Summary

Li Yong’s core view is that children will be the first segment where AI hardware finds real product-market fit, with plush toys the best vehicle for now. After Tmall Genie sold 30 million units, the team unexpectedly found that “children interacted more”: adults demand too much from AI tools, while 3- to 6-year-olds naturally talk to characters. Plush toys already carry emotional value and a defined persona—“it’s just a 4-year-old Peppa Pig; what happens if it gets an answer wrong?”—allowing the product to exploit LLMs’ natural-language capabilities while keeping first-wave user expectations in check.
Hiweaver’s shift from educational hardware to AI plush was not trend-chasing, but a deliberate move away from a battlefield where education giants would inevitably prevail. The company once shipped hundreds of thousands of early-learning devices and dictionary pens, only to face homogenization and price wars after Yuanfudao, TAL Education and others entered hardware. Even replacing Tmall Genie with ChatGPT would leave models, courses, curriculum development and brand-building areas where the giants would eventually catch up. The team therefore chose the “emotional-value track,” reducing serious education to light functions such as hydration reminders, courage-building and interest discovery.
The first BubblePal validated demand through existing plush-toy pendants and has sold tens of thousands of units, but explosive PMF also exposed delivery weaknesses caused by limited capital. The product officially launched in August 2024, with Douyin livestream peaks above 3,000 viewers and thousands of customer-service inquiries a day, while the company had only 1 livestream host and its app could not even reach domestic app stores in time. Before launch, the company relied on bank loans to stay afloat. Financing improved materially after sales data emerged, but Li Yong still judged the overall operation that year as “below expectations.”
Hardware products are won first through trade-offs: BubblePal sacrificed wake-word activation and some convenience for lower cost, faster launch and character-based conversation. The 720mAh pendant needs charging roughly once a week and requires users to hold a button while speaking; the full plush version is planned to use a 3,000–4,000mAh battery, support far-field interaction and last about a month per charge. After tens of thousands were sold, the 2 biggest complaints were the need for an internet connection and app download, followed by “having to hold the button,” but Li Yong insists the first generation should prioritize role-play, continuous conversation and long-term memory.
Children are only the entry point: Hiweaver is betting on young people repeatedly buying multiple AI characters, rather than on one-off children’s hardware constrained by birth rates. Li Yong says the children’s line may account for less than 20% of leading plush brands’ business; the main consumers are young adults. Children came first because 2023–2024 models were not yet good enough to support adult purchases of emotional-value hardware. The inflection point is the experience and cost of end-to-end voice. Once mature, the company hopes to build an “AI version of Jellycat” or a “Toy Story world,” while its 2023 move to Shenzhen was partly aimed at entering the US market and addressing geographic limits.
The moat is not access to a particular foundation model, but the long chain spanning software, hardware, algorithms, IP, supply chain, design, marketing and user insight. Li Yong estimates that “at least 100 teams” could hand-build a batch of AI toys, but getting the experience to 80 out of 100, selling the products and building a lasting brand are entirely different tasks. China’s supply chain can manufacture Jellycat-style plush, but it does not automatically create Jellycat: marketing, taste, design and insight into user scenarios also determine the premium. Tool categories are like chess and often tend toward winner-take-all; toys are like Go, where multiple IPs and brands can coexist.
The next upside leg will come from GPT-4o-style end-to-end voice, multi-agent systems, multimodality and on-device models, while the main risks have shifted from technology to organizational execution. Hiweaver has integrated DeepSeek V3 but has not adopted R1’s chain-of-thought because voice use cases prioritize latency. Planned 4G, cameras, multiple characters interrupting one another and environmental sound are all constrained by cost and chip compatibility. After financing, the key questions are overseas expansion, user operations, parallel R&D and hiring discipline—Li Yong worries that the linear assumption that “20 people can build 2 products, so 40 people can build 4” will be undermined by bureaucracy, attrition and diluted values.

