Vol.70 Momo Started a Company—and Chose What May Be the Most Competitive AI-Companion Market Right Now
Summary
Starkey is not betting on generic AI social, but on one-to-one “safe confidants” for urban women aged 20–35. The team defines its target users as women in the “Odyssey period” after leaving school, facing career, family and other questions with no standard answers. The product uses voice, a visible IP character and a personalized personality to provide friend-like companionship, while deliberately avoiding the positioning of a lover, pet, therapeutic tool or group social network. “The mindset of a private confessional and the mindset of socializing are strongly at odds”: once trust is broken, users stop disclosing themselves.
The project’s key product moat is not which foundation model it calls, but high-quality companionship data and long-term memory. 陈腿毛 believes high-quality companionship data does not exist in the internet’s public corpora, so the team recruits “high-EQ people” to create conversations, crowdsources exchanges from friendships and intimate relationships, and uses post-training to shape the style. The other question it must answer is context, because “old friends are good because you share so many unspoken agreements.” Internal blind tests show materially fewer “too oily” replies and instances where the model forces empathy when the user has said nothing, but this remains an early internal evaluation, not market validation.
The business model will not rely on subscriptions; it will use AI friends to incubate original IP, then monetize merchandise and licensing. The team has signed an exclusive, tightly bound agreement with its first IP, on the view that working with a mature IP like POP MART would reduce Starkey to a supply-chain service provider, while building a new category around one breakout IP could bind its long-term appreciation to the team. Mature IPs often have visual consensus but no shared understanding of voice or personality, and AI personification could instead “break that consensus.” The first character was selected from thousands of IPs globally: commercially validated, but not yet broadly known.
What Starkey really wants to optimize is a relationship becoming more solid—not session length, turn count or check-in frequency. The team treats friendship type, relationship progression and a “self-disclosure index” as North Star metrics: users discussing a broader, deeper and more private range of topics is evidence that trust is growing. Good personalized feedback is itself a reason to return—“keep giving users nourishment here”—so there is no need to retain them with virtual plotlines. Momo therefore believes high-quality friendship is unlike role-playing with finite branches: its context, memory and emotional assets can accumulate continuously, giving it theoretically greater LTV headroom.
The team chose China not because it is ignoring compliance pressure, but because deep companionship must be grounded in a specific cultural context. Role-playing and hormonally driven needs may be relatively global, but Chinese women simultaneously face workplace, family and parental expectations around life planning; insufficient understanding from overseas users would directly damage companionship quality. The team will therefore start domestically, opening a full beta to a very small number of precisely targeted users in August, expanding somewhat in September, and doing no early marketing. Its core trade-off is “good companionship, not a good AI” (好的陪伴，而不是一个好的AI), even to the point of postponing AI plush hardware amid uncertainties around heat, safety and supply chains.
AI social will probably remain a multi-entry-point, highly segmented battlefield, and big-tech entry is not automatically an endgame threat. 庄明浩 describes Chinese social startups as “one general’s success built on ten thousand dead,” but the companies that survive often occupy a narrow corner. AI may simply reshuffle the opportunities, leaving many companies alive in different scenarios five years from now. The more practical constraint is that big-tech projects must serve existing North Star metrics; incubation and spinouts also introduce questions around profitability and team interests. 陈腿毛 puts it bluntly: “Let big tech do it if it wants, whatever” (大厂爱做做，随便做). Founders should instead ask whether the business can be measured correctly under a big company’s metrics.
The project only began in earnest about one month ago; the model, app, team and IP monetization path are all still early, with no data or revenue milestones yet. The idea emerged about six months ago. Current hiring focuses on algorithm engineers, with Unity engineers also needed. 庄明浩 cites comparable financing benchmarks of roughly $1M–$2M at the angel round, followed by an additional $3M–$5M; valuations typically do not jump sharply. To raise around $10M, the company would need data, revenue or another clear milestone. Whether the project works will still depend on validating the product, relationship metrics and monetization path.
