GENiEX makes a fictional world the reason to return

GENiEX turns an AI conversation into a return visit to SEIVOID, METAPEX's original science-fiction world, through a physical device whose characters, unlocks, and shared encounters extend beyond a single prompt. Its Hangzhou-based creator built GENiEX as the first physical product in that IP after Zhou Baixiang spent nearly 10 years developing three core novels and more than 4 million Chinese characters of setting material.

The return loop has stakes inside the fiction. In single-player mode, conversation builds affinity with characters and opens upgrades, skins, and gacha draws; when two devices meet face to face, each user can access content unavailable alone. Three or more devices can also enter an offline murder-mystery or TRPG-style adventure.

That design shifts the question from how long someone chats with AI to what gives them a reason to come back. GENiEX offers one answer: progress in a world, with other people able to change what the experience can reveal.

Return to a world, not a chat

GENiEX is built as an entry point into SEIVOID, an original science-fiction IP, rather than as a standalone chat device. According to reporting, Hangzhou-based METAPEX created the product after Zhou Baixiang spent nearly 10 years developing the universe: three core novels and more than 4 million Chinese characters of setting material preceded the hardware.

That backlog gives a return visit somewhere to go.

In single-player mode, conversation can build affinity with characters and unlock upgrades, skins, and gacha draws. The loop also moves beyond a private session. When two devices touch face to face, they reveal exclusive content unavailable to either person alone. Three or more devices can enter an offline murder-mystery or TRPG-style adventure. The relevant retention question is therefore not simply how long someone speaks to an AI, but whether SEIVOID has another encounter worth returning for.

Companion-product metrics show the limitation of treating session time as the answer. In an August 2025 WIRED interview cited by woshipm, Character.AI reported about 20 million monthly active users and about 75 minutes of average daily use. Yet its internal self-reported data put companionship at less than 20% of use cases, while its CEO characterized the service as role-playing and entertainment rather than solely companionship.

Time can indicate attention. It does not identify the reason for return.

The 2025 Journal of Consumer Research study cited by woshipm found that AI companions could offer immediate relief from loneliness, with feeling heard an important mechanism. That is meaningful within a session. GENiEX instead makes continued access to an authored world, and encounters with other owners, the product's distinct promise.

A public channel needs rules around the prompt box

Infinite Slop turns a chat prompt into a public programming queue. Viewers request scenes during a continuous livestream, but requests are reviewed before AI generates the next segment, according to woshipm. The page also rejects duplicates already waiting in the queue. That rule prevents the channel from treating every message as an automatic production command.

The control point matters because the output is public and monetized. Infinite Slop offers ready-made categories including banana, cat, space, news, and game show. As of September 1, its frontend showed branded pre-rolls, sponsored segments, and Chatbase among its sponsors, woshipm reports. Review can therefore protect a recognizable channel format while giving operators a gate between audience improvisation and material that appears beside advertisers. The creator has also resisted paid bidding for the next video, arguing that it would make the communal chaos less fun.

Bilibili's AI Creation Open Competition shows the opposite boundary: contribution can be much freer when people are making separate works rather than programming one shared broadcast. As of August 20, it had attracted 13,400 submissions, according to ifanr. About two-thirds of participants had no professional development background. More than 80% were one-person teams, and nearly 90% were ordinary Bilibili users with fewer than 10,000 followers.

Low-cost participation widened the funnel. More than half spent 30 hours or less and kept costs below 500 yuan.

The design choice is not open versus closed in the abstract. A shared public channel needs editorial control at the prompt box; a creator competition can let many people build, then let the audience encounter the resulting work. Teams should decide which contribution enters a common feed and which can remain an independent experiment.

Three AI product loops: world progression, co-presence and collective programming

Measure agency and progress before message volume

Message volume can describe habit, but it cannot explain what the habit delivers. In GENiEX, conversation builds affinity with characters and unlocks upgrades, skins, and gacha draws, according to a report. Those are observable markers of voluntary play: did a user advance a character relationship, pursue a chosen unlock, or return to an unfinished fictional thread? The product can measure those actions separately from raw turns in chat.

