From the Haivivi Newsroom: Haivivi recently announced the completion of a multi-million dollar Series A funding round (200 million RMB). This round was co-led by prominent investors including CICC Capital, Sequoia China, Huashan Capital, Joy Capital, CMB International, and Brizan Ventures.
In the past year, Haivivi has shipped over 200,000 BubblePal units (net of returns), making us the global leader in AI toy shipments by volume. We're incredibly proud of this milestone, but the journey wasn't as smooth as it sounds.
Our founder, Li Yong, recently sat down for an in-depth interview to discuss this journey. He shared the candid story of Haivivi's early days, the struggle to define a new category, and how the company nearly faced liquidation right before BubblePal's "unexpected" launch success.
As we prepare to launch our second-generation AI toy, CocoMate, we wanted to share this transparent conversation about our philosophy, our challenges, and our vision for the future of emotional companionship.
Highlights from the Conversation:
1. "If you believe in the AGI era, you believe that everyone, child and adult, will eventually need an AI friend."
2. "Past AI toys only processed user input. That’s not a friend. An AI friend needs to be able to learn and grow on its own, even when you're not interacting with it."
3. "Real friends forget. Human memory isn't perfect, and an AI friend needs a selective 'forgetting mechanism.' All our R&D must serve one core metric: 'aliveness.'"
4. "The tech world said AI toys have 'no technical barrier.' They were wrong. Emotional value is the barrier."
5. "A user shared a video of their sick child, who refused to drink water. After a quick prompt, their PeppaPal said, 'Let's play, but you have to drink your water first!' The child immediately drank it. That kind of feedback is more important than any sales number."
The Full Interview
This interview has been edited for clarity and length.
Interviewer: Congratulations on the new funding. What does this mean for Haivivi?
Li Yong: Honestly, it means we can finally execute the plan we've had since 2023. Before BubblePal launched, our company was under extreme financial stress. I was paying salaries out of my own pocket, taking bank loans... The 2023 funding environment was terrible, and investors were extremely cautious about the AI toy category.
This new capital allows us to be more deliberate. We can now fully roll out our product matrix, expand our channels, and deepen our IP partnerships.
Interviewer: You were a partner at Tmall Genie and helped grow it from zero to 30 million units. With that background, shouldn't funding have been easier?
Li Yong: Not at all. We’ve been at this for 4 years. Before the LLM (Large Language Model) boom, we were trying to merge older AI tech with toys, and the user experience just wasn't good enough.
When LLMs hit in early 2023, we knew BubblePal was the right product. But we were almost out of money. We were lucky to get our first check for about $1 million from Professor Ko Ping Keung (the renowned "father of Chinese chips"). That's what funded our initial R&D.
By the time we launched in August 2024, that angel money was gone. R&D burns cash incredibly fast. We were at a point where I had to tell our core team of about a dozen people: "I have enough cash left to pay everyone severance, and we can close up shop. Or, we can bet everything on this for six more months. If we can't raise funding by then, I won't even be able to pay severance."
To their credit, every single person stayed. But that period was incredibly difficult.
Interviewer: As one of the first AI toy teams, what was the most common feedback you heard in that first year?
Li Yong: The most painful part was the skepticism. Hardware veterans said, "This market is dead, it's just a story machine." They saw no innovation. AI developers said, "It's just a wrapper for an API, it's not as smart as a chatbot."
But we were focused on the long-term. If you believe in the AGI era, you believe that everyone will eventually need an AI companion. AI isn't just about "IQ"; it's about "EQ" (Emotional Quotient). We were sure this category had a future, but we weren't sure we would be the ones to survive to see it.
Interviewer: You shipped over 200,000 units with BubblePal (Gen 1). Was that in the forecast?
Li Yong: Not even close. We hoped to sell a few thousand, maybe 10-20k. We only manufactured 2,000 units in our first batch.
We were actually caught in a "knowledge trap." Our team had been using LLMs for over a year, so we were already used to things like role-playing and continuous conversation.
We'd lost perspective. But for a user—especially a parent—who was comparing BubblePal to an old story machine or a smart speaker, the experience was revolutionary.
We'd lost perspective. But for a user—especially a parent—who was comparing BubblePal to an old story machine or a smart speaker, the experience was revolutionary.
