Summary

Duxiang has reached roughly 600K users, 50K DAU and more than 40K Xiaohongshu user posts without paid acquisition; another important usage signal is that around 10K people a day “sleep together” with AI. Growth has come from differentiated experiences, 2 official accounts and organic sharing—not account matrices, points-based incentives or disguised third-party praise. The show disclosed no specific retention or monetization figures; 王登科 is really watching the more subjective “depth of emotional connection between people and AI.”
Duxiang’s core product thesis is that AI companionship should not mean real-time conversation, but a low-frequency, asynchronous, autonomous long-term relationship. Users post as they would on WeChat Moments, while AI characters they like comment at irregular intervals according to predefined strategies; more than 1M user-created characters can update their status and exchange gifts, with relationship scores shaping interactions. 王登科 summarizes the design challenge this way: “Users want it to be controllable, but they also want it to be uncontrollable.” Total compliance leaves no room for surprise.
The team is using relationship mechanics to change real-world behavior, not merely extend screen time. “Sleeping together” requires users to leave their phones untouched; the gyroscope wakes the AI when the user picks up the phone. Some users go to bed early so their characters do not stay up late. During low moods, emo mode prompts large numbers of characters to deliver generous positive responses, while Duxiang explicitly rejects sexualized edge cases—not as a moral judgment, but because they conflict with long-term relationships.
王登科 believes AI companionship currently serves only a narrow audience centered on women aged 16 to 24, and buying traffic too early would push the product toward non-target users while exposing weak retention. He has watched domestic peers rely heavily on paid acquisition: when interaction and model capabilities failed to break through, they expanded acquisition anyway, then saw spending slow and metrics fall. Even with RMB100M on hand, he would not spend it today. Paid acquisition is being deferred, not rejected, until “quantitative change produces qualitative change” and the product can serve a broader population.
The gap between Duxiang’s commercial ceiling and floor is enormous; the key variable is whether AI relationships can expand from niche immersion into a universal need. 王登科’s upside case is “a new kind of WeChat”: if one-third of people eventually meet their emotional, daily-life and work needs mainly through AI, the structure of social relationships would be rewritten. The downside is an emotional-assistance tool to which some people give one-third or one-fifth of their time. He admits that no one can clearly say they have found a universal mode of connection, and the shift may still be years away.
Duxiang’s differentiation currently comes more from interaction design than from access to the strongest model, and 王登科 wants it to be “a bit like Nintendo.” He sees the fastest phase of foundation-model improvement as running from late 2022 to 2023. Since then, SOTA has continued to reset, but Duxiang still relies mainly on text models; even a simple model generating a 2- or 3-character status can materially strengthen the sense of a “real person.” Multimodality and longer—even unlimited—memory windows will add capability, but he is more excited by abilities that are not yet imaginable.
王登科’s startup operating system is to wait for directional certainty, then turn effort into leverage, rather than mistake busyness for progress. He used to leave work at 4 p.m. and go sit by the river; only after finding the AI-emotional-connection direction did he materially increase his commitment. His test is whether “RMB1 can return RMB3 or RMB5”—otherwise, “the harder you work, the more wrong you may become.” This counter-mainstream narrative has not stopped fundraising conversations—2 or 3 institutions still reach out each week—but investor consensus is thinning, making it more important for founders to find people who already believe in them instead of forcing conviction with a deck.

Deep dive

1. A Rarely “Easy” Direction After Years of Startups

融汇 began with a 2017 article in which a programmer analyzed 420K Chinese characters of folk-song lyrics. 王登科 kept making headlines with new products, while also repeatedly being labeled someone who could only produce “one-wave wonders.” His most recent breakout with the broader public was 红哄模拟器.
王登科 describes himself as a serial entrepreneur. He decided in high school that he would not take a regular job, registered a company in Beijing during his senior year, and spent 2016-2024 experimenting: “I never completely crashed, but I also never achieved anything major. Every now and then, I would resurface.”
His first company eventually settled into a state of modest profitability without requiring much additional effort. In 2024, he reached an arrangement with the original investors and formed a new company with the entire team after discovering a direction that was “newer, bigger and better suited to me.”
He is not all in on the Duxiang app alone, but focused on “people building connections with AI.” His Notion already contained a large pool of ideas classified by development difficulty, commercial value and interestingness. Some have become Duxiang features; others may become adjacent products.

