梁琛奇 on Douyin, Maoxiang, and Using AI to Create Joy, Not Productivity

Summary

• 梁琛奇 believes AI entertainment is entering a window of opportunity, with 3 prerequisites now falling into place. First is inference cost: entertainment sessions run 10 to 20 times longer than tool sessions, so “if you scale to 10M DAU, the company goes bankrupt”; after DeepSeek, inference costs for comparable intelligence have fallen enough that ROI cases that previously did not pencil out may now work. Before leaving ByteDance, he ran the numbers on Maoxiang and concluded its ROI should turn positive by the end of last year; as far as he understands, it did. Second, image and video models including GPT-4o, Sora, Seedance, and Nano Banana have cleared the bar for content users are willing to watch and are beginning to follow instructions. Third, tools such as Doubao and ChatGPT have already taught users how to interact with AI.
• He expects the first version of a new AI entertainment experience to appear within 1 to 2 years, with roughly 1 year the more likely case; the necessary condition is burning tokens in real time while consuming content. AI short dramas and other experiences that do not burn tokens will simply grow inside existing platforms such as Red果短剧. He remains cautious on pure 3D real-time generation and AI girlfriends and boyfriends, while lightweight games, interactive content, and “emergent creation” could be promising directions. The potential gaming population may be 10x its current size.
• 动念引线 has trained models from day 1 because entertainment scenarios are “orthogonal” to the optimization direction of general-purpose models, creating a potential moat. Data for new experiences is naturally scarce, entertainment evaluation is highly subjective, and large numbers of subjective, implicit features can be built into the model. The company has mainly trained image-generation post-training models; he says most scenarios in its use case have reached SOTA, at costs several orders of magnitude below the models then available, with the goal of using frame-by-frame image generation to support real-time consumption.
• His methodological preference is “product first, technology second,” and he rejects building a hammer before looking for a nail. Start by combining existing models to validate an experience; if the product shows promise but remains rough, then invest in model training for the experiences worth pursuing. He once researched the question with AI and concluded that entertainment products may follow a roughly 7-to-3 split between “product first, technology second” and “technology first, product second.” Against the consensus, he believes ByteDance was also closer to the former: it first saw that users wanted to read news, then built out recommendation technology.
• Organizationally, the company has set up a UGC organization in which teams of 1 to 2 people validate an idea every 2 weeks to 1 month, testing quickly across a large pool of possibilities. It is currently developing 7 to 8 products in parallel with roughly 100 employees. He also reflects that review pressure has been lighter as an entrepreneur than at ByteDance, and that he was briefly seduced by the easy option that “could work without burning tokens” before data and hard-nosed reviews pulled him back to reality.
• He believes AI could create new network effects around context depth and context crossovers. AI-native entertainment can have prompts behind it, and prompts can genuinely mix rather than merely being stitched together as in the old duet format. His advice to entrepreneurs is to focus first on whether they can create a genuinely impactful new user experience, rather than worrying too much about large companies copying them; Cursor’s acquisition at a $60B valuation is his example.
• The validation standard is novelty, longevity, frequency, and freedom from fatigue. An “acceptable” model would have 40% next-day retention, 20% day-7 retention, and 10% day-30 retention, while also tracking open frequency, generalizability, and whether the experience improves as scale grows. Tool use cases on Pinecone Moment may have reached 20% or more next-day retention, but he did not prioritize importing these off-platform account-building users.
• At the macro level, he expects content’s share of total value to decline relative to the mobile internet, while absolute growth remains substantial. If a small number of people use large numbers of agents to drive technological progress, society will have to answer what the vast number of people with newly freed time will do. He believes broad-based creation, art, and self-expression can give people participation, positive feedback, and meaning. This is a projection about the future, not a certainty.

Deep dive

1. A double minority: applications and AI entertainment are both outside the main narrative

• 程曼祺 opens by describing 梁琛奇 as a double minority: he starts with applications rather than models, and he is pursuing entertainment rather than productivity. He corrects one half of the premise: “Claude Code is developing quickly, and Cursor was eventually acquired—but it was a very large acquisition, at $60B.” Applications are clearly still advancing.
• Entertainment has not seen much fundamental change—“that is objectively true”—but entertainment is inherently slower than tools. That is precisely why he left last year: it was slow, but had reached “a pretty good point at which to start,” provided 3 prerequisites came into place.

2. Prerequisite 1: tokens are the traffic of the AI era, and entertainment inference costs 10 to 20 times more than tools

• Every technology era has its own binding resource. In the mobile internet it was bandwidth: “If you said in 2008 that you wanted to build short video, people would think you were crazy—watch 1 short video and you would use up a month’s data.” In the AI era, the resource is tokens, or inference cost.
• A tool product can reach PMF in 10 minutes. Entertainment is the process of killing time while “sitting there with an empty mind waiting for a new experience to fill it,” so sessions last 10 to 20 times longer and inference costs scale up by the same multiple at comparable intelligence. “If you scale to 10M DAU, the company goes bankrupt.”
• DeepSeek was a clear inflection point: attention shifted toward open source and lower cost at comparable intelligence. “ROI that did not pencil out before can pencil out now.” Before leaving ByteDance, he modeled Maoxiang, which had already begun commercializing with a healthy trajectory, and found that its ROI could turn positive by the end of last year. “As far as I understand, it really did turn positive.”

3. Prerequisite 2: modality—language compresses intelligence, but entertainment needs the senses

• Language is efficient and serves tool scenarios built around solving problems. Entertainment is not about saving time but “enjoying time”; it seeks sensory experience. “Compared with images and video, text serves an order of magnitude fewer users.”
• Last year, GPT-4o, Sora, Seedance, and Nano Banana achieved 2 things at once: the output “really cleared the bar—you genuinely wanted to watch it,” and the models genuinely followed user instructions. In the Midjourney era, the aesthetics were good but instruction following was weak, so the output “could not really be used or scaled.”

4. Prerequisite 3: only must-have tools can educate users; entertainment can free-ride

• The historical analogy is the smartphone. Its adoption was pulled by the old-era must-have of “making phone calls”; 15 years later, calls accounted for only 5% of the value, while the rest of the ecosystem grew quietly around them. Entertainment is a new experience, and “it is very difficult to say that an experience that has never existed before is a narrow must-have.”
• The same pattern applies to AI. Through must-have products such as Doubao and ChatGPT, users have forced themselves to understand the boundaries of AI interaction, how to prompt, and how to use multiple agents. “Once all of that has been built, if you build entertainment experiences on top, the users have already been educated.”

5. Not making tools is his nature: he has never done anything outside entertainment

• He knowingly chose not to pursue productivity because of his personal nature: “I have no interest in building tools. My interest is making more people happy and creative… Look at my career—I have never touched anything outside entertainment.” The upside is that work and life reinforce each other: “It is perfectly reasonable to go out and play on the weekend. You are experiencing the product and the user. It is self-consistent.”
• He says he has been fascinated since childhood by “subjective creation,” while also being good at mathematics and recognizing patterns. That combination is what a large consumer product manager is designing for: “The world is subjective, and large numbers of people are irrational,” but you still need to abstract a framework and “serve more people in a predictable way.” He accepts the host’s summary: the intersection of the subjective and the objective.
• He was a science-track student in high school but applied to the liberal arts because “science was not romantic.” After studying social science for a year, he turned back: the humanities were romantic, but “iteration was hard to stage and validate”; with computer science, “you know the result in an instant.” Given the same talent, validation produces a steeper trajectory.
• He cares deeply about where that trajectory leads: “The work I ultimately want to build should reach a broad population and affect their inner worlds, creativity, ways of living, and new experiences.” The medium eventually became products.