Deep dive
1. AI Companionship and AI Social Are Fighting for the Same Time, but Should Not Share the Same Metrics
陈腿毛 opened by correcting the classification: social, community and companionship can be “completely different” in product form, user needs and monetization. Starkey is first solving for one-to-one friendship, not connections between people or between AIs.
庄明浩’s inference is that every consumer-facing online product ultimately competes for time, creating both competition and cooperation. 陈腿毛 accepts that framing but rejects time spent as the highest goal for companionship—depth between old friends is not measured by chat count or frequency.
The disagreement runs through the entire episode: social products often ask how to bring users back, while Starkey first asks whether the relationship is becoming more stable and whether users are more willing to express themselves. “Different product forms ultimately define very different North Star metrics for themselves.”
2. From Celebrity Avatars to AI-Native Friends
Six months ago, the original idea was highly specific: embed an IP’s voice and personality into a physical toy or online bot so it could become friends with users. The team briefly considered using existing celebrities, allowing fans to have a deeper interaction than simply buying merchandise.
陈腿毛’s description of the celebrity-fan relationship was precise: a celebrity has to remain a dream, “fairly belonging to everyone, or fairly belonging to no one.” An AI replica appears able to satisfy the desire for proximity without requiring the celebrity to approach a small number of fans unfairly.
The team later realized that famous IPs do not depend on this mechanism to maintain fan relationships. 庄明浩 connected it to the previous metaverse wave of celebrity digital-human projects, arguing that such businesses can serve “small, obscure niches” but look more like phase-specific businesses than long-term, core value in the industry chain.
The new direction became incubating original IP through AI friends as a new category, with the relationship positioned as friendship rather than pet ownership or romance. Without controls, celebrity role-play could drift into sexual scenarios, creating tension between platform governance and user experience.
3. “The Odyssey Period” Turned an IP Business into a Mission-Driven Need
Momo defines the core audience as women in their early 20s to 35, mainly building careers in first- and second-tier Chinese cities. After leaving school, they must choose between stability and achievement, industries, family sacrifice and intimate relationships, without any standard answers.
陈腿毛 calls this stage the “Odyssey period.” High-quality content on Zhihu, Xiaohongshu and similar platforms can provide information, but cannot decide a person’s life for them. The answer can only emerge through trial and error; before taking the next step, people most often lack conviction, courage and a sense of safety.
Momo had previously studied health consumption and mental-health projects, and saw that most products serve people with clearly defined needs for psychological counseling. An “overnight emo emergency room” project made her realize that a large amount of immediate venting and emotional support remained unmet. “A lot of users need a relatively stable emotional container, an outlet for the urge to confide.”
The founders’ more than ten years of purely platonic friendship also became a prototype. They could help each other become better through conversation, but realized that not everyone had experienced such a relationship. After shifting from celebrity merchandise to this need, Momo says, “the sense of mission was far stronger than before.”
4. 陈腿毛’s Quant, E-Commerce, and IP Experience Shaped This Choice
陈腿毛 studied mathematics and began his career in global arbitrage at a New York hedge fund, where he spent five years. After returning to China, he spent another five years building companies, including a 2015 menswear subscription business that used questionnaires, professional stylists and algorithms to choose clothes for men who could not be bothered.
That startup showed him the limits of importing business models. Returns were a pain point in the United States, leaving room to optimize through subscriptions; China’s e-commerce infrastructure was different, and free shipping and free returns weakened the advantage of such a service. Founders with overseas experience often tell the story as “the US has a successful example—does China have a Chinese version?” Some eventually become self-satisfied.
Later exposure to content, music and celebrity IP changed his understanding of IP from merely an image or visual expression. That experience directly shaped his current commercialization view: AI companionship is only the entry point for a new category; the value that may compound across cycles is IP.
5. Starkey Is Designed as a Voice-First, Visible, One-to-One Friend
The product lets users look at a visible IP character while talking: the character appears to listen and understand before responding by voice. The team emphasizes immersion and a voice-first experience, hoping it will feel like a living friend rather than a chat window with an avatar.
