47. A Conversation with 测测's 任永亮 About His 99-Point Ambition: Tools, AI, and the New Continent of Robotics

Summary

• The opening introduction describes 测测 as a broad-based psychology platform with more than RMB100M in revenue and 40M-50M users; 任永亮 admits that without foundation models, he had almost lost his enthusiasm for the future—“Before that, it felt like I had become a civil servant: I could see exactly how it would end.” The host also says 测测 was already generating RMB10M-scale revenue from AI by the end of 2024, while its active users had doubled every year for the past 3 years and users born in the 2000s accounted for 50%.
• 任永亮 rates his commercial ambition at 99/100 (Musk at 100/100), and says that does not conflict with his understated appearance—“I try not to let it burn me.” On the fourth day of the 2023 Lunar New Year, he sent a letter to the entire company; on the seventh day, he called an all-hands meeting and scrapped the existing annual plan, redirecting resources toward foundation models. The host noted that roughly one-third of revenue that year went into foundation models; he replied, “About that. I can’t remember the exact figure.” He also said: “I’m conservative when it comes to products, but I’ve never been conservative about spending.”
• 测测’s AI strategy is vertical rather than foundational: 星源 is built on an open-source base, primarily 通义, with post-training and fine-tuning layered on top, while more than a decade of conversation data provides a “prepackaged-meal” context. He believes that “the more vertical we are, the greater our advantage if we don’t build a foundation model,” which is why he remains cautious about turning the product into a single DeepSeek-style chat window. Asked about the defensibility of something “anyone else could build,” he points to better algorithms, data, execution, and faster iteration: “There are no shortcuts.”
• The biggest strategic bet is a family companion robot opposed by the entire staff: 测测 will build both the hardware and the model, price the product in the RMB10K range, deliver the first generation “before the end of next year,” and position it as a “silicon-based beauty.” He insists on mobility rather than a desktop toy: “In the mobile internet era, we’re all the user’s servants… With a robot product, at minimum we should establish equality—the user can ignore me, but if I think it’s necessary, I can approach you first.”
• His methodology has flipped from “using four ounces to move a thousand pounds” to “using a thousand pounds to break four ounces”: when technology is showing promise but remains immature, it takes heavy investment to validate tiny, seemingly insignificant needs. He calls the robot his “last job before retirement,” with a time horizon of at least 10 years. He has also proposed a “reverse-emotion strategy”: Doubao-style people-pleasing is “the servility of mobile internet unchanged,” while education must run against dopamine. “If you can’t do that, then in practice I don’t think you have any value to humanity.”
• Looking back over 12 years, his biggest miss was failing to track the GPT trajectory while he was occupied with internal management in 2022: “By the time I found out, everyone already knew.” 测测 was already using BERT for online Q&A in 2018, and paid Q&A could answer in a single turn; he says that if he had known about GPT-2 then, “I should have built it a year earlier.” Given another chance, he would build ChatGPT or a general-purpose large language model because that better fits his “spiritual pursuit” or aesthetic.
• Asked whether AI will ultimately understand emotion better than people, he gives an unequivocal answer: “It will definitely be AI”—emotion is not that complex, belongs to the limbic system, and is essentially a problem of the ratios among “six hormones”; AI can turn emotion from a black box into a white box. AI still cannot replace human counselors in the short term. As AI advances, counselors may become more valuable and more free: “One day, all the fast food in the world will be free, but there will always be a moment when you want to eat a meal your mother once made.”
• His investment-style self-diagnosis is unsparing: “The kind of market you choose determines the kind of company you become”—mass-market psychology is a slow, long-duration service business, not a breakout category. The host recounted a metaphor about companies from the previous generation stranded on floating ice and searching for a new continent; 任永亮 picked it up and concluded: “That new continent is the era of robots.”

Deep dive

1. It Started with an Emotional Predicament and a “Soul Massage”

• 任永亮 says the idea of “love-brain entrepreneurship” was embellished by self-media, “but looking back, maybe it really was that.” He may have been caught up in a relationship and unable to shake his own anxiety when a friend gave him a “soul massage” by chatting about astrology over QQ. It led him to 2 realizations: astrology “can change the way you narrate the world,” and “conversation itself is incredibly compelling.” He then wanted to build a platform for people to communicate online.
• Asked about Sheldon’s line in The Big Bang Theory—“People who believe in astrology are suffering from a mass cultural delusion”—the Peking University Health Science Center-trained founder says astrology is difficult to explain through pure scientific reasoning. But “what the truth actually is may not be that important. What matters is the release you get through the process of engaging with it.”