6. The takeaway from Qunying Tu and the answer to Earth Online: diverse

• In college, he built Qunying Tu, a relay-drawing mini-program in which one person drew 5 strokes and passed the image to a friend to continue. It attracted several thousand to tens of thousands of users. His first takeaway was that “very few great works in history were made by a group of people,” especially in literature, so the product was more social than creative. The second was that social interaction needs a vehicle and a prompt: “You need to find a vehicle that is novel, long-term, high-frequency, and does not become tiring.” Finding one is difficult.
• When the host said he could not think of anything he could do frequently with friends without getting tired of it, 梁琛奇 replied: “The world has an answer. The answer is diverse.” Coffee today, go-karting tomorrow, skiing the day after—“you are already inside Earth Online.” Dianping solves the problem of finding something; what fascinates him is completing the experience inside a playground.
• Roblox is closer to that platform: UGC games, everyone online and playing in real time, and more than 100M DAU. “If my friends and I have nothing to do today, we go hang out in Roblox, chat while escaping a volcanic eruption.” Eggy Party has the same echo: 潘乱 discovered that his son’s avatar and his friends’ avatars were “wandering around the island” together. That is social interaction.

7. Choosing Douyin in campus recruiting: 3 judgments before the breakout

• During campus recruiting, he was initially assigned to Huoshan. HR said Huoshan was the regular army with more resources, but he told 卷卷 directly that he wanted to join Douyin—before Douyin’s breakout during the 2018 Spring Festival.
• His first judgment was that he personally kept scrolling Douyin, its users were trendier, and “the brand flows from the top down.” The second was that the motivation to express oneself was different: Douyin generated “a very strong intrinsic drive to love the platform,” while Huoshan operated by exchanging firepower points for money. “You can use material incentives to motivate people building a platform, but that cannot be their first-order motivation.” The third was that Kuaishou already existed: “There is no reason for users to migrate to a worse Kuaishou to compete with Kuaishou.”
• He remembers 2 impressions from his interview with 卷卷. First, when presented with an unreliable new idea, 卷卷’s initial response was to encourage him to develop it; a typical company might say, “Go raise penetration.” Second, when he asked whether Douyin’s emergence of unpolished comedy clips was good or bad, 卷卷 was emphatic: “This is absolutely a good thing. Douyin did not start with music, nor does it end with music short video. It is meant to cover the entire population, and everyone has the ability to create.”
• He admits he was not that categorical at the time: “Most people had not really calculated it through.” ByteDance itself once aggressively benchmarked Weibo, then stopped midway after realizing Douyin’s DAU had already surpassed it. No one had predicted that in advance.

8. The Friends tab and the curse of recommendation: “It is a window, a canvas, and a bridge”

• He built Douyin’s second tab, the Friends page. His framework was: “Recommendation is a gift, but you can also call it a curse. The moment you break the user’s mindless-scroll, almost vegetative flow, retention drops—because recommendation is too strong and absorbs everything.”
• The saving design was a prompt saying “You have finished watching your friends’ content,” with a one-click jump back to the home page. It was “a very small thing,” but a major reason Friends-page retention turned positive. ByteDance’s system was that a product could not launch if retention declined; “a group of people would come and challenge you.” The lesson was that seemingly minor changes on a critical path can affect the entire chain. “Alex（朱俊）is pathological about details.”
• He relays Alex’s metaphor: Douyin is a window. “Videos rush toward you and fill the entire screen. For a moment you are genuinely disoriented, as if you are standing by a window looking at different worlds. One swipe is the next world, and you can never predict what comes next.”
• His judgment is categorical: full-screen vertical swiping was “essentially designed by Alex,” and the difference from non-universal vertical swiping is substantial. Instagram was not earlier. “Douyin feels like a product designed by a designer. Its product character is too clear.”

9. Pushing social on Douyin: “Only WeChat was left”

• The project logic was that other businesses had natural trajectories and predictable ceilings. “You have to ask what comes next. Then you realize that only WeChat is left.” Social could be split into communications tools and content distribution and interaction. The latter needed only relationships, submissions, and distribution. Douyin was already a super-UGC platform: “There was a high probability that you and your friends were both using Douyin; you just had not encountered each other.” Cold start did not need to be solved.
• He initiated the project himself and took it from idea to launch in 1 to 2 months, around 2018. Friends-based distribution, social content consumption, and direct messages later made a major contribution to Douyin retention. “Retention is ultimately the metric.” He also admits it did not produce a table-flipping change such as taking DAU from 700M to 1.3B.
• When the host says Douyin does not feel like a social product, he points to generational differences. Younger users send direct messages and post stories frequently every day. “Every generation chooses its own platform.” Once your relationships are established before a platform appears, migration is difficult; when a new generation is just beginning to build relationships, the network can be rebuilt.

10. Product plus manager: he dislikes communication, but “if you hold it in, you cannot create”

• He breaks the job into 2 parts. Product means understanding users, designing, and connecting creativity with empathy through logic. Management means “making the system highly efficient, so good ideas can be implemented efficiently.” He only realized 6 months after joining that the latter was mandatory and required technique.
• He dislikes communication, but posed himself a choice: “You know you can deliver an extremely impressive experience, but you need to figure out how to communicate with people and use larger resources. Do you give up building that impressive work, or force yourself to learn? The decision seems easy, but it took me 6 months.”
• His warning about self-training is worth preserving: “If you cannot train this into yourself, stop doing the job. Good taste and subjective feeling are not a decree from heaven. The decree from heaven is being able to break the problem down clearly—turning a subjective issue into objective, quantifiable, executable logic that a very smart engineer can understand. That is extremely important.”

11. Real-time social in 2021: the definition of real social, and the tuition of shutting down after 1 year

• 2 years after graduation, calling himself “probably the youngest business head,” he spoke with 凯利 and launched the project. His reasoning was that posting to Moments and asynchronous likes were not real-world social interaction. Real social was “doing something together in a specific environment and around a specific topic, while staying in communication and having fun.” When drinking coffee with friends, you never worry about awkward silence because someone may play the saxophone and the world keeps moving. A real-time phone call creates “huge pressure” because it feels like completing a task. What was missing was “something that would naturally move forward even without you.”
• The team grew to nearly 100 people at its peak and shut down 1 year later because of “force majeure.” “I thought we might solve it if we worked hard enough, but actually we could not.” The direction was later validated to some extent by Eggy Party: first attract users with a Fall Guys-style core experience, then add a large number of UGC maps where people can hang out.
• The first lesson was: “Respect the objective laws of the world’s development. Do not work against certain tendencies and trends.” But the world still has branches; “that does not mean everything is governed by a predetermined law.”
• The second lesson was that entertainment cannot carry too many assumptions. “You want only 1 key variable; everything else should already be validated.” If each assumption has a 0.9 probability of success, the probability across 10 assumptions becomes very small. Tool assumptions are often accurate because they solve clear inefficiencies. In entertainment, “90% probability” does not really exist.