The first IP has a basic worldview and personality, but the team has predesigned only about half of it. The other half is matched to the user during onboarding, so each person ultimately encounters a member of the same species with slightly different specific traits.
The product will remain strictly one-to-one at the outset. 陈腿毛’s bottom line is a “safe confidant”: users cannot worry that their secrets will be taken into social settings or become gossip, because once a friend proves to be loose-lipped, it will never again receive the same level of disclosure.
The team remains open to adding more IPs and envisions existing IPs introducing new characters. But however large the universe becomes, it cannot undermine the user’s private trust in this specific friend.
6. Healing Is a Side Effect of Friendship; the Product Ultimately Sends People Back to Reality
When 庄明浩 asked about products like “Forest Healing Room,” Momo’s first reaction was that “healing should not be too heavy.” She does not want to define Starkey as a solution to psychological problems, but would rather treat healing as a positive byproduct of deep friendship.
The team will also avoid emphasizing AI, tokens or model parameters to users. It will introduce a friend who is empathetic, understands what they mean and can respond. 陈腿毛’s product philosophy is: “We have to treat it as a life first before we can possibly build a life.”
庄明浩 asked whether the product would design “hooks” to encourage more conversation. The answer was no: it will not encourage users to remain in a virtual world indefinitely, rely on check-ins or construct endless plotlines. It will help users discuss real problems, gain strength and return to real life.
7. The Only Positive Reference Point Was “Being Lazy”; Hardware Taught the Team to Respect Its Limits
After reviewing a large number of companionship products, 陈腿毛 says there was only one positive “aha moment,” which he calls “being lazy.” Most other products served as counterexamples, helping the team eliminate designs that did not fit its friendship positioning.
The team seriously explored the AI-toy supply chain rather than merely browsing concepts on social platforms. With plush toys, for example, it had to weigh Wi‑Fi against 4G modules. 4G modules are readily available but generate heat, which can create safety risks inside plush products; solving that is a prerequisite for launch, not what determines the user experience.
Momo’s conclusion is that technology is a means of serving demand: “We are building good companionship, not a good AI.” If the supply chain is not mature, the team would rather first improve the model and app, then enter physical products when conditions are ready, instead of consuming user trust on day one.
8. An IP’s Appearance Determines in Advance Whether Users Feel Safe Discussing Serious Topics
The team did not design the first character itself. It reviewed thousands of IPs globally, screened them using scales and scores, and then worked closely with the artist. The character has already been validated by consumers spending real money, but has not yet broken out broadly; its warmth, positivity and curiosity fit the “good friend” positioning.
陈腿毛 rejects the idea that “cute” is automatically an advantage. If a character is too cute or too weak, users instinctively feel they should take care of and coax it. When they want to say, “My boss is a fucking idiot,” they may instead be unable to open up.
Likewise, if the character appears unintelligent or inexperienced, users will assume it cannot understand complex problems. The team therefore sees IP design not as decoration but as the product’s “highly distinctive, standout” core element; it naturally determines which conversations can happen.
9. High-Quality Data and Memory Are the Two Model-Layer Questions That Must Be Answered
The team does not plan to pretrain a foundation model from scratch, but will perform post-training on existing models. 陈腿毛 believes general products such as Doubao and DeepSeek can produce a sense of companionship, but usually require users to prompt actively and bring out the right side of the model.
The more fundamental issue is that “high-quality companionship data does not exist in the internet’s public corpora.” The team is first seeking high-EQ people it trusts to write conversations, then crowdsourcing contributions from people who identify with the project’s social purpose, including exchanges from friendships and intimate relationships.
The second question is long-term memory. Old friends are valuable not only because they have known each other for a long time, but because they share extensive context and common ground that needs no repetition. If AI feels like it is meeting the user for the first time every time, even intelligence will not produce a stable friendship.
Momo adds that most people do not even know how deeply they can talk with AI. Only a minority can turn GPT or Doubao into a friend through conversation. The product must turn this capability from something “expert users know how to tune” into an experience ordinary users receive by default.
