Summary

筱宁, a former UGC game producer at a leading game company and the host of 405游局, maps AI interactive entertainment across 4 layers: AI as a productivity tool, a creative gateway, an interactive counterpart, and, at the deepest layer, a force reshaping entertainment and social relationships. He draws the boundary with just 2 words: “interactive” and “fun,” because “a game’s real competitor isn’t another game—it’s TikTok”—users do not distinguish games from anything else when spending their free time.
The timing call is the episode’s anchor: we are roughly 1 to 1.5 years before Angry Birds—some model infrastructure and distribution are in place, but the product that makes the entire industry agree the direction deserves more resources has yet to appear. The test is simple: without AI, it would not work. And it is more likely to emerge outside the major platforms: “If I were at Tencent, I might not even be able to get it approved or shipped.”
The 3 major misconceptions double as a diligence checklist: one-line generation is overestimated, natural language is over-weighted in interaction, and infinite choice is mistaken for fun. “Generating a prototype takes 1 sentence; tuning it into something I can actually play might take 100 sentences and 100 rounds,” with context and instruction-following already breaking down by the 5th or 6th round. Battle royale had existed for years; PUBG became a hit by using the shrinking zone to impose structure. “Does anyone really like open worlds? Don’t we like open worlds that have been designed?”
For the “TikTok of the AI era”—Loopit, AIPPY, and Refold—筱宁 leaves 2 questions unresolved: how creator value becomes visible, and who the consumer actually is. You can watch 10,000 people perform the same dance on TikTok because “it’s you” performing it; an AI-generated Snake game lacks that identity value, and supply could dry up. For an ad business to work, DAU needs to be at least 30M-40M, by his back-of-the-envelope estimate—a DAU logic, not an MAU logic. “The next TikTok definitely won’t be TikTok.”
At the major platforms, Tencent and ByteDance have model-training capabilities, Tencent is more conservative while ByteDance moves faster, NetEase has changed the most, and miHoYo is pushing down into the model layer. miHoYo’s suspected 《星布谷地》 agent NPC 娜洛 may be in a league of its own among AI NPCs, evolving from an answer engine into “social infrastructure” that rescues dead air and manages the room. Basic-role headcount is also genuinely falling.
World models are “too early”: they could make explorable, influenceable worlds like Westworld and the Oasis possible, but remain far from usable, while the biggest opportunity may be games as AI training environments rather than games themselves. The only application-side consensus is companionship.
One underappreciated effect is the speed of prototyping. AI short dramas may be able to “make an infinite number of first episodes,” while the corresponding gaming bottleneck is 3D. Once AI’s 3D pipeline—topology and rigging—works, small teams can iterate without limit. “Even if AI can only make prototypes and demos, the impact could exceed the demos themselves.” And when 405游局 guests are asked what they have been playing lately, more and more answer Cloud Code: creation becomes consumption, and the point where businesses charge users moves with it.

Deep dive

1. The boundary is just 2 words: interactive and fun—the competitor is TikTok

筱宁 was previously a producer on a well-known UGC game at a leading game company. After leaving last year, he launched the podcast 405游局, which releases weekly and has produced nearly 50 episodes in a year. He broadened the scope from “AI games” to “AI interactive entertainment” because the boundary of the category comes down to 2 things: interaction, and whether something is “fun to play”—people exchange time, attention, and money for a positive emotional experience.
His core judgment: “A game’s real competitor isn’t another game—it’s TikTok.” Users do not separate games from other forms of entertainment when spending their free time; they only care whether something delivers satisfaction in the moment. Plot interaction depth on a single axis and the threshold “maybe is a TikTok filter,” with games at the deepest end.

2. The 4-layer map: from productivity tool to reshaping entertainment relationships

The first layer is AI as a tool that improves efficiency in the production workflow; AI coding is closer to infrastructure. The second is AI as a creative gateway: “The coding itself is hidden; I care more about the result”—generated video, images, and games. The third is AI as an interactive counterpart, starting with character AI and extending to companionship products, the recently launched Eve, and in-game AI NPCs.
The fourth layer is the least explored: AI reshaping entertainment and social relationships. Who creates and who consumes are both changing. AI is no longer the object of direct interaction; it facilitates social interaction and manages the atmosphere within a social setting.
One example is the suspected miHoYo title 《星布谷地》. Its agent NPC 娜洛 is a café manager with both 1-on-1 private chat rooms and a public room where “me and 3 other players” are present. In a multiplayer setting that combines people and an agent, the agent’s role changes.

