She Wants to Build the AI Era’s “Super Playground” | A Conversation with Roi, Founder and CEO of 幕间

Summary

• 幕间 is not betting on more human-like chatbots, but on AI simulated worlds made up of Agents that can be restarted over and over. Roi divides role-playing products into companionship, interactive narrative, and simulation. 幕间 chose the third path: compressing real-world systems into rules so that dating, time travel, stock trading—even “being a cockroach”—become “lives you cannot experience in reality.” The core promise is: “Earth OL is a terrible game because you only get to play it once,” whereas here, users can restart indefinitely.
• None of the three AI interaction paths has fully reached the consumer market, but text simulators have already produced the strongest monetization proof. Traditional AI-generated games remain simple and formulaic, while world models are furthest from being both playable and cost-effective. Text simulators, by contrast, are already selling for RMB30 on Xiaohongshu, with some products moving several thousand to more than 10,000 copies—even though buyers still have to configure their own APIs. That is why 幕间 does not treat token consumption as its core metric: “Tokens represent cost.” Its North Star is closer to the number of dialogue rounds users complete.
• If the so-called “TikTok of the AI era” is only an endless feed of mini-games, its moat is probably within TikTok’s reach. Roi’s objection is that this is not a fundamentally new interaction model: TikTok already supports H5 games, has roughly 100M DAU using its chat function, and owns a mature social graph. Once a use case proves out, a large platform can reproduce the product format with little difficulty. An independent platform must build AI infrastructure, a distinct understanding of what is fun, and a base of early power users—not merely wrap the experience in an information feed.
• The most counterintuitive design principle for AI-native entertainment is to create open-ended experiences with convergent systems, rather than pursue unlimited freedom. “Complete freedom is guaranteed to be unfun,” because unconstrained input raises the bar for expression and offers no assurance that the output will be interesting. A good simulator limits the boundaries of interaction while increasing the odds of positive feedback and the impulse to take the next step. A stock-trading simulator can make users “profit no matter what,” or lose “no matter what” and turn self-mockery as a bagholder into part of the fun—virtual losses can be designed as entertainment too.
• 幕间’s key UGC assumption is that AI will turn people with taste and something to say into “one-person armies.” The platform currently has roughly 10 full-time employees and completed its angel round last year. It has no revenue or profit in the cold-start phase, but its burn is also low. Early creators are mainly post-2000 students and young white-collar workers in first- and second-tier cities, along with kindergarten teachers and stay-at-home mothers—not traditional geeks. Romance is the entry point, but the product is evolving from “dating the guy who runs a café” to “running the café itself is more fun.”
• The creator economy has emerged ahead of platform monetization, and 幕间 wants to bring underground transactions, distribution, and token costs into one system. Creators already trade prompts, interfaces, artwork, music, and code; a complete work can generate several thousand to 10,000 transactions. The platform may eventually bundle paid works and token usage into a subscription, but Roi acknowledges that the business model is still being explored. She wants to build “Xiaohongshu for AI interactive experiences”: identities shared across worlds, works distributed through the app, links, or mini-programs, while the account and spending relationship remain on the platform.
• World models may cross the productization threshold within a year, but a mass-market hit will depend on stability, low latency, and real-time multimedia—not the model demo itself. Roi compares them with Midjourney and video models, which went from “ghost images” to usable products in 1-2 years, and argues that today’s slow, expensive, ugly world models could also improve rapidly. Koji notes that products such as Google Genie 3 currently involve more WASD exploration and observation than genuine interaction; Roi agrees that the distance from a model to a genuinely playable product remains substantial. The real opportunity belongs to a platform that is both AI-native and fluent in game design, though she still sees meaningful risks: the romance entry point, the simulator direction, and the timing of any pivot could all prove wrong.