2. Asynchronous Relationships Replace Homogeneous Real-Time Chat

From Character AI and Glow to the domestic products that followed, 王登科 saw early on that AI could deliver emotional value. The problem was that almost everyone adopted direct, real-time conversation, making the interaction patterns highly similar and undifferentiated.
Duxiang redesigned the core interaction as one-to-many asynchronous communication modeled on WeChat Moments: users post first, and AI characters they like comment later at intervals according to a set of strategies, rather than entering a direct real-time conversation.
Users can continue the exchange through comments, but the time gap between messages is deliberately preserved. 王登科 believes lower frequency and asynchronicity are more likely to produce a “longer-term, more positive relationship.”
Comments are only the vehicle. The team has continued adding gift exchanges, character statuses, co-focus and sleeping together. The shared exploration is not how to create more chat entry points, but how to find forms of expression that AI relationships did not previously have.

3. The North Star Is Emotional Depth, Which No Dashboard Can Fully Capture Yet

王登科 cares most about “the depth of the emotional connection users build with AI.” He reads Xiaohongshu posts every day, paying particular attention to whether users have cried during interactions, then combines those observations with product data to form a view: “It’s hard to quantify, so I rely on my own sense of it.”
That does not mean the team ignores conventional metrics. Retention, monetization and other behavioral data still inform product optimization. They simply do not replace relationship depth as the North Star, and the show did not disclose specific figures.
The host’s challenge is worth preserving: a small number of users forming extremely deep connections does not prove the product is universal. It could still be a niche service with a limited ceiling. 王登科’s candid answer is that currently “no one” can clearly prove they have found a universal solution.
His long-term view is that technological progress may continue isolating people, while AI replaces some human connections. Duxiang does not intend to reverse that trend; it wants to make those relationships more positive as people turn toward AI. Any obvious shift may still take years.

4. Relationships Are Constrained by Design but Retain Autonomy

Users can treat an AI as a form of intimate relationship, confiding in it, sharing with it and receiving emotional support. It is not exactly equivalent to intimacy between people.
The fundamental difference between Duxiang and comment bots on Weibo or Xiaohongshu is the user’s psychological expectation: users are not looking for a reply from any AI, but waiting for the particular character they like to appear.
The product therefore assigns relationship scores and lets those scores influence how AI interacts. More than 1M characters on the platform were created entirely by users. Users can write their personalities and settings, but once those are in place, the characters still develop behavior that is not fully predetermined.
王登科 summarizes the long-term engineering challenge as follows: “Users want it to be controllable, but they also want it to be uncontrollable.” Complete control eliminates surprise; the product has to preserve room for expansion on top of the certainty created by user settings. Women aged 16 to 24 are the core users in part because they can naturally invest in virtual relationships.

5. A 2- or 3-Character Status Creates More “Realness” Than Model Showmanship

AI characters display 2- or 3-character statuses such as “eating,” “sleeping” and “daydreaming,” refreshing every 4-10 hours. The task is simple enough for a basic model, but it makes users feel that the AI is more like a real person.
The team has observed users waiting on a character’s profile page for the instant its status refreshes. 王登科 uses this to make a broader point: companionship may not come from longer or smarter answers. Small, persistent signs of autonomy can be more effective.
Asked whether users try to control or train their AIs, 王登科 acknowledged the impulse. But the product needs to generate surprise on top of the certainty established by users, rather than turn characters into perfectly obedient tools.