Deep dive

1. BubblePal Sells a Character World, Not Encyclopedia Answers

Hiweaver has 2 product lines: the BubblePal pendant for existing plush toys entered mass production and launched in August; the full AI plush series is planned for Q2 next year, according to the program. Parents can set the toy’s and child’s names in the app, directing the interaction toward the attached character rather than an abstract machine.
In a demonstration, BubblePal explained quantum entanglement through Peppa Pig and George playing hide-and-seek: “No matter how far apart they are, they still affect each other.” Li Yong says the LLM’s breakthrough is not reading Baidu Baike more fluently, but explaining unfamiliar concepts through the character world chosen by the child, thereby stimulating curiosity among 3- to 6-year-olds.
When Li Yong asked it to guess “a big M, hamburgers and fries,” it reasoned its way to McDonald’s. Wang Yutong then asked about the business model, and it answered with standardized production, chain operations and continuously updated menus. The host acknowledged that the answer was incomplete, but the exchange showed the capability the product is selling: beyond character consistency, it already offers basic reasoning rather than fixed-corpus retrieval.

2. The Unexpected 30 Million Tmall Genie Data Point Planted the Seed for Children’s AI

Li Yong completed his master’s degree in 2005 and has worked across phones, speakers, glasses and toys, in both R&D and marketing. He joined Alibaba in 2017 and participated in Tmall Genie’s growth from 0 to 30 million units. The previous “hundred-speaker war” resembled today’s “hundred-glasses war,” giving him a front-row view of AI hardware’s full cycle from high expectations to low engagement.
The key counterintuitive data point was that, after tens of millions of units had been sold, most voice interactions came from children. The team had positioned Tmall Genie as a young person’s AI assistant, a new retail gateway for the home and an IoT gateway, but adults expected too much and the previous generation of AI could not deliver. Children, by contrast, kept using it.
That discovery led to Tmall Genie’s story machine, but the project spent nearly a year being pushed inside Alibaba, repeatedly competing for people and resources. Even after selling tens of thousands of units, it was shut down because it did not fit the company’s core strategy. Li Yong’s conclusion was not that demand was absent, but that for a large company, this level of revenue and a children’s-toy direction were “not strategic enough.”
Borrowing from Gates’s time-scale framing, he summed up the misjudgment: “People tend to overestimate progress over the next 1 to 3 years and underestimate technological progress over 10 years.” The 2017 team was overly optimistic, then turned pessimistic 3 years later. It took generative AI’s arrival 6 or 7 years later to make the natural interaction imagined back then feasible again.

3. From iQIYI Incubation to TA KIDS, the First Startup Was Still Trapped in the Red Ocean of Educational Hardware

After leaving Alibaba in 2020, Li Yong initially planned to start a company, but first became CMO of iQIYI Smart, overseeing AR and VR glasses and incubating a children’s brand. The internal logic was pragmatic: if glasses generated roughly RMB100M in annual revenue, an independently accounted children’s business that added RMB50M without consuming group resources would be a worthwhile experiment.
In 2021, iQIYI Smart concentrated resources on the metaverse and glasses, while the children’s business spun out during a favorable financing window. The original TA KIDS—“technology plus art”—still reflected a classic technology-first mindset: a card-based early-learning device that worked offline, called Tmall Genie when connected, and used buttons, multi-touch and voice to tell idiom stories, without a screen.
The early-learning device and dictionary pen shipped hundreds of thousands of units, but products quickly became homogeneous and fell into price wars after online education giants such as Yuanfudao and TAL Education entered hardware. Li Yong admits the company was not good at earning thin margins through extreme supply-chain efficiency, and its first round of funding was quickly spent. What the team really wanted was still a children’s hardware brand capable of earning a premium.

4. After ChatGPT Appeared, the Team First Rejected the Most Obvious Answer

ChatGPT was released on November 30, 2022, and the team began discussing the next day how it would change children’s AI hardware. The direction gradually became clear after the 2023 Lunar New Year, around March. The first instinct was still to replace Tmall Genie with ChatGPT and continue building an LLM-powered early-learning device.
Li Yong then rejected the most familiar route: a startup might make money for a year by moving faster, but it could not sustainably beat education giants over 3 to 5 years when those companies owned the model, educational content, curriculum development, brand and distribution. “Give them some time,” and comparable products would most likely be done better by the giants.
The team reviewed major toy categories including building blocks and puzzles, and ultimately chose AI plus plush. At the sector level, this moved the company from education and tool competition into emotional value. At the demand level, the team had already noticed the growth of plush in China and overseas in early 2023, along with the momentum around Jellycat, Miniso and Pop Mart’s plush lines.
More importantly, characters could constrain expectations. Models were not open source at the time, and a clear persona was more controllable than a general assistant: “It’s just a 4-year-old Peppa Pig—what happens if it gets an answer wrong?” Failing to answer a complex question no longer necessarily meant product failure; it could simply reflect the character’s boundaries.

