Amid Rapid Technological Change, Seizing the "Constants" | A Conversation with Two AI Entrepreneurs

Introduction

从Next Token到Next State。

Now that generative AI can produce images that are virtually indistinguishable from reality, write fluent scripts, and even generate coherent video clips, the next question arises: Why haven’t we seen an interactive entertainment product that truly belongs to the AI era?

The technical elements seem to be in place, but they are more like parts scattered on the ground, not yet assembled into a machine that can run. The entire industry seems to be in a stage akin to “the blind men and the elephant”—everyone has felt a part of the elephant, but no one has yet truly grasped the complete picture.

The two guests recently invited by the 405 Gaming Bureau happened to be standing on either side of this elephant.

Dai Bo, founder of Feeling AI (Yunwu Intelligence) and assistant professor at the University of Hong Kong.He has been immersed in the field of generative AI since beginning his doctoral studies in 2014. He completed his Ph.D. at the MMLab at the Chinese University of Hong Kong under the guidance of Professor Lin Dahua. His work has been cited over 10,000 times on Google Scholar, and he has been consistently ranked among the top 2% of scientists globally for the past three years.His best-known work is AnimateDiff—an open-source video generation model that took the entire Stable Diffusion community by storm. In 2024, he turned down offers from several major tech companies to found Feeling AI, with the goal of building the infrastructure for world models to give virtual worlds a “sense of lifelike presence.”Within six months of its founding, the company completed two rounds of funding totaling over 100 million yuan.

Hu Xiuhan, founder of NieTa (See Concept). He holds a dual degree in Intelligent Science and Economics from Peking University and a master’s degree from Columbia University. He served as Head of Short-Form Video Technical Architecture at Meta, became the youngest P8 expert in Alibaba’s Data Technology Business Unit, and built a product line from scratch at Tezhan that generated over 100 million in revenue.In late 2022, just four months after the release of Stable Diffusion, he concluded that “the building blocks of content had changed” and immediately left his job to start his own company. NieTa now has over 12 million registered users, with an average daily engagement time exceeding 110 minutes. In March 2026, the company completed a Pre-A+ funding round of over $10 million.

Their connection predates their business partnership. They were old friends from their days competing in computer competitions, and what brought them back together was AnimateDiff—as soon as Dai Bo finished developing the model, he sent it to Xiu Han to see, and Xiu Han’s reaction was, “I’ve always wanted something like this.”Since then, one has delved into the underlying architecture of world models, while the other has built an AI-native creative community at the application layer. Though their paths differ, they both point toward the same ultimate goal.

This conversation revolves around several core questions: As the basic unit of content evolves from pixels to concepts, how will the boundaries between creation and consumption be redrawn?Why must AI-native applications be “active” and “real-time,” rather than pre-generated and then redistributed? How exactly should we understand the world model—Dai Bo breaks it down into a three-layer architecture comprising the idea layer, the dynamic layer, and the static layer, while Xiu Han focuses more on whether it can become a self-sustaining, dynamic system.They also delved into an often-overlooked distinction: “creation” (professional creation) versus “recreation” (mass re-creation/recreational creation). Truly mass-market UGC behavior is, in essence, closer to the latter—it places high demands on the quality of the process but not on the quality of the outcome, deriving enjoyment from the flow of the creation process itself.

The two founders share a common belief: the greatest opportunity lies not in someone creating an “oasis,” but in someone providing the building blocks to create one. When users can build worlds, define rules, and create characters using concepts rather than code, the forms of interactive entertainment will no longer be limited by today’s imagination, but only by when technology catches up.

The following is a conversation between 405 Game Studio and Dai Bo and Hu Xiuhan. Game Teahouse has compiled this transcript with permission; the content has been edited.

01

Content building blocks are being redefined

“In the past, the building blocks of content were pixels; in the future, they will be concepts. When concepts are directly linked to prompts, that represents the most fundamental change at the building-block level.”