Deep dive

1. 幕间 Sells Not Chat, but an Ideal Life You Can Restart Indefinitely

• Koji played a 16-person speed-dating game that randomly generates 16 potential matches based on the user’s identity inputs. After talking to each person for 2 minutes, the player chooses one for a deeper relationship. Other works on the platform include traveling back to the Song dynasty, receiving tips from wealthy older women, playing an emperor, and experiencing “a cockroach’s life.”

• Roi summarizes the common demand behind these experiences as “enjoying a life you cannot experience in reality.” In the real world, a date may not bring 16 ideal matches at once, and life cannot be restarted. Fictional worlds let users repeatedly test different identities, relationships, and fates.

• Her most distinctive analogy is: “Everyone in the real world only gets to live once… Earth OL is a terrible game because you only get to play it once.” 幕间 is trying to make every user the protagonist, with AI playing the NPCs orbiting that protagonist.

• The company currently has roughly 10 full-time employees and completed its angel round last year. Roi says the cold start is still in its early stages: “We have no revenue and no profit,” but “we also haven’t spent much money.”

2. Role-Playing Is Splitting into Companionship, Narrative, and Simulation Systems

• Roi sees role play, represented by Character.AI, as a sector that could generate more API calls than programming, and organizes the product opportunity into three extensions rather than one unified “AI companionship” market.

• The first route is about the role itself. If models can simulate real people, products can substitute for part of users’ real-world social and relationship needs. Early product Replika experimented with 3D avatars, daily conversation, reminders, and alarms; some newer products are going further by pulling in real-world data to provide both emotional and utility value.

• The second route centers on narrative content generated during interaction, extending from interactive fiction into multimodal interactive dramas and games. Roi cites 猫箱, whose multiple characters, rich backstory, plot progression, and “fake heiress comeback” themes make it resemble an interactive version of 番茄小说 or an AVG game.

• 幕间 chose the third route: not chatting with a single Agent, and not merely having a model continue a story, but having Agents form a functioning system that simulates a social system or small-scale real-world structure. The structure is compressed and filtered to retain “all the fun parts.”

3. World Models Supply the World’s Frame; the Product Still Has to Explain How to Play

• Roi classifies Gamble and newer AI game engines such as TapTap as tools: AI generates conventional genres such as platformers and bullet-hell shooters, essentially evolving traditional game engines with AI rather than redefining games.

• World models such as Google Genie 3 and Odyssey are “interactive, playable models,” but their productization so far is primarily a showcase for model capability. Gameplay remains limited. As Koji puts it, users move with WASD, change viewpoints, and observe rather than truly interact. Roi believes there is still a wide gap between these models and genuinely playable products.

• When Koji asks whether 幕间 competes with world models, Roi says it will absorb rather than replace them. 幕间 wants to provide the framework that “turns models into fun products,” compatible with language models, multimodal models, and mature world models in the future: “I’m a frame. I can put every model inside it.”

4. Text Simulators Are Closest to Consumers, but Still Not Irresistible

• Koji uses his own engagement to calibrate the product’s pull: the 16-person speed-dating game is the work he played longest and it successfully drew him in, but it still did not reach the point where he kept thinking about it after logging off and wanted to play nonstop. Roi concedes that all three routes remain some distance from mature consumer products.

• The problem with AI-generated traditional games is not a lack of supply, but a lack of good games. Steam already has tens of thousands of new games every year, yet very few break out. Automatically generated content remains simple and formulaic, and is “still one final push short” in polish and experience.

• World models are the furthest away because the experience, cost structure, and interaction layer are all too early. Text simulators—especially text-based ones—are the closest of the three categories to consumer readiness.

• Roi’s market evidence comes from Xiaohongshu: creators have sold “small phone” and similar simulators for RMB30, reaching several thousand to more than 10,000 orders. A shell she built herself in 3 days through vibe coding also drew unsolicited purchase inquiries, though buyers still had to supply their own API—the exact barrier a platform could remove.