3. Suspected 《星布谷地》: AI NPCs become social infrastructure

娜洛 does 3 things: shows newcomers how to chat in the café, steps in when conversation stalls, and manages the atmosphere. “It’s no longer really an NPC role; it’s a social infrastructure role.” The point of the conversation is to facilitate social interaction between players, not to conduct a human-machine exchange.
That reverses the design problem. With character AI, a good experience might be “I say 1 sentence and it says 10.” In a social setting, “the question may instead be when the AI should shut up.”
筱宁 believes the suspected 《星布谷地》 is in a league of its own among AI NPCs. Most games simply attach a language model to an NPC, give it roughly a prompt, and stop there—a plus-up to the experience. miHoYo appears to be taking it seriously; multiplayer settings are a different level of difficulty, and it is “treating this somewhat like an IP.”

4. Fragments are everywhere; complete products are scarce

There are plenty of flashes of promise: an interactive version of Along the River During the Qingming Festival that you can walk into, the Chongzhen historical simulator, a game that builds sentences by eating words, and a Snake game in which AI decides whether the resulting string counts as a sentence. “That was the hardest Snake I’ve ever played.” But products with a complete experience capable of bringing people back repeatedly and keeping them playing longer remain rare.
Koji pointed to Nintendo’s Tomodachi Life: create yourself and your friends, put up to 8 people in one house, set up drama—“all 8 people fall in love with the same older woman”—and watch the characters perform the melodrama themselves before it is repackaged into short video for secondary distribution. “You are the creator, the director, and the audience at the same time.”
筱宁’s comparison is Patrick’s Beside. The direction is similar, but “the first 1 or 2 hours are especially good; the repeat-use loop still isn’t strong enough.” Town-like games from last year, including Civilization-style experiments, also stalled at the same point: they were not fun enough.

5. We are 1 to 1.5 years before Angry Birds

The analogy is about consensus formation. Angry Birds was neither the endpoint of mobile gaming nor its most commercially successful product. Its paying user base appeared small and its heat short-lived, but it made everyone agree that the direction deserved more resources and exploration. It arrived roughly 1.5 to 2 years after the iPhone launch and 1 year after the App Store opened: infrastructure, distribution, and user habits were all in place. The model-side conditions are partly in place today, but the consensus moment has not arrived.
The test for AI’s Angry Birds moment is that the product would not work without AI, or that AI would amplify the experience by several levels. Angry Birds was not fun on a computer with a mouse because “a mouse is good for precise clicking, not dragging.” The breakthrough need not come from a new mechanic or from what we traditionally think of as a game.
The only area with a tangible focus today is AVG-style content, but that is a very narrow field. “Most people don’t play games through natural language. They play through the multimodal stimulation of sights, sounds, and effects, along with click feedback and feel.”

6. World models are the frontier; games may become AI’s training ground

The progress in agents’ long-horizon planning and coding in 2026 has changed how seriously game companies take AI. In July and August last year, people still thought AI was “somehow just a productivity tool.” Now the prevailing view is that AI can do a great deal more: “The AI of last August and the AI of this April aren’t even the same AI.”
World models are “too early.” Their most compelling feature is making explorable, influenceable worlds—Westworld or the Oasis in Ready Player One—possible, but usable products remain a long way off. The biggest opportunity may not be games at all. He mentioned what may have been DeepMind, though he was not certain, using an offline version of EVE Online for AI training. Game worlds could become environments for training robotics and embodied intelligence, as well as sources of synthetic data. “That is a much bigger frontier, but getting from there to games may still take a long time.”
The same applies to real-time multimodality. The vision is attractive: enter a world, have its story unfold in real time, and gain a sense of control. But games fundamentally deliver “a stable, continuously interactive system that has to be designed.” It is “a beautiful future, but not something we can call deployed today.”