Xiaoning: You both have technical backgrounds—why did you both choose to start businesses in the entertainment and content sectors this time?

Xiu Han: I used to compete in computer science competitions with Dai Bo. Since graduating, I’ve been working on product projects. At a company like Facebook, where product culture is a top priority, I’ve conducted numerous design A/B tests and growth experiments, and I’ve gradually grown accustomed to thinking about the impact of technological innovation from an engineering and product development perspective. With every new generation of AI technology, my first instinct is to figure out how to solve problems in concrete, real-world scenarios.

Dai Bo: I’m pretty straightforward.When I was a kid, my mom bought me a Xiaobao educational computer, and ever since then, I’ve been very interested in interactive content. When I was choosing a research direction for my Ph.D., I thought: being able to use the technology I develop to create interactive content would be something that would truly satisfy me. As I watched the technology gradually mature, I figured I might as well try it myself—because only by doing it can I find the answers.

Xiaoning: What signs led you to decide to start a business back then?

Xiu Han: I’ve always been focused on the atomic nature of content—how to deconstruct and reconstruct it—and I’ve also been thinking about how models influence the way creators collaborate. Of course, I’m not approaching this from a technical perspective based on underlying principles, but rather from an engineering perspective—how it affects human collaboration and how it influences creators in their content creation.

When I first encountered Stable Diffusion in July 2022—even before large language models had emerged—I clearly realized that the “Lego blocks” of content had changed. In the past, they might have been a single frame, a layer, or even down to the pixel level; but in the future, these “building blocks” will likely be concepts—some kind of visual concept.This is the core shift at the building-block level. Throughout this process, we’ve conducted extensive exploration and gone through phases of continuous evolution with various building blocks, such as LoRA.

I feel that recent breakthroughs like Nano Banana and Seedance 2.0 are gradually bringing us closer to a state where “a world can be built using just a concept.” This is a gradual process of understanding.

Xiaoning: As we discussed last time, some things haven’t changed at all—for example, your initial definition of this project was “GitHub + IP.”

Xiu Han: When we consider the evolution of content atoms, there have been roughly three to four generations throughout history: The first generation centered on pure pixels; the second generation introduced templates, layers, and 3D models;the third generation was the component-based approach pioneered by Figma; the fourth generation is the current concept-centric atomic structure. Each generation has given rise to great products: the first and second generations of Photoshop, the third generation of Figma, and the fourth generation may very well be unfolding right now.

The deeper you look, the more these building blocks resemble code. And the inherent nature of code means it can enable GitHub-style collaboration. Previously, collaborative creation was limited to text-based collaboration, such as on Wikipedia; when it extends to the visual realm, it opens up opportunities for a shift in core patterns. Among these, the most frequently reused building blocks are the elements of intellectual property.

Dai Bo: I first became interested in generative AI in 2014 when I started my Ph.D. program, and I’ve been following its development ever since. In 2022, after seeing how mature Stable Diffusion had become, I felt the technology had reached a tipping point.

In 2023, we developed a video-generation project called AnimateDiff. As soon as it was finished, I sent it to Xiu Han right away, and he told me straight away, “I’ve always wanted something like this.”After we open-sourced it, the entire Stable Diffusion community was captivated within about two or three months. There’s still a dedicated subreddit on Reddit, and even with Seedance now available, people are still using AnimateDiff to create all kinds of content.

This made me realize that I, too, seem to have mastered this mature technology. At first, we tended to move forward through collaboration—I would handle the technical side, either by open-sourcing the work or partnering with others. But after trying this approach in the second half of 2023, I felt overall that the progress wasn’t fast enough.AnimateDiff was completed in May and sent to Xiuhan in July, but it didn’t really gain traction until October—three months later. If you don’t do it yourself, progress will be slow.

At that juncture, you’ll start asking yourself: Is this luck, a coincidence, or an inevitability? The best approach is to address user needs head-on—that’s the only way to know where to accelerate technological iteration. So we decided to roll up our sleeves and build a solid AI infrastructure ourselves.We believe that this infrastructure will give rise to more native applications, and this requires users to evolve their ideas alongside us throughout the process.

