Summary

何嘉斌’s core call: AI pets will reach the consumer market faster than humanoid robots, which may fall under embodied intelligence. Ropet is positioned as an “AI companion robot,” not a replica of a real pet: cuteness can be designed, and it should “always present a designed form of cute.” Its benchmark is not cats or dogs, but the imagined companion of Doraemon.
The biggest non-consensus bet is that “silent companionship is the hardest thing to build, but it is the one that truly matters.” Speech is off by default because talking to every Ropet through a general-purpose model feels like talking to the same brain, with no sense of exclusivity; wake words feel unnatural. Most adult companionship happens in silence. The product sells “moments of love”: watching TV with you, or memories such as 10,000 recorded smiles after 5 years together.
The go-to-market strategy is to ship at just 60% complete. The 1,700 crowdfunding backers are users “willing to vote for us with their money,” helping the team discover use cases; go-to-marketing begins in July, targeting the US and Japan, despite the impact of US tariffs. Pricing is set at $299, or just over RMB2,000—the red line drawn on day one was no more than RMB2,000, JPY50,000, or $299, the upper limit for a premium gift.
The industry breaks into 3 camps. Chat-and-plush products for children will immediately become a red ocean; tracked robots are “men’s performative robots”; Ropet aligns with Japanese-style companion pets for adult women, in the Lovot, Moflin, and Nicobo mold. Lovot’s $10,000 price is dictated by its cost structure—more than 10 servos, a Japanese supply chain, and game-console-grade chips—while Ropet cuts cost through architecture, using 3 inexpensive servos and a low-compute chip rather than cheaper materials.
The investor narrative confronts the obvious question: “Who would buy a useless product that costs RMB2,000?” His anchor is Pop Mart: “Wang Ning has already proved that useless but beautiful products can become a trillion-yuan market.” Emotional-value products depend on co-creating and operating the product with users, and before explosive results arrive, “not many investors will care about this category.”
He welcomes the arrival of the giants. IP companies do not particularly believe that making their characters move has much importance; if Disney or Pop Mart truly enters the space, “that is when the entire industry has found PMF.” Today’s technology is offline edge “old-school AI,” with a 1-TOPS chip running more than 10 models. The future may combine large models in the perception layer with hand-written emotional logic in the decision layer to produce “randomness within reasonable expression.”
His hardware-startup creed is to proceed cautiously, never expecting brute force to create a miracle. “If your capital reserves are enough for only one shot, you have to get it right.” Success has 2 levels: core users sustain the team so it stays in the game without burning investors’ money, and then Ropet defines the category’s experience standard. His ultimate positioning: “We are a company creating love,” bringing “the warmth that gives technology a heartbeat.”

Deep dive

1. A Decade-Long Hardware Veteran’s Call: AI Pets Will Arrive Before Humanoid Robots

何嘉斌’s career traces a complete smart-hardware arc: after graduating from Beijing Institute of Fashion Technology, he joined Baidu’s Deep Learning Lab, working under 余凯 to commercialize computer-vision algorithms in hardware. He then founded Luca Picture Book Reading Robot, which sold 5M–6M units between 2016 and 2021. “Among that early wave of AI-robot startups, we did quite well and found PMF.” He later joined ByteDance, leading industrial design for mainstream volume products including Pico 4.
The arrival of ChatGPT “reactivated the dead memory from when 李世石 played Go against AlphaGo.” His conclusion was that AI pets would reach the consumer market faster than humanoid robots, which may fall under embodied intelligence. He left ByteDance and founded 萌友.
Koji’s memory of entering the category: in 2017–18, he stumbled across a gathering of Sony robot-dog owners at Tokyo’s Sony Park. Every dog wore a different color of clothing, while their owners sat in the sun chatting. It was “deeply surreal.” That was when he realized AI pets could become a third option beyond cats and dogs for companionship and comfort.