7. Big-platform scan: Tencent cautious, ByteDance fast, miHoYo moving into models, NetEase turning the most

Tencent and ByteDance both train models in-house, but “Tencent is still relatively more conservative,” while ByteDance reacts fastest. NetEase has changed the most—from showing little enthusiasm for AI applications to publicly embracing it aggressively this year. miHoYo’s ethos is “tech nerds save the world”: 蔡哥’s experiment actually calls a model, an unusual level of effort for a pure game company, while 大伟哥 said in a recent speech that the company would make major technical investments. Gigabyte gave everyone 1 week off from regular work to enter an internal web-coding competition.
The Angry Birds moment is more likely to grow outside the major platforms. “If I were making a very, very small experience today and I were at Tencent, I might not even be able to get it approved or shipped. Outside, I might barely have started before it showed up on Xiaohongshu.” He is betting on uncontrolled growth from the edges.
At the headcount level, “a lot of basic roles are genuinely disappearing; everyone can feel it quite deeply.” His advice to industry workers is to use AI more: “The more you use it, the less anxious you actually feel.” That is not because the tools are weak, but because you learn what they can and cannot do. “I only see it as a threat when I constantly feel it is competing with me for my job.”

8. A primer for AI builders: games are precision instruments, and generation is only the starting point

Blind spot 1: the word “generation” understates the industrial complexity of games. A generated video or image is already a finished product; a game is a system that must be delivered reliably to players. Debugging, design, feedback, mechanics, and systems architecture do not appear in a single generation step. “A game is a precision instrument.”
Blind spot 2: fun is not designed once. Making a demo that dazzles at first glance and making a product that brings players back repeatedly for a long time are entirely different levels of difficulty. The latter requires the demo-test-feedback loop and sustained iteration; it is not a one-time design exercise, and even hiring someone who has made games does not guarantee they can design it.
Blind spot 3: a platform’s core metric is retention, not first-glance impact. Why do users keep coming back? Why do creators keep creating? How is content distributed? The sharpest line in the episode is: “For many content products, being generatable may be the destination they pursue; for games and interaction, being generatable may be the starting point from which they can be designed.”

9. The 2 unanswered questions for the “AI TikTok”

The excitement around Loopit, AIPPY, and Refold is rational because the creative gateway has changed. Building an interactive experience used to be “the privilege of an internet product manager”; now the barrier has fallen dramatically, and the result really is fun at first glance.
The first unanswered question is how creator value gets pulled into the equation. “If the 2 of us both generate a Snake game, how does anyone see the value of either of us?” On TikTok, 10,000 people can perform the same song and you will watch it 10,000 times—not because the dance is inherently great, but because “it’s you” performing it. If identity value cannot be solved, supply is finite.
The second is who the consumer actually is. Once supply is amplified, “why is this not simply a subset of TikTok?” Since the mobile-internet era, the 2 strongest business models have been games and advertising: games rely on sufficiently immersive, repeat interaction, while advertising requires “DAU north of 30M-40M” by his back-of-the-envelope estimate. It is a DAU logic, not an MAU or total-user logic. “The next TikTok definitely won’t be TikTok”—though founders may say that because capital markets find the idea easier to understand.

10. An infinite number of first episodes: the underappreciated change in production relations

AI short films have already produced a breakout, 《丧尸拾荒者》: a young person from a small town made a 3-minute film in 10 days for RMB3,000, earning strong reviews online in China and abroad. The absence of a breakout in AI interactive entertainment is simply “a matter of timing”; video content is clearly ahead. The most impressive part of AI manhua is not the breakout itself but the speed of improvement: “A month ago the visuals were still a little rough; 2 weeks later it had already updated 4 or 5 seasons, with very good quality.” Seedance, pronounced “C Dance,” also only caught fire around Lunar New Year.
Koji relayed an insight from an AI short-drama founder: shooting the first episode of a live-action short drama consumes more than half the production budget—flying people in, locking schedules, and starting production. AI live-action short dramas have no physical constraints and can “make an infinite number of first episodes.” Only after the first season works do you make the 2nd and 3rd; the old film-and-TV production system could not do this.
The corresponding bottleneck in games is 3D. The ceiling for 2D in China is low, while 3D blockbusters are beyond the capacity of many major and even mid-sized studios. “Today, AI’s 3D pipeline is not solved”—topology and rigging still need breakthroughs. Once 3D is no longer the bottleneck, “we can have an infinite number of first episodes, and maybe people will compete their way into something else.” That leads to another underappreciated point: “We always say AI can only make prototypes and demos. Even if it can only make prototypes, the impact could exceed the demos themselves.”