Shared progress deserves its own measure. A report states that face-to-face contact between two GENiEX devices unlocks content unavailable to either person alone, while three or more devices can join an offline murder-mystery or TRPG-style adventure. Count completed shared sessions and voluntary re-entry into the same world, rather than treating every repeated interaction as equivalent.

The ethical contrast is clearest with companions. According to woshipm, Common Sense Media surveyed 1,060 U.S. teenagers aged 13 to 17 in 2025: 72% had tried an AI companion, and 52% used one at least several times per month. Entertainment and curiosity were the leading motivations. Yet average daily time can mislead; woshipm reports that Character.AI described average use of about 75 minutes per day in an August 2025 interview.

A stronger companion score asks what changed away from the product. A 2025 Journal of Consumer Research study found immediate loneliness relief after use, with feeling heard as a key mechanism, woshipm reports. Track a user-reported resolved need, such as feeling calmer or taking a step toward a real-world relationship, alongside return use. Treat self-reported claims cautiously: Tolan's survey found 85.6% believed it helped through difficult experiences and 72.5% believed it improved real-world relationships. Headspace's Ebb states that it cannot replace therapy or a mental-health coach. High frequency without progress should trigger review, not celebration.

Persistent experiences carry an operating bill

A persistent AI world has recurring costs even when the player sees an open-ended interface. According to ifanr, the main development cost for World Gate is tokens, at about 1,500 yuan per month. Its players can choose locations, create entrances and scene maps, and meet randomly appearing NPCs through AI-defined interactions. Each of those encounters makes inference part of the product's operating bill.

Generated output is not free because it feels immediate.

The world also needs material that gives returns a direction. Before GENiEX, Zhou Baixiang spent nearly 10 years building an original science-fiction universe with three core novels and more than 4 million Chinese characters of setting material, according to a report. SEIVOID's animated series has a completed first-season script; the season contains 11 episodes, each 23 minutes long, while animation production continues. That is an authored base for a continuing experience, not an expense eliminated by generation.

Video makes the variable-cost problem especially visible. woshipm reports fal's published standard price for H3 Max at 768p as about $0.08 per generated second, or about $4.8 for one minute. A channel playing entirely new material for 24 hours would therefore carry a theoretical generation cost of $6,912 per day and more than $200,000 per month. Speed does not erase that bill: fal's comparison chart lists H3 Max average latency at 3.49 seconds. Nor can moderation be treated as an afterthought; Infinite Slop requires prompts to be reviewed before they enter its queue. A return loop needs room in its plan for inference, review, authored continuity, and the work of keeping participation usable.

Design for the return loop, not AI time spent

• You are building an AI game or fictional experience and need players to return after the first novelty interaction. Give the world persistent progression and new reasons to explore. In "World Gate," players can create entrances and scene maps at any location, meet randomly appearing NPCs, and use AI-defined interactions; GENiEX pairs character conversations with affinity, upgrades, skins, and gacha draws. Do not assume an open-ended chat alone will create a durable loop.
• Your product becomes better when participants meet or contribute together rather than consume alone. Make collective participation unlock something unavailable to a solo user. GENiEX unlocks exclusive content when two devices touch face to face, while three or more devices can join offline murder-mystery or TRPG-style adventures. For a broadcast format, Infinite Slop turns chat requests into a reviewed queue for future segments.
• You want to turn audience feedback into the next piece of content. Build a visible request-and-review mechanism rather than promising instant generation. Infinite Slop confirms that prompts require review before entering the queue and rejects duplicate requests already in it. This can make participation legible, but it also creates an expectation-management problem: one observer saw about 400 users online without seeing a submitted request enter generation or air.
• You are deciding whether a rapid, low-cost prototype is enough to validate a product. Use a small prototype to test whether users want to create, share, or extend the experience-but do not confuse initial attention with retention. More than half of participants in the build in bilibili - AI Creation Open Competition spent 30 hours or less and kept costs below 500 yuan, and "Wan Geng Nie" lets users share creations. Yet Infinite Slop did not publicly disclose next-day retention, average viewing duration, or the number of users returning after one month.
• You are considering an always-on AI video channel. Treat generation economics and moderation as core product constraints. fal's published standard price for H3 Max at 768p is about $0.08 for each second of generated video; a channel playing entirely new content for 24 hours would have a theoretical generation cost of $6,912 per day and more than $200,000 per month. Infinite Slop therefore reviews requests before queueing them and, as of September 1, used branded pre-rolls and sponsored segments.