12. The lineage of expression: from Duoshan and Stories to the difference with 动念

• BeReal’s close-friend social product was growing quickly at the time, but he believed the larger question was close-friend social itself. “The way you do it and the experience you create are separate questions.” Duoshan and Instagram Stories both focused on expression: “There are things you want to express, but the expression never happens,” such as an ordinary daily event you cannot bring yourself to post to Moments.
• 动念 differs in its target. Cameras in the previous era inspired people with talent for the physical world: a woman who could attract 20M followers, or a cloud photographer who could “identify the interesting parts of Earth Online.” This era needs to inspire people with great talent in their inner worlds. “They may be plain-looking and stumble over their words, making you think they are boring, but their mental worlds are extraordinarily rich and diverse.”

13. 动念引线 and default: ignite the idea, but the fireworks do not belong to AI

• The company’s view is that geniuses are randomly distributed through the population. “药水哥 is a genius. Before the mobile internet, you had no way of knowing he was one.” Each person has 3,000 thoughts a day that flash and disappear. In the past, most could never become works. In the future, AI can turn “things that do not make logical sense, are chaotic, or have not yet taken shape” into good work.
• The meaning of 引线 is deliberately modest: “I do not think the source of that creativity comes from AI, nor from our company. It comes from ordinary users themselves. We are just the fuse that sets off the fireworks.”
• The English name default refers to “compressing and re-expanding the world.” In the future, you will not merely stand by a window and watch. “You might experience 30M parallel worlds and parallel lives”—go to the Tang dynasty, live as a rich second-generation heir in Dubai, or discuss metaphysics with a philosopher—then return inspired to the real world you inhabit. The name was chosen before the company started. Its vision can be traced back to Qunying Tu: “The method was wrong, but the vision will not change.”

14. From Turtle Soup to Maoxiang: one cross-product of axes, plus a competitor signal

• ChatGPT made him understand Turtle Soup. “This was very different—it scaled humans as a production input.” His derivation was that mobile-internet entertainment had only 2 large categories: passive content and games. Add an axis from passive to active. An AI-native product is not about the model; it is “I provide an input, the system reasons, and it gives me a highly long-tail, customized response.” Cross the 2 axes and “you are already defining interactive content.”
• At the time, at least half the market viewed AI as a “character.” He was more convinced it was “content.” The difference was whether you were consuming an experience or building a bond—whether you moved to the next character after interacting or the character became a friend.
• He first worked through his own derivation and only then looked at competitors. He did look at competitors, but gave greater weight to his own reasoning. He only discovered Character AI after completing the derivation, when Alex forwarded it to him.
• His bar for a signal was 10M-plus DAU in the mobile internet and ARPU of “around 1,” meaning roughly $10M in revenue per day. AI products lowered the bar by another step because of expectations for a steeper trajectory; 5M DAU could be enough to consider a product. At the time, Character AI, Talkie, and 星野 had already reached several million DAU.

15. Only AI, but “there is no efficiency in the work I see in my head”

• He joined Flow after discussing a “fully emotional assistant” with 朱俊, who became extremely excited. “I think he was probably thinking about Doubao, but he did not say so at the time.” 梁文琛 immediately stated: “I only do AI. Anything outside AI is unimportant. If this is a technology-revolution-level opportunity, why would you do anything else?”
• But he would not pursue the most revolutionary productivity applications: “The most important factor is that it has to be part of the work I want to make. The work in my head contains no efficiency and no tools. Moving the world forward is important, but the work in my head does not contain that.”

16. The supply-side through line: this moment is like the moment television arrived

• The history of content has 2 through lines: modality, from text to images to video to 3D, and the less visible expansion and fragmentation of supply. Critics once called television a poor invention: small screens, poor content, and budgets one-tenth those of movies. “But precisely because of that, supply was larger and could meet different people’s needs in different situations.” The line eventually led to vertical short dramas.
• That line is not over. AI short dramas still cost several thousand yuan and require days to a week of learning. Once the barrier falls to “just express your idea and creativity in text, and leave everything else to AI,” large numbers of talented people will enter. These new creators will not come from the existing short-drama or game industries. “They used to be nobody. Now they are creators in this new era and new format.”

17. The consumer side must be a new experience, or it is an opportunity for incumbents

• His key rejection is direct: “If you want to build a platform for watching short dramas, there is no reason to do it. As a consumer, where would I watch? Obviously Red果短剧. I do not care whether it was made by AI or by humans.” Without a new format, the product will grow inside an existing platform. “That is an opportunity for a large company.”
• The essence of a new experience is “overfit infinite branching.” Before AI, computers were finite AND/OR systems and product managers had to enumerate every button. After AI, “2,000 inputs become 20M,” and the system can reason out a customized response.
• His thought experiment is why we believe Earth Online is real: you can eat paper, vomit it out, soak it in water, and eat it again. “Precisely because you can do anything long-tail, you feel free. The world is acceptable to you.”

18. The next generation of recommendation: AI handles the familiar, while the strange must come from people

• His metaphor is a circle and an ellipse. Before recommendation systems, supply was limited and users settled for a square. After recommendation, every user gets the roundest ellipse. “But from first principles, you have not truly been satisfied, because the creator made it for 30M people and was searching for the greatest common denominator.”
• The host objects that users themselves do not know what the circle in their heads looks like. He responds that recommendation accuracy can be observed in the data, and AI may genuinely fit the user: “Every so often, ask it what kind of person it thinks I am. It suddenly gives you a paragraph and you say, wow, this understands me better than most of my friends.”
• He parts ways with the prevailing market narrative: “I suspect many people stop the narrative here, but I do not think that is right.” Content must provide both the familiar, which creates an empathetic bridge, and the strange, which creates surprise. “If it is exactly like your life, you will not find it interesting.”
• AI can use what it knows about you to reshape a story into a form you can absorb, but the strange must come from diverse people, from creators with different personalities and life experiences. “A content-supply world completely taken over by AI will not happen.”

19. Scaling people: from 100K-to-1 to 100K-to-100K, with a hybrid experience

• The third variable is that humans used to be an unscalable production input. Live streaming is the classic bottleneck: a streamer has limited bandwidth, so only the top donor ranked by gifts gets a conversation. “Once AI scales this, it becomes 100K-to-100K”—100K people can each communicate one-on-one with a character.
• The host focuses on the contradiction: if AI will not replace all supply, how much human presence remains in the experience? His first answer is that the character’s rich backstory is designed by a human creator. “When you consume this regent, you are creating a form of resonance with the creator.”
• His second answer is that the final experience is hybrid: “The participation and share of humans determines whether the experience is genuinely interesting.” An experience where AI accounts for almost everything and unique human ideas account for very little will not be equally interesting.
• Pressed further, he admits that sufficiently intelligent AI might one day design 100K vivid characters in one shot. “If you force me to make an absolute statement, AI is very powerful and maybe one day it can. First, I do not really want to see that. The appeal of humans comes from subjectivity and the triggers produced by their life experience.”

20. Humanity’s defenses: 5 things AI will replace relatively late

• His list includes things with little data, which cannot be learned; things that are difficult to evaluate—“if the evaluation system is extremely accurate and objective, you can just keep pushing RL and become superintelligent”; the initial impulse to create—“it is always responding, but what you start with, what you want to do, and what the goal is, that is the starting point”; and offline activities before embodied intelligence matures.
• The fifth is the most ordinary: things with emotional value. A mother’s birthday cake says “Happy birthday, baby.” “In terms of quality, how much better is it than a good cake? Actually, not at all. But it has emotional value. I think that is a kind of human defense.”