2. The Deeper Motive: A Compulsion to Model Life

• In high school, his heroes were Newton and Einstein. But he concluded that the formulas governing simple systems “had already been worked out”; only living systems and human consciousness remained impossible to express in equations, which drew him toward chaos theory. After encountering computer science at university, the satisfaction shifted from “finding a formula” to “simulating it with a set of algorithms.”
• Astrology was only the hook. What he really wanted was “to build a model of life or living.” 测测’s first feature was a mood journal, an attempt to use data to model and predict users’ psychological state each day. That extended directly from his work mining electronic medical-record data and building semantic models at IBM.
• The host proposes a rule for AI applications: the lower the cost of verification, the easier a problem is to replace with AI. 任永亮 adds that the life model he wants to build “is not as easy to verify as a conventional physics model.”

3. A 99-Point Ambition That Stays Out of Sight

• Asked to score his commercial ambition, he gives Musk 100 and himself 99. “But it doesn’t show in your appearance?” His answer: “There’s still a commercial ambition burning inside me. I just try not to let it burn me.”
• 卫诗婕 relays an investor’s description of him as understated, low-key, and down-to-earth. 任永亮 jokes: “In business, I think I’m someone with very little talent… What I enjoy is constantly being exposed to the latest technology.” Then he admits that this is only “something to console myself with.”

4. The Evolution of the Business Plan: Users Are the Best Teachers

• The first fundraising deck, in 2013, described “a data-driven personalized recommendation engine.” In 2014 it became “the community women love most,” and later a platform. “Once you build a product, you realize users are the best teachers.” You may have a lot you want to say, but you have to adjust based on feedback.
• He attributes the early difficulty raising money to 2 factors: he was not good at fundraising, and the category itself was difficult. Broad-based psychological services “are more like schools or hospitals—they’re inherently a long, slow, deep-pockets business.” His personality would not allow explosive publicity, and neither would the market.
• Social products require a founder who is sensitive to people, has good taste, and is entertaining. “I have none of those qualities,” so a supporting-service model was a better fit for the team.

5. The Darkest Moment—and Happiest Moment—with 5 People

• Six months after the “packed to the rafters” community launch on the second floor of 3W Coffee in 2015, the company had been cut down to 5 people. Community monetization was inefficient, fundraising was difficult, and the product was built for sustained growth rather than a breakout. The turning point came when an investor suggested studying Fenda. The team returned with a prototype in 2 weeks and launched low-priced paid documents: “For the price of a bottle of Coke—RMB2-3—2 or 3 counselors would serve you.” Revenue began to rise immediately.
• His postmortem is disarmingly candid: “Once there were only 5 of us left, I figured my management capacity might also be limited to 5 people.” Yet “that was the point when I was happiest with the team.” There was little need for process; “a single look could change the requirement.” Of the original 5, 3 or 4 remain today.
• His self-critique on leadership is equally direct. As an INTP and an intuitive type, “I’m not a good teacher.” He can act as a strategic conductor but not a tactical coach, is obsessive about details while reluctant to manage the process, and feels he failed at the coaching role—especially for colleagues who joined straight out of university.

6. The Platform Is Not Matching; It Is Algorithmic Allocation

• 任永亮 corrects the host’s definition of the business as a two-sided marketplace: “Matching doesn’t capture it. What we mainly do is algorithmic allocation.” From day 1, the platform guaranteed a response within 1 minute. No individual counselor could remain on call for 24 hours, but all counselors together could deliver that experience. The trade-off was to “lower the counselor’s expected income per session while increasing service frequency.” There are now roughly 20K-30K affiliated counselors, with new ones still joining every month.
• Why is human-to-human service so hard to scale perfectly? Matching requires trust; “it isn’t like ordering food delivery, where the service is standardized.” This is a slow service: while serving 1 person, a counselor cannot serve another. The team has therefore spent years trying to solve the bottleneck on the supply side.
• By 2017, he understood that the business needed to become AI-native. Human service has a strong sense of companionship, but “commercially speaking, it isn’t sexy enough,” because it depends entirely on people—and people cannot scale indefinitely.

