Summary

测测 is a profitable broad mental-wellness platform that has kept its head down: about 40M users, 75% women, typical users are urban women aged 20-40, profitable continuously since 2017, and using consulting and user payments to fund its AI team and robotics projects. Ren Yongliang’s thinking mirrors 梁文锋’s: “I lack top-tier fundraising ability… so we use business cash flow to support an algorithms team.” The company now has fewer than 300 employees, more than half of whom joined within the past year, and growth is accelerating.
DeepSeek ended the AI war—the hardest judgment in the entire episode: “It squeezed out the last bit of room for my imagination.” (“把我最后一点想象力的空间都给我挤没了”) Before DeepSeek, every AI startup might still have retained the fantasy of “building China’s OpenAI”; now that fantasy is gone, and everyone has been forced to “put their egos aside and look at applications.” Embodied intelligence has become the new hot sector, intensifying competition for him. “I just didn’t expect the AI war to end so quickly. It would have been better if the competition had lasted a few more years.”
The core new bet is a pet-sized home companion robot, with a sellable product—not a demo—expected in the first half of next year. The only successful consumer robotics category is the robot vacuum, which should have 100M-level users; Japan’s “Lavot” and China’s toy-like robots may each have shipped fewer than 50,000 units. His goal is to ship more than 100,000 and eventually over 1M units within several years and “create a category”—explicitly not a toy, humanoid robot or industrial manipulator. Multimodal interaction and mobility are the two non-negotiables.
Li Xiang characterizes AI on the application side as a cost center: every inference burns tokens, and Ren Yongliang says, “Of course.” “The best business model for delivering emotional value… I don’t believe it’s selling tokens.” He admits he does not yet have the answer and is running several experiments. AI is inherently personalized, but “it cannot generate economies of scale; the cost of every service it provides is rigid.” Vertical players’ moat lies in scenarios, real-user feedback and a professional psychology-data team: “Building a simple demo is easy. The key is observing long-term results in real-world scenarios.”
The macro backdrop is deflationary: “AI does not create demand; it destroys demand.” Unlike the Industrial Revolution, it reduces jobs without creating enough new ones, breaking the loop between specialization, income and consumption. “The chances that AI startups grow into midsize or large companies will also be limited.” Psychological and emotional needs, however, are rising gradually and benefiting from the lipstick effect—he cites Hong Kong-listed POP MART, Laopu Gold and Ne Zha. “Our platform is basically a massage.”
Human counselors cannot be replaced “within 100 years.” As AI spreads and educates more users, demand for human counselors may instead rise, like organic vegetables. The moat is the dirty work associated with 朱啸虎: managing tens of thousands of counselors, resolving disputes, and building certification and pricing systems. Taobao and Baidu both tried entering the space, but “once they saw growth was too slow, they said, ‘Forget it, we’re not doing this.’”
The tuition bill is worth recording: missing GPT-2, parting ways with a chief scientist after one year, spending tens of millions in 2023 on an AI team, buying GPUs and labeling data—all led to the conclusion that “the standard answer was DeepSeek” and everything else had been wasted effort. The fantasy of general-purpose foundation models lasted only 2-3 months: the funding and talent density requirements were extreme, “domestic giants may all struggle to close the gap with the US,” and chip compute was the bottleneck.

Deep dive

1. 40M Users, 75% Women: A Broad Mental-Wellness Platform Built from RMB200,000

Ren Yongliang, founder of 新颜集团, defines 测测 as a company that “provides users with broad mental-wellness services through tools, platforms, AI and, eventually, robots.” Its typical user is an urban woman aged 20-40, around college graduation, “facing a lot of challenges in life—in relationships and careers”; women account for 75%. Li Xiang’s verdict: this is the demographic every consumer brand wants.
The starting point was almost naively simple: in 2011, he left his job at IBM with RMB200,000 in savings, rented an office and hired 2 Peking University interns—one studying computer science, one psychology. They started with H5 cross-platform development, “then discovered the experience was far too poor.” His self-critique is blunt: “I relied too much on myself… and had a bit of a personal-heroism mindset. I was a programmer; I didn’t like communicating or bragging.”
He explains the 2013 inflection point in 4 words: “Poverty forces change.” He found a co-founder, went straight to 3W Coffee’s Demo Day, and secured Wang Xiao as the earliest investor. “Sometimes I really miss that era—you’d just go out and demo, and people would come over to strike up a conversation.”