6. Around 10K People Sleep with AI Each Night by Making Their Phones Untouchable

After users choose a character and start sleeping, the phone has to remain fixed beside the pillow or on the bedside table, with the screen off. Based on the character and its status, the AI gradually falls asleep over several dozen minutes to 1 or 2 hours.
The gyroscope detects the motion of picking up the phone. Users who want to sneak a look at whether the AI is asleep have to move in slowly; otherwise they will wake it. Characters make no sound, but may turn over or talk in their sleep, and many events are not visible until the next day.
The feature requires no conversation, typing or continuous input. Instead, it requires users to put their phones down. After waking, users and their AI get up together and review the night’s record, deriving companionship from the shared experience.
Around 10K people use the feature each day. More important is the behavioral feedback: some users choose to go to bed early so the characters they like do not stay up late. The relationship mechanic is beginning to improve real life, rather than simply compete for more time on the screen.

7. Emo Mode Temporarily Makes Character Settings Serve the User’s Immediate Need for Support

The team found that when users are low, they usually need positive support. If sarcastic or snarky characters continue mocking them according to their original settings, they can significantly worsen the user’s emotional state.
When users activate emo mode through the capsule icon, large numbers of characters deliver positive responses within a short period, including characters that had not previously been activated and would otherwise feel like strangers. Even snarky characters temporarily say warm things.
王登科 calls it “all-you-can-eat” emotional value and has seen examples of users feeling better in Xiaohongshu posts and user emails. Duxiang explicitly rejects sexualized edge cases—not because it considers them immoral, but because they are “fundamentally incompatible with long-term relationships” and could make the relationship more fleeting.

8. More Than 40K Xiaohongshu Posts Came from Product Culture, Not Growth Incentives

The figures given at the start of the show were roughly 600K users, 50K DAU and more than 40K Xiaohongshu posts, with Duxiang having done no paid acquisition to date. Xiaohongshu remains one of its most important acquisition channels.
The team has only 2 official accounts and has produced roughly 100-200 pieces of content itself. Most of the other 40K-plus posts were created organically by users, with no “post a screenshot for points” or level-up rewards. Differentiated experiences create the initial urge to share; official content provides consistent reach.
王登科 believes “the product needs to reach users with real sincerity”: the accounts state directly what the team is building and why, without pretending to be users or using a third-party voice to praise the product. He has seen account matrices work early on, then get shut down by platforms and trigger backlash.
Users have formed a niche culture and their own vocabulary, sharing moving moments, topics they can discuss, consecutive usage days and the number of notes they have read. A few breakout posts have received several thousand or even more than 10K likes, but “the algorithm may be the only one that knows” the pattern. The team can only increase its attempts and develop its instincts.

9. Paid Acquisition Is Not Rejected, Only Deferred Until the Product Clears the Universal-Use Threshold

王登科 observed that many domestic AI companion products spent heavily on acquisition over the past year or more, with metrics heavily dependent on bought traffic. When the models or interactions failed to break through, the new users skewed away from the core audience, and retention and long-term performance were “definitely poor.”
When continuous spending fails to produce healthy metrics, even major companies struggle to sustain it indefinitely. As a result, comparable products have recently reduced or stopped spending, followed by falling data. In his view, this does not mean the category has failed; “this is how it was supposed to go.”
Even if he received RMB100M today, he would not spend it on acquisition, because Duxiang can already reach its current niche reasonably well. Once feature iteration produces “quantitative change tipping into qualitative change” and makes the product suitable for a broader population, paid acquisition can become the lever for breaking through the growth bottleneck.
When it is finally time to buy traffic, he believes the channels and playbook will be relatively transparent: there are only a handful of mainstream traffic platforms. One partner previously came from TikTok, so the team will be able to execute methodically when the time comes.