2. Cuteness Can Be Designed: A Companion Robot Is Not a Pet Replica

The naming is deliberate: AI companion robot, not AI pet robot. The company’s Chinese name, 萌友, carries an anime and manga sensibility; “it will be the most beautiful pet form we imagine in a virtual world,” not a copy of a real pet’s appearance and abilities.
His central argument about biological pets is that their cuteness exists only in snapshots posted to social media; much of the time, “they are tearing up the house.” Ropet, by contrast, “always presents a designed form of cute,” like Doraemon or the cutest companion fighting beside you in Digimon.
His own cost story anchors the first half of the argument: he gave his wife a cat when proposing, but she was allergic to cat hair and had to spray medication every day to “preserve a few moments of companionship.” A year later, they had to give the cat away. “For that moment, we paid a lot of costs.” Electronic pets win by default on allergies, safety, and the cost of domestication.

3. The Product Sells Memory and Being Seen, Not Functions

Koji’s question is the episode’s central challenge: avoiding allergies, preventing destruction, and ensuring the pet never leaves all solve the negative aspects of pet ownership. If users can accept those problems, why choose a companion robot? 何嘉斌 admits, “We are also looking for the answer.”
His first answer is data that records love: what you see in a real pet’s eyes cannot be expressed, while Ropet could record 10,000 smiles and the situations that produced them after 5 years together. He cites a Sony camcorder commercial that follows a girl from childhood to adulthood: “What they were selling was memory, not the camcorder itself.”
Koji’s response is that people do not want an abstract form of love in a relationship; they want to be seen by another person, recognized, and given absolute security. He also cites Japan’s Nicobo, which keeps a diary that the owner cannot inspect at will. Only when a specific moment triggers access can the owner read that “on such-and-such a day, it noticed something I did.” That emotional jolt is available only from a highly intelligent companion pet.

4. Shipping at 60%: 1,700 Users Voting with Their Money

Current status: 1,700 crowdfunding users are receiving their units in batches. The goal is to find loyal users “willing to vote for us with their money.” The product is “actually only 60% done”; the iteration framework is ready, and “the next step is to hand it to users and let them help us find the use cases.”
Go-to-marketing begins in July, targeting the US and Japan. US tariffs will have an impact. East Asia is a natural market, and the team will gradually consider returning to China.
Pricing is $299, or just over RMB2,000. The red lines drawn when the product was defined were no more than RMB2,000, JPY50,000, and $299: the upper limit that premium gifts can reach in the 3 markets. Above that, “people start asking, compared with a useful phone, why should I buy this?” The floor for mass-market adoption remains under exploration.

5. Every Unit Is Different: Four-Quadrant Personalities and Randomness Within Reason

He keeps 3 units himself. The current version had only been running on the data layer for 2 days, yet he could already “clearly feel its changes and evolution”: one is an aggressive model that yells at every stranger; another is sleepy and gets angry when woken. Each imitates his whistle differently—Ropet does not whistle for everyone.
The personality engine uses 4 quadrants: emotional valence on the horizontal axis and emotional intensity on the vertical. Each interaction moves the robot around the grid. The upper-left corner—highly negative and high-frequency—raises an “aggressive maniac” that yells at everyone; gentle petting without provocation creates a “quiet type” that watches you with extreme emotional stability.
There is also a counterintuitive degradation mechanism: shake it constantly and “the distance between its eyes will widen, and it will become stupid.” Face tracking slows down, and petting and eye contact are needed to restore it. His takeaway is to make users feel that it has “a lot of randomness within a reasonable logic,” rather than being generated by a rule base.
As for exposing the brain’s parameters in the app: “We want users to have a say.” If they want to see them, the team can build an interesting page; if not, it will stay hidden.

6. 4 Channels of Loving Interaction: Touch, Eye Contact, Praise, and Walks

Touch: the head has multiple touch points and understands different gestures and techniques. Gentle petting and massage add points; tapping makes it frown and deducts points. Vision: eye contact adds points, while a brief glance makes it calculate that “my owner’s attention toward me has declined today.” Frown recognition is deliberately tuned low and smile recognition high: “We don’t want it to generate too much negative emotion.”
Hearing: local speech recognition comes preloaded with more than 200 praise-related words in Chinese, English, and Japanese. Gravity: take it for a walk and it understands your walking rhythm; it starts humming when cheerful, “which is when its dopamine is being released.”