5. An “AI TikTok” Built Around Mini-Games Would Struggle to Escape TikTok’s Reach

• Faced with products that call themselves the “TikTok of the AI era” but look like feeds of mini-games, Roi’s biggest concern is not whether the content feels satisfying, but whether “this experience needs an independent platform.”

• Her reasoning is straightforward: TikTok can already take users from ads into H5 games, and downgrading a more complex H5 game into a lightweight interactive experience would not be difficult. If the platform validated the use case, ByteDance could probably build a similar capability “in about a month.”

• Roi also points to a social variable: roughly 100M DAU use TikTok’s chat function, while mechanisms such as extending the “Little Flame” streak have already created high-frequency relationship chains. An independent AI mini-game platform may struggle to build sociality; TikTok can place chat, video, full-scale games, and lightweight interactions in the same product.

• That leads Roi to the fundamental question for this type of startup: “Why must users play in your place?” If the answer is merely a new content format rather than a new interaction model, the product may still be a natural extension of a large platform.

6. Unlimited Freedom Expands Expressive Space while Stripping Away Playability

• Dropping a premise into a frontier model and proceeding one exchange at a time is closer to an unbounded interactive novel. The user must keep supplying input and actively co-create the plot, while the rules and outcomes remain vague: “You can do whatever you want.”

• Roi acknowledges that power users may use products this way, but ordinary users struggle to “interact casually and still have it produce fun.” A single prompt cannot reliably handle control, rules, feedback, and pacing, leaving the experience to a small group of highly invested players.

• She states the principle in absolute terms: “If something is completely free, its experience is guaranteed to be bad… Complete freedom is guaranteed to be unfun.” Real life is, in a sense, a completely free system, but it often produces too many choices, uncertain outcomes, and “the vertigo of freedom.”

• 幕间 therefore uses systems and rules to constrain the user’s boundaries, increasing the probability of interesting outcomes and reducing the odds of boring ones. The goal is not to eliminate freedom, but to create a sense of safety and continuous feedback between freedom and control.

7. Good Interaction Must Reward This Step and Induce the Next

• Roi says interactive fiction typically gets stuck at 2 hurdles: the model’s response may not be interesting in itself, and even if it is, the user may not know what to input next. Without either element, the result is merely “a piece of text,” not yet an experience.

• Creator tools can provide more control over prompts and interaction, but tools cannot replace methodology. 幕间 will teach creators to offer bounded choices, including AVG-style options beyond text, then use coding to reduce operational friction and make the interaction logic clearer to players.

• A stock-trading simulator demonstrates 2 forms of positive feedback within the same system. One controls the market so the user keeps making money, rising from nothing to become a market wizard. The other ensures the user “loses no matter what,” turning a competition to see who loses more into self-mocking fun among bagholders. Because there is no real loss, failure itself can be a reward.

• Roi later condensed the approach into: “The system should be convergent, but the experience should be open.” Designers do not need to guarantee a fixed happy ending; they only need to keep supplying surprising next steps and “aha moments.” When the experience loses momentum, the player can switch directly to another world.

8. Historical Simulation Shows the Imagination—and Exposes System Complexity and Token Costs

• Koji cites Parks Historia, where players can alter historical conditions—prevent the Soviet Union from dissolving, make Singapore part of Malaysia, create a Soviet-German alliance, or have Einstein invent the atomic bomb—and then observe how the counterfactual world evolves. As he recounts it, Garry Tan called it one of the highest-token-consumption and highest-retention games Y Combinator had seen.

• 幕间 already has a work set in the Song dynasty. Users are born randomly, enter the world as time travelers who know subsequent history, meet figures such as Su Shi and Li Qingzhao, and alter events. Creators currently tend to adopt the perspective of “being one member of the world,” because a complete historical-simulation architecture from a god’s-eye view is more complex and has not yet been fully built.