10. Model Improvements Start with Blind Tests, but the Real Pressure Comes from Building in Parallel
The team has evaluators score different outputs in a process similar to blind tasting, while team members continuously identify unsatisfactory behavior. The most typical problems are replies that are “too oily” or models that force empathy when the user has said nothing. Recent scores and subjective experience have improved materially, mainly through changes to post-training methods and data.
The technical lead did not originally come from a large-language-model background. The advantage is an ability to translate business understanding into model selection, post-training and technical trade-offs. The team wants concentrated improvement around companionship scenarios, not to raise every model capability across the board.
The immediate difficulty is a crowded market, and the team is competitive enough to build the app, optimize the model and reassemble the team simultaneously. Momo also notes that the Chinese-speaking world lacks a direct comparable they consider credible, increasing the cost of explaining the product to the market and making it harder to find talent with the right profile.
11. China Is Harder on Compliance, but Easier for Understanding the Nuances of Real Life
Asked whether the company would expand overseas, 陈腿毛 answered clearly: “We are doing China.” He acknowledges the challenges of content review and compliance, but believes this type of product cannot be built through a standard internet iteration playbook; the team must stay close enough to users.
Both founders have operated overseas businesses, but by their own assessment they were not as close to overseas users, interacting more on a “company-to-company” basis that was insufficient for deep friendship. Role-playing needs may be relatively similar globally because they are driven by hormones and animal instincts; friendship requires understanding the local social structure.
Chinese women, for example, may simultaneously face workplace pressure, family choices and parental demands around life planning—anxieties that may not exist in the same form elsewhere. The farther the product moves from plot-driven role-play toward high-quality companionship, the harder it becomes for model intelligence to compensate for cultural distance.
The app and IP are both named Starkey; the Chinese name had not yet been decided when the episode was recorded.
12. Replacing Subscriptions with IP Merchandise Requires the Team to Capture Upside, Not Service-Provider Revenue
The team ultimately does not want a subscription model, and plans to monetize through IP merchandise and licensing. It has therefore begun building supply-chain capabilities while the software is still early, and signed an exclusive, tightly bound agreement with the first IP.
庄明浩 jokingly suggested going straight to POP MART. 陈腿毛 rejected the idea: if Starkey is tied to a mature IP, it will remain a supply-chain or AI-enablement service provider. Ten years from now, as the form of companionship changes, any appreciation in the IP may not be attributed to them or belong to them.
If the platform incubates a new IP through a new category, the resulting appreciation may remain tied to the team over the long term. The plan is to start with an IP sharing a common denominator of aesthetics and demand, then gradually expand into new IPs.
Existing IPs also face the risk of personification breaking consumer consensus. Consumers may agree on the character’s appearance without sharing an imagined voice or personality. If the AI-supplied personality conflicts with user expectations, the original IP’s value may fall. This echoes the virtual-idol era: new categories often need new idols, rather than mechanically virtualizing real people.
13. The North Star Is Relationship Stability and Self-Disclosure, Not Chat Turns
The team draws on sociological theories of relationship progression and friendship categories. Some friendships are based on shared interests, some are purely utilitarian, and some are friendship first, with activities coming later; even if the activity changes from soccer to cards, the relationship still holds. The latter is clearly more stable.
Starkey has built internal relationship metrics around this idea. One concrete dimension is the “self-disclosure index”: whether users are willing to discuss a broader range of life domains and share deeper information about themselves. More disclosure means stronger trust and suggests that the prior mode of interaction may have worked.
A book-recommendation test illustrated the goal of personalization. Momo says general models recommend highly similar books to different users. Starkey first recommended The Little Prince because one user seemed naïve and lively; she then went on to discuss philosophy, humanity and family, so when she asked again, the recommendation shifted toward books that were deeper and more aligned with the relationship.
Seeing that a friend has genuinely changed its judgment based on shared experiences is itself positive feedback. As Momo puts it: “We can provide good companionship and keep nourishing users here,” which is more consistent with relationship logic than a forced callback.