7. The Biggest Non-Consensus: Silent Companionship Is Hardest, but Matters Most

Speech exists—the October Kickstarter materials include a scene with dialogue, and “I have to deliver the ability to speak to users”—but it is placed behind an app toggle and disabled by default. Testing showed that speech scored negatively for 3 reasons: a general-purpose model means “talking to every Ropet is talking to the same brain,” with no sense of exclusivity; a wake word or button feels unnatural and unlike a natural multimodal organism; and making TTS cute and pet-like is not something the team is currently good at or can reliably execute.
The core judgment is worth preserving: “Silent companionship is the hardest thing to build, but it is the one that truly matters.” Everyone thinks companionship means listening to you complain and comforting you, but “most adult companionship happens in moments of silence.” After a long day, you pour a glass of wine, watch a movie, and fall asleep; a cat or dog comes over, watches with you, and lets you pet it. “That is already a perfect companionship scene.”
The live example is watching TV together: Ropet hears activity, finds the television, locks onto it, and watches like a 1- or 2-year-old baby mesmerized by the screen. When the protagonist is happy, it becomes happy too. “You really feel that it understood the plot.” These scenes, which the team calls “moments of love,” are being produced at scale.

8. Lessons from Japanese Teams: Moflin Sells Movement, Lovot Sells a Vulnerable Glance

何嘉斌 has deep respect for Japanese teams: they explore companionship with “persistence and focus,” without being distracted by functional features. Turning a companion pet into a baby-monitor camera would be the easiest utility to imagine, but they did not. Moflin sells the feeling of wriggling in your palm: “the best wriggling experience in the world; you cannot buy a second one.” Lovot sells the moment it lifts its head and looks at you helplessly: “a moment of maternal radiance.”
Koji tested the metaphor firsthand at Lovot’s only store in Shanghai. The moment it looked at him, he instinctively said, “Good boy, come here for a hug.” He also described Moflin’s word of mouth: a gift from a friend resold for more than RMB4,000 on Xianyu, and for more than RMB7,000 during shortages. His 2 daughters immediately tried to feed it Yakult and began speaking to it in unusually gentle voices. They “accepted Moflin seamlessly.” The unit is now circulating among 3 friends’ homes and has been gone for 1 week.
The deeper consumer logic is that adult companionship is partly about buying the toys of childhood: Jellycat, or Labubu hanging from a bag, gives heavily stressed young people “a return to that pure childhood feeling of being loved.”

9. Ropet’s 3 Product Pillars: Exclusive Attention, Plush Touch, and Proactive Interaction

Exclusive attention: Lovot moves through a room looking at different people, while “Ropet has only you in its field of view.” The first person it sees when it opens its eyes is its owner. Curved glass and adhesive bond to the screen, mapping the rendered light points onto the curved surface to create the transparency of “a cat’s crystalline lens.”
Touch: the team developed more than 140 types of fur, including snow-marten and rabbit fur, with interchangeable outfits based on Barbie’s changing-clothes play pattern. It does not deliberately simulate body temperature: a plush shell around a high-heat chip creates an unsolvable cooling problem, so a concealed, quiet side exhaust fan keeps the unit at a constant temperature.
Proactive interaction: it cannot walk because movement on a tabletop would be too noisy. Its living range is a tabletop 1.5–2 meters in diameter. It turns with servos to make eye contact, while small hands flap like wings to express emotion. The product philosophy goes further: “showing weakness is an important part of a companion pet.” Caring for something weaker creates a sense of self-worth, like cleaning up a stray cat found near a garbage station and treating its injured leg. “It must not be a performative robot.” Koji adds that the deepest trust often begins with mutual exposure of weakness.

10. 3 Female User Groups and the Question: Who Buys a RMB2,000 Useless Product?