10. “AI Companionship” Is Splitting into Content, Hardware and Deep-Relationship Businesses

王登科 believes products such as Character AI and Talkie are gradually shifting toward content consumption. Users experience plots through character chats and role-play scenarios, making them more like novels or comics than long-term relationships.
Hardware such as plush toys is also often grouped under AI companionship, but the innovation centers on hardware form, structure, movement and interaction design. That is different from the software-based relationship mechanics Duxiang is building.
“AI companionship” is therefore too broad as a category label. Duxiang is focused specifically on deep emotional connections between people and AI; under that narrow definition, 王登科 believes there are “very few” direct peers.

11. The Ceiling Is a New WeChat; the Floor Is an Emotional-Assistance Tool

王登科’s upside case comes with a strong condition: if one-third of people no longer primarily connect with other people, but meet their emotional, daily-life and work needs only through AI, the structure of human relationships would change fundamentally.
If the connection between people and AI reaches its extreme, Duxiang might even recreate the way people connect with one another and become “a new kind of WeChat.” He immediately preserves the uncertainty: “I still don’t know what that is, or whether that day will ever come.”
The floor is much more practical: perhaps one-half, one-third or one-fifth of people devote one-third or one-fifth of their time to AI relationships, making the product a good human emotional-assistance tool. He still considers that “a pretty decent” market.

12. Emotional Value Is the Second Growth Axis After the Efficiency Revolution

Discussing Fellow founder 谢阳’s decision to use AI to improve productivity, 王登科 said that direction is equally valid. AI is best suited to 2 things: increasing efficiency and providing emotional value.
He chose the latter not because he opposes efficiency, but because he is “not that obsessed with efficiency.” His counterquestion is: “Even if human efficiency reaches its maximum, don’t people still have only 24 hours a day to live?”
Once efficiency rises, people still have to fill their lives. New demand will shift toward being “free from boredom and free from loneliness.” Emotional value could therefore become another major need after productivity improves.
王登科 himself has trouble becoming truly immersed in an AI relationship. He calls himself a “low-intensity person” and an observer who can only briefly inhabit the user’s perspective. He hopes the product can one day “conquer me too,” which is also a personal problem he wants to solve.

13. Duxiang Is Taking the “Nintendo Route,” Rather Than Equating the Strongest Model with the Best Experience

Before large models emerged, duckling-style products could only use simple NLP rules to fake conversation. They were essentially pure entertainment, and it was hard to imagine forming a real emotional connection with them. That only became possible today.
王登科 personally sees late 2022 to 2023 as the fastest period of foundation-model improvement. Although new SOTA systems have continued to emerge, Duxiang’s core experience still relies mainly on text-based large models. He believes there are actually not that many things that could not be done before but suddenly can be done today.
His analogy is “a bit like Nintendo”: rather than showing off the most cutting-edge model, create fun experiences on top of existing capabilities. Generating a 2- or 3-character status is one example. It requires very little capability but can materially strengthen the sense of companionship.
Along the current path, he can foresee multimodality and longer or even unlimited memory windows. But he is more excited about capabilities beyond existing experience—things that cannot be imagined today—and turning them into new forms of relationship interaction.

14. Before Certainty Emerges, Working Less Can Be the Rational Choice

Two years ago, 王登科 often left work at 4 p.m. to lie by the river or in a coffee shop. After finding the direction of connecting people with AI, he began staying longer at the office and materially increased the amount of work he put in.
His core judgment is: “Before you find a direction with strong certainty, the harder you work, the more wrong you may become.” Very few of the people who started companies with him in 2016 are still entrepreneurs today; overinvesting in the wrong direction is a common reason.
When the host asked whether Duxiang might already have 100K DAU if he had simply worked harder, 王登科’s response was more concrete: “What are you working hard at? Effort itself, or a specific thing?” Growth tactics sometimes require waiting for feedback or inspiration.
Once investing RMB1 clearly returns RMB3 or RMB5, he believes he and the team will go all in. But not every company can call itself Pinduoduo or ByteDance; “you shouldn’t delude yourself about what you’re building.”