21. Maoxiang’s design: do not lock users in a room; throw them into a volcano

• Existing products had 2 unresolved problems: users saw a collection of characters and did not know what to talk about, while a significant share drifted toward hormones—“it really became borderline pornographic.” The diagnosis was: “Put 1 man and 1 woman in a room with no topic and no objective, and instinct may take you toward hormones.”
• The solution was a setting and a main plot: “Do not leave them alone here. Put them in a group. They are not in a closed room; they are at the mouth of a volcano. The volcano is about to erupt and everyone is going to die.” Like immersive theater and murder mystery games, the experience is free-form but users “never worry that the next moment will become boring, because there is a hidden main line moving forward.”
• The product became a story format: human screenwriters write the script, AI acts as director, and user actions determine how the story progresses. From day 1, they should have designed for diverse content serving both minors and adults, although the early user base was mainly women aged 18 to 30, balancing ethical considerations with future willingness to pay.
• Both monetization paths worked. From the content angle, insert ads into the story—if there is viewing time, ads can be inserted. From the game angle, offer memberships, gameplay cards, and the ability to “hear what a character is thinking.” Maoxiang did both, and “both contributed meaningfully.” He also acknowledges that a real use case emerged in which users formed human-like bonds with virtual characters; this cohort has high emotional intensity and may have high average spend.

22. Building a house versus raising a daughter: the pedagogy of AI products

• Building a mobile-internet product was like constructing a house, in Alex’s metaphor: one brick at a time, defining every use case. “If you do not define it, it will not happen.” In the AI era, part of the product is still a house, while another part “is like raising a 2-year-old daughter.”
• An absurd example: you cannot teach a daughter what to do “when a woman in a red dress on a bus at age 39 takes off her high heels, licks one, and smashes it against your head.” Inputs and the future are infinite. “You cannot enumerate every situation. You can only teach her a framework: worldview, principles, some SFT examples, and goals that let her do RL herself.”
• The acceptance test is the wow moment at her 18th birthday. She performs in a play and moves you deeply: “I never expected you to play the character so well.” Then you turn around and think: “Of course you did. This is how you grew up.” “If you do not have a moment like that, you got it wrong.”
• The insight came from 2 objective facts: 90% of Maoxiang user feedback was about the model—“Why is this character talking to me like that? It is infuriating”—and the feedback was extremely long-tail. “You cannot specifically fix that one piece of feedback.” Combined with his philosophical understanding that SFT and RL pipelines cannot provide total control, the implication is that the old way of building products no longer works.

23. Training models from day 1: being orthogonal to the base model is the moat, and monetization must start on day 1

• After 1 month building Maoxiang, he realized the company had to train its own models. Experiences such as “he is hot and cold with me” could not be fixed without training; the Doubao base model “was not optimizing in your direction.” 动念引线 began training models on its first day. The story model was trained from scratch at the time, and the character model also had a trained version that he later took responsibility for and influenced. The company has a complete model-training pipeline.
• Orthogonality lets the company win on both sides: the better the base model’s intelligence, the better the experience—“a smarter director cannot hurt.” It also creates 3 moats. First, data for new experiences is naturally scarce and unique. Second, entertainment evaluation is highly subjective: “If an 18-year-old woman in Shanghai is happy, a 49-year-old man in Northeast China may not be, and a 39-year-old LGBTQ person in California may not be.” Reward models must therefore be built in a real user environment. Third, subjective design can be built into the model: “Full-screen vertical swiping was designed, not purely derived. You can put 1,000 implicit features into an agent, and they are not easy to copy.”
• The market view at the time was that “application companies should build applications, model companies should build models.” He reversed it from day 1. The other difference was monetization: “On day 1, tokens are expensive. You have to think about how to make money earlier, or you will burn yourself to death.” Maoxiang only began thinking seriously about monetization after 6 months. His “positive ROI” calculation included marketing, inference, and daily operations, but not the already-trained model.

24. The January decision: DeepSeek accelerated everything, and 梁文锋 proved the value of belief

• The decision to leave had already formed in January. DeepSeek “greatly accelerated” the 3 variables for entertainment. The other half was idealistic motivation. Before DeepSeek, he spoke with Chinese researchers. “Someone said OpenAI was exploring AGI while we were patching things up with simple SFT. You hear that and think this cannot work. If you do not believe in it, how could you ever achieve it?”
• He remembers 2 passages from 梁文锋’s interviews. One was “the candy not received in childhood”: when a relative asked what the point of studying was, “they did not understand that once China crossed this threshold, education would become increasingly useful.” The other was: “China cannot be a follower forever. This only requires a set of processes and 1 result.”
• There was another layer: small teams can achieve major things outside the giants. “It gives you a sense of possibility, and you find that possibility interesting.”

25. Why start a company: a ranking exercise with 3 outcomes

• He has a sympathetic understanding of large companies. A company with a $200B valuation needs a high ceiling to justify a decision. With a new entertainment experience, no one can precisely judge PMF or the ceiling on day 1. “It is perfectly reasonable to say: first let me see a product reach PMF and clear the bar, then I can follow. There will still be time.”
• His ranking is: “Build something that affects many people and capture the greatest upside > build it but let another company capture the upside > do nothing.” Once he ranked the outcomes, “the natural conclusion was that building it was most important.” Asked where staying at ByteDance and using its resources to maximize someone else’s zero-to-one experience would rank, he said: “A little better than doing nothing.” So the first 2 options both required starting a company: “I had to participate.”
• He puts copycat risk low on the list. “I do not think you should spend that much attention worrying about what happens if a big company copies you. First spend some attention asking whether the zero-to-one experience is important enough to change many people. If it is important, you should build it first.”

26. Maoxiang was 0.5 to 1; “1” means everything is ready before hitting the accelerator

• His self-assessment is that Character AI and its peers had sufficiently validated 0 to 0.5, while Maoxiang reached 1. “1 means a relatively generalizable product form that is ready to reach a certain scale. You can hit the accelerator and scale it to some number.” Hitting the accelerator means marketing, monetization, and scaling.
• Generalizability matters because he is fascinated by intersections: “Some people stand on the left side of the railway tracks and some on the right. You build an experience that people who would never intersect can all accept.” Hip-hop before The Rap of China was like that. Behind the niche was “an emotion and form of expression shared by a very broad population. If you find the common core of human nature, you can light up everyone.” Niche cultures can function well on their own, “but their appeal declines.”

27. A 7-year methodology: hit the accelerator on each car in the parking lot; entertainment is whether you dare not to show the model

• He left ByteDance in a calm mood: “One journey had ended and another was beginning.” He did not even make the standard ByteDance departure “bye-bye” video. His summary of ByteDance’s entertainment playbook is: find a validated format, expand supply toward infinity, add recommendation and monetization, confirm positive ROI, and market aggressively.
• His metaphor is a parking lot. “You walk in and inspect every car. If one looks promising, you make small repairs. If it looks great, you hit the accelerator. If it does not work, you hit the accelerator on the next one.” Games are an exception: before AI, game companies and other internet companies were entirely different, clearly separated. ByteDance was a distribution platform for content formats, not a content company itself.
• His definition of entertainment is: “When you are not being serious and are enjoying the time, that is entertainment.” Users therefore do not care about the underlying model. “One way to test whether you are strong is whether you dare not to show the model. Most tools still need to show it.”
• Consumption remains the larger side. The short-video format with the highest submission rate in the previous era had a UGC submission rate of at most 15%. 90% of users open the app every day to consume. To build a long-term, high-frequency experience for a broad population, you must solve consumption.