2. Why Astrology and MBTI: High Frequency, Built-In Virality and an Obsession with Recommendation Engines

His product-selection logic was a process of elimination followed by intersection. His IBM to-B experience convinced this self-described introvert that “to B demands too much in terms of resources, management and communication.” His undergraduate medical training led him to conclude that “healthcare is too specialized and serious, with too low a frequency for to C.” Astrology and psychology tests were high-frequency and spread organically after use: “When I talk about these topics, it’s often in a social setting.”
The first BP was titled “A big-data-based personalized recommendation engine.” Li Xiang asked whether that was not essentially what ByteDance later built. Ren’s self-analysis is worth preserving: “Everyone sensed the technology trend at the time. It’s just that the recommendation supply in the area I chose—daily food, clothing, housing and transportation—was too limited. ByteDance chose content with an infinitely rich supply, allowing algorithms to show their power to the fullest. That’s why they were more successful.”
The zero-budget cold start was to search Weibo for everyone whose tags mentioned astrology. “From then on, organic users kept following that trend without stopping,” and organic traffic made up the majority. Does he believe in astrology himself—he is a Gemini and an INTP? “I didn’t at first. Anyway, every time I got dumped, I believed a little more.”

3. “Fanta” Opens His Eyes: Two Infrastructure Shifts Create the Platform Model, Profitable Since 2017

In 2015-16, the data from the tool-based community “wasn’t enough to support another round of fundraising.” “Those were all explosive opportunities. A product like ours, with growth but not explosive growth, simply wasn’t sexy enough.” An investor who rejected him handed him a product called “Fanta,” created by Guokr. After studying it, he identified 2 timing variables: voice technology had matured and mobile payments had become widespread. “Everyone had a few yuan in their wallet and was willing to spend a few yuan on novel services,” while the old BBS bounty model reflected “the inertia of PC thinking.” Revenue began growing from there; the company became profitable in 2017 and has remained profitable ever since, positioning itself as a “broad mental-wellness online platform.”
His reflection on the product’s later fate is restrained. Broad mental-wellness consulting “is precisely not an industry defined by influencers or centralized distribution. Every service has to be delivered minute by minute.” There is no moment when “everyone wants to hear Wang Sicong’s one reply.” “It’s a process of gradual, quiet influence, without much explosive force, but with durable stickiness.”
A footnote from another path: the once-popular astrology influencer 同道大叔 learned all his astrology concepts from Ren’s co-founder and was later sold to a listed company. The co-founder had wanted to build up social media, but “after all, I’m an engineer… I never expected that even today I’d still have to do social media. We went all the way around.” He also misses the 7-8-person phase: “You’d just shout and everyone would move; no meetings needed. That was peak efficiency.”

4. AI Started Before ChatGPT, but Missing GPT-2 Was a “Very Serious Mistake”

The original motivation for AI was a supply-side bottleneck. Broad mental-wellness services were constrained in both “quantity and quality”; one person could serve only one user at a time, and the lack of scale had always capped growth. Planning began in 2017, and the first AI service went live in 2019—based on BERT, with roughly 100M parameters and only single-turn dialogue. “Nobody believed in the GPT route at the time, so we designed it with BERT.”
He first used GPT at home during the 2023 Lunar New Year and, “like everyone else in the industry, was hit extremely hard.” He returned and scrapped the annual strategy that had just been finalized, announcing an all-in move to AI with the mission of “serving the human mind through technology.” What still bothers him is not acting slowly, but perceiving slowly: “Not paying attention to this technology at the GPT-2 stage was a very serious mistake for me… by the time I knew about it, everyone already knew.”
The deepest shock was personal. In high school he wanted to “use mathematical principles to explain consciousness”; at university he worked on predicting RNA sequences in bioinformatics. His choice of broad mental wellness also carried that thread. “After seeing ChatGPT, I realized it had already solved more than half of that problem—what we once thought of as mysterious thoughts and consciousness may really be determined by mathematical probabilities.” He added a marker of the era: many people started building Chinese model companies because “their ambitions were born under that kind of shock.”