• Koji also mentions 崇祯, a simulation made by 智谱 with GLM-5. The company says one interaction consumes 30M tokens. Roi considers that amount manageable, noting that “one interaction” could mean a single turn or the tokens consumed to complete an entire first playthrough. One user on 幕间 consumed 100M tokens over several months, although it is unclear whether that came from one work or multiple works.

• She rejects token consumption as a standalone business metric. High usage may signal immersion, or simply inefficient prompt engineering. “Tokens represent cost.” The platform should deliver a better experience at lower per-session consumption, so for now it focuses more on the number of dialogue rounds.

9. Romance Is the Cold-Start Entry Point, Not 幕间’s Preset Destination

• Most early power users entered through Character.AI-style character interaction, and the clearest use case for enjoying a relationship with a single character is romance. Female users in particular care about tension, mechanics, and flirtation—not simply chatting or receiving direct validation, but “actually being in this relationship.”

• Roi makes a careful distinction: men also invest deeply, but may lean more toward direct feedback. Many young women want an intellectual romance with a “high-IQ, high-EQ virtual man,” and are therefore willing to study models, prompts, and interaction design.

• After playing romance for long enough, users naturally expand the system. They first date the guy who runs a café, then ask to run it together, and finally realize that “running the café itself seems more fun than dating this guy.” The relationship mechanic then grows into a standalone management simulator.

• Her answer to the ceiling question is not that romance is necessarily large enough, but that users will not play just one category. As long as non-romance systems provide stronger feedback, supply will evolve from romance into management, history, stock trading, and more forms of life simulation.

10. UGC’s Core Bet Is Turning One Person into an Army

• Koji asks whether ordinary creators can produce work with the polish of Dispatch. Roi remains optimistic because the platform already has highly multimedia works incorporating AI-generated music videos, images, and interactive video; some early versions are “genuinely stunning.”

• She believes coding is the hardest barrier for ordinary creators to cross. Once they learn a language and coding for the sake of a project, generating images, video, and other assets becomes comparatively easier. Richer works are not a distant possibility; the evolution is already happening.

• Although Roi comes from PGC and game production, she sees choosing UGC as the most important current assumption: AI will not make everyone perfectly equal, but will let people with ideas and taste “turn one person into an army.” A sufficient number of super-creators could produce a scale of creativity larger than a handful of organized teams.

• Koji says he had assumed the more fundamental question would be whether simulators could work. Roi’s distinction is that AI will inevitably create new interaction experiences—that is a belief, not an optional assumption. What actually needs to be validated is whether UGC can turn that capability into scale.

11. 幕间 First Found the People Most Eager to Play with AI, Then Built a Platform Out of Teaching

• After leaving her previous company, Roi took time away from work to play with different AI products. That led her into open-source communities, where she met large numbers of hardcore creators and made small works herself. She was “converted” both as a player and, as a content professional, by the long-term potential.

• The early community was not transactional. People shared tools, works, and methods. Because Roi learned quickly and had semi-professional production experience, members began asking her to teach classes. She later taught on Xiaohongshu and for a period became a “mysterious KOL known in the circle for teaching.”

• A class on building a small-phone simulator drew 90 people at 9 p.m. on a weekday, before 幕间 had even been founded. The willingness to spend time learning a complex tool became important evidence for her belief that a UGC product could cold-start.

• The core user profile is mainly post-2000 university students and early-career white-collar workers in first- and second-tier cities, but not exclusively anime fans or traditional mobile-game players. Any novel, TV series, or anime with fantasy and immersion can be an entry point. Kindergarten teachers and stay-at-home mothers also form a highly diverse part of the user base.

12. The Next Generation of Super-Creators May Be Heavy Fantasy Consumers, Not Geeks

• Roi observes that many geeks initially focus on turning AI into a productivity tool. Another group of young women starts from relationship and fantasy needs, invests heavily in model personalities, narrative, and interaction feedback, and gradually learns complex technology as a result.

