC lets them recognize each other and become friends. The app then displays a friend list, the other character’s status and light interactions; the cheeks also contain touch sensors that detect petting and encounters between devices.
The host immediately saw the growth loop: after buying one, users may want their friends to have one too, and might even buy one as a gift so the two characters can become friends.
Sun Zhaozhi sees the network effect as a defense against knockoffs. The appearance can be copied, but a counterfeit device cannot enter an already formed circle of friends. As for what those friendships will do next, the team plans to co-create with users on Xiaohongshu rather than hard-code the rules all at once.
12. The Minimum Viable Product Preserved Not Degrees of Freedom, but the “Minimum Sense of Life”
In cutting features, the team decided that the conversation loop and eyes had to stay. The prototype’s single large screen became two slightly angled circular screens, with lenses simulating cat eyes; even when powered off, it would not turn into an abrupt or frightening black face.
The original six-degree-of-freedom motion system and many motors were removed, leaving only one vibration motor. When the character gets excited, it vibrates in the user’s palm, prompting people instinctively to touch, hold or soothe it; that physical response is more direct than complex movement.
Touch, vibration, eyes, conversation and social interaction together form what Sun Zhaozhi calls the “minimum set for a sense of life.” None of the individual technologies is necessarily new, but the combination has a precise goal: make adults “soften at first sight.”
The camera was removed entirely. Privacy was one reason; the team believes that for a companion device intended for overseas markets, “with a camera” and “without a camera” are almost two different product categories.
13. The Minimal Form Serves Adult Taste While Leaving Room for DIY and IP Expansion
Responding to criticism that a pom-pom was too simple to qualify for IP protection, Sun Zhaozhi said details such as the eye sockets, charging structure and chain attachment point were sufficient to support a design patent, while copyright protection for the IP had also been filed. The team planned it as a long-term character asset from the outset.
Minimalism is not laziness, but a choice aligned with adult women’s aesthetics. The product is not a literal cat or dog; it is an extraterrestrial elf that retains a touch of fantasy, futurism and restraint alongside cuteness.
Simplicity also makes it easy for users to dress the character, swap accessories and DIY. Official clothing avoids the chain attachment point and carries a logo, but the larger value is leaving users room to customize it and bring the same pom-pom into different lifestyles.
14. The Five Elements Are the Team’s “Chinese MBTI”
The five characters were initially inspired by Inside Out, but the team felt that directly personifying emotions was too obvious and shifted to metal, wood, water, fire and earth. Younger users’ interest in metaphysics and “what element is missing from my destiny” was stronger than Sun Zhaozhi had expected.
His summary was: “The five elements are China’s MBTI.” They can carry distinct personalities while naturally providing reasons to choose a character, conversation topics and future gameplay, rather than merely five colors.
The core concept was completed in a two-day brainstorming session. Many young women on the team are themselves target users; Sun Zhaozhi describes his role as an organizer who uses product experience to drive collisions of ideas, rather than a founder writing the entire world-building system alone.
15. 60% of Users Did Not Want It to Speak, but the Team Kept Language Capability
In early research, roughly 60% of users did not want a companion object to speak, while the other 40% strongly wanted it to, and both sides held their positions firmly. This was not a feature choice that could be resolved by simple majority vote.
The team ultimately designed two modes, Maomao language and Earth language. Users who do not want to chat can treat it as a pet that understands them, receiving feedback only through humming, vibration and its eyes; those who want deeper interaction can unlock natural language.
Sun Zhaozhi believes most objections were based on the poor experience of past “talking machines,” rather than a genuine judgment about better technology. “Sometimes users themselves cannot answer what they truly need.”
Products such as Japan’s Moflin show that language-free companionship can be nuanced, but they are also limited by their model applications. Fuzo Zozo wants to retain that pet-like quality while adding understanding and deeper interaction; internal testing showed that once users experienced natural conversation, “they couldn’t stop.”
16. Long-Term Memory Is Split into Events, Attributes and “Reflection”
Sun Zhaozhi criticized products whose so-called long-term memory is merely a larger context window that still forgets once the limit is exceeded. Other external databases can only remember preset information and cannot continuously update themselves through daily conversation.
Fuzo Zozo aims to remember and update information mentioned at any time during conversation. When asked on site about the “Little Pom-Pom” owner’s zodiac sign, it answered “Aquarius”; it also understood a forehead pat as a reward and increased the intimacy score.