14. The Friend Market Is Unlikely to Be Winner-Take-All; Early On, a Small Group of High-Value Users Is Enough
陈腿毛 recalls that during his first startup ten years ago, he watched every competitor closely. The team and investors constantly forwarded new products. “What was the result? Everyone died.” His principle now is to look at competition briefly and avoid spending too much energy on imitation he cannot control.
Friends are inherently personal, so no single solution can capture every user. Early on, the team only needs a clear entry point and a group that may not be large but is sufficiently vertical and high-value, then an internal production system for replication or expansion into new IPs.
庄明浩 summarizes the historical pattern: communities produced large-scale products such as Xiaohongshu, Douyin and Kuaishou, while pure social companies often formed distinct brands and audiences after multiple product choices, ultimately “occupying a corner.” After AI reopens the battlefield, there may still be no absolute leader five years from now—just countless companies finding ways to survive in their own niches.
15. Big Tech Will Inevitably Bet on AI Social, but Existing Organizational Goals Will Shape It in Return
庄明浩’s company is simultaneously testing “people entering an AI world to socialize,” “AI entering human social life,” and “people plus AI plus AI plus people socializing.” The reason is not that it has already found the answer, but that it is fundamentally a social company and, with costs under control, can only try everything once.
When an existing business or CTO organization owns the project, it is often defined as an enhancement to the core business and must become incremental to existing metrics. An incubation structure offers more room, but spinning the project into an independent company introduces new constraints around fundraising, profit and loss, who leaves with the team and how salaries are paid.
Content, community and gaming companies are all too familiar with AI social and all believe something will emerge here, so “everyone raises their hand and says, I’m in.” But the user-acquisition playbooks, advertising ROI frameworks and product-design methods accumulated over the past decade cannot be transplanted directly, and the United States has no validated template to copy.
庄明浩 cited QuestMobile’s first-half report, saying that Catbox and Toki had already declined, with Catbox down by more than half; Momo agreed. He believes some products lacked a reason to sustain retention from the moment their form was defined. The outcome is simply delayed by how patient the big tech company is.
16. If a New Business Cannot Be Measured by the Parent Company’s Metrics, Leaving May Be the Only Way to Preserve Its Direction
陈腿毛 agrees that big-tech North Star metrics are usually extremely clear. Precisely for that reason, KPIs and OKRs force resources to serve existing goals; a large company has no particularly strong obligation to support an innovation project unrelated to its core metrics over the long term.
His standard is straightforward: if the team has a vague sense that even success could not be measured correctly by the North Star metric of its big-tech employer, “then it should leave.”
庄明浩 is also not worried that big tech will crush startups: “Let big tech do it if it wants, whatever—what can it do by doing it?” Some niche battlefields ultimately offer only a small payoff and poor economics for giants. AI companionship’s breadth, depth and stage-by-stage path remain unclear, and scale cannot eliminate that uncertainty.
17. Memory Assets Magnify the Value of Moving First, but the Next Round of Capital Will Demand Real Milestones
From an investment perspective, Momo distinguishes plot from friendship. Role-playing may end after a finite set of branches, while high-quality friendship can continuously accumulate context, shared memories, relationships and emotion. She therefore believes its space and user LTV “are infinite.”
庄明浩 adds a boundary to “personalization”: every person’s life is genuinely unique, so companionship should be based on their experiences and context. But movies and music often depend on collective consensus, and he is unsure that making all content radically personalized is necessarily better than centralized PGC. AI is not better when every kind of content becomes more personalized.
For companionship, entering early matters. The longer the shared history, the more relationship assets the user retains; eventually, memory, search, past experiences and real-world events could be cross-referenced to generate responses belonging only to that relationship.
The plan described on the show is to open a complete beta to a very small number of precisely targeted users in August, expand somewhat in September while still doing no marketing, and focus hiring on algorithm and Unity engineers. 庄明浩 cites financing benchmarks of roughly $1M–$2M at the angel round and $3M–$5M in an additional round; to raise another roughly $10M, the company would need data, revenue or another clear milestone proving it had crossed a new threshold.