5. The 2023 Tuition Bill: Tens of Millions, One Chief Scientist and a Standard Answer Called DeepSeek

The year’s checklist included hiring a chief scientist, buying GPUs, collecting labeled data and debating whether “general-purpose AI was an opportunity or a threat” to the company. They invested tens of millions. A scientist recommended by a friend worked for a year before they “parted amicably.” “We had been cultivating our vertical field for years, but when it came to frontier innovation, we hadn’t thought through how to manage a team or collaborate with partners, which led to gaps in our respective understandings.”
His summary is almost self-mocking: “There was no standard answer for AI at that point. In the end, we discovered that the standard answer was DeepSeek. All the ideas the rest of us had—we worked for nothing and paid the tuition.” They had not even figured out clearly “what AI native actually meant.”
The difficulty of hiring had 2 layers. First: “How much money will this thing really require, and how do I ensure correct decisions under massive investment?” Second: separating genuine expertise from surface familiarity. “Everyone who works in NLP thinks they understand it, but if they haven’t actually trained a large model, they only understand the surface.” He highlighted the structural gap with DeepSeek: “They had always had senior algorithms engineers from their previous quantitative work. We had always built technology applications.” The fantasy of general-purpose models lasted only 2-3 months: “Models demand extremely high concentrations of capital and talent. Domestic giants may all struggle to close the gap with the US, and chip compute is also a bottleneck.” The company therefore retreated to applications and post-training.

6. Funding AI with Cash Flow: A Post-80s Founder Cannot Raise Model Money, So Vertical Models Live on Scenarios

His reason for not raising money for models in the market was blunt: “I lack top-tier fundraising ability, so it’s simple—we use business cash flow to support an algorithms team.” Li Xiang noted that this was the same logic as 梁文锋’s, and Ren accepted the comparison. Two other constraints mattered: the window when investors were willing to get carried away was short, and it was difficult to position new and old businesses in the fundraising narrative. More broadly, “people also felt AI should be built by younger people. For people like us born in the 1980s, it’s hard to give investors confidence—I simply didn’t have the bandwidth to try.” External funding did arrive: Tencent invested, and a local government invested in 2023. “We used that money to buy servers.” GPUs were still affordable that year, “though we paid a little more than we should have.”
The 2 ambitions for the in-house 星源 model are emotional intelligence—building humanlike emotional interaction—and expertise in broad mental wellness. “How to listen is supported by a corresponding body of knowledge in psychology.” The decisive factor is “mainly data,” combined with training and engineering systems. General-purpose models do not have this advantage in the short term, “but that may be hard to say in the future.”
On the threat of others using Qwen or DeepSeek to build similar models, he neither hypes nor dismisses it: “The difficulty isn’t that high, but it isn’t simple either.” The company’s advantage is its scenarios: once the model goes live, it can quickly obtain real-user feedback, supported by a professional psychology-data team that keeps optimizing it. “Building a simple demo is easy. The key is observing long-term results in real-world scenarios.”

7. Human-AI Coexistence: Humans “Cannot Be Replaced Within 100 Years”; The Dirty Work Is the Moat

His ecosystem is fundamentally different from an AI-native product. AI at 测测 handles education and service delivery at the lower end, while a human-service network sits behind it. “Our experts are service experts, not knowledge experts. AI and humans are much more complementary than substitutable.” Although “we are also fairly aggressive and are working on replacement from a capability perspective,” his underlying judgment is that “in this industry, AI cannot replace the value of humans for at least 100 years.” His analogy is organic vegetables: ordinary vegetables dominate the market, but some users will still choose organic. In fact, “the more AI develops and the more users it educates, the demand for people may rise”—like eating fast food every day, then suddenly wanting the taste of home.
Trust is an insurmountable link: “I can’t tell users a counselor is AI while claiming it’s a human.” They are also developing virtual counselors, but “their user groups, consumption habits and perceptions are distinct… talking to people is a fundamental human need with thousands of years of history. As long as human nature does not change, this market will likely persist”—though “how large it can become and how explosive it can be may be limited.”
Quoting 朱啸虎, he describes the moat’s substance: “The dirty work is the moat.” In practice, that means improving product details every day, managing every counselor, resolving conflicts between users and counselors, and operating the full system of certification, progression, ratings and pricing. Giants have all tried: “There were similar modules inside Taobao and Baidu, and some people even came to talk under the guise of investment, just to understand the situation. Once they saw growth was too slow, they said, ‘Forget it, we’re not doing this.’”