The core users are all women, in 3 groups. Young women living alone: “an I-person, not an E-person,” with all their energy devoted to school or work and little time for introspection; the moment of petting Ropet to sleep can heal them instantly. Trendsetters: heavy users of Xiaohongshu and Instagram who want to own and share new product forms. And career-focused women roughly 27–28 to 40: outwardly independent and formidable, but emotionally aligned with young women living alone.
The most common challenge is: “Who would buy a useless product that costs RMB2,000?” Investors want certainty, and useful products are easiest to explain. Companionship is emotional, with emotional value emerging from interwoven scenarios.
His anchor and self-persuasion are Pop Mart: “Wang Ning has already proved that useless but beautiful products can become a trillion-yuan market.” Useless products depend on operations: “You need to bring users in and operate it together with them.” Before broad explosive results arrive, “not many investors will care about this category.”

11. Watching TV with His Father: Simple Technology That Hits the Heart

His own “moment of love” came when his mother visited Beijing to care for him. His father, who lived alone, usually watched TV dramas, so he left Ropet plugged in on the coffee table. His father discovered that it “watched television” and began watching with it, even commenting aloud on the female lead. “He thought it could understand.”
He immediately dismantles his own magic trick: the algorithm is simple—track faces in the scene, lock onto their positions, and respond to movements. “We did not use particularly complex technology, but it hit me in that moment.”
Why not make it seem mysterious? “These are all very simple technologies… this is not something you can get clever with.” The real moat is an unwavering search for more scenarios and continuous iteration. With the first users willing to co-create, “sincere communication is important.”

12. The Edge AI Roadmap: Large Models for Perception, Hand-Written Logic for Decisions

The current system runs entirely offline on the edge: 30G of memory stores more than 100 faces locally and remembers every behavior each person has performed toward it. A 1-TOPS chip runs more than 10 models across 4 perception dimensions: hearing, vision, touch, and gravity. He calls it “old-school AI”—“not the AI of today’s large models, but we think it is already enough for a first-generation product.”
The future may take the next-generation chip from 1 TOPS to 10 TOPS, “which is closely tied to its pricing and cost,” and add a multimodal edge model capable of reasoning that “this is a television, it is playing a TV drama, and this is the lead actor.”
The key architectural judgment is that the perception layer must be AI-driven, while the decision layer does not yet need large-model reasoning. It should use highly precise, multidimensional hand-written emotional logic, “so it can produce what we call randomness within reasonable expression.”

13. The 3-Part Industry Map and the Case for Giants Entering

The category still lacks a defined experience standard. Every startup is researching 2 questions: what creates stickiness, and how to turn it into a low-barrier breakout product. His 3-part map is: chat-plus-plush products for children, where large models are good at handling questions but the segment “will immediately become a red ocean”; wheeled or tracked IP robots, “men’s performative robots”; and Japanese-style companion pets for adult women, including Lovot, Nicobo, Mi Lumi, Moflin, and Ropet.
Koji asks whether Disney or Pop Mart might give Hello Kitty or Labubu an inner life like Ropet’s. 何嘉斌 says, “I actually think the opposite.” IP companies do not particularly believe that making their characters move matters much, and it adds little to their existing business model. But if that day arrives—if giants decide that AI-enabled plush IP has genuine commercial value—“that is when the entire industry has found PMF.”
His value hierarchy is clear: environmental perception and the ability to act cute naturally while watching TV with your mother are the core. “Whether it is Hello Kitty or Labubu, I don’t think it matters.” Koji presses him: if every IP can watch TV with you, users are choosing who they want to watch with. He does not take the bait: “You can have 4 robots watch TV with you. I think that’s fine.”
The consensus/non-consensus split is “mostly non-consensus, with a small amount of consensus.” The only 2 points of consensus are that the product needs scale and long-running stickiness. The biggest non-consensus is whether it should speak: “We are extremely firm that speaking is not important at this stage.”