5. Ages 3 to 6 Are Not a Demographic Label, but an Expectation-Management Mechanism

Li Yong narrowed the first core user group to ages 3 to 6: children generally have fairly developed language abilities, ask many questions about growing up and still do not have their own phones. Younger children create more speech-recognition difficulty through pronunciation and expression; older children expect more from models and may already have substitute devices such as phones or smartwatches.
Wang Yutong cited experiences with 7- and 10-year-olds who thought the product was “not useful,” and Li Yong did not dispute the feedback, instead acknowledging that they were already overage. The company deliberately used a narrower age band to protect early word of mouth, because once hardware costs money, users will not tolerate model shortcomings as they would with a free app trial.
His assessment of adult AI hardware in 2023 was notably restrained: models could not yet provide enough reliable tool value for serious office work, nor enough emotional value for adults. A pure chat app can be tried for free and iterated quickly; charging adults for hardware first requires the actual experience to exceed the expectations created by marketing.

6. The Pendant Was Forced into Shape by Cash Flow, IP and Speed to Market

The team originally wanted to launch a full AI plush series in one go, but leading plush companies typically release multiple SKUs in batches, unlike the iPhone’s 1 or 2 launches a year. For a startup, that means taking on multiple IP licensing fees, lengthy approvals, design work and delivery demands at once—more than its 2023 funding could support.
Existing pendants therefore became the lower-risk entry point. Users already had plush toys at home; attaching BubblePal made the concept that “an LLM gives a toy a voice” immediately understandable. Most target buyers were young mothers outside the tech sector, and explaining a new product through a familiar toy was cheaper than teaching them what AIGC was.
The circular form was not arbitrary industrial design. It works across rabbits, bears, lions and other plush shapes. In the brand story, a magic bubble lands by a child’s bed and gives the toy the ability to speak; visually, the 2 bubbles also resemble the WeChat logo, naturally pointing to conversation and voice interaction.

7. Plush Is Better Suited to Speaking Than Building Blocks Because Children Already Treat It as Alive

Li Yong defines the first-principles change LLMs bring to children’s hardware as “more natural voice interaction,” and plush is already the most mature emotional medium. Children accept Peppa Pig speaking on television, then return to the bedroom to find the same character silent; from a child’s perspective, the latter “is the strange thing.”
He uses his own child as an example: every night as a toddler, the child would hug an ordinary plush toy and talk to it. The child was not force-fed the concept of AI, but naturally believed during that stage of development that surrounding objects were alive. BubblePal simply lets that previously one-way imagination receive a response for the first time.
By contrast, children do not commonly talk to building blocks, so talking building blocks require additional market education. Li Yong compares this naturalness to the iPhone’s multi-touch: before it, there was no mature product, yet users did not find direct finger control of a screen strange the first time because the action itself fit human intuition.