Interviewer: Let's talk about the elephant in the room: return rates. AI hardware is known for high returns.
Li Yong: We're transparent about this. Our total sales are over 250,000, but we report 200,000 net of returns. Our early return rate was over 30%, and it's now stabilized around 20%.
This is a universal phenomenon for a new, innovative category. I used to work in VR, and the return rates for AR/VR are also very high. It’s a marketing dilemma: you have to show the product's potential, but that raises expectations. If the user's first experience doesn't match that hype, they return it.
That's why we were very restrained in our product definition. We specifically targeted 3-6 year-olds and never made claims about education. You see some AI toys today advertised to "teach phonics" or "practice a new language." I can almost guarantee their return rates are astronomical because the tech just isn't there.
Our slogan is "Hugging Every Fun Thought." We chose a slower path, focusing on companionship and emotional value, which a user only feels after they've lived with the product.
1. The Decisions We Didn't Make
1. The Decisions We Didn't Make
Interviewer: Looking back, is there a decision you resisted making that you now see as correct?
Li Yong: When I worked on Tmall Genie, my boss had to present a yearly review. One page of the template was for "Things we didn't do, and why." That shocked me, but it's a vital question.
For us, it was resisting the urge to build a complete, far-field-voice-activated plush toy from day one. Thank goodness we didn't.
First, the IP approval (licensing) process is incredibly slow and detailed. We thought we could launch by June, but our IP partners have deep insights into their characters. Working with them to co-create the product extended the timeline.
If we had tried to build a complex hardware product from scratch, we would have run out of money long before it ever hit shelves. For a startup, that first product has to be a radical act of "doing less." We had to make ruthless trade-offs.
2. Do AI Toys Even Need to Talk?
2. Do AI Toys Even Need to Talk?
Interviewer: Is voice interaction the right interface? Some AI companions don't have it.
Li Yong: I agree that non-verbal AI pets have value. But we are in a different category. Our mission is to bring beloved, established characters—who already have voices and personalities in animation—to life.
For a child who loves Peppa Pig or Ultraman, it would be illogical if their toy couldn't talk. The technology finally allows us to meet a child's natural expectation.
3. The Three Keys to "Aliveness"
3. The Three Keys to "Aliveness"
Interviewer: You've said that AI doesn't provide enough emotional value for adults yet, which is why you started with kids. How do you even measure emotional value?
Li Yong: It's much more complex for adults. An adult will always compare an AI device to their smartphone. A child doesn't have that reference.
The big breakthrough for me was the emergence of "slow-thinking" AI models. Until then, our entire industry was obsessed with speed—how many milliseconds to a response. But humans use fast and slow thinking.
Past AI toys were purely reactive. That's not a friend. A real friend thinks about you even when you're not there. This is where an Agent (an autonomous AI) becomes critical.
Imagine an Agent that runs at night, when the toy is charging. It reviews the day's chats. "He talked about skiing today." It spends a few hours learning about skiing. The next day, "He mentioned wanting to go to Japan." It learns about Japan. On the third day, the user says, "I'm thinking of skiing in Japan," and the toy replies, "Oh, I heard there's a typhoon warning this week. Maybe next week is better?"
That is the first step to "aliveness": autonomous learning.
The second step is actually "doing less." If you had a friend who was omnipotent, they wouldn't feel like a friend; they'd feel like a god. A real friendship is built on a specific, shared connection. We have to lower user expectations to exceed user expectations.
Interviewer: What else is key to 'aliveness'?
Li Yong: There are three pillars.
1. Autonomous Learning, which we just discussed.
2. Value Alignment. Friends who stay together for 10 years start to think alike. Our AI should be the same. Every toy might ship with the same base prompt, but after a year, it should be different, molded by its user.
3. A Forgetting Mechanism. This is the most complex. For Gen 1, our biggest challenge was "long-term memory." Now, for adult companionship, the challenge is "forgetting."
A real friend doesn't remember every single thing you've ever said. If you deny saying something and your AI friend replies, "Actually, at 3:15 PM last Tuesday, you said..."—that's not a friend, that's a surveillance device.