11. The 3 major misconceptions: one-line generation, natural-language share, and infinite choice

Misconception 1 is one-line generation. “Generating a prototype takes 1 sentence, but turning it into a version I can actually play might take 100 sentences and 100 rounds.” By the 5th or 6th round, context and instruction-following often start to break down. Misconception 2 is the overestimation of natural language’s weight on the interaction side.
Misconception 3 is equating infinite choice with fun. Choice and control are designed into the experience. Battle royale already existed; PUBG became a hit through the shrinking zone. The point was not that players could go anywhere or fight anywhere, but that directional objectives compressed the experience. “Does anyone really like open worlds? Don’t we like open worlds that have been designed?”
Taken to the limit, could every person’s TikTok feed become completely unique, with everything generated specifically for that person? “Possibly. But is that actually fun?” People already worry about information bubbles; “would we really choose to create an information bubble for ourselves?” The strongest stimulus is often a common denominator: “People are still more alike than they are individual.” The popularity of mini-games and tiny TikTok-like products comes down to a predictable dopamine hit, like vocabulary drills. “When life disappoints, do vocabulary drills.” Small, certain gains offer a renewed sense of control over life.

12. The editor is not the deciding factor: Tap Tap Maker’s split verdict

Whether a platform works does not depend on its editor. “An editor is necessary foundation, but not the deciding factor.” The operating questions are who the service is for, why users return, and how content is distributed. “It is hard for me to imagine any company in this era following the same growth path as Roblox—and Roblox itself grew very slowly.”
The most interesting result from Tap Tap Maker was its polarization. 筱宁’s first reaction was “Wow, this is unusable,” and he put it down. He recommended it to friends who had worked in development; several came back saying, “It’s incredibly usable, and the ceiling is high.” Some users recreated Minecraft worlds—work that previously took a team of several people 2 to 3 months, but can now be done by 1 or 2 people in 1 month or even 2 weeks. Others built complete roguelikes. People without a game-design background felt it was “worse than Cloud Code,” while experienced designers discovered after a few rounds of dialogue that “it understands game language extremely well.” His guess is that creator content from the 星火编辑器 era was fed into the AI during training.
A neighboring direction is AI interactive film and games; he previously spoke with Lina Game. Historically, there have been very few interactive films because branching storylines make them even harder to shoot than short dramas. AI generation could cut costs by 100x. There is no small breakout yet, but as video models heat up, interactive video will become an active area of exploration.

13. The only consensus left is companionship; creation is consumption’s deep water

The industry has only 2 genuine areas of consensus. In the long run, AI will change productivity and production relations. On the application side, “the only thing that can really be called consensus is still companionship”—language models make language-based experiences the most intuitive. Nothing else qualifies as consensus.
The split over companion robots is worth noting. Koji relayed that “many people think game companies will build them,” because robotics companies focus on utility while emotional design, IP, and content are game companies’ strengths. 筱宁’s pushback is that the 2 needs may need to be separated: “I can’t really imagine a robot that helps me mop the floor also being the robot I fall in love with.” China already has desktop-scale experiments, including a small desktop device released last year by 恺英, pronounced phonetically.
The closing thread is the most thought-provoking. 405游局 asks guests what game they have been playing lately, and since Lunar New Year, more and more answer Cloud Code. So is AI really just a tool for us? Once creation becomes consumption, “where do you charge the customer?” Creation itself can be fun: selling smartphone-photography courses to mothers is the most lucrative business, while people photograph their food because 3 meals a day are the closest thing most people encounter to a work—or even a piece of art—on an ordinary day.
The methodological conclusion is that when both technology and products are early and non-consensus, “don’t try to analyze everything rationally.” Intuitive signals matter more. When his child first saw life-size Mickey and Minnie at Disneyland, the reaction was “3 seconds of stunned silence, then a trembling scream.” “That kind of more instinctive experience is what people in interactive entertainment are willing to pursue.”

🔬 The Bitter Lesson is Coming for Proteins - Alex Rives, BioHub The $1 Trillion Firm That Refuses The Private Equity Label | a16z