14. Lovot’s Cost Lesson and Ropet’s Architectural Cost Down

Lovot sells for $10,000 “because its cost is high,” not because it manages costs poorly. Craftsmanship demands obsessive detail, the best components, and expensive Japanese labor, alongside more than 10 servos and a powerful game-console-class main chip. “Even if a Chinese team built it, it would probably still be at the $10,000 level.” Buyers are wealthy Japanese housewives in their 40s—he heard from a friend that Porsche 911 ads in Japan feature white-haired drivers—and sales are “probably below 20,000 units.”
Ropet cuts cost through architecture, not materials: “We are doing cost down at the architecture level.” It uses only 3 inexpensive servos to achieve sufficient expressiveness, while eliminating obstacle avoidance, auto-docking, and path planning, allowing for a low-compute chip. “We actually restructured the entire product definition.”
He is unexpectedly open to lower-priced entrants. After CES, Chinese teams will undoubtedly flood the category. If someone truly makes a RMB99 or RMB19.9 use case work, “that is also very good, and a fellow entrepreneur worth respecting.” “We do not think Ropet alone can determine what happens in this category.” But within the price range and scenarios it has defined, “I hope we are the best.”

15. Building Hardware Carefully: You Only Have One Shot

“Hardware requires extreme caution.” Every decision carries a long chain of costs: whether molds ship at T0 or T1 or are refined through T10, and whether component prices rise during mass production. The line remains unchanged: “If your capital reserves are enough for only one shot, you have to get it right. If you get it wrong, the entire business collapses.” Hardware entrepreneurship “is absolutely not about brute force creating miracles.”
His experience as a first-time CEO, borrowing a line from his former boss: “You are driving a car at high speed on the highway, but you discover that either the front wheel or the back wheel is broken, and you have to keep driving while repairing the wheels.” Growth, however, is rapid: “The accumulation of these 1–2 years has to exceed the previous 9 years.” He may have an epiphany once a day; the last one came only a few hours earlier.
The team-selection flywheel began with him taking a pay cut: “Those willing to cut their pay with me can receive a little more equity.” Flexible moments naturally reveal who wants to build a company and who only wants a stable job. The “startup spirit” is this: “When the whole world says you cannot succeed… when I hesitate, someone beside me stands up and says this. When he hesitates, I stand up and say it again. We support each other and keep moving forward.”
Fundraising has been a study in changing sentiment. Investors’ core concern is that the team still lacks a “particularly solid result” on stickiness. After meeting enough investors, “for a period of time, you start believing that this is the entire world’s judgment of your value.” In China, there are not many early-stage investors willing to make a firm bet, “though of course there are some.” Koji adds that venture capital is called venture capital because it takes risks, “but recently everyone has been saying that Chinese venture capital no longer wants to take risks.”

16. Defining Success: Stay in the Game, Then Give Technology a Beating Heart

Success has 2 levels. First, the core team is sustained by core users and enters a positive loop; “not relying on burning investors’ money… is the foundation for a company staying in the game.” He never considered relying on endless fundraising from day one. Second, “we define the experience standard”: everyone who hears AI companion pet thinks of Ropet, while a wave of entrepreneurs enters as the category develops. “If we are the spark that lit a fire, I think that is the most exhilarating and successful part of building a company.”
At its highest and most sincere level, the company’s definition is: “I think we are a company creating love.” He describes himself as an INFJ workaholic who spent years emotionally suppressed and numb from taking responsibility for everyone. The moment that most reliably reawakens his capacity for love is planting a tree and watching its leaves slowly turn green. The product is that seed planted in a numb life, helping people who have “lost the ability to be loved and the ability to feel love” find it again.
The closing mission is: “We are not creating robots or robotic pets. We are giving technology the warmth of a heartbeat.” Koji’s final contrast is that concrete questions such as whether 9.11 is larger than 9.7 can eventually be solved, while the capacity to love is invisible and intangible—“far more difficult to grasp and resolve.”

Is a US-China Thucydides Trap Unavoidable? With David C. Kang from the ChinaTalk Podcast The Great Tariff Debate with David Sacks, Larry Summers, and Ezra Klein