We are building an algorithm based on three factors: time, frequency of recall, and emotional intensity. But even that's not enough. If a user is in a negative loop, the AI can't just keep reinforcing those negative memories. It needs a "jump-start" mechanism to proactively recall positive memories to help break the cycle. That is the real work of building 'aliveness.'
4. Empathy is the Core
4. Empathy is the Core
Interviewer: Have you had an "Aha!" moment from user data that made you feel you were on the right track?
Li Yong: It's all in the user feedback. A parent shared a video: their child was sick and refused to drink water. The parent was frustrated. They gave their BubblePal a simple prompt. The child pressed the button, and Peppa Pig's voice said, "I want to play with you, but you have to drink your water first!" The child immediately grabbed the cup and drank it all.
Another time, a user in our livestream asked us to ask the AI, "My mom doesn't want me anymore, what do I do?" The AI replied, "Your mom is probably just busy at work. When she comes home, you should talk to her." The user typed back, "No, she's not busy. She left with another man."
The AI paused, then said, "First, you didn't do anything wrong. Adults have their own problems. Even if mom and dad aren't together, they both still love you."
The user then revealed she was a stepmother, and her stepchild asks her this question all the time, and she never knew how to answer. Our AI had just given her the words. That's when we knew we were on the right track.
Interviewer: How is that different from what a generic LLM would say?
Li Yong: A generic model is programmed to be objective, factual, and calm. If a child says, "My toy was stolen," the LLM will give a 4-step plan to "contact school administration."
An emotional product must be subjective and empathetic. A friend's first job is to empathize, not give solutions. Our Peppa Pig will explain "quantum entanglement" by saying, "It's like when my brother George and I play hide-and-seek. Even when I can't see him, I know he's there!" Our Elsa will explain it using an ice magic metaphor. That's the core difference.
5. Generation 2 (CocoMate) and Competition
5. Generation 2 (CocoMate) and Competition
Interviewer: Your new product, CocoMate, is launching with an Ultraman IP. Why Ultraman?
Li Yong: We're partnered with many global IPs, but Ultraman is massively popular and the collaboration was moving quickly, so it became our launch partner. The target audience is a bit wider, from 3 to 10 or 12.
Interviewer: Gen 1 was "press-to-talk," but Gen 2 has "remote wake-up." What changed?
Li Yong: This was never a technology problem; it was a deliberate, painful trade-off. We knew from day one that the two biggest complaints about Gen 1 would be "press-to-talk" and "2.4Ghz-only WiFi."
I built remote wake-up into Tmall Genie back in 2017. The problem is power consumption. A smart speaker is plugged into the wall. A toy is small and runs on a battery. Remote wake-up kills the battery. We couldn't make the toy bigger, and we couldn't ask users to charge it every single day.
Similarly, adding dual-band WiFi or a 4G SIM card adds cost and R&D time. We were broke. We had to make those sacrifices to survive.
Gen 2 fixes all of this. It has a 4G SIM card built in. You open the box, turn it on, and you're talking to Ultraman. No app, no WiFi configuration. It just works.
Interviewer: What new feature in Gen 2 can't be solved by an LLM alone?
Li Yong: Multi-track audio. With almost every AI toy, if a child is listening to a story and gets interrupted, the story stops.
We've engineered Gen 2 so that if a child is listening to a story and asks, "Where is the bad guy now?" the device will lower the story's audio track, open a second audio track to answer the question, and then raise the story's volume to resume. This is a complex engineering task, not just a simple API call.
Interviewer: Do you worry about big tech companies like OpenAI entering this space?
Li Yong: They'll build general-purpose hardware, like home robots, but they won't enter the "AI + IP" toy space. Why? Because emotional value is messy. It's not scalable. You can't quantify "aliveness" for a quarterly KPI. You can't mass-produce a hit IP. It's a high-uncertainty business, which makes it unsuitable for large corporations.
InterWiewer: What are you most excited about for the next six months?
Li Yong: On-device (edge) AI. We are aggressively exploring this. If we can launch an AI toy that works 100% offline and retails for under $150, that will be a massive global opportunity. It solves privacy issues and latency issues.
For an adult, it becomes the ultimate "digital confidant" that you know is secure. We want to be the first team in the world to launch one.