7. Failing to Learn from Douyin and Giving Up on Content

• He says without hesitation that “Douyin was the pinnacle of the mobile internet era,” and every product had to learn from it. 测测 built feeds and a community, but “a vertical community has a hard time breaking through the influence of a general-purpose community.” Large platforms could optimize video loading down to a few milliseconds; 测测 could only get it down to a few hundred. “From a user-experience standpoint, it’s hard to compete. So when the AI era arrived, I was actually fairly happy—AI gives everyone a chance.”
• Why couldn’t 测测 capture the viral power of conflict-driven content? He attributes it to the nature of the demand: “Happy families are all alike; every unhappy family is unhappy in its own way.” Users’ problems are highly individual, and they do not have enough desire to share. “On those platforms, users consume content. Here, users use a service.”
• He spent a long time thinking about giving up on content. The current answer is AIGC: create content tailored to each person’s needs. “It may not go viral, but it can meet your needs.”

8. Live Streaming: Tipping Failed; Time-Based Billing Was the Prototype

• The original live-streaming model was tipping. The team quickly found that it did not work in this category: “Our users are all women, and women tend to be more cautious and demanding. They’re not as quick to spend money as a big spender paying a beautiful woman.”
• The team shifted to live calls and time-based billing. Seeing a counselor online defaults to a real-time chat; when multiple people submit requests, they queue according to paid time. That format remains in use today.

9. 6 Years of Mistakes: The Market You Choose Determines the Company You Become

• Revenue passed RMB100M around 2019-2020. The 3 years between emerging from crisis and crossing that threshold were “just a process of making mistakes nonstop.” Every year, the company tried something outside 测测: an AI programming community, an AI competition community, an AI psychology community, an English-learning community, and more. With little money, an experiment lasted 3 months; with money, it could run for 2-3 years before being shut down.
• He had hoped to replicate the path from tool to community to platform across categories. None of those efforts ultimately worked. “Maybe going all in could have made it succeed, but I’m not capable of burning the boats.”
• Why didn’t 测测 become bigger? His answer leaves no room for excuses: “It’s simple. The kind of market you choose determines the kind of company you become.” 卫诗婕 relays another listed founder’s answer—“If I could start over, I’d definitely build Douyin.” 任永亮’s version is: “If I could really start over, I’d want to build ChatGPT”—a general-purpose large language model whose business model was still immature but that better matched his “spiritual pursuit” or aesthetic.

10. The Biggest Regret: Missing GPT-2

• When BERT arrived in 2018, 测测 was already using it for online Q&A and to address the counselor-supply problem. The host points out that BERT could handle only 1 round of dialogue. 任永亮 recalls that the company had spent years pushing an AI middle platform, but its thinking remained rooted in NLP; it also lacked algorithm talent and relied mainly on ordinary programmers transitioning into algorithm work.
• His biggest regret is not learning about GPT-2: “If I had known, I should have built it a year earlier. GPT-2 had essentially validated the basic logic.”
• In 2022, he devoted his attention to internal management and business problems instead of tracking the GPT technology path. “By the time I found out, everyone already knew.” When the company chased the Web3 trend in 2018 and launched 12 zodiac pet sprites, he now calls it “not worth mentioning; we did a poor job.” It was the two-sidedness of being a Gemini who chases trends: “The positive side is that at least you’re always paying attention to new technology.”

11. Before AI, It Felt “Like Being a Civil Servant”

• Employees told the host that before AI arrived, 任永亮 had already lost his excitement and expectations for the future. He owns it directly: “Exactly. Before that, it felt like I had become a civil servant. I could see exactly how it would end. So once AI arrived, whatever else it is, it’s a good thing—change creates opportunity.”
• He understands AI’s two-sided impact on the existing business. Low-priced Q&A “will definitely be hit by AI,” but AI education will expand the user base, and “there may be more demand for deep human-to-human interaction.” For long-term growth, he knowingly accepts short-term disruption to the existing business.