02

Characteristics of AI-Native Applications:

Proactive, Real-Time, Consumed as It Is Generated

“If it were entirely a model of pre-generating content and then distributing it, I wouldn’t have much confidence that it could beat TikTok. That’s why I strongly support ‘consumption during generation.’”

Xiaoning: You all say you’re working on something “new.” To put it more concretely, from an experiential standpoint, what does the future you envision look like?

Dai Bo: I can’t describe its specific form, but I can outline the characteristics it should possess. The most significant characteristic of generative AI is its uncertainty—uncertainty within controllable limits.When the term “AI” was coined in 1956, it was defined as possessing creativity: Can it make educated guesses? You have control, you are educated, yet there is still uncertainty.

This uncertainty means that AI-native applications must perform inference in real time. Content should not be generated offline for you to consume later; rather, during real-time interactions, the model should work with humans to perform inference and generate content.

The second feature is the shift from passive to active. Rather than waiting for you to take action, it proactively provides you with feedback and stimulation. Therefore, the two most significant features of this new app are: proactivity and unlimited real-time functionality.

Xiu Han: From a business and user-demand perspective, we’ve also arrived at a similar conclusion. When users view others’ content in the community, it’s more like watching an advertisement; the true consumption occurs the moment they create the content themselves.

Looking further ahead: The ultimate long-term form is, in a sense, humanity’s shared fantasy—virtual space, or a *Westworld*-style experience. However, it will encompass aspects that traditional content struggles to cover, so the role of IP will become even more significant. Since content is constantly changing, you need some constants to aid recognition; a system where everything is in flux is essentially unusable.The “long tail” nature of this content will become more pronounced, as will its personalized nature, because the issue of “generic high-quality” content has already been well addressed. The next challenge is how to produce “non-generic high-quality” content. The dynamics of supply and demand will cause the distribution of economic benefits to shift in this direction as well, leading to a devaluation of generic high-quality content.

We’re already seeing this trend—on many platforms, AI-generated videos of beautiful young women dancing still have value in terms of traffic, but it’s become very difficult to drive sales through them. This is the process of appreciation and depreciation currently taking place.

Xiaoning: You’ve mentioned the term “IP” repeatedly. In the age of AI, how should we understand IP?

Xiu Han: IP in the AI era may not be what we currently understand as IP in a commercial sense. The concept of intellectual property itself emerged with the printing industry; it is, in essence, a system derived from advancements in printing technology. To a certain extent, AI has completely overturned the fundamental assumption that “the production of a single piece of content requires human labor.”

What I fundamentally believe is this: an IP is a shared understanding of a certain type of content. Whether it’s a shared understanding of a character or the setting of a story, people share a common perception of it. Sometimes it takes the form of a meme, and sometimes it takes the form of a running joke. These are essentially “repeaters” that people use over and over again, and in the long run, they all constitute IPs in a certain sense.

03

Creation与Recreation，

The True Logic Behind UGC and Opportunities at the Application Layer

“It’s important to clearly distinguish between ‘creation’ and ‘recreation.’ Recreation—as in leisure—is a very broad form of UGC behavior that places high demands on the quality of the process but not on the quality of the outcome.”

Xiaoning: Our generation has been very well served by PGC content. But do we really have that much demand for UGC? If consumption fundamentally relies on users participating in the creation process themselves, doesn’t that place higher demands on people?

Dai Bo: Going back to what Xiu Han said, the true consumption process takes place during the creative process, and the content produced is more like a trailer. This space still exists because everyone has a desire to express themselves. By using the camera to capture moments of daily life, Douyin provides a platform for user-generated content (UGC).

The second issue is that creativity has two aspects: motivation and ability. Generative AI provides the ability. And why do I emphasize that the application should be proactive? It’s precisely to address the issue of motivation.You may have inspiration, but the creative process is not easy—for example, many writers say that creating is actually quite painful. An interactive creative model allows you to start with small sparks of inspiration and gradually keep the momentum going. Generative AI is precisely what resolves this paradox.