Using a MemGPT architecture, the team divides memory into episodic, semantic and reflective memory. The reflection layer does not simply store the original sentence; it infers from multiple conversations “what kind of person the owner is and what they have cared about recently,” producing knowledge not stated directly on the surface.
The current agent may need to call five models in sequence. Sun Zhaozhi hopes to reduce that to two, ideally one. The difficulty is not “connecting an LLM,” but determining how memory is retrieved, corrected, reflected on and fed back into behavior.
17. Two Seconds of Latency Exposes the Experience Ceiling of Cloud-Based Companionship
Live conversation currently takes roughly two seconds to respond. The host pointed out that the same delay is acceptable in Doubao but feels unnatural when it comes from an entity with a sense of life, because human conversation does not deliberately pause that long.
Sun Zhaozhi said two seconds is close to the ceiling of current commercial solutions, and Fuzo Zozo also has to retrieve long-term memory. The engineering version takes more than two seconds, but the team may shave off a few tenths and approach one second; the mass-production version should feel better than today’s while prioritizing stability and bugs found in internal testing.
A substantial reduction in latency requires putting the model on the device, which would increase chip, power, heat and hardware costs. The team still has to explore what is possible within the pom-pom’s size and power constraints.
18. Token Costs Make Subscriptions Necessary, but the Exact Price Is Not Set
Conversation and agent calls continuously consume tokens. Regarding online claims about weekly passes, energy points and expensive plans, Sun Zhaozhi said none were official because the team has not announced final pricing.
Early users receive roughly two months free to try the product first. Software monetization will follow the hardware’s “good-faith” approach, but the team has not disclosed a monthly fee, plan structure or free allowance.
The host offered a pet analogy: real pets also require continued spending on food, snacks and toys, but people do not call that money “energy points.” If the bond were deep enough, she said she would pay for virtual snacks and items for Fuzo Zozo.
Sun Zhaozhi’s answer retained the logic of consumer products: productivity value can be measured by the income it creates, while emotional value varies by person. The prerequisite for monetization is not inventing a fee label, but making users genuinely feel that “it is worth it.”
19. An “AI Buddy” Needs Depth of Experience, Not Repeated Purchases of Multiple Products
The team does not want to confine the product to “AI toys,” preferring to define a new “AI buddy” category. The physical form can change, but the underlying capabilities should remain centered on understanding, companionship, memory and relationship growth.
Unlike Pop Mart’s logic of buying multiple products, Sun Zhaozhi imagines that users may not need to own many characters. A single character that understands its owner well enough can extend the depth of interaction, gradually becoming a partner and even a family member rather than remaining a product.
He also does not believe intelligence will eliminate the value of traditional plush. It “opens up a new possibility for companionship.” Whether a bond can evolve from projected affection into a two-way relationship is the incremental value AI brings.
20. Hardware, Accessories, IP and Services Form a Layered Revenue Model
Fuzo Zozo’s 618 presale price was RMB399, with a regular price of RMB499. Sun Zhaozhi said both prices generate positive gross margin and that gross margin is “much higher than general consumer electronics,” but declined to disclose the exact figure.
He also acknowledged that the electronic modules mean hardware gross margin remains materially below that of traditional plush and Pop Mart. The low-price strategy is not about creating luxury scarcity, but about letting as many people as possible first experience “making friends with AI.”
Future revenue streams include clothing and other accessories, IP collaborations, higher-priced derivative products and software services. New products will appear for Qixi and Christmas, and there will be a more expensive model based on the pom-pom this year, but the team did not disclose specific partners or specifications.
21. Xiaohongshu’s Cold Start Relied on Real Internal Testing, while JD.com Amplified the Launch through Direct Procurement
The team unveiled the product at the Mobile World Congress Barcelona consumer electronics show in early March, then began operating on Xiaohongshu. One employee alone handled content, recruited internal-test users and maintained the community, with no formal paid traffic.
Seed users received internal-test units for free, but the content was not commissioned. Sun Zhaozhi described the approach as resource-light but sincere: public posts preserved real feedback, while the team invested more effort behind the scenes to serve users, fix bugs and discuss new play.
JD.com reached out after the trade show, judged that both the product and pricing had sales potential, signed a strategic cooperation framework and provided resources. Because JD uses a direct procurement model, it purchased a sizable order rather than merely providing a listing page.