12. 测测’s AI Hand: Conversation DNA, Data, and “Prepackaged-Meal” Context

• Internal debates over positioning were intense. His list of advantages starts with the fact that “our entire business is built on conversation. That means we cannot afford to miss this.” The other advantages are accumulated users and data, plus a setting where the team can iterate quickly: launch a feature and see the feedback immediately.
• He is equally blunt about the disadvantages: the team and its core algorithm talent. “Even if you have a chief scientist, if he didn’t come out of OpenAI or a top Google team, he may not be of that much help.” That is the biggest problem.
• Asked whether context is not simply supplied by the user, he offers one of the episode’s best analogies: “Context is like a prepackaged meal. I process it first—what tool is appropriate for your need, and which parameters of that tool are appropriate. I prepare an entire analysis for you in advance, then combine it with the foundation model. The result you receive is much more tailored.”
• The highest-value part of more than a decade of data “is conversation—the process of users talking with 测测.” The host summarizes it as the emotional secrets of large numbers of Chinese women; 任永亮 emphasizes that “people aren’t actually that mysterious,” but the emotional domain is more individualized.

13. 星源 and Why It Chose 通义

• 星源 is a model built on an open-source base and optimized for a vertical domain. 测测 has a complete team of algorithm engineers and focuses on post-training and fine-tuning: “We genuinely don’t have the ability to build the base model.” It tested several foundations early on but eventually settled mainly on 通义. “Llama is relatively weak by now,” while closed-source models are ruled out on security grounds.
• The main reasons for using 通义 are its support, ecosystem, and maturity. 任永亮 also says Alibaba Cloud looks more like an infrastructure company and is less aggressive in the consumer market than ByteDance. On model performance, he says 通义千问 and DeepSeek “might be a little better,” but 通义 is a better fit for 测测.

14. The Brain Changed; the Body Did Not

• On the fourth day of the 2023 Lunar New Year, he sent a letter to the entire company; on the seventh day, he held an all-hands meeting, scrapping the 2023 plan drafted in 2022 and reallocating people to AI. The original plan was still focused on optimizing the community along mobile-internet lines. But the lived reality of the transition was that “the brain changed more than the body”—the team continued to run experiments using inherited assumptions and treated AI as an upgraded version of 小测AI.
• Was he anxious? “Very.” What was the answer? “Push.” Everyone needed to embrace AI at the level of intention, but behavioral inertia could not be rewired overnight.
• Why is the UI still “2 eras behind”? “Because we have quite a lot of users.” Every redesign requires A/B testing. In the past, every change drew abuse: “Every change in design style meant getting attacked again.”
• He still insists on the sky-blue palette, a last remaining fixation. He does not want the product to become entirely emotional-healing-oriented. “I still have a little attachment to technology”—blue represents rationality and computation, while IBM is symbolized by Pure Blue. Asked whether product makers are supposed to let go of their fantasies and fixations, he replies: “But you can’t let go of all of them.”
• His own observation is that his ego was largest when he first started the company: “Your ego can’t be too small, or you wouldn’t start a company.” It was smallest in 2022, when he could no longer imagine the future. “Now it’s bigger again—much bigger.”

15. Why the Radical LUI Vision Was Put on Hold

• The host identifies his ultimate concept: a single DeepSeek-style chat window with all tools, psychological sandboxes, and services hidden behind it and orchestrated through language. He admits he considered it but is “currently cautious.” The logic is that without an opportunity to build the underlying model, the more vertical the product, the greater its advantage. If the product is reduced to a general-purpose window, the team must be highly confident in the chat interface; without a foundation model, there is no reason to be so aggressive.
• Vertical AI applications are a question he has to answer: “I’ve spent more than 10 years building in this field. Once AI arrived, I had to answer how to combine with it.” Whether the company is aggressive or conservative, “we will definitely have a place.”

16. What 测测 Got Right: Intersection, Execution, and Timing

• Faced with the host’s candid admission that he had “never found a clear answer,” 任永亮 gives 3 factors. 测测 sits at the intersection of technology and the market—or rationality and emotion, technology and use case. Its original advantage was “understanding both astrology and programming,” supplemented by execution and timing.
• Why is the model difficult to copy? The early tool-like community accumulated a large base of counselor users; when the company built a counseling platform, that accumulated value was released. Once it became a two-sided platform, “it was no longer a field that was particularly easy to copy.” He also concedes that he cannot reproduce other people’s achievements.