Xiu Han: Let me add to that from another perspective. The listeners of this podcast may tend to view things more from a gaming perspective, and we don’t deny that PGC dominates the gaming industry as a whole. But when viewed from the broader perspective of entertainment-oriented gaming experiences, UGC is everywhere.

Service-oriented experiences are characterized by creators who go above and beyond to design for their audience, which sets the bar very high and therefore naturally limits participation to a small group of people. However, much UGC is essentially self-indulgent and social in nature, with self-expression at its core.From the filter features on Douyin to offline activities like knitting, bead art, and building with LEGO, the focus is first on enjoying the flow of the creative process, followed by the fantasy of producing something impressive to showcase. Showing off is a byproduct, but having that space for imagination is crucial.

I’ve never heard of anyone taking up photography who doesn’t want their photos to turn out well. But even among those who buy equipment just for their own enjoyment, very few may actually share their work—yet it still constitutes a highly engaging form of entertainment.

Recreation—activities that, in their early stages, aren’t centered on pure self-expression—places a high premium on the quality of the process but not so much on the quality of the outcome. It’s enough to be happy as long as you have a small circle of people who consistently support you. This aligns with the content value of the long-tail effect. Moreover, there’s no such thing as purely UGC or purely PGC content.When you create something using a TikTok template—which is likely PGC—the final content is a combination of your own expression and the template. It’s sometimes hard to determine whose contribution is greater. AI actually provides a universal guarantee of high-quality baseline content.

Based on this, any form of entertainment has four key elements in the long run: First, the character—your own representative and your point of view; second, the setting and world; third, the rules—which correspond to the resulting gameplay mechanics; and fourth, the methods of interaction.

These days, everyone seems to be looking for a consistent form of interaction to create stable platform simulations, but Vibe Coding has really challenged that assumption. The world and the rules are the elements that are least likely to change. Characters are the easiest to establish at first, but they’re also the most likely to be discarded once they’ve served their purpose; the environment and rules, on the other hand, tend to persist, allowing for the creation of a series of dynamic systems.

These elements can be used to create a “Creative Harness”—how can we turn AI into a wild horse that can be tamed? It isn’t locked in a cage, so it still has the power to surprise you, yet it’s harnessed to gallop in a single direction, bringing joy to everyone.

Xiaoning: If visual consistency is achieved, what will be rewritten?

Dai Bo: The key is real-time reasoning capability. By reducing the latency of a single reasoning step to real-time levels while maintaining long-term reasoning stability, you’re no longer just using a tool—you’re interacting with an agent within a complete world model. In a sense, it’s like the Oasis in *Ready Player One*.

Xiaoning: Where is the biggest opportunity in the application layer? In *Ready Player One*, the Oasis was also created by someone, wasn't it?

Dai Bo: The most significant thing isn’t that someone created an oasis, but that someone provided the building blocks to create one. The final oasis is made up of individual planets, and each planet is created by the users themselves. Because they have these building blocks, they can construct different worldviews and sets of rules.As Xiu Han said, the fun comes from the process of building. Once the construction is complete—or even before it’s finished—countless people can join in and continue building together. It all starts with someone providing the space and the building blocks, letting you build whatever you want.

Xiuhan: Let me add one point: Why does this need persist in the long term? People’s perceptions and experiences of their environment are, at their core, a way for them to build an understanding of the world in their own minds. You cannot truly understand something you’ve never experienced firsthand. So, essentially, it offers multiple lives.

Human needs boil down to extending life, increasing the time available for experiences, and maximizing the number of experiences per unit of time—and infrastructure that enables more life experiences is, in itself, the greatest opportunity. These experiences cannot be fully constructed and delivered to you by others—to a certain extent, that is a false premise.

