After Leaving ByteDance AI Lab, a 6-Person Team Built an "AI Cat Island" | Creation Notes

This article is republished with permission from "Game Artist Game Art".

​Played Meow Island — it’s hard to imagine that the team behind this AI-powered game consists of just six people, and none of them come from a traditional game development background.

This is a simulation game developed by Miao Jituo Studio, combining AI cat emotional companionship with island management gameplay. Compared to traditional games like Animal Crossing, the biggest difference is the addition of “AI cats”—cats that can talk to players, understand the player’s intentions, respond to their needs, and feel truly “alive”.

Producer Li Chi previously explored the AI + games direction at ByteDance’s AI Lab. Back then, he encountered many games and became deeply interested in how AI could genuinely change the gaming experience and surprise players.

After founding Miao Jituo Studio, Li Chi and his team developed several games. For them, the most meaningful outcome was not so much an increase in development capability, but a more professional and deeper understanding of what “game experience” really means.

This is crucial for finding a truly valuable way to combine AI with games. And the simulation management genre that Meow Island has chosen is one of the most mainstream directions for AI Native to reshape traditional game mechanics.

This summer, Meow Island was nominated for the second “Shulong Cup” Best AI Game Award and won the gold prize in the AI Game Track of the 2026 Photon Game Competition.

In this game and the team behind it, we can see a progressive path from AI technology to game experience. The singularity moment for AI games has not yet arrived. Besides the limitations of AI itself, the harder part is how to turn the technology into an experience that players can feel, immerse in, and stay with long term.

​Game Artist spoke with Li Chi about how Meow Island has become a valuable exploration in the current era of AI reshaping the game industry, using simulation management as a starting point to find more possibilities for combining AI and games.

In June this year, Meow Island launched its first playable demo. With a cute Q‑style art, light simulation management gameplay, and story simulation experience, players could already enjoy a relaxed and fun island life: adopt AI cats with rich personalities and emotions, chat with them, play games, fish, farm, and freely decorate their own home.

Compared to traditional simulation management games like Stardew Valley or Animal Crossing, Meow Island aims to use AI to break the boundaries of pre‑designed content, so that characters, stories, and even the entire island can change with the player’s choices, delivering a more personalized experience.

The core change comes from the AI cats, each with distinct personalities, emotions, and memories. Driven by agents, the AI cats can act autonomously, interact with players, remember what players like, and turn the player’s ideas into tangible in‑game objects.

Li Chi says one of the positive feedbacks for him in making games is that the AI cats really surprise players.

Three years ago, when Li Chi was still working at ByteDance’s AI Lab exploring the AI + game direction, he encountered the field of AI NPCs. He quickly realized that using AI just for NPC chat was boring and not the kind of experience game players actually enjoy.

Just that year, the “Stanford Town” paper caused a sensation. Inspired by it, Li Chi’s team also ran a similar AI town experiment, placing more than 30 AI NPCs modeled after colleagues into the same virtual town, letting them move around and socialize autonomously.

During the experiment, the AI kept generating unexpected stories: colleagues who were close in real life often argued in the game, while a male and female colleague who barely spoke in reality developed a romantic relationship in the game.

This became the team’s source of joy and showed Li Chi a new possibility for combining AI and games—players could act as observers, watching the natural interactions between NPCs.

Li Chi told Game Artist that the team even reported the idea to Zhang Yiming, and ByteDance’s founder also found it interesting.

But at that time in 2023, AI technology wasn’t ready for a formal project: on one hand, the cost was high—running an experiment could consume hundreds of US dollars in tokens per day; on the other hand, they hadn’t found a gameplay that perfectly matched the core framework. Eventually, the experiment didn’t turn into a real project.

The same year, ByteDance adjusted its game business. Li Chi, who still wanted to pursue game development, gave up the opportunity to transfer to Douyin and instead chose to start a company in the AI games direction he had been researching, founding his own studio, “Miao Jituo”.

Today, Miao Jituo has eight members: six people and two cats acting as mascots. Starting from the AI town, the team wanted to preserve the autonomous interaction of AI characters while making the experience differentiated. So they chose an “island” as the game space—it is naturally isolated and separated, suitable for both single‑player experiences and easy addition of multiplayer features.

​Breaking Pre‑written Narratives，AI Games Can Be “a Thousand People, a Thousand Faces”

Before Meow Island, Miao Jituo had already explored many directions for combining AI and games. Li Chi admitted that the team mainly came from a technical background and initially lacked game design experience. Over the past three years, they made many mistakes, and several projects were abandoned because the combination of AI and gameplay wasn’t ideal.

How should AI truly enter games to bring new gameplay and experiences that are hard to achieve with traditional methods? This is a key question for the industry today, and the Miao Jituo team’s understanding has gone through several iterations.

Li Chi initially had a framework: AI could create content that is difficult to produce through traditional means, such as generating game assets and storylines in real time. The team’s earlier games—Mengzhua Pai Dui, Mengzhua Da Luan Dou (originally titled Yanling Jihua), and Bao Wai Jiu Yi—all started from this direction.

Through practice, they found that AI generating large amounts of content in real time could indeed bring a different game experience, but without optimization, the content could become redundant and variable in quality, making players feel bored.

So in the second stage, the team started to “subtract”: they would rather delete low‑quality content to ensure that the AI‑generated game experience was not inferior to manually crafted games.

This process initially relied on human judgment. The team accumulated a large amount of game experience, structured and tagged cases, and further recorded “what was good and why it wasn’t good”. As data accumulated, they used it to train models, letting agents assist in judgment to improve efficiency.