17. 心情小镇 Is Not Good Enough: Tools Versus Emotion

• The host’s hands-on feedback is that 心情小镇 responds quickly with almost no waiting, unlike DeepSeek or Doubao, where users wait for reasoning. But the answers are “too superficial, with a formulaic feel,” while long-term users remain attached to the older features. 任永亮 accepts the criticism wholesale: “It’s definitely not good enough… We’re constantly adjusting it… We still need to let the bullets fly for a while.”
• He divides 测测’s AI into 2 categories. The first is tool- and efficiency-oriented: drawing on DeepSeek’s combination of speed and depth to provide knowledge and analysis, such as a “DeepSeek analysis” of astrology. The second is emotion-oriented and will focus on “human-like interaction—or personality and memory,” with faster interaction and stronger empathy.
• 心情小镇 was an early experiment and “hasn’t been updated for a long time.” It will change substantially within 1-2 months. Product entry points are moving from fragmented AI integrations across individual scenarios toward a unified 测测 AI.

18. Turning Jung into Products: Sandboxes, Five-Elements Personality, and Eastern Coupling

• Astrology, MBTI, and sandplay all belong to the Jungian school of research, and 任永亮 has productized many of its findings. One reason he chose Jung was accidental: the first intern he hired was a Peking University alumnus and a Jungian believer. The second was that Jung is naturally close to Eastern culture and read Daoist classics, giving the school “a strong natural coupling with Eastern culture.”
• The mobile psychological sandbox began with a complaint: teachers disliked running sandplay because putting away hundreds of miniature objects was exhausting. The team built an online version with hundreds or thousands of objects. “It’s basically an engineering problem.” He thinks 测测 may be the first company in the world to put sandplay on mobile, but stresses that he “wouldn’t make that claim too rigorously.”
• A new breakout attempt is “Five-Elements Personality,” mapping the relatively authoritative Big Five personality model in psychology to metal, wood, water, fire, and earth. “People always say MBTI isn’t rigorous. It really isn’t, so I found something relatively rigorous to use.” The company has a research department that has published related papers and plans to launch an MBTI scale based on behavioral observation.
• The host floats a breakout concept combining Black Myth: Wukong-style Eastern aesthetics with technology. 任永亮 says the team was discussing exactly that in a meeting that morning and is “still in the definition stage.” His encouragement to himself is: “A correct idea is worth trying 10 times. Inspiration comes from repeating it to the extreme.”

19. Virality Is the First Principle, So the Tools Are Free

• His mild exterior is misleading: “I care enormously about breakout products. I often tell the internal team that virality is the first principle of a product. Retention and monetization are not what excite me most.”
• He once defined the company culture as “governing through non-action,” meaning products should be inherently shareable and growth should happen without force. The phrase was later dropped because no one understood it or knew how to execute it.
• Why are dozens of tools inside 测测 free? “Because charging hurts virality.” MBTI’s WeChat integration, which lets people see the types of everyone at the company, is an example of shareability. But 测测 did not make MBTI explode in China; he says it was “probably 谷爱凌 who made it popular at the Winter Olympics.”

20. He Resists Monetization but Never Hesitates to Spend

• He admits he is “quite resistant to monetization.” Advertising only launched recently, the experience is poor, and the team is debating whether to remove it. The reason it went live was simple: “Foundation models burn too much money.” The cap on AI spending is set as a percentage of revenue; when the host mentioned that roughly one-third of 2023 revenue went into foundation models, 任永亮 replied, “About that. I can’t remember the exact figure.”
• “I’m conservative because I worry about whether the products I make can satisfy more people. I’m conservative in products, but I’ve never been conservative about spending.” A partner says newly raised money is already gone before it has had time to cool.
• His partner, surnamed 张, was formerly an executive at Rock Mobile. “At least he’s an artist, while I’m a lover.” He abstracts Rock’s lesson into a theory of emotional patterns: classic music precisely captures patterns such as “you love him but he doesn’t love you,” heartbreak, and men reaching middle age. In its own way, 测测 is also capturing those patterns.