8. AI Is a Cost Center: Naturally Personalized, but Without Economies of Scale

Li Xiang summarized the problem: every inference consumes tokens, making AI primarily a cost center. Ren said, “Of course.” His conclusion comes from the company’s own ledger: either apply AI to improve efficiency and add value in a profitable field, or be extremely rich and bet on the future. “We happen to be a company with cash flow, so we can build applications around it and capture the application-layer upside.”
Personalization is both AI’s nature and its commercial weakness: “A different prompt gives a different answer, and even the same prompt can produce a different answer each time—it is inherently personalized. But the drawback is that it cannot generate economies of scale; the cost of every service is rigid. From AI’s perspective, it is still at an early stage.”
The question consuming half his time is: “What is the best business model for delivering emotional value in the AI era? I don’t believe it’s selling tokens.” Does he have an answer? “Not a final answer yet, but I’m running several experiments.”

9. DeepSeek Squeezed Out the Last of His Imagination: 3 Lunar New Years Gone Bad

His reaction to DeepSeek was entirely different from his reaction to GPT: “It squeezed out the last bit of room for my imagination—it wiped out all the fantasies still left inside us.” He explained how widespread that fantasy had been: before DeepSeek, every AI startup might have thought, “Do I still have a chance to build China’s OpenAI?” Even those who had already given up retained a little self-consolation.
The second-order effect was what he had not expected: “I just didn’t expect the AI war to end so quickly—it would have been better if the competition had lasted a few more years.” Once the war ended, everyone had to first “put their egos aside and look at vertical applications,” while embodied intelligence became the new hot sector, intensifying competition in his area. As for companies still refusing to give up on general-purpose models, Li Xiang asked whether they had to keep going. Ren’s implicit answer was yes: “Otherwise the whole thing would lose its meaning.”
The timeline’s trauma is worth preserving: “The Lunar New Years of 2023, 2024 and 2025 were all terrible—GPT in 2023, Sora appearing around the 2024 Lunar New Year, and DeepSeek arriving in 2025, each delivering an enormous shock. Every Lunar New Year was spent in a state of extreme anxiety.” His current focus is VLA in embodied intelligence: “Progress is extremely fast. It feels like it is going to unify the whole field.”

10. Betting on Emotional Robots: Pet-Sized, a Member of the Family, Sellable Next Year

The shift into embodied intelligence began in mid-2024, before the subject rendered as “Yushu” in one subtitle—possibly referring to 宇树—took off. 3 motivations converged. After having a child, his own needs shifted from heartbreak to companionship and education; many of 测测’s deep users had been with the platform for 8-10 years and moved their focus to family after marriage and children. The company also needed to “extend its life cycle.” And “general-purpose foundation models are definitely not the right dish for a company like ours—the funding and talent costs are too high, and there is also the China-US chip barrier.” AI revenue covered costs last year, “but the imagination was still not big enough. Getting Chinese people to pay for a virtual AI still has a fairly high barrier.” The answer was to combine the virtual with the physical.
The product has 2 hard rules: “I definitely won’t make a toy—my product should not look like a toy”; and multimodal interaction plus mobility are mandatory. The target form is pet-sized, so “he genuinely feels it is a member of the family, not a tool and not a toy.” The company is also avoiding humanoid robots and industrial robots that manipulate objects, focusing only on family emotional and interactive scenarios. He describes the route choice with 2 distinctly period-specific terms—likely “trade first, technology later” versus “technology first, trade later”: “It’s like climbing a steep slope versus climbing a gradual slope.”
The timeline is concrete: “We expect to make a product in the first half of next year—not a demo, but something we can sell.” The robotics team has dozens of people, the 测测 side has more than 200, and the entire company has fewer than 300 employees. He also disclosed the company’s hardware history: a brainwave-detection headband and VR glasses for psychological therapy both failed. The earliest seed was an Intel-sponsored hackathon in 2014, where they built a smart doghouse overnight and won $1,000.