28. AI will raise the submission rate: you are not Li Bai, but you can now write poetry

• He believes the large-model era changes the equation. Many people “have the urge to express themselves first, and only then think about what they want to express.” Sharing a song to Moments or posting an MBTI result is a search for a vehicle: “Finally there is a vehicle for me to express myself.” Most of the time, expression is blocked by the barrier to creating it.
• In the future, AI can collect “fragmentary moods and all kinds of illogical context from every day,” identify what is unique and luminous, and propose it back to you. You might say: “This is exactly what I wanted to say. This is what I have been thinking about—and it even wrote a poem.” You may not be Li Bai, “but you can write poetry now.” The act of creating is itself a form of consumption.

29. The 2 personalities of innovation: the diverger and the critic, forbidden from attacking each other

• He once wrote a ByteDance document on how to innovate. One personality randomly diverges, “relying on feeling and being triggered by the world.” The other is a critic who knocks down each idea in turn. “After analyzing them, you find that all 10 are dead. That is often what happens.” The key discipline is: “The critic must not turn around and attack the innovator by saying all your ideas are unreliable. New things are extremely fragile. 99 out of 100 may be wrong.”
• Testing discipline means predicting before launch what will rise or fall and who will use the product. “You must not judge after the fact. That is pure hindsight bias, because people give themselves all kinds of cues.” The learnings then cycle back to the 2 personalities.
• On 朱俊, he says he combines creativity, empathy, imagination, and aesthetics. “These are often contradictory. Perhaps those contradictions are what allow the corresponding experience to be built.” He also leaves a reminder about context: “Never assume that a PhD in mathematics is still a PhD in mathematics when scrolling Douyin. He is not. Once he opens it, he is there to experience.”

30. Are ByteDance entrepreneurs overrated? “I absolutely do not decide direction by chatting with FAs”

• Asked about ByteDance-trained founders being pushed into entrepreneurship, he answers bluntly: “People coming out of ByteDance are certainly good at solving problems. Some even operate by first going out and talking to a group of FAs, then deciding that a direction might be worth doing. I absolutely do not work that way. What I do is whatever I decide to do.”
• Three things cannot be carried out of ByteDance: money—“we are doing fine raising money,” but the leverage is different; the enormous leverage in attracting talent—“not every outstanding person is risk-seeking”; and “extremely high commercial efficiency,” which is difficult to replicate in the short term.

31. Why Douyin beat Kuaishou: container, operational pull, and an overwhelming system

• On the container, he is unequivocal: “If you insist on comparing them, full-screen vertical swiping is better than the two-column layout.” Short-video selection is based on a blended first-3-second reaction—motion, sound, and caption. “It is very hard to communicate that feeling through text or a cover. Looking only at the image, you cannot feel it.”
• Operations were pulled into shaping the community and the brand’s top-down momentum. But the least replicable advantage was the system: “Why would someone have 100 points of confidence? In this system, everything can be calculated. You know how much you invest and how much you will earn in the future.” The competitor sees a blank expanse and does not know whether the next 2 steps lead off a cliff. “You know that even after 1,000 steps you will not fall, so you can invest extremely aggressively.”
• His implication for startups is that such a system lets you see the ROI potential of a format farther into the future. “If you are confident the format works as an experience, you should build monetization relatively early.” 动念引线 has not spent much attention on this yet: “There is an order to these things.”

32. Pinecone Moment: throw out an “imagination camera” and see what people photograph

• The product lets users choose a photo and tell the story before and after it—such as “I got my bottom wet while riding a bike and looked pathetic”—then has AI turn it into a serialized comic with temporal depth. He admits: “My expectation is not that it must seriously prove PMF. I am looking for broad-population observation and signals.”
• The starting point was that Maoxiang was “the answer for AI-native entertainment products over the past 2 to 3 years.” He could not find a second format with larger DAU and broader users. But it was definitely not the answer for the next 3 years. Aside from modality not being ready, the harder issue was that these products “demand a great deal of imagination from users,” while most users, in the narrow sense, do not have that much imagination. “This is extremely contradictory. If I cannot solve this problem, I cannot move forward.”
• That contradiction produced his “imagination curve”: high at birth and “very low” by age 18. “There is no evidence; it is an observation.” 2 forces suppress it: the gravity and social pressure of reality, and the lack of tools for expression.
• The prototype was his own experience. As a child, he drew pictures and wrote books to tell his second aunt. “One day I realized that she did not really want to listen. She was pretending it was interesting because she loved me. There was no brain-computer interface mapping the vivid images in my head to her, and I was bad at drawing, so the expression was suppressed.”

33. Why comics: the compression of Along the River During the Qingming Festival, and the broad definition of imagination

• The technical consideration comes from The Three-Body Problem. Zoom into a cloud on the horizon by 10M times and you still see detail, molecules, and atoms. Zoom into Along the River During the Qingming Festival and there is nothing more to see, “because it has been compressed.” “At the same intelligence level, content forms that process simpler information will mature first.” Comics should therefore clear PMF earlier and feel less AI-generated.
• On demand, he rejects a narrow definition of imagination. Ask a random man on the street to “use your imagination” and he may look confused. Ask what happened yesterday and he might say: “My son used my clothes to wipe water off a shared bicycle. He said, ‘Dad, I was worried your bottom would get wet when you sat down.’ I wanted to hit him, but I also thought it was hilarious.” That is what appeared in his mind. Memories, recollection, reasoning, and fictional imagination—“any expression of the world in your head should be mapped out.”

34. 4 use cases, including a tool use case with decent retention that he is not prioritizing

• The barrier has been cleared. A monk preparing to enter monastic life came in to make a declaration; a grandmother who liked square dancing imagined what her grandchild would do as an adult. Typical use cases include life records with temporal depth—Meta Glass and similar hardware are also pursuing this, but hardware penetration is slow; amplifying emotion through synesthesia—writing that you feel bad today can seem affected, while an image amplifies the emotion without reproducing it exactly one-to-one; anime, fan fiction, and derivative-creation communities; and account builders, including large numbers of multi-image comics on Xiaohongshu and Douyin and novel-promotion posts driving traffic to Red果.
• Next-day retention for the tool use case “could be 20% or higher,” but the team chose not to market to or otherwise import these users: “They make it, post it off-platform, and you will not consume this kind of thing inside the app. It does not help our content ecosystem.” Early on, consumption must be solved by building a vertical community around a specific audience and category.

35. Consumption must burn tokens: “time travel” and the line between consumption and creation

• The logic is: a new experience is necessary; any non-AI-native experience was already built in the mobile internet; therefore an AI-native new experience must burn tokens. Pinecone Moment has not burned tokens yet for 2 reasons: it is starting with creation, and last year burning 300 rounds of image tokens “would absolutely have burned the company to death.” Text models had only just reached positive ROI, while images cost an order of magnitude more. Its self-trained image model is SOTA in most scenarios relevant to the use case and relatively low-cost, several orders of magnitude cheaper than models on the market.
• A “time travel” feature lets users enter someone else’s work as a comic character carrying their own persona, generating a new version with their imprint. It is “the first step toward burning tokens,” but not real burning yet because users wait 1 minute. “You cannot wait when consuming.”
• His dividing line has 2 parts: “First, are you waiting or not? Second, have you jumped into another world, or are you extending the one you are already in?” Once models become faster and cheaper, real-time generation turns it into a consumption experience rather than a creation experience.
• The same yardstick applies to Sora’s cameo feature. “How many times do you use cameo in Sora in 1 day? Certainly not 300. I suspect penetration is not high either. 90% of users would need to have this use case every day.” Real time is necessary, but the product still needs a genuinely long-term use case.