Turn community response into the next build

A viral response is a starting signal, not proof of a return loop. GENiEX had raised $134,500 from 429 backers against a $12,700 Kickstarter target, exceeding 1,000% of its goal, according to a report. That establishes attention and willingness to support a premise. It does not show which backers will return once the initial promise becomes a product.

Small teams should make continuity visible in the next build.

Show what changed because people participated: a new area, an altered character, or a tool shaped by a submitted creation. Then capture feedback close to the moment of use and connect it to a named next-release decision. The useful test is simple: can a participant identify a contribution that changed the experience, and can the team explain what will not change?

Ownership needs the same precision. Define who can reuse a user contribution and who can delete it. Separately, state what the product remembers between visits, how long it remains available, and whether the contributor can revise that memory. These boundaries prevent community material from becoming an unexamined input pile.

ifanr offers a warning against treating view counts as durable demand. Dian Kao Huang Daiyu's 2 videos, totaling 5 minutes and 10 seconds, received 4.96 million views, while the independently developed World Gate remains in small-scale testing. The creator expected hundreds of thousands of views, yet the result exceeded that expectation by a factor of 10. A spike can reveal curiosity before it reveals a repeatable reason to come back.

Track the people who return after their feedback appears in a later build. Compare that group with the initial audience, and ask what they do next. If contributions keep producing visible consequences, community response becomes product direction rather than applause.

Set the return contract before selecting the revenue surface. Nomi sells unlimited messages and voice calls, yet those are access benefits, not proof a user's need is being met. For emotionally sensitive uses, add a visible boundary: Headspace's Ebb says it cannot replace human therapy or a mental-health coach. Measure whether users resolve an offline need, rather than message volume.

Separate commercial signals from user value. Character.AI's annualized revenue run rate exceeded $30 million and it introduced ads and rewarded ads, while its paying-user count was not disclosed, according to woshipm.

Treat Tolan's self-published figures-85.6% on emotional difficulty and 72.5% on real-world relationships-as hypotheses to test with your own outcome measures.

For readers outside China

• Availability: Availability outside China is not disclosed in sources. "World Gate" is in small-scale testing and has not been opened to all players. GENiEX was crowdfunded on Kickstarter, but the source material does not say where it will ship or whether its services will operate outside China.
• Pricing: "World Gate" is reported to cost about 1,500 yuan per month in tokens to develop; no player price is disclosed. GENiEX's lowest Kickstarter early-bird price was 1,471 Hong Kong dollars, approximately $187; its Founder's Set was priced at 5,191 Hong Kong dollars, approximately $661. Character.AI offers basic chatting for free, while c.ai+ adds better memory, no ads, updated models, removal of slow mode, unlimited voice calls, and more customization features; its subscription price is not disclosed in sources. Nomi offers monthly, quarterly, and annual subscription plans, but prices are not disclosed in sources.
• Closest Western equivalents: Character.AI is the clearest reference point for the conversational role-playing layer: it describes itself as role-playing and entertainment rather than solely an AI companion.; Nomi is a reference point for a subscription companion product, with paid features including unlimited messages, voice calls, multiple Nomis, group chats, longer backstories, and image and video capabilities.; Tolan is a reference point for a companion app positioned around emotionally difficult experiences and real-world relationships.; Infinite Slop is a reference point for the participatory livestream model, where viewer prompts are reviewed, queued, and turned into future AI-generated video segments.
• Data residency: The source material does not cover where user conversations, prompts, images, device data, or generated content are stored or processed. Data-residency, cross-border transfer, retention, and deletion policies are not disclosed in sources.

Sources

• woshipm 爆火之后，Infinite Slop 靠什么活下去？ https://woshipm.com/ai/6458384.html
• 36kr 杭州小团队做了台「AI对讲机」，3天众筹了1000% https://36kr.com/p/3967445813468681
• woshipm 当"让用户离不开"不再是好指标：AI陪伴产品如何区分留存与依赖？ https://woshipm.com/operate/6453080.html
• ifanr AI 降低了创造的门槛，B 站则想放大创造的回声 https://ifanr.com/1678567