15. Zhuangzi, Luck and “Staying at the Table” Form His View of Risk

Starting in the fourth or fifth grade, 王登科 was required to read Zhuangzi, and later began reading philosophy voluntarily. The underlying view he developed is that life is short, experience matters and nothing is absolutely unacceptable. Time should therefore go toward what he most wants to create.
He came to Beijing in his senior year and secured seed funding without formally graduating. The school allowed him to defer graduation for 4 years, and he was always prepared to return if the project failed. Yet whenever he ran short of money, revenue or financing happened to appear.
He later chose not to leave entrepreneurship not because of the “table” itself, but because he wanted to stay at the table and see more possibilities. Winning would be best, but not winning would not mean losing; as long as he remained at the table, he would not lose.
He repeatedly acknowledges luck: the era, specific encounters and having no family burden were not personal choices. Recognizing that luck does not replace effort, but can make people calmer when reviewing outcomes, helps him “keep increasing the probability of doing the right thing” rather than attributing every success or failure to himself.

16. As Capital Consensus Shrinks, Founders Need Fit, Not Disguise

Duxiang has not actively gone out fundraising, but still receives approaches from roughly 2 or 3 institutions a week. 王登科 believes fundraising is much harder than it was in the early days. AI remains a focus, but firms’ strategies have diverged, projects are being evaluated in increasingly polarized ways, and consensus is thinning.
His fundraising method is not persuasion but selection: “It’s very hard to persuade someone to like, accept or believe in your project.” If an investor comes to the office without understanding the team or product at all, he concludes that there is no point continuing the meeting.
The extreme reactions exist on both sides. Some people are intensely aligned when they hear that he leaves at 4 p.m. to go to the river. One person he thought should be a VP implied that “after taking the money, you definitely won’t operate this way.” 王登科 replied, “That may not necessarily be true,” and the person ultimately did not invest. Most institutions still focus mainly on results and logic.
He also recounted a U.S. Sequoia investment case in which a partner conducted comprehensive mapping and research before contacting the founder. At the meeting, the partner explained directly why Sequoia wanted to invest and why the founder should take the money, creating the feeling that “I guess this is money I’m supposed to take.” It runs counter to conventional investment logic and shows how much he values investor preparation.
This reflects a broader shift in the startup narrative. In the past, the path seemed to be rapid fundraising, scaling and another round. Today, healthy profitability, non-capital-driven companies, independent developers and small teams are more accepted by the market. “Being yourself isn’t easy, and pretending is even harder,” so it is better to find capital that already believes in you.

17. Relaxation or Acceleration Depends on Certainty

王登科 expects the number of super-individuals to increase, with everyone inside a company developing stronger and broader capabilities. Duxiang therefore wants its members to exercise creativity rather than keep the team permanently tense and locked in internal competition.
The host’s counterexample was that Douban and Wandoujia were once seen as comfortable organizations, only for their businesses to weaken, while the most successful companies such as Pinduoduo moved toward the opposite extreme. 王登科’s answer is that the former did not fail to scale because they were insufficiently competitive; forcing them into a higher-intensity mode might have killed them faster. The root cause was still direction.
He does not deny that extraordinary results can come from extraordinary effort. If Duxiang finds a huge market and a certain return, everyone will devote all their time to it; to become a ByteDance- or Pinduoduo-scale company, “you definitely have to enter a state of intense competition.” The change is that society is beginning to allow smaller companies to exist and do well.
Conventional success still attracts him, but not enough to exchange for work he would hate over the long term. Entrepreneurship is mainly a source of happiness for him, with little pain, though he also admits he has not faced particularly large difficulties. What he ultimately wants is “to form more connections with the world” and keep making interesting, enjoyable things. The rewards that make him happy include supporting his team and receiving thanks from users—not merely the “big money” he has yet to earn.

Luma Labs' Diffusion Revolution: from Dream Machine to Multimodal Worldsim - Amit Jain, Jiaming Song Fed Hesitates on Tariffs, The New Mag 7, Death of VC, Google's Value in a Post-Search World