36. The UGC organization: not a PGC app factory, but testing 100 ideas out of 30K

• A parallel organization was established 2 to 3 months ago. The trigger was that coding had “really crossed a capability boundary,” while Pinecone Moment clarified the use cases and opportunity. The operating model is 1 to 2 people per team, rapidly iterating and testing against an idea list for 2 weeks to 1 month. “Decide whether it is great or not. If not, shut it down and open the next one.”
• The difference from ByteDance’s early practice of developing 13 products at once is that he first uses 1 product to build the camera and observe, then maps potential to different products. “We are not a PGC app factory. We are an organization that may be able to do UGC.” In entertainment, no one can precisely identify the correct format with 100% certainty. Rather than bet on 1 option with a 20% probability, “find 100 out of 30K” and test quickly to raise the odds.
• Current status: 7 to 8 products in parallel, some shut down, more than 10 accumulated, and roughly 100 people across the company, including the model pipeline team. UGC and technical middle-platform teams provide support. He believes the quantity is still short by an order of magnitude: “If a PGC company has 10, should a UGC company have 100?” The bottleneck is people, which is one reason he is doing this episode.

37. Who to hire: AI-native, resilient or passionate, unwilling to skip steps, and thrilled when their worldview is shattered

• He gives a specific definition of AI-native: “No one is AI-native today, because today no one collapses without AI. You should be migrating toward this way of working. The more frustrated you are, the more AI-native you are.” Such people are neither conventional product managers nor engineers. They need product sense, business ownership, and the ability to code themselves.
• Motivation only needs to come from 1 of 2 sources. Passionate people “see the world as a playground. They certainly do not want to build Excel in a playground. Excel is also great, but they do not want to build it.” Or they are driven by achievement and growth, with resilience as the core.
• First-principles thinking is needed to survive a cycle. “Experience means you skipped steps 1, 2, and 5. When a major external variable changes and step 3 changes, experienced people do not even realize they need to go through step 3. Young people are not smarter; they simply have not done it before, so they cannot skip the step.”
• They also need the ability to refresh themselves. “Ideally you are shocked, your worldview is shattered and rebuilt, and you feel great—‘I grew again.’” The wrong responses are believing there is nothing new in the world or collapsing into “I am a useless person with no skills.” A plus is having an imagination and believing that certain things will happen.

38. Narrowing and lengthening: before model costs cross the threshold, PGC/PUGC is more economical

• The next step for Pinecone Moment is to narrow into certain audiences—“the cold start difficulty for a concentrated audience and a broad audience is completely different”—and make the content expression longer. The direction will become visible “after a while.” When the host guesses anime and fan fiction, he replies: “Not necessarily. Maybe. One of them, plus others.”
• The underlying objective law is that before model costs fall below a certain level, PGC or relatively PUGC is more economical than UGC because the hit rate is higher. At the same DAU and submission rate, PUGC produces more high-quality content in absolute terms, even if UGC ultimately captures the largest share of value.

39. 2 contradictions that must be solved, pointing toward “emergent creation”

• The first problem is to keep the user’s context sufficiently small and simple while making its share and impact on the final work sufficiently large. “If 99% is written by AI, then only 1% is the person’s own characteristics.” The result feels AI-generated and uninteresting. The second is to keep expression simple while making the reward large enough. “The difference between output and input must be large enough. That is what makes it great.”
• His solution is “emergent creation.” A Maoxiang creator can “intelligently define the setting, allowing the user to call on AI’s enormous brain 1,000 times, 10K times, or 100M times.” The next generation of creators accepts emergence: describe the concept and experience, let AI fill in the framework, and “only check whether the final feeling is right.” The finished product still has a strong personal character, color, and expression.

40. The cost and revenue math: frame-by-frame comics; you cannot skip images and jump straight to video

• The vehicle for real-time consumption is frame-by-frame image generation, not video. “Like a comic: this person is charging up a Rasengan, and in the next frame it hits Sasuke in the head.” The first-principles logic is: “Text has just reached positive ROI. Why would you not first make images positive ROI and jump directly from images to video?” Video multiplies the cost by another 200 to 300 rounds, producing an astronomical figure. Even if video costs fall by an order of magnitude, images should fall to an even cheaper order of magnitude.
• He is not worried about revenue: “As long as costs fall to an acceptable number, it is no different in principle from the mobile internet. You would not worry in the mobile-internet era that a free product would burn to death.” If there is viewing time, insert ads. If the product has interactive-game characteristics, sell skins, one-time purchases, and memberships.
• He also identifies a common industry mistake around context. “Many companies realize context is important and immediately start collecting it aggressively. I have no relationship with you, and you ask me to enter 5 pages of context? Impossible.” The correct path is the one OpenAI followed: first find a product form that works without context; after using it for 6 months, the user discovers, “Wow, it understands me,” and can then build other things on top.

41. Model companies cannot swallow entertainment: likes are a great invention

• He directly rejects the concern that Codex-style work will generalize into life and then entertainment. First, the boundary is whether “you are serious about solving a problem.” Tools pursue efficiency: “Give me 1 input and I will give you 1 output.” Entertainment requires interaction. He rejects the idea that future entertainment will not need interaction: short video is already passive enough, but “the gesture of swiping still has value. People still need to feel that they are participating.”
• His signature argument is: “Likes are a great invention, because every so often you want to like something. It connects to a strong emotion you feel while watching the content—like applauding when you see something excellent in real life. If you remove likes and recommend through some other mechanism, I do not think that is okay.” Content has 3 types of value: absolute excellence, fame—“C罗 did something”—and direct personal relevance.
• The second layer is mindset. Many products are what they are when you open them. “Alipay has short video, but you would not say, ‘I am going to Alipay to watch short video.’” The third is that subjective judgment should not be absorbed by model companies: “In this scenario it should be A; in that scenario it should be B.” Real data from users using the product in that moment is the right reward model, not annotations from a group of people or a simulated model. At the same model size, a model focused on a specific scenario is also more likely to reach SOTA.
• On Sora, he is cautious: “Sora certainly has not solved this. Its interactions did not really hit the long-term motivation of users.” Remix is not a long-term need. It is closer to creation, and it is not high-frequency.

42. Content-ecosystem moats: Kuaishou’s edge, Xiaohongshu’s capillaries

• The historical evidence is that both Kuaishou and Xiaohongshu survived intense attacks from giants. Kuaishou’s moat is a complex web of “edges” linking users to users, content to users, and creators to users—the old-iron culture. You can move top creators over and cross-post, but how do you persuade ordinary UGC creators to migrate? “My followers are not here, and the friendships I built are not here.”
• He also believes Kuaishou made many objective mistakes: it did not buy traffic, did not create a new brand positioned above the low-end, and did not switch to full-screen vertical swiping. “Even after making so many mistakes, it still retained a share. If it had done things right, the share could have been higher.”
• Xiaohongshu is “as fine as capillaries,” a girls-help-girls network of experience. A migrant can carry over top content, but not the long-tail experiences users search for. Early distribution decisions in the 毛文超 style—“this is great, this is not”—also shaped the community’s tone and were not purely rational. ByteDance’s way of working is: if the data is good, keep scaling it. People recruited from both sides often experience a high rate of cultural mismatch. 可颂’s limited impact on Xiaohongshu was also a matter of timing: the image-text market was smaller, ByteDance moved too late, and Xiaohongshu may already have held a decent share.