• These users know model versions “like the back of their hand.” They constantly compare Gemini 3.0 with 2.5: which has the more distinctive visual style, which is more oily or more stereotypically straight-male, and which version fits a particular relationship experience. Model evaluation here is not benchmarking; it is character-consumption expertise.

• Creation usually starts as “doing it for love”: making a character, playing god, or building an ideal world. Once a work wins recognition, creators quickly ask whether “this thing can make money,” and may even push the platform to open up monetization.

• When Koji asks what she would do with $3M to invest in 3 people, Roi would rather split it into 300 investments of $10,000 each, backing super-creators and future consumer-facing AI KOLs. She believes individual studios already have productization, social-media operations, and monetization capabilities—and that “it’s especially cheap to invest now.”

13. The Moat in Creator Tools Comes from Solving Specific Pain Points over Time

• 幕间’s tools grew out of the problems Roi encountered using open-source solutions, building small-phone simulators, and debugging issues for students. Some problems came from users not knowing how to do something; others existed because the original tools simply could not do it. The team has addressed control, interaction, and creation capabilities one by one.

• She is even willing to add features whose long-term value may be limited but that core users urgently need today. Solving these small, concrete problems continuously is meant to create the perception that “this is the place that understands me best and can satisfy my needs best.”

• Creators can of course sell shells themselves, build standalone apps, or distribute on Steam or Roblox. But vibe coding is still constrained by quality and distribution, and Steam is not the primary traffic venue for AI gameplay; before last year, it even restricted listings that used AI-generated images.

• Koji describes these people as outsiders on legacy platforms, rejected and exiled, “slowly lighting their own campfire in a place of exile.” Roi extends the analogy: after exile, they build a new country. 幕间 wants to bring together this population of AI-model consumers and the audiences around them.

14. An Underground Creator Economy Is Taking Shape, but Settlement Still Happens Across External Platforms

• AI is turning individuals into near-generalists, but it has not eliminated specialization. Some people are good at prompts, others at visuals, code, or music, and taste varies as well. As a result, the community has long traded interfaces, assets, music, and technical services.

• Complete works can be packaged as software and sold like single-player games. Roi cites a range of several thousand to 10,000 transactions for a single work. She does not consider that small: many indie games without publishing resources fail to reach that scale on Steam.

• The common model today separates the work from token usage. The creator collects a one-time software fee, while the user pays the API cost. If 幕间 monetizes, both pools of value could stay on the platform, though they may not be charged separately and could instead be bundled into a subscription.

• Roi is explicit that the business model remains unresolved: “It is definitely still in an exploratory stage.” The goal is not to copy a buyout model or traditional in-game purchases, but to find a structure that covers ongoing token costs, creator income, and user comfort at the same time.

15. 幕间 Wants to Be Xiaohongshu for Virtual Life, Not a Mini-Game Shelf

• Roi’s near-term vision is concrete: creators should stop selling simulators on Xiaohongshu and instead create, showcase, trade, and consume them on 幕间. Xiaohongshu shares beautiful real lives; 幕间 shares beautiful lives created by AI that cannot be experienced in reality.

• The platform’s foundational design is identity portability across worlds. The user remains the same “me” and can enter different works using either a real or virtual identity. That continuity means the works are not merely unrelated mini-games, and it is one reason she does not want to split the product entirely into standalone apps.

• Distribution can remain flexible. Roi is comfortable turning popular works into standalone links, mini-programs, or some kind of app shell. The non-negotiable is that the account system remains connected to the platform; user assets and spending relationships cannot disappear with the work.

• She believes the real breakout will not be a single simulator, but a group of people from different backgrounds turning the interesting parts of their lives into simulations, so every user can find an “imagined beautiful life and world.”

16. The Question Is Not Whether People Will Get Addicted, but What Kind of Play the Platform Offers

• Responding to concerns that AI interaction may be more addictive than traditional games, Roi first rejects singling out games for stigma. Novels, mahjong, and other “unproductive” activities can also be addictive; games are simply the easiest target.