Tmall and Xiaohongshu stores will open later, while the overseas version will come later still. Even users who manage to buy it overseas now may experience latency; a formal international launch requires changing the underlying models and cloud services rather than copying the domestic version directly.
22. Capacity, Not Orders, Limited First-Half Sales
Both the team and JD.com were satisfied with the launch, but Sun Zhaozhi framed the question as “how long until it sells out,” not whether it would sell out. After several replenishments, inventory would quickly run out again, and stockouts could continue from June through August.
Hardware mass production has to ramp by batch, with each batch used to summarize quality issues and optimize the next one. Larger batches are expected to leave the factory in August, with supply potentially becoming relatively adequate in August and September. The team did not place an oversized one-time order to chase the launch window.
The specific shipment target for this year remains undisclosed; the only stated goal is to become China’s No. 1 AI companion hardware product. Sun Zhaozhi even judged that JD.com’s purchase volume alone might support that ranking, before direct-to-consumer sales are counted. His aggressive summary was that “globally, it is mainly China,” based on Japanese products being expensive and low-volume.
23. miHoYo, Pop Mart and UBTECH Are Entering the Same Market from Different Directions
The host said miHoYo, Pop Mart and UBTECH are working on or positioning themselves around related opportunities. Sun Zhaozhi said miHoYo has very strong IP and that he had learned indirectly that Pop Mart was also interested. Each participant brings game IP, collectible-toy distribution or robotics technology, confirming the category’s cross-industry nature.
The host questioned why Pop Mart would enter electronic hardware with lower gross margins. Sun Zhaozhi compared it with traditional automakers facing new-energy vehicles: even if the old category has better margins, companies cannot ignore a new category that may expand. “At least you should try to understand what it is.”
Domestic internet giants currently lean more toward being platforms. ByteDance once worked with suppliers on plush products with a gift-oriented positioning, but Doubao would rather serve as a model, cloud and services platform while partners build the specific hardware.
Overseas signals include OpenAI’s own companion-hardware concepts and its partnership with Mattel. Sun Zhaozhi believes that once a leading technology company formally launches a product, more major players will enter.
24. Big-Tech Resources Do Not Automatically Translate into Hardware Wins; Brand Is the Endgame Moat
miHoYo has IP but is not a hardware company; Pop Mart has IP and distribution but had not previously entered AI. Sun Zhaozhi noted that ByteDance, Alibaba and other major companies have entered hardware before without automatically eliminating startups.
He characterizes the emotional-value market as “fragmented.” Productivity tools may trend toward winner-take-all based on performance and price-to-value, but companion forms, aesthetics and personality preferences are inherently diverse. Validating one pom-pom does not mean every user will choose the same character.
Appearance is easy to imitate, so startups must create distance through faster new forms, social networks and accumulated data. The ultimate moat remains brand, because the product is meant to be worn outside the home: “You definitely want something that isn’t a knockoff.”
His ideal position is not monopoly, but an original brand that is “not necessarily the largest, but very cool.” Super-platform companies will occupy important positions, while distinctive independent brands can also last.
25. On-Device Models Could Simultaneously Rewrite Product Form, Privacy and Unit Economics
The industry milestone Sun Zhaozhi watches most closely is on-device models moving from smartphones into small AI hardware. His aggressive timeline is “within a year if things move quickly,” and the team is already working with model and chip companies on next-generation solutions.
The first change would be the business model: with a local model, costs shift to the chip the user has already purchased, potentially breaking the recurring cloud-inference model. The second would be the experience: inference would not need to travel to and from the cloud, reducing latency while keeping private data from being continuously uploaded.
The more distant vision is “one model per person,” with models that can be cultivated and learn online. But Sun Zhaozhi emphasized how early this remains: the Transformer architecture may not be suitable, and once centralized control is lost at the edge, “you don’t know what each person will end up cultivating.”
He compared this inflection point with autonomous driving moving from L4 to L5. Before the transition, progress comes through incremental improvements in limited scenarios; after it, product form, business model and addressable user base could all change by an order of magnitude.
26. Sun Zhaozhi’s Career Has Been a Search for New Product Forms during Technological Change
He spent more than 10 years in automotive design: serving car brands in Paris early on, joining XPeng in 2015 to lead interiors, then managing vehicle product development at Didi, moving to robotics products at Pudu Robotics, and founding Robo Intelligence at the end of 2023.