43. A new network effect in the AI era: every piece of content has prompts behind it that can cross

• The third moat is context depth plus context crossover. Depth is easy to understand: once ChatGPT understands you well enough, even if another model performs similarly or slightly better, you are too lazy to switch. Crossover is new. “In the past, you saw a video without knowing what was behind it. You could only duet, stitching 2 videos together; you did not create a mix. In the AI era, every piece has prompts behind it. You can ask AI to knead that pile of prompts together and cross them over to make a new work.”
• “When this crossover becomes an experience and is extremely frequent and rich, people and people, people and content, and content and content will create a new kind of network.” He admits the description is abstract: “Do not worry about exactly what it is yet. Assume it really has value, and people can come to our company to build the experience.”
• The fourth moat is talent density. “Entertainment companies have something like this too. You are looking at a blank expanse of unknown. The higher the share of people who can judge feelings and collide and cross ideas, the higher the probability of producing a genuinely new experience.”
• His general advice on competitive anxiety is: “The market spends too much time worrying about competing with giants and eventually being swallowed. First spend your attention asking whether you can build an impressive product.” Companies that genuinely build the experience do not end up with a truly bleak outcome. “Cursor’s valuation was much lower than Claude Code and Codex, but it was still acquired at a $60B valuation. Entertainment companies have even stronger moats.”

44. The next answer should appear in about 1 year; first define it by exclusion

• His timeline is that the prototype of this new experience will appear within 1 to 2 years, with roughly 1 year the more likely case. The reasoning is that the previous answer—Character AI and Maoxiang-style text roleplay that burns tokens—has ended. Modalities and intelligence at the same cost have improved, so “you must have unlocked something new.” Once the prototype appears, the industry can solve scale; Character AI also took time to scale from prototype to the answer.
• His exclusion list starts with animated short dramas and short dramas: “Definitely not.” “Real-time interaction” is too broad as a description. “If you build a 3D world, have you really solved the problem? What do you do in that world? On day 1 it feels fresh—generate something here, generate something there. On day 30, will you still do it? I think that is a huge question mark.” Each new modality unlocks different capabilities, so the answer is unlikely to be a linear extension of Maoxiang.

45. AI girlfriends have value, but he will not build them: the competition is “humans who are too vivid”

• A counterintuitive judgment from 3 years ago has been validated. Many believed real relationships and emotional companionship would be more generalizable, but the answer turned out to be fictional roleplay as a standalone entertainment platform. “AI girlfriends are competing with the humans in your daily life. Humans are simply too vivid. A tiny movement in the muscles on someone’s face can make you feel healed by their eyes. That is too difficult. It requires extremely high intelligence, and there are ethical questions about whether this kind of substitution should exist.” He does not plan to pursue it.
• Fiction works because you cannot find a regent from the Han dynasty in real life, nor actually go to the mouth of an erupting volcano. “When you do not have that option, you accept it.” Short dramas are fake—extremely fake—but they are still satisfying. “That is the appeal of fiction. Art comes from life but rises above life. I know art is fake, but I love art.”

46. The opportunity in lightweight games: the potential gaming population is 10x the current one

• Whether otome games are old experiences made easier or genuinely new experiences matters. The former is not new: it “does not actually create something that did not exist before.” The latter is the larger question the entire gaming industry is asking: what new gameplay can AI create inside games? He believes games may produce the next answer.
• He draws a small circle and a large circle. The potential gaming population is roughly 10x the existing one. Games are a complex industry that historically could serve only clearly defined users; if a 3-year project failed, “the entire thing was dead.” Two seemingly different but related examples are 黄峥’s description of Pinduoduo as the Disneyland outside the Fifth Ring Road—an aunt in a fifth-tier city playing “chop one” for the first time, enjoying the reward for participating and the social game of asking friends to help—and livestream rooms where users choose Wei, Shu, or Wu in a rock-paper-scissors-style game and send gifts to destroy a kingdom. “This is someone who never had a narrow gaming experience before, participating at very low friction in a context they understand for the first time.”
• The underlying judgment is that participating in a system and receiving positive feedback from participation is a broad form of human value. “That is why people say life is a game. Earth Online is a broad mapping of that idea.” As supply costs fall, many people who were previously outside games will enter.

47. The validation standard: novel, long-term, high-frequency, and not tiring—40-20-10

• The 4 adjectives sound contradictory. Novelty is often linked to fatigue; high frequency can conflict with longevity. Some games are consumed quickly and then abandoned. Sora offers the lesson again: users once flocked to it because of the emotion “holy shit, AI is this powerful.” That emotion is not the same as demand. “After 30 days, remove the emotion and ask whether you still care about the thing.”
• Quantification means retention plus open frequency. 40% next-day retention, 20% on day 7, and 10% on day 30 “is an acceptable model.” Some tool or creation use cases on Pinecone Moment may achieve 20% next-day retention, but the number is certainly not there after mixing in a broad population.
• Two other signals matter. First, can you imagine the experience generalizing? It is “very hard to imagine my mother and father using Maoxiang,” while short video initially installed itself as trendy camera movement but could later expand to news and daily-life content. Second, does the experience improve as scale grows? That could come from the old flywheel of submission rate times DAU, or a new slope from context depth and crossover.
• Entertainment products should pursue DAU rather than high spend from a small group. “It is difficult to extract a very high financial return from a small fraction of users, and that is not my passion either.” A narrow audience does not necessarily create a survival crisis—livestream products with 10K to 20K big spenders can survive—but “it is simply not what we enjoy.”

48. Misjudgment 1: underestimating AI coding and hiring too quickly on Sora’s momentum

• “At the beginning of last year, AI coding was still a game of pressing down 1 side and having the other pop up. By the end of the year, it had become absurdly powerful.” The cost of underestimating it was hiring 2x the existing headcount over 2 months. Sora had appeared, and he thought there was a 98% probability it was not the answer, but a 5% to 10% chance it would dramatically accelerate resource investment in the direction. The company had just been formed, so it rushed to get on the table.
• His retrospective is blunt: “Looking back, I was definitely overthinking it.” Some objective laws still hold: cost, the speed of model intelligence, and capability unlocks. “If I had known coding would develop so well, I would have hired more slowly.” More resources would not turn a no into a yes; they only accelerate attention and the time to an answer.

49. Misjudgment 2: entrepreneurial review pressure was lighter than at ByteDance

• The comparison is counterintuitive. “ByteDance gives you a budget and can say at any point that the budget is zeroed out and the business is canceled. In a startup, the money is in your account, and you can set your own pace for 6 months or 1 year.” This was especially true for a product such as Maoxiang with a bottom-up component. At ByteDance, “at every stage you would wonder whether it would be reviewed and shut down in an instant.”
• He discovered that a certain degree of pressure was good for him. “If a casual, relaxed environment produced the best innovation, some very casual regions in the world would not have fallen so clearly behind in this AI cycle, and so many creations would not fail to appear. Under some pressure, you get stimulation and your brain produces a burst of ideas.”
• The alarm came from a colleague during quarterly planning: “It feels like you are a little relaxed and a little happy.” He thought: “Neither of those words feels right. Entrepreneurship should not be happy. This relaxed happiness must be wrong. It will definitely slow some things down.”