11. The Robot Risk Ledger: PMF Is Hardest, Internal “Reluctant Agreement” and Only Apple and Tesla as Benchmarks

He sees the market coldly: the only successful consumer robot is the robot vacuum, which “should have 100M-level users.” Japan’s “Lavot” and many toy-like robots in China “may each have shipped fewer than 50,000 units.” His ambition is to ship more than 100,000 and eventually over 1M units after several years and “create a category.” The hardest part is not the supply chain but establishing PMF: “It’s a completely new category, so you have to do a lot of market education. For a small company like us, there’s an element of biting off more than we can chew… if it really doesn’t work, I can at least become a martyr.” The product definition has been overturned repeatedly: “You discuss an idea with one group of friends, then someone else comes along and everyone overturns it again.”
The internal reaction, in his exact words, was “reluctant agreement.” “At meetings, everyone keeps asking: Why should this be your turn? We’re a company that has never made hardware.” Which was more presumptuous than their earlier attempt at a large model? “About the same. The large model was a question of resources, which you can quantify—you can see immediately whether it requires tens of billions of dollars or several hundred million yuan. Building a robot today is presumptuous about our capabilities: the ability to define a product and determine whether it can sell.”
Responding to Li Xiang’s friend’s analogy that hardware is like chips and should not be started without strategic commitment, he offered a middle position: building robots “is harder than software but easier than chips.” The investment is below chip scale, but flexibility is lower than software’s; you cannot build a demo and test it in a week. It heavily tests product-definition ability, which is precisely what interests him. “ByteDance would have more advantages in speed, capital and iteration. We are a product-driven company and like craftsmanship.” The model shifted from free first and paid later to “a large upfront hardware payment plus a subscription.” He knows the benchmark is brutal: “Very few companies can do software, hardware and services well—Apple and Tesla. Sometimes I wonder whether I’m biting off more than I can chew. The benchmarks are too high.” To hedge the hardware risk, they have recently resumed contact with external capital and “secured some financing.”

12. The Macro Backdrop: AI Does Not Create Demand, the Lipstick Effect Provides a Floor and the Lords Know Their Place

The episode’s most important macro judgment is: “AI gives me a sense of danger—it does not create demand; it destroys demand.” Unlike the Industrial Revolution, which destroyed old jobs and created new ones, AI mainly reduces jobs without adding enough. The Industrial Revolution began with specialization: work creates money, money creates consumption, and consumption creates more work. “AI has somewhat cut that loop.” The implication is directly usable: because AI does not create demand, AI startups will have fewer opportunities to grow into midsize or large companies. Mobile internet created food delivery, ride-hailing and WeChat, allowing $10B and $100B companies to emerge; AI satisfies existing needs, and existing companies can simply use AI to strengthen themselves. Li Xiang asked whether the presence of demand in psychology would not intensify competition. Ren conceded: “That is indeed true.”
The counterweight on the demand side is rising emotional consumption. Investors who approached him often began from the business angle, but “in spiritual consumption—POP MART and Laopu Gold—you can clearly see from the leading Hong Kong-listed companies that everyone is watching emotional consumption.” He places himself within the lipstick effect: “When the economy is under pressure, cultural products and small luxuries can see strong growth. Why did everyone go see Ne Zha? They identified with it. There was a kind of shouting, a desire to resist. Our platform is basically a massage, helping you feel less bad.” But he does not overstate its explosiveness. Psychological services are like healthcare: “The number of patients Peking Union Medical College Hospital receives each year is basically fixed; only infectious diseases are explosive.” Education can explode because exams create a standardized hard need. Psychological services are “high-frequency but not a hard need; they become a hard need only in certain states.”
His self-positioning remains clear-eyed, like a minor lord among warring states: “Using the Spring and Autumn and Warring States analogy, we are in Bashu. There is some room to survive, and the giants cannot be bothered to attack me, so we can recuperate here.” It is not the final answer—Cao Cao’s order to withdraw from Hanzhong was “chicken ribs,” but Liu Bei was delighted. The once-unrecognized positioning as “the engineer who understood astrology best” gave him strategic breathing room. Now that general-purpose models and embodied intelligence are both hot sectors, “everyone is an expert, and the pressure is much greater than before.” Li Xiang noted that everything he had done in the past 2 years was in a hot sector. Ren admitted, “That is what worries me most about myself.” The answer is differentiation: “What we do is combine the latest technology with a specific need. From mobile internet to building platforms on mobile payments to AI services, it has really been the same thing.” One final observation about his generation is worth keeping: looking back at the experiences of people born in the 1980s, “it turns out they were not universal; they were exceptional—a once-in-a-lifetime set of circumstances.”

Sovereign AI: Geopolitical Strategy & Industrial Policy for Countries 3-193, with Anjney Midha, a16z Type a Sentence, Get a Playable 3D World in 3 Seconds - Shlomi Fuchter & Jack Parker-Holder