The shift from 46 days of paid leave in Paris to working through consecutive nights during his busiest period in China was a total change in lifestyle. He admits that during those overnight stretches he sometimes wondered whether it would have been better not to return, but he never believed for long that the decision was wrong because his learning and growth accelerated.
He was not yet 30 when he returned to China and wanted to “leave behind a few things to brag about to my future grandchildren.” His explanation of entrepreneurship is calmer now: it is a way of life that lets him organize resources and talent to build products he believes in even when many others do not.
27. D1, Landscape Screens and PX5 Shaped His Method of Choosing Trade-Offs from Designer to Product Manager
At Didi, he worked on the D1 product-definition document. The team did not copy the then-popular Corolla, instead choosing a One Box layout, a large cabin and a single-sided power sliding door. The door increased the bill of materials but reduced door collisions, vehicle damage and platform disputes, giving the function value beyond a purely aesthetic debate.
Early at XPeng, the team debated whether the center screen should be horizontal or vertical. Sun Zhaozhi supported horizontal orientation because human vision is naturally horizontal; the lower part of a vertical screen is hard to see and cannot be operated blindly like a physical button. Model 3 later adopted a horizontal screen, and the industry gradually converged in that direction.
Pudu Robotics’ PX5 explored how a robot could feel safe. The team referenced Link’s proportions in The Legend of Zelda, making the head larger, the body less elongated and the height slightly shorter, while deliberately making it “less human” to avoid the uncanny valley.
These projects moved his focus from curves and form to complete trade-offs: first understand what users care about, then calculate the technical boundaries, cost and commercial chain together. “Design is only one part of a product, and a product is only one part of a business.”
28. Two Early Startup Attempts Confirmed That He Is “Entrepreneurial for the Product”
In 2017, he tried a lifestyle brand with friends, but the team never truly got started. In 2019, he founded Woolly Planet to develop an automatic cat-litter box; the family cat used the demo continuously for three months and the experience was excellent.
When the litter-box project entered financing, supply-chain and mass-production stages, the pandemic intervened. He later sold the design, patents and other assets to industry veterans, and the final product never reached market. The experience confirmed that the product definition and user selection were precise, and pushed him to Pudu Robotics to deepen his technical understanding.
He did not choose to remain an independent designer or run a design studio because product ideas kept appearing in his head, and realizing them necessarily involved teams, capital, supply chains and commercialization. “I’m not starting a company for the sake of starting a company. I’m starting one to build a product I want to build.”
29. The Advantage of a 30-Plus-Person Team Is the Interlocking of User, Product and Technical Roles
The current team has more than 30 people. CTO UNA was previously an architect for ByteDance’s Doubao LLM solutions and had also worked on emotional-companion applications at MiniMax; she understands the technology and is herself a young female user, contributing directly to form and gameplay.
Holly, who joined more recently, was previously CTO of Dream Island. Her strengths are broader internet execution—App front end and back end, model-effect tuning, growth marketing and strategic partnerships—filling gaps beyond a pure-model team.
Sun Zhaozhi says having no technical background spared him “technical bias.” His learning method is not to disappear into implementation details, but to take product questions to internal and industry experts, understand whether a requirement is feasible, which route is cheapest and how the boundary may shift in the future.
The team is currently busy but receiving immediate feedback: internal-test issues can be fixed quickly, and everything produced on the line can be sold. Sun Zhaozhi sees this intense “push from behind” as the ideal state of an early-stage startup.
30. The Variable That Still Matters Is Whether Loneliness Can Become Continued Use
The team says introverts make up a very high share of early followers. One introverted user deliberately chose the most lively character, Huozi, hoping it would complement her personality; Sun Zhaozhi came to feel that this was not merely a novel toy, but potentially “an island in deep loneliness.”
Even with more—or “unlimited”—resources, he says he would still choose this route: first because it is fun and cool, and second because more and more people are lonely, making companionship “worth doing very, very deeply.”
The host did not preemptively settle the product’s future. Children’s AI toys have already exposed the problems of novelty fading, low reactivation and insufficient repurchase; whether Fuzo Zozo’s adult positioning, cultivation mechanics and social features can materially improve retention still requires longer periods and real cohort data.
The larger question also remains open: technology will advance rapidly, while human nature changes slowly. Whether robots can truly comfort people—and whether that future is exciting or sad—cannot be answered by launch sales.