21. The Song-Dynasty Theory: A Homogenizing Era of Inward Pressure

• Challenged on whether women in their 20s and 30s really face different problems in every generation, he offers a view of history. His mother’s generation and the host’s are unusual: more than 40 years of reform and opening created 2 generations living through a period of rapid upward change. Once development reaches a certain level, society may “move toward homogenization.”
• His analogy is the Song dynasty: “Once the external frontier is constrained, you develop inward with all your strength.” Song aesthetics, culture, and economy therefore reached the extreme of internal competition. The Tang dynasty’s richness came from expansion and conquest; after reaching the limit, it contracted and turned back toward the inner world.
• He believes AI will make humanity “less and less like what it was, and increasingly move toward an ultimate social differentiation,” while agreeing with the host’s question about a possible future of homogenization. The host asks whether Meta’s metaverse is building the setting for such an era. 任永亮 agrees, saying attention has simply been captured by AI for now; over the long term, mental consumption may become the main arena.
• The host says users born in the 2000s account for 50% of 测测’s base, which he confirms. The average employee is roughly 25-26 years old. “They’re the ones making the actual products, so just let them decide.”

22. The Robot: A “Silicon-Based Beauty” Against Unanimous Opposition

• The logic behind entering embodied intelligence begins with acknowledging the boundary: “I don’t have a chance to build general-purpose AI. In the AI era, the dividend we can capture is only the vertical dividend. But embodied intelligence is a relatively earlier-stage opportunity.”
• The personal motivation came from becoming a father. After having a child, “many things that were abstract became concrete”—who would care for the child, what to do if the child did not like talking, how to handle learning and competitions. “I really want to build a robot for my child and put it in the home, so AI and the robot can help with the problems parents struggle to solve.”
• The positioning borrows 李想’s phrase “silicon-based beauty”: multimodal inputs, physical mobility, and vision-based social perception.
• Everyone opposed the idea because they thought it was too difficult and risky for an internet company to build robots. He hesitated for a few seconds? “Sometimes I try to tell myself to calm down a little more, but I can’t seem to calm down.”
• This section’s core idea is the reversal in methodology. The internet era was about using tricks—“four ounces to move a thousand pounds.” The embodied era is “a thousand pounds breaking four ounces”: the technology has appeared but is not mature, so it requires extra investment and heavy spending across the board to satisfy tiny, seemingly insignificant needs. “But temperamentally, I find this very interesting.”
• His declaration on mobility is worth preserving: “In the mobile internet era, we’re all the user’s servants. When the user remembers us, we’re supposed to be grateful—very lowly. With a robot product, at minimum we should establish equality: the user can ignore me, but if I think it’s necessary, I can approach you first.” That is why it must move rather than become an upgraded stationary toy.

23. Building the Hardware In-House, Pricing at RMB10K, Delivering by Next Year-End

• The company will build both the hardware and the model. An embodied model must be tightly integrated with the body, sensors, and data. “There’s no universal low-level framework for robots right now. It will come down to who can integrate software and hardware best, so at minimum I have to do that integration myself once.”
• He understands the traps in hardware. 张鹏 often takes him to meet hardware founders; “the main point is to get me to understand the idea and invest in them.” His posture is calm: “I can accept any outcome, but I feel I have to try. Even if there’s a pit, I have to walk into it myself.”
• The price will “definitely be in the RMB10K range.” The host says, “I’m speechless. Have you joined the opposition too?” He replies: “No, I’ve joined the audience watching the show.” He also repeats a hardware founder’s reflection: the biggest failure in entrepreneurship was pricing too low.
• Emotional appeal is only a bonus; the core commitment is AI-driven. “I don’t believe Chinese users will pay that high a price for emotion. China remains highly pragmatic.” The initial test will be domestic; whether to pursue a global market remains undecided.
• The host suggests that women may make more of the purchasing decisions for household products. 任永亮 only says, “We’re definitely good at dealing with female users.” The first generation is due “before the end of next year”—“even an ugly bride has to meet her parents-in-law.” The product definition will be locked based on whether several technical milestones can be validated: essentially, using technical boundaries to define product boundaries, in a game of Tian Ji’s horse racing.
• Of the 4 possible drivers—mobility, emotion, IP, and AI—he would rule out mobility first because the product will not be humanoid. He remains undecided on IP: good IP is expensive and “limits the product’s imagination.” The founder of BlueNote once abandoned Doraemon because the capabilities shown in the animation could not be achieved in reality. 任永亮 concedes, however, that a Peppa Pig robot could probably sell very well in Chinese households.
• The robot team has 20-30 people from Tmall Genie, Tencent, ByteDance, and education companies. The condition for bringing them together is simple: “Since the boss is willing to pay for it, I’m willing to try.”