Dai Bo: Often, what holds people back in the real world is the idea that “I have to make a choice—I can only take one path.”A world model should be able to create this kind of possibility. At its core, it’s essentially the same thing as a game—“游” means wandering around, and “戏” means interacting by touching this and that. Some aspects of earlier games took their current form due to technical limitations, but this isn’t a limitation of human imagination—it’s a limitation of technology. This aspect should be liberated.

Xiu Han: Let me offer a slightly different perspective. “Games are a constrained art form”—from the perspective of traditional, finite games, this holds true, and it’s the core reason I have such deep respect for the gaming industry.We rarely say, “Users can create games directly,” because I have great respect for the narrow definition of games. It is a form of experience design—much like a wedding, which a person might experience only once in a lifetime. The way it has come to establish current norms and requirements is incredibly sophisticated; it’s not just a matter of coming up with an experience on a whim.

However, gameplay elements can be broken down and engineered. For example, many open-world games today are described as “mongrels,” incorporating a wide variety of gameplay mechanics—and some even go so far as to try to weave in the experience of watching a movie. This represents a trend toward environment- and world-driven designs that integrate a multitude of gameplay mechanics.

However, to get the most out of AI, we should build it up from the ground up, rather than breaking it down from a professional experience. Breaking it down from a professional experience limits the AI’s potential.I often say, “From a gaming perspective, assets that aren’t substantial always feel like they lack respect.” This perspective leads to a certain indefinable sense of unease when breaking down a game into UGC.

So what we’re thinking is: from an AI-native perspective, what are the best small modules to build? Let humans handle the parts that are difficult to build. Many genuine solutions for UGC actually emerge from a more native, bottom-up approach.

04

World Model,

From “The Blind Men and the Elephant” to the Three-Tier Architecture

“A world model should feel alive. That sense of life comes from the fact that it moves.”

Xiaoning: If we want real-time capabilities, reasoning abilities, and controlled creativity, what gaps still exist today in achieving these goals?

Xiu Han: The building blocks are already in place; they just haven’t been fully integrated to deliver a great overall experience for everyone. That doesn’t mean you can’t start building characters, creating AI-driven worlds, or establishing rules for future gameplay right now—for example, tavern-style games primarily focus on establishing gameplay rules to a certain extent;We might focus more on environmental and visual consensus elements from the perspective of characters and the world. However, if issues like the model’s real-time performance and sustained capability remain bottlenecks or are too costly, the transformative experience that the established rules can ultimately deliver will still be limited.

Dai Bo: The current state feels both very close and very far away to me. All the elements are in place, but in a sort of “flood-like” way—they cover every direction, which means you can begin combining them. As Xiu Han mentioned, the worldview and rules that have been established won’t have to start from scratch just because technology evolves.

However, the technology falls short in three areas: depth, breadth, and dimension. Depth—for example, in a tavern scene, problems begin to arise as the complexity of the plot or the length of the performance increases.In terms of breadth—there is always a distinction between what the technology excels at and what it struggles with across various genres. As for dimensions—text-based, language-centric approaches tend to be more comprehensive; however, there remains a significant lack of development in more visual, three-dimensional world models.

Overall, you can use this “not-zero” state as a starting point for combinations, but if you really want to reach the endgame, you’ll at least need to take the time to carefully consider the technical aesthetics—and then keep moving forward at a steady pace.

Xiaoning: How do you understand the concept of a “world model”?

Dai Bo: Many people are talking about world models these days, and the idiom that best describes this situation is “the blind men and the elephant.” Depending on their different application scenarios and backgrounds, different people are feeling different parts of the elephant.

My own understanding is quite simple: it’s a game similar to the metaverse, and it should incorporate all the elements found in the real world. We’ll break it down into three layers.

Level 1: Logic/Thought Level—Each dynamic entity’s thought: Why am I engaging in these activities?

Level 2: The Dynamic Level—Once I have an idea, how do I put it into action?