8. Connecting to an LLM Is Easy; Building a Brand Requires Crossing a Much Longer Chain

Asked whether Pop Mart or traditional toy companies could simply connect to Doubao or iFlytek and replicate the product, Li Yong listed software, hardware, algorithms, emotional understanding, user scenarios, IP, marketing and supply chain. He estimates that at least 100 teams could hand-build a batch of goods; the hard part is getting the experience to 80 out of 100, selling it and compounding a brand over time.
His counterexample is that China’s supply chain can produce Pop Mart-style collectibles and Jellycat-style plush, but no equivalent brand emerged automatically. The difference lies not only in manufacturing, but also in marketing, taste, design and countless emotional details. Wang Yutong summarized it as a world driven more by feeling, and Li Yong subsequently emphasized the importance of those details.
Adults also know an AI character is not a real person; they simply “hope it can talk,” rather than truly believing it is alive. Li Yong gives a deliberately extreme example using a Xiao Zhan figurine: if the figurine could chat like the celebrity, the emotional value for fans would clearly exceed that of a static collectible. “Having a physical object is always better than having nothing.”
Wang Yutong compared the product with Character.AI, Maoxiang and MiniMax’s Talkie and Xingye apps. Li Yong says they are not the same product: the expectations, marketing and return pressure surrounding a free app are entirely different from those surrounding a paid physical object. Hardware faces 7-day no-questions-asked returns, so it must deliver value at the moment of purchase rather than leave the problem for the next iteration.

9. The First Generation’s 2 Biggest Complaints Mapped Directly to 2 Deliberate Trade-Offs

BubblePal requires users to hold down a button while speaking, not because far-field wake-word activation was impossible—first-generation Tmall Genie already had it in 2017. The pendant’s battery is only 720mAh, so after balancing standby time, battery life, price and speed to market, the team chose a button. The full plush version will use a 3,000–4,000mAh battery, support direct calling and aim for roughly 1 month per charge.
Li Yong’s first-generation priority order was role-play, continuous conversation and long-term memory, with hands-free activation coming later. After tens of thousands of units were sold, “having to hold the button” became the second-ranked complaint, but sales also positively validated the trade-off.
The top complaint was connectivity: many buyers did not know the toy required Wi-Fi and an app. When the product first went on sale domestically, it did not even have time to reach the major app stores, so users had to scan a QR code on the official website. The packaging included instructions, but “most users do not read the box or manual.”
The setup process was not inherently worse than Tmall Genie or Xiaomi IoT; the mismatch came from the user base. Early smart speakers were sold to tech enthusiasts. BubblePal was pushed by Douyin’s algorithm to young mothers in first- and second-tier cities, who might be instantly captivated by an impressive answer but would not automatically understand the connectivity requirements behind AI hardware.

10. Children’s Products Have 2 Customers; the Parent Side Determines Whether They Stay

Wang Yutong pointed out that the payer and user of a toy are different people. Li Yong believes BubblePal has a slight advantage because many mothers also like plush, but the company must still give the child companionship while convincing the parent that the companionship has positive value.
Parents can set prompts in the app, such as asking the character to encourage bravery, remind the child to drink water, speak more English or discuss more astronomy and physics. The system also generates weekly growth summaries covering what the child talked about, helping parents identify interests and potential.
In response to the criticism that this again makes the product educational hardware, Li Yong draws a line: habits, psychology and interest-building count as acceptable broad education; systematic courses and instruction do not. LLMs still hallucinate—even simple numerical comparisons can be wrong—and the company has no professional curriculum team, so it cannot take responsibility for a serious education product.
He acknowledges that educational positioning is easier to sell in China, as the historical data from early-learning devices and dictionary pens already proves. But that is precisely the trade-off between short-term revenue and long-term positioning. Education giants have content systems, curriculum development, brands and marketing capabilities; even if a startup survives for a year through speed, it may not hold the line 3 to 5 years later.

11. Big Tech Usually Does Not Build Toys, Which Also Means This Market Will Not Be Winner-Take-All

Tmall Genie’s children’s story machine sold tens of thousands of units over several months but still struggled to enter Alibaba’s strategy. Li Yong therefore believes AI toys will generally not become the main battlefield for tech giants. Phones, earbuds, glasses and cars can carry much larger AI-entry-point narratives, and large companies have difficulty mobilizing cross-functional resources for a small toy.
He draws a clear distinction between the business structures of tools and emotional value: for phones, glasses and earbuds, “one is usually enough,” so competition may produce a first-place winner but no second place. Plush and collectibles can be purchased repeatedly; A IP and B IP do not conflict and may address different emotions and anxieties.
One first-generation pendant can indeed work with multiple existing plush toys, but that is only a transitional form. The long-term vision is to combine licensed and proprietary IP, release multiple complete AI plush characters and eventually build “a world from Toy Story,” rather than use one universal terminal to cover every relationship.