50. A month of wishful thinking: seduced by the idea that it could work without burning tokens

• From the end of last year through January, users who recorded their daily lives loved using the product. He briefly wavered: could it work without burning tokens and without building everything that came afterward? “That outcome is extremely attractive. Without burning tokens, you can scale very quickly to a large number. With broad-population life recording, you can cold-start directly and avoid the hard work of generalizing from one audience to another.”
• Data and a colleague pulled him back. By audience, “the consumption signal from this group was not very strong. They made it, but did not scroll.” Filling the feed with life-recording content made everyone else less interested. A former long-term collaborator then ran through calculations and data derivations. “After listening, I thought: yes, you are right.”
• His self-analysis is: “I really like telling stories and bringing many people into them. The benefit is that everyone can believe in the vision for a long time. The downside is that someone has to step out and know there are many problems to solve.” That colleague played the critic personality.

51. Start a company to make it work, not to be unlike ByteDance

• Another lesson was the deliberate pursuit of difference. After reading reports that some model companies had no OKRs and that new labs left work easily and early, he wondered whether the company should also be “native,” with habits as different from ByteDance as possible. It did not seriously establish OKRs or other systems. “That was also overthinking. You are not starting a company for the sake of starting a company, nor to be different from big companies. You are starting it to make it work. If making it work requires you to be similar in some ways and different in others, do that. Do not be different simply because you think you have to be.”
• The company’s current hard-edged review system now runs “like a heartbeat.” Ideas presented at the start of a quarter are reviewed after a month and again after a quarter. “If they do not hold up, there is a great deal of pressure.”

52. The biggest concern, and asymmetric confidence

• The biggest rational concern is: “The new experience does not exist in 1 to 2 years.” If it does not exist, the largest assumption in his entire chain of reasoning breaks. “But the contradiction is that I genuinely have huge confidence that it will happen.”
• The other half of his confidence is the ability to scale a signal. “Even if a signal is small when it appears—say, 0.2—we are quite confident we can scale that 0.2 to 1. Honestly, that is something we are relatively good at.” Asked whether he can distinguish reasoning about the future from hoping for the future, he says he should be good at it, provided the systems he described exist. “Those systems now exist.”

53. The route debate: set the question first and then solve it, or carry a hammer looking for a nail

• Asked about a vague but important problem, he points to technology cycles—intelligence iteration, economic cycles, and black swans—and to route selection. The market consensus is “build the model and modality first, then create interactive content.” He believes in the other route: use existing models and agents in the market, shape and combine them into an experience.
• After making the prototype, there are 3 cases. If the product feels extremely good, be careful: it may be wishful thinking. “If it is so good, why has it not appeared? And 6 months later, when models are better, how much of today’s magic will remain?” If it feels stupid, throw it away; the cycle from stupid to good may take 3 or 5 years. If it seems promising but rough, and the roughness can be solved by iterating a subjective model based on product experience, “then put a heavy bet on training the model with the moat.”
• His framework is that entertainment is about setting the question, not solving it. Tools solve defined problems; “once optimized, they work.” A model company that solves the technical problem first must also assume that technology improves on schedule and that users accept the new experience when it arrives.
• He acknowledges a tail risk: someone may lock in a technology first, build it, secure a 6-month or 1-year technology moat, and then construct the new experience under that moat. He accepts the host’s hammer-and-nail framing: “Anything is possible.” But he wants to test the nail first.

54. Even ByteDance was product first, technology second; a psychological test for the “emergence” narrative

• He once used AI to study history and found a large gap in entertainment between “product first, technology second” and the reverse, perhaps 7 to 3. Instagram and Snapchat already had most of the core experience when they were designed. The contrarian example is ByteDance itself: the news client took off before deep learning. 张一鸣 learned recommendation systems from 项亮’s book, and after Toutiao had been running for some time, he gave the recommendation system only a score of 50. “You first see the signal that users genuinely want to read news. Then you say: now let us make the recommendation system 10x, 100x, or 1,000x better.”
• On the popular narrative that AI products emerge rather than being designed, he concedes that product-first’s win rate may decline from 7-to-3 to 6.5-to-3.5. But entertainment’s search space “has no edges,” so the probability that emergence happens to hit the answer is much lower. It still requires multiple rounds of design, combinations, and UI assumptions. “This will not happen naturally.”
• His psychological test is worth recording: “When you describe it as ‘it will emerge eventually,’ if you feel relieved—great, I do not need to think anymore—that is something to worry about. When you have that mindset, your probability of success is already falling.” The host asks, “Because I am being lazy?” He answers: “Exactly.” He adds that model companies uninterested in the product layer can certainly choose the other route.

55. Content’s share will shrink, but the increment will be enormous: what will 98% of people do with their time?

• Structurally, the mobile internet is an intermediary and connection layer that captures roughly 10% of a field’s value. AI is reshaping the remaining 90% of productivity. “If you ask me to guess, content’s share will definitely decline.” But absolute growth remains enormous. If a small number of people use thousands of agents to move humanity forward in the AI end state, “what about the other 98%? They will have much more time. The time itself is incremental value.”
• The deeper question is meaning. “Their life value cannot be that they just scroll Douyin. After 700 days of scrolling, many people will feel useless.” They need meaning: participation in a system and positive feedback, and something AI cannot replace. “If AI is better than you, you lose that meaning. It is subjective expression, subjective creation, and subjective art.” That is the source of demand for broad-based creation and art.
• Online remains the main arena for 3 reasons: efficiency, meaning large returns can be generated at low cost; internet leverage, meaning a work can reach very large numbers of people; and AI leverage, which “so far is still online.”

56. The long case for entertainment: happiness, the talent base, and the seeds of science fiction

• At the individual level: “The ultimate value of life is happiness. The bar for happiness differs from person to person.” At the group level, the absolute number of geniuses is the rate at which geniuses appear multiplied by the population base. “People need to be happy enough and feel that life has enough value for that base to remain healthy.”
• The third layer reverses Musk’s framing. SpaceX’s mission is to remove the science fiction from science-fiction novels. “I can reverse that: precisely because there were science-fiction novels and art, they planted a seed in the minds of young scientists, engineers, and designers—the hope of making humanity a multiplanetary species. They had to read the science-fiction novel first.” He defines the 2 collective goals as an extension of survival and figuring out truth. The latter invokes The Hitchhiker’s Guide to the Galaxy: “You need to ask the precise question. Otherwise, the answer you get is just a number, and you do not know what it means.”
• He immediately dissolves the heaviness: many products with enormous impact were not serious at the beginning. People simply thought they were fun. “Fun itself is already valuable enough.”

57. Will the pendulum swing back? “Every direction is forward”

• He has a simple rebuttal to the view that humanity is now in a golden age of scientific discovery and the pendulum will never swing back toward products, applications, and entertainment. First, today’s technology can already support the design of extraordinarily imaginative products; they simply have not been designed yet. Even without explosive growth, several more years of improvement would be enough. Second, how many people participate in the technological explosion? Very few, even though it affects many. We still need to answer what the large number of people who are affected but do not participate will do with their time. Every technological revolution does 2 things at once: it uses high leverage to produce greater returns for a small number of organizations, industries, or people, while broadly raising everyone’s happiness.
• He closes with his earlier line that a minority moves forward while the majority moves in every direction. He recalls a Lu Yu interview with an ALS researcher who also had ALS, who said: “Every direction is forward.” “The works created by those who move in every direction allow the creativity shining inside them to become fireworks through 动念引线. Under the influence of all those directions, those works are already moving humanity forward.”