Layer 3: Static/Rendering Layer—Loads dynamic actions into an environment for presentation. This environment can be a 3D environment built using a traditional game engine, a 3D environment generated by AI—such as Professor Feifei Li’s World Lab—or a video-based environment like Google’s Genie 3.

Depending on the situation, different people may not need to implement all three layers. But a true world model is one that can be mapped one-to-one to reality; it’s a living, dynamic entity where every entity has its own thoughts and plans.

Xiu Han: I hope it will at least be a sustainable system with its own intrinsic momentum.The problem right now is that when a single character moves between multiple scenes, there are disconnects; or the character isn’t proactive—it won’t act unless you prompt it. In contrast, in the real world and in offline experiences, the environment itself often helps you achieve your goals. As long as we can solve the issues of continuity and proactivity, many of our previous assumptions about high-quality content can actually be revised.

Xiaoning: When it comes to training visual models, do you face challenges that are very different from those of language models?

Dai Bo: The paradigm of language models is next-token prediction, while world models emphasize next-state prediction.In fact, language models can also implement a world model within the linguistic domain—you use text to maintain the state of the world. This is also why coding ability is so important: it provides a constant, stable, and consistent way to verify the model’s thought process and logical accuracy. If you rely solely on natural language descriptions, you’ll find yourself caught in the dilemma of wondering, “Does it really not understand, or is it just not explaining itself clearly?”

Visual content also requires consistency—but not logical consistency, rather consistency that aligns with physical intuition. I want to emphasize that, from an application perspective, what we’re looking for isn’t “physical consistency,” but “consistency with physical intuition.” People don’t stop to calculate whether the physical laws underlying visual content are consistent; as long as it doesn’t feel out of place, we perceive it as consistent with physical intuition.

Xiaoning: When it comes to interactive entertainment, what are the key challenges to overcome in the visual aspect?

Dai Bo: The traditional approach is to build the game world using a game engine. The way game engines incorporate the laws of physics is quite “hardcore”—it’s based on traditional simulation, where formulas are actually hard-coded into the engine for calculations.Neural networks, on the other hand, are only given input and output; they figure out the entire process from input to output entirely on their own, without any guidance from rules or formulas—a very “soft” approach. However, this raises the problem that they may take shortcuts, and you cannot determine whether the patterns they derive align with the ones you intend.

So the challenge we now face is this: How can we combine the strengths of both approaches—the “soft” flexibility and the “hard” guarantees? Only by doing so can we achieve long-term consistency in visual content.

In terms of the structure of visual content, the model should understand what it is representing. It’s not about “what color this pixel is now and how it should change in the next moment,” but rather about understanding that “this pixel belongs to a person’s hand; in the next moment, the entire hand is changing, and the model should know how that change should be projected onto the pixel.”With this understanding—which better aligns with real-world structure—we can fundamentally resolve the issue of hallucinations, which essentially stem from the model not knowing what the pixels it generates actually represent.

Xiaoning: The models you’ve recently open-sourced focus on memory and planning. Why is that?

Dai Bo: We’re evolving synchronously across three layers—the conceptual layer, the dynamic layer, and the static layer. Even within the dynamic layer, we’re building a dynamical world model, and we’ve just completed an innovation in the model’s architecture.To use an analogy—it’s as if the Transformer architecture had already been established, and all that remained was the process of evolving from GPT-1 to GPT-2 to GPT-3.

The three layers evolve in parallel, with each layer having its own level of technical maturity and pace. The conceptual layer moves at a slightly faster pace because the related agent technologies are developing rapidly and are already quite mature. We consider the dynamic layer to be extremely important—it’s an area that’s currently being overlooked. Although it’s overshadowed by the wave of embodied intelligence, it’s equally critical for interactive entertainment.

Xiaoning: How should we understand the “dynamic world model”?

Dai Bo: As part of the state of the world, dynamics must follow the laws of physics. The dynamic physical laws that humans have derived from reality are collectively referred to as dynamics: human kinematics, the dynamics of clothing, fluid dynamics, aerodynamics, and so on. What we do is embed these a priori physical laws of dynamics into the Transformer architecture, making next-state prediction highly stable.