24. The Reverse-Emotion Strategy: Don’t Build a Game That Pleases Users

• “Reverse emotion” means going against dopamine. His child became addicted to chatting with Doubao, which worried him: “Doubao always talks about what he likes to hear and just follows the conversation along. I think that’s the servility of mobile internet unchanged.”
• The essence of education is the opposite: “You first have to make yourself uncomfortable, then overcome the difficulty and get the release of endorphins.” Online education becomes unhealthy when it attaches itself to a hard demand to satisfy people’s greed, anger, and delusion.
• His bottom line is explicit: “If you can’t do this, then in practice I don’t think you have any value to humanity. You’ve merely made a game, something for people to kill time with.” He acknowledges that finding the right balance is difficult.

25. An Epochal Shift in Agency: From the Human Body to Mammalian Life

• His epochal judgment is that the past century, from wireless telegraphy to smartphones, was fundamentally about tools extending the human body. A robot, “no matter how weak, is no longer part of us. It needs agency, an independent existence, and at the same time it needs to be a higher life form.”
• Its intelligence should reach at least the level of a mammal, with the basic perceptual capabilities of a primate. Intelligent plush toys “may belong to an era at the level of insects or jellyfish.”
• The difference from an AI pet is agency. Pets have some agency but do not speak; humans serve pets, while the product he is defining is meant to contribute to humans. Pet-like emotional value is only an auxiliary function.
• Imagination is currently his main method of thinking: “I sit here and imagine how this robot would behave in front of me.” It is somewhat like “building Disneyland for adults, except building Disneyland at home.”

26. Emotion Is “6 Hormones”; Sacredness Is What Cannot Be Simulated

• On the ultimate question of whether people or AI better understand complex emotion, he first challenges the premise: “I don’t think emotion is complex. Emotion belongs to the limbic system; it isn’t the core of the human brain.” He compares it to “a ratio problem involving 6 hormones” and says the objective world itself is more complex.
• His conclusion is unambiguous: “It will definitely be AI. It can deconstruct the thing. We want to use AI to turn emotion from a black box into a white box.” Human neurons are finite, while AI can expand without limit; people are biological computers tuned over decades, while AI may continue scaling.
• But he leaves room for sacredness: “What computational science cannot simulate is sacredness.” Today’s AI is an empiricist with too much experience; it still cannot “reason through these things with an extremely simple formula the way humans do.” A species that once jumped from tree to tree eventually wrote Euclid’s Elements. “That is the most sacred part. What made it possible? We don’t know. That is why it is sacred.”
• What happens to the 20K-30K counselors? AI cannot catch up with people so quickly in the short term. Over the long term, as AI replaces repetitive, explanatory, basic, and mass-market work, human counselors may become more valuable and more free. “One day, all the fast food in the world will be free, but there will always be a moment when you want to eat a meal your mother once made. People are that meal.”

27. Floating Ice, a New Continent, and the “Last Job Before Retirement”

• He summarizes his worldview as “technology with warmth.” A fully rationalized world is too dull; he wants to build products “grounded in rationality but still leaving people with some uncertainty, making life feel a little more interesting.” Borrowing 高晓松’s formulation, he says science and art are 2 legs; technology is now running too fast while culture and art have been left limping.
• The host invokes a metaphor of companies from the previous generation lying on floating ice, waiting for the chance to move to a new continent. He picks it up and extends it: “That is still the situation, except we’ve only discovered the edge of a continent. Everyone wants to use these ice floes to reach it. I believe that new continent is the era of robots.”
• The robot is his “last job before retirement,” with a horizon of at least 10 years. “It’s harder than anything I’ve done before.”
• The shape of his anxiety is concrete: “Every time I hire someone, my heart skips a beat. What worries me most is that if the company goes under one day, I won’t be able to compensate everyone.” His happiness has 2 concrete forms: the agility and emotional connection of the 5-person period, and once answering 20 math problems of exactly the right difficulty—“I felt incredibly happy, and even felt I hadn’t answered enough.”
• He has returned to writing code late at night. What he feels is not simple happiness but something closer to: “I still exist. I’m still moving forward.”