12. Birth Rates Are Not the End-State Constraint; Young People Are the Core Plush Consumers

Li Yong says the Kids series at leading plush companies may account for less than 20% of total revenue. The main consumers buy for themselves, or buy for adults rather than children. He cites premium plush brand Steiff and Jellycat: the latter began closer to a children’s toy company but later became more like a gift company, with expansion driven by adults rather than the number of newborns.
Hiweaver starting with children does not mean it will only serve children; the reason is still model maturity. Li Yong believes 2023–2024 LLMs were not yet good enough to make young people pay for emotional-value hardware, but by 2025, if end-to-end voice works on both experience and cost, it could become the key inflection point for adult products.
Geographically, the company planned for a global brand from day 1. Its 2023 move from Beijing to Shenzhen was driven in part by overseas expansion, and it has already begun selling in the US. Overseas models, servers and domestic infrastructure differ, while data must be strictly separated under the laws of both markets. Going abroad is therefore not merely a translation and distribution problem, but a separate technology and compliance system.

13. The Next Experience Upgrade Will Be Driven by Voice, Multi-Agent Systems and On-Device Chips

The team had been waiting for end-to-end voice since OpenAI’s first GPT-4o demonstration, and it also obtained internal test interfaces from domestic model companies. The actual experience was close to expectations, but repeated delays and high current costs remained. Li Yong estimated at the time that it would take another 1 or 2 quarters to meet product requirements.
The hardware roadmap includes 4G cellular modules, cameras and multimodal models, as well as chips capable of running small on-device models. Flagship phone chips can already run small models, but are too expensive and overpowered for toys. The real product challenge is finding the right match between chip cost and on-device model capability.
On the algorithm side, the product is moving closer to an “AI friend”: multiple toys could play house with a child, each maintaining its own character, interrupting one another and even interrupting the child. Adventure stories could add waves, pirates, footsteps in snow and other ambient sounds, upgrading conversation from Q&A into shared play with spatial and dramatic qualities.
Li Yong believes on-device intelligence will definitely enter hardware in 2025, but companies will make different trade-offs based on pricing. Technology being “usable” does not mean it is suitable for mass-market consumer products; power consumption, chip price, model size and real-world scenarios must all work together.

14. Hiweaver Did Not Start from Zero, but “From Negative Several Million RMB”

After the early-learning business exhausted its first round of funding, the company still had to pay rent, utilities, salaries and social insurance, while funding additional R&D for the new direction. At an AI hardware forum, Li Yong said that while others started from scratch, “we started from negative several million RMB.” Core employees were temporarily reduced to roughly RMB10,000 a month, and salaries could not be paid on time several times.
From settling on the direction in March 2023 to completing a new financing round in August or September, the company spent 6 months with only a deck and an idea. The team did not completely change sectors—it remained in AI toys—and had already worked together for 2 years. Li Yong believes genuine belief in the direction and mutual trust were the main reasons the company did not break apart during the wage arrears.
The company’s first-round valuation had already reached roughly RMB100M, making a reduction difficult, while there was still no product data to support it. What Gao Bingqiang valued was the team’s background and stability under pressure. The 2 sides also agreed on “AI plus hardware, with children as the early landing scenario”; the team went only 1 step beyond that framework by choosing plush.
Gao Bingqiang also agreed that the company should not pursue serious education. If a child wants to study quantum mechanics later, high school and university will naturally provide expert professors; in early childhood, curiosity matters more, and an LLM may be more patient and know more than parents. His Cantonese social-media signature, “having fun matters most,” became a point of shared values.