Xiaoning: Why are dynamic layers particularly important in interactive entertainment scenarios?

Dai Bo: A static environment cannot be considered interactive. Without dynamic spaces and dynamic mediums, there can be no talk of proactivity or interaction. With static elements, you can only engage unilaterally—there is no feedback. Only dynamism provides feedback, and it is through that feedback that new possibilities emerge, continuously giving rise to richer experiences and content.

To put it simply: A world model should feel alive. By "feel alive," I mean it should move—because something that doesn't move at all can be pretty scary.

Xiaoning: Videos are, by their very nature, dynamic. Why are you specifically highlighting the motion? What are you trying to say?

Dai Bo: The video is indeed in motion, but the continuity of the pixels does not mean that the underlying physical laws are consistent—which is why errors accumulate and illusions arise. Either we use more data to force the model to learn a unified set of physical laws on its own, or we embed the physical laws that humans have accumulated from real-world experience as prior knowledge into the model. Both approaches are valid.

Additionally, video generation involves modeling both the static environment and dynamic elements together within a single model.I believe these should be kept separate. Here’s an intuitive example: Imagine we’re sitting in a room, and I wave my hand. From a reasonable perspective, only the states related to the person should be updated; the environment shouldn’t refresh along with me. It’s like in a video game where a character takes a single step forward and the entire environment is regenerated—that doesn’t make sense.

05

Overvaluation and undervaluation,

How Entrepreneurs Can Navigate Technology Cycles

“When things are moving slowly, it’s actually quite hard for a company to fail, but when things are moving fast, it’s very easy for a company to fail. Many people have the misconception that opportunities arise when things are moving fast.”

Xiaoning: How do you tell the difference between something that just looks technically impressive and something that can actually lead to innovation?

Dai Bo: We analyze a technology based on objective, qualitative principles—setting aside the terminology to examine exactly what the breakthrough entails.

What tends to be overestimated? First, the speed of iteration; second, human adaptability.It’s hard for people to imagine things they haven’t experienced themselves. Even those working on these technologies have their limitations—once AnimateDiff is developed, the people who ultimately use it to create great content may not necessarily be us. That’s why I’ve always been cautious and reluctant to define exactly what the native applications of the world model are.

Xiu Han: From a technical perspective, it’s easy to overestimate the speed at which an application can be rolled out and underestimate the cost of changing user mindsets.The most common mistake made in recent years has been an excessive focus on the end goal. If the underlying prerequisites are not in place, this is the core reason why it has historically been difficult to “develop AI products from a purely product manager’s perspective”—the result is either an unsustainable user experience that fades once the novelty wears off, or a single-point solution that fails to deliver enough excitement. These two factors have always been at odds with each other.

Xiaoning: The pace of development in underlying AI technologies has been constantly changing—it was rapid in the first half of 2023, relatively stagnant in 2024, and has been particularly fast over the past six months. Which phase do you prefer?

Xiuhan: Each has its own challenges. The key is this: when things are moving slowly, think about what the company is building up; when things are moving quickly, know how to immediately connect those accumulated resources to the rapid changes—that way, you won’t panic when things move fast.

Another paradox is this: when things move slowly, it’s actually quite hard for a company to fail, but when things move quickly, it’s very easy for a company to fail. Many people are under the illusion that opportunities arise when things move quickly.

Xiaoning: When faced with a groundbreaking new technology, what’s your first reaction?

Dai Bo: Trace the origins right away: where did it evolve from, what makes it different, and why is it causing such a stir?From my perspective, new technologies rarely feel like they’ve come out of nowhere. There are two key points: First, cut through the noise to understand the essence of the change; second, reflect: Why wasn’t I the one to bring about this change? Only by doing this consistently can one become a leader.