• She believes East Asian culture often attaches a negative label to “play,” even though people need to relax and cannot remain tense forever. If AI raises productivity and frees up more leisure time, “how to play better” will become a larger social question.

• The responsibility of an entertainment product is therefore not to eliminate play, but to invent forms of play with “more value and more meaning,” so that positive experiences outweigh negative effects. Interaction, creation, and actively changing a world may provide stronger psychological rewards than purely passive stimulation.

• Roi also rejects unlimited stimulation. Fragmented fun can be acceptable, but continuously keeping the brain in a high-stimulation state and letting time disappear requires restraint. She prefers the room for personal choice on Xiaohongshu to becoming “purely an NPC” controlled by the system.

17. Liblib’s Darkest Period Forced Her to Recalculate AI’s Timeline

• Roi entered Peking University because Hetalia made international relations seem “fun,” then studied art at NYU before dropping out because of the pandemic and visa issues. Professionally, she first worked as a producer of women-oriented games, then built social and education products at ByteDance. Her interests consistently centered on interaction and creation.

• At Liblib, she worked on community cold start, the creator ecosystem, and a viral AI-art campaign before moving into creator tools. Many early AI-image users were anime creators making fan works, an experience that closely resembles 幕间’s community path.

• The company once went 6 months without paying salaries. Team members with high-paying big-tech backgrounds inevitably felt shocked, worried, and doubtful. Roi had watched Silicon Valley repeatedly and already understood that entrepreneurship would not be smooth, so she saw the setback as part of the expected path—“it just came a little early.”

• Around October 2023, she was still immersed in designing a freeform canvas tool and believed a creative workstation should not be just a chat form. She was also thinking about Agents, but the capabilities were too early and many approaches failed. Asked about the relationship with Lovart, she said the interaction layer drew substantially from it and represented a form of inheritance.

18. Moving from Number Two to Number One Meant Repricing the Commitment from 5-10 Years

• The trough forced Roi to revise her extreme optimism about AI in 2023. She had initially expected an answer in 2-3 years, then concluded that the real payback period might be “5 or even 10 years.” Once the horizon lengthened, the choice became a “once-in-a-lifetime” commitment.

• She began asking whether she wanted to spend that period as the number two or product lead at a company, or take the wheel herself and commit to the AI entertainment and interaction direction that felt most mission-driven. She ultimately chose the latter.

• She did not leave by retreating at the company’s most difficult moment. Roi believes “you absolutely cannot leave during the company’s darkest hour,” so she stayed until the business improved and someone could take over. That allowed her to honor her obligations to colleagues without giving up the long-term direction.

• Her need to be number one also comes from her personality. She has extremely strong initiative and agency, “really dislikes being passive,” and tends to take the initiative in relationships as well. Entrepreneurship lets her decide the team, direction, and product vision in the way she believes is right.

19. The Courage to Start Comes from Wanting to Win, Not Calculating How to Lose Less

• A female investor once advised Roi that women are better suited to being the number two than the number one. Her first reaction was not anger but curiosity about the reasoning. She eventually understood that the investor may have experienced gender discrimination or a career ceiling, and was therefore focused on maximizing money, status, and the probability of success after 5 years of commitment.

• Roi does not deny that the world is unfair, but refuses to optimize around it: “Just be yourself.” Gender is only one of the many difficulties in entrepreneurship. As long as she retains an “independent spirit and free thought,” she has the confidence to challenge unreasonable rules.

• Koji brings the question back to his 2 daughters and asks how to raise them to take on greater responsibility. Roi believes family education should encourage independent thinking, self-motivation, and genuinely loved goals, rather than merely calculating wins and losses.

• She summarizes the psychological difference this way: “A lot of people in this world are just afraid of losing; they don’t really want to win.” Fear of losing leads people to seek a break-even solution. Strong enough passion keeps a person pursuing the goal: “I’m not in a state of being afraid to lose. I’m in a state of wanting to win.”