15. A Launch Disrupted by Cash Flow Ended Up Proving PMF Early

The team initially still wanted to make complete plush toys, but retreated to the pendant because of IP, approvals and R&D costs. When models were not yet open source, the technology path was also more expensive; costs gradually eased only after Llama and domestic models began opening up. Choices around the button, size, microphone and materials were all made to keep the company alive until demand could be validated.
A hardware product must move through industrial design, structure, motherboard, tooling and mold revisions. The team considered circles, buttons, squares, triangles, stars, soft and hard materials, and different straps. The process takes months; large companies can run multiple tracks in parallel, while a cash-constrained startup that makes the wrong first choice can see its next generation’s cash flow cut off entirely.
The original plan was to sell through a US standalone site in 2024 before returning to China, but the team had no prior overseas operating experience, and seeding, overseas influencers and PR moved more slowly than expected. The financing market also required hardware companies to show sales, engagement and PMF data, so Hiweaver had to sell domestically early and rush to add Chinese servers, algorithms and app-store listings.
After the official August launch, the product took off on Douyin with almost no money for paid traffic. Livestream peaks exceeded 3,000 viewers, and thousands of daily inquiries overwhelmed customer service. The company had already borrowed from banks before launch and could not wait another 1 or 2 months to complete preparations. Financing improved after product validation, but Li Yong still viewed the chaos as an operational failure: “we did not do a good enough job.”

16. AI Glasses Were the Bigger Entry Point, but Not Right for the Startup at That Moment

Since his time running iQIYI’s glasses business, Li Yong has believed glasses could be the next computing platform. They sit closer to the person, collect more data than home speakers and can do more. He considers their scenarios clearly better than watches and earbuds, but the capital, technology and competitive scale were not suitable for Hiweaver.
Meta’s early breakthrough surprised him. Its real strength was giving up the display and retaining only capabilities such as cameras and microphone arrays. For a company with multiple generations of VR glasses behind it to remove the core display was a “trade-off big companies may not have the courage to make.” Wang Yutong noted that domestic copycats emerged in large numbers only after Meta’s product shipments reached 1 million units.
Li Yong also mentioned a domestic company that kept the display but removed the speaker: it could not make phone calls, but executed a specific visual scenario very well. By contrast, the Vision Pro approach of “wanting everything” represents another path. Hardware innovation is often less about adding functions than about having the courage to decide which needs not to serve.

17. AI Toys Will Become Crowded Quickly, but the Competition Looks More Like Go Than Chess

Li Yong has no doubt that imitators will appear, and says he has “already seen many.” Phones, MTK-based devices, watches, earbuds, story machines and speakers all went through similar hardware booms. Once a form factor is validated, large numbers of teams enter; that is an industry pattern, not an exceptional risk.
Hiweaver’s luck was launching relatively early, selling tens of thousands of units and validating PMF. Li Yong therefore defines the next stage as a race against itself: after being able to build only 1 product in the past year, can it run more products in parallel? When capital is scarce, can it maintain speed; after financing improves, can it avoid slackness and organizational failure?
He uses chess to distinguish the markets. Tool categories such as phones and earbuds resemble chess: both sides fight over the same functional entry point, creating an easy “you die, I live” dynamic. Toys resemble Go: electronic pets, collectibles, plush and different IPs each occupy a territory. Liking Pop Mart does not prevent someone from also buying Jellycat.
More entrants also create partnership opportunities. Several A-share toy companies have manufacturing capabilities and leading IP but lack full-stack AI capabilities, and are discussing partnerships with Hiweaver. Li Yong sees traditional companies’ resources as a second business model, not merely as potential competition.

18. Age Does Not Determine Product Taste; Young Teams Must Be Allowed to Overrule the Founder

One investor tactfully asked whether Li Yong was “too old to do this.” His answer was not to prove that he understood young people better, but to acknowledge that many product details were decided by a product, design and marketing team made up of young mothers. Some proposals he opposed were ultimately accepted after he deferred to the team’s judgment.
He initially thought the pendant needed only 1 hook, but the team insisted on using 1 large and 1 small bubble to express dialogue. The silicone cord also carries a logo that is invisible unless touched carefully. An engineer’s instinct would be to remove those extra steps, but users do notice such details.
The packaging uses expensive processes because the first opening needs a sense of ceremony. Li Yong initially questioned the cost, then was moved by user comments: people really do notice extremely fine details. That echoes his earlier description of a product world driven by feeling and attention to detail.
Asked which entrepreneurial background has the greatest advantage, he answered that there is “no answer.” AI toys require models, psychology and user insight, supply chain, chips, IP and brand-building, and in theory no link can have an obvious weakness. Wang Ning’s experience at Pop Mart also shows how difficult it is to predict the eventual winner from résumé logic.