However, game experience is hard to fully quantify. The team has a set of basic principles: AI‑generated content should be concise enough to avoid information overload, leaving room for players to think and imagine; more importantly, it should bring a sense of surprise.

This surprise could come from AI’s connection to the player’s personal experiences. For example, if a player owns a cat in real life, they might encounter a very similar cat in the game. Such content is not pre‑written but generated by AI based on the player’s background.

In traditional simulation management games, the storylines and NPC interactions are mostly pre‑designed, so when players replay, the dialogue and events are largely fixed. Meow Island aims to use AI’s “emergent design” to break this preset, making the game a “living” world: characters, events, and stories all change based on the player’s choices, preferences, and experiences.

What the player has done before, what they like or dislike, all become part of the content generated later. Each player and each island can develop different stories, and they can extend infinitely.

In the team’s view, this might be the greatest value of combining AI with games.

“Dawei from miHoYo once said he wanted to make a truly ‘thousand people, thousand faces’ game, where every player has their own exclusive experience,” Li Chi said. “We are doing something similar.”

In Li Chi’s view, Meow Island cannot be just a “pure research” game that demonstrates AI capabilities; it must be genuinely fun, so that players will appreciate it and pay for it.

And to make players stay, the first problem is whether the AI NPCs can “come to life”.

In Meow Island, the team hopes that the cats not only have unique “cat personalities”, but also change appearance, state, and personality as they interact with players, growing through shared experiences, becoming “digital cats” that accompany players long term.

To ensure these cats are smart enough and say interesting lines, Meow Island has built a “Model Mesh” on top of a multi‑agent framework with mixed models, matching different models to different tasks: stronger models for complex tasks, lighter models for simple role‑playing and chat, to solve performance, cost, and latency issues when running 30–50 agents simultaneously.

The choice of “cat‑form” NPCs rather than “human‑form” was also deliberate.

Li Chi believes that pet images are more suitable for AI companionship at this stage. Human emotions and cognition are more complex, making it difficult for AI to simulate a real person stably over time. In contrast, players’ expectations for pets are lower; a cat that is occasionally forgetful or looks a bit “dazed” might even be endearing.

Of course, from the studio name to earlier titles, it’s clear that the team really loves cats.

Li Chi thinks that besides attracting cat lovers, Meow Island targets two other types of users: players who enjoy simulation management, building, and social interaction; and younger users who are interested in AI itself and already use various AI products. Especially young people growing up with AI may want a “second world” where they can express themselves freely, beyond real life.

Based on this, Meow Island wants to further expand the freedom of AI + UGC in the game. Players can use AI to generate buildings, flowers, and grass on the island; they can also create social mini‑games like “Truth or Dare” or “Draw and Guess” to play with others.

Previously, this kind of play style was limited by token costs. But now, Li Chi says the cost of running AI in games has dropped very low. The technical model is somewhat like ChatGPT and DeepSeek without calling APIs; a player might play for twenty hours and consume only a few cents. So the team’s focus is less on cost and more on how to retain players long term.

From the feedback after the demo launch, players gave many positive comments. Besides finding it “fun”, they also developed an emotional attachment to the island and the cats, feeling that they gained companionship and spiritual comfort.

Li Chi himself enjoys casual simulation games, using them to escape the hustle and bustle of city life. “Just hearing those cat sounds, farming and fishing—it recharges you.”

In the future, the team will gradually improve the multiplayer social system, exploring moving from “AI companions” to an “AI society”.

In Li Chi’s view, the “AI society” is already taking shape. In Meow Island, this “AI society” will take a more concrete form: AI pets and NPCs gather on the same island, chat, play games, become friends, and generate new stories through interaction. This is the “AI life” that AI games can create.

Miao Jituo Studio was founded in 2023, precisely when large language models became the trend, and discussions about AI restructuring the game industry swept the field. Three years later, Li Chi still finds it hard to judge when the real turning point for AI games will come.

“Maybe we’ve only explored the tip of the iceberg; 90% of the domain is still in complete fog, and the full map hasn’t been opened yet.” It’s too early to determine when the entire AI game track will take off.

But some possibilities have already emerged. For example, History Simulator: Chongzhen — Li Chi thinks it captured a core element: the “emergent design” that Meow Island is also working on.

Although high freedom comes with many problems—such as how to maintain AI consistency, NPCs forgetting what they said before, disconnected story contexts, logical fallacies, etc.—that doesn’t stop it from being loved by some players, and the direction for optimization is already clear.

Li Chi believes that AI’s value will differ across game genres. For example, highly mature genres like MOBA already have well‑defined core gameplay and art systems, and don’t rely heavily on AI to generate new mechanics or content, so the incremental value AI can bring may be limited.

Furthermore, because AI‑native games have not yet formed a clear user base or mature community, the distribution logic of traditional genres is hard to apply directly. New ways may be needed to reach potential players who are interested in AI‑native experiences.

For now, Miao Jituo’s most important goal is to finish Meow Island completely, and then move on to a truly free open world. In terms of business model, Meow Island leans toward a long‑term operation game that continuously supplies content, referencing monetization methods of mature products like Heartbeat Town.

In fact, over the past three years, what truly brought the team to break‑even was not a game, but a self‑developed AI game creation tool—Meowa.

While developing Meow Island and other games, the team built tools and technical pipelines for different needs and gradually integrated these capabilities into Meowa. As more users wanted to use this tool, the team decided to open it externally, and it now has over 30,000 users.

Recently, the team also plans to start a new round of financing—several million US dollars—which will be used mainly for two things