20. Accepting an Ordinary Floor Makes It Possible to Pursue an Unordinary Ceiling

• NYU’s art school exposed Roi to many lifestyles outside the technology career track and loosened the East Asian tendency to bind money and status together. She realized that as long as she avoided things she should not touch, the worst case was simply having little materially without starving. “I can completely accept that floor,” which gave her the freedom to pursue the ceiling.

• She does not believe failure locks anyone in place. Every situation offers a unique experience, and routes can always change, so people should choose the path with “the best scenery and the highest probability of experiencing something wonderful.” An ordinary life is not frightening, and “no one’s life is ordinary.”

• Roi pays almost no attention to sunk costs: “What happened in the past is past. Move on. Next.” The past cannot be changed, and continuing to invest energy in it is wasteful. Her version of return maximization is to put energy into the next worthwhile target and try several more times.

• 《Dance Dance Dance》 taught her to keep moving and stay connected with people. 《The Hours》 reminded her to “love specific people and specific lives” and look for happiness within finite hours. 《Kaili Blues》 suggested that the unknown and frightening can still be worth exploring. Together, they point toward active engagement and turning beautiful moments into works to share.

21. Creator Tools Should Pursue Power First, Then Unfold Complexity Layer by Layer

• Roi rejects making “extreme ease of entry” the first priority for a new tool: “The first priority of a tool should be power, not a low barrier to entry.” Ease should come from removing meaningless steps and providing progressive guidance, not from sacrificing the capability ceiling.

• She rejects the analogy that because CapCut is simpler than Premiere and Figma simpler than Photoshop, new AI tools should also be shallower. Mature paradigms can optimize for usability, but AI is still defining its paradigm. If capabilities are cut first, creators end up with low ceilings and little differentiation.

• The first generation of creative products that generated content from “one prompt” has already exposed the problem. If everyone writes one sentence, works quickly become homogeneous. “Make Snake in one sentence” proves generative capability, not the ability to create a consumable work.

• Koji uses the 0.1-second and 0.2-second delays and subtle feel of 贪吃蛇大作战 to illustrate the value of design. Roi adds that the game also took off because most opponents were bots, giving every player strong positive feedback inside the virtual system. The key is not one-line generation, but the creator’s ability to identify and control these details.

22. An AI-Native Hit Must Clear Three Gates: Model, Gameplay, and User

• Roi does not believe interactive content will linearly replace novels, films, and short video. Passive consumption is naturally easier, and TikTok may still have the larger audience. But some people are not satisfied with being spectators and want to enter a world and actively change things. Even if only roughly 50% of people enjoy interaction, that is already a huge market.

• Roi does not expect all of TikTok to become real-time generated for every user. Real-life records, pre-produced content, and real-time generation will coexist. Real content serves social connection; generative content mainly serves fictional needs. Koji illustrates the distinction by comparing a group discussion of what Faye Wong sang at the 2026 Spring Festival Gala for the Year of the Horse with each person generating a private song: the consumption value is different.

• World models are currently slow, expensive, and visually unstable. But Roi compares them with Midjourney’s evolution from failing to render 5 fingers correctly to becoming product-ready, and with video models’ progress from garbling somersaults to their current level. She believes they could reach application-grade consumer use within a year. The prerequisites for a mass-market hit are system stability, low latency, sufficiently impressive real-time multimedia, and interaction simple enough for anyone to use.

• She breaks AI-native capability into 3 layers of accumulation: engineering infrastructure for stable, fast model calls; a methodology for understanding how free input and personalized generation become “fun”; and a data and insight flywheel formed by feedback from the earliest users. She still worries about missing the timing, the romance entry point being too small, or the need to shift—but her entrepreneurial posture is: “I’ve never thought about the possibility of failure… Mission accomplished.”