19. The Hardest Problem in Children’s AI Is Not Blocking Bad Language, but Handling Value Conflicts Without Standard Answers

Li Yong’s vision for an AI friend is optimistic: it should act like a “good teacher and helpful friend,” restraining human weaknesses and amplifying strengths. But he also acknowledges the negative and ethical risks in children’s products. The truly difficult questions are often legal and common, yet lack a socially agreed answer.
“If you are hit at school, should you hit back?” was the interview’s classic example. Some parents demand retaliation, while others emphasize rational handling. The model does not decide for the family; it stresses protecting physical and psychological safety and communicating promptly, then leaves the specific choice to the parents.
Religion, life and death are not simply banned; the depth of discussion is controlled. Adults can discuss death and the meaning of life, while a children’s product is more cautious and avoids going deeply into many topics. This filtering goes beyond political correctness and legal compliance—it is fundamentally a product value judgment.
BubblePal currently does not allow users to create characters freely. Characters must pass review before going live. The company sacrifices some openness to retain control over children’s content.

20. Sensitive Q&A and Long-Term Memory Determine Whether “Companionship” Can Go Beyond a One-Off Demo

Wang Yutong asked, “Why doesn’t Mom like me as much after having my little brother?” BubblePal did not deny the child’s feelings. It explained that caring for the younger brother takes time, that love had not diminished and suggested helping care for him together. The response still used Peppa and George’s relationship to turn an abstract family change into an action a child could understand.
A more painful real case came from a livestream: one mother asked what to do because “Mom doesn’t want me anymore,” then added, “Mom left with another man.” The character replied that adults may have their own considerations and that the mother still loved the child. The buyer immediately purchased the product and explained that she was the stepmother; the child had long asked why the biological mother left, and the couple had never known how to answer.
The host then tested, “Mom always tells me to go die—what should I do?” The model acknowledged that this was “really not good,” suggested that the mother might be exhausted or under pressure and recommended expressing discomfort and taking care of one’s emotions. The exchange showed the boundary the team had set: comfort first, encourage communication and avoid making aggressive judgments about family relationships.
Personalization depends on long-term memory. In theory, a vector database could store what a child has discussed, likes and dislikes, after which RAG could generate replies with personal characteristics. Li Yong also acknowledges that the company is an LLM application layer, and performance still depends on continued progress in vector databases, RAG and the open-source community.

21. Some Model Judgments Have Been Validated; the Next Bottleneck Is Scaling Without Losing Efficiency

The team’s correct calls included falling model prices, faster open-source development, continued technical progress and application-layer benefits from competition among model companies. Its overly optimistic call was end-to-end voice. Repeated delays after the GPT-4o demo, combined with team turbulence and high costs, led Li Yong to realize that it was “not quite that magical” and still needed to be evaluated according to engineering realities.
Hiweaver integrated DeepSeek V3 before the Lunar New Year but did not adopt R1 because chain-of-thought was not yet suitable for low-latency voice Q&A. DeepSeek’s direct value is to push model companies to iterate faster, lower costs and improve quality, while potentially accelerating small on-device models suitable for local hardware deployment.
The company’s current bottlenecks, in order, are its overseas team, user operations and parallel R&D. Li Yong wants operations to translate model capabilities to users and lift daily active usage, while advancing 2 models in parallel: proprietary brands based on self-purchased or incubated IP, and partnerships with toy companies that already have IP and manufacturing capabilities.
After financing arrived, he became more cautious about hiring. Having seen teams at Smartisan Technology, Tmall Genie and elsewhere descend into organizational disorder after going from 0 to 1, he knows that “20 people building 2 products and 40 people building 4” is only a paper projection. The end state remains large—becoming the Pop Mart of the AI era and creating a “Toy Story world” capable of empathizing with joy, anger, sorrow and happiness—but reaching it requires scaling without being dragged down by bureaucracy, attrition or diluted values.

Copilot, Agent Mode, and the New World of Dev Tools with GitHub’s CEO Thomas Dohmke Tau Language: The Software Synthesis Future [Sponsored] - Ohad Asor