Xiu Han: Our first instinct is to rush into testing to see what the actual level of performance is—there’s just too much “cherry-picking” in research papers these days. Once we’ve assessed that, we focus on whether it changes the user experience along the main flow.Many technologies are just icing on the cake; they don’t have a significant impact on the main user journey. A key criterion is whether users are willing to come back and spend more tokens. If it brings a major improvement or change to the user’s core loop, then it’s a change we absolutely need to capitalize on.

Xiaoning: You guys have been talking a lot lately. What’s the topic you’ve been discussing the most?

Dai Bo: One reason is that technological progress is still too slow.

This is especially true for visual content. Although video generation models are making good progress, they’re still moving too slowly in key areas—they aren’t being developed entirely with interactive entertainment in mind, as other factors are driving their development. The second point is to explore which technologies are just hype and which are the real deal.Our hope is that, since the technology is still in the “blind men and an elephant” phase, we can find common ground regarding our vision for application scenarios.

Xiaoning: When starting a business focused on app development in this environment, what is the most important skill for a founder?

Xiu Han: From my own perspective, I think we rely on resilience in what we do.I’ve always believed that different approaches can all yield results at various stages—some people surf, while others climb mountains. What’s more important is having a clear understanding of yourself and your team, and not forcing yourself to do things that aren’t a good fit. If expectations are set too high, it becomes difficult for the team to leverage their greatest strengths and produce results that exceed those of others.

Dai Bo: Set aside your ego, take an objective view of technological evolution, deduce the ultimate architecture and elements based on first principles, and then, drawing on the team’s strengths, choose the path with the steepest gradient descent and move forward. Above all, avoid letting technology fuel your ego.

Xiaoning: What do you most want to achieve in the next two to three years?

Xiu Han: We are gradually building a “Collaborative World-Building” framework where people and agents work together.The most critical question for the next two to three years is: Will we see experiences of truly high enough quality—ones that can compete on equal footing with other types of consumer experiences? Once the first few such models emerge, the system’s efficiency will improve significantly, and after that, it will be very difficult to slow down.

So, essentially, we’re addressing the mechanism behind cold starts—the problem as defined, or rather, the protocol issue. Next, we may conduct more experiments overseas, and the ultimate impact will be reflected in the emergent results that arise from these combinations.

Xiaoning: Over the past year, has commercialization been a source of pressure and challenges?

Xiu Han: There’s both good news and bad news. The good news is that, globally, users’ willingness to pay and payment rates for interactive gameplay are better than expected. The bad news is that token costs and the prices of new models haven’t gotten any cheaper, so cost pressures will persist. So right now, I’m inclined to push our capabilities as far as possible without burning ourselves out.Revenue will grow naturally as we move forward—perhaps more naturally than people expect.

Dai Bo: Over the next two to three years, we’ll continue to build the world model framework based on a three-tier architecture, and on that foundation, we’ll explore native applications—we have a pretty clear vision for the entire technology landscape.We’ll also be paying close attention to the development of agentic AI—it’s both a component of the conceptual layer of the world model and a driver of productivity. If it truly boosts efficiency, many things will come to fruition sooner. Two years might be a bit too long—in the AI field, two years is a very, very long time.

Xiaoning: Are you planning to develop your own apps, or won’t you be satisfied with just building models?

Dai Bo: It’s essential to refine your understanding at the application layer. Straying too far from real-world use cases can cause technological evolution to veer off course; moreover, we’re talking about native applications powered by world models—just as language models power ChatGPT—so the gap between the model and the application won’t be that wide.

Xiaoning: Will this app be a tool or a content platform?

Dai Bo: It’s definitely not a tool; it’s an interactive entertainment app, nor is it content in the traditional sense—as soon as users log in, they’re both consuming and creating content at the same time.

Xiaoning: Can we think of it as a platform?

Dai Bo: I think it’s a platform, but I’d be cautious about describing it that way right now. Just like ChatGPT—once the model matures, it quickly becomes a widely accepted concept. So for now, I’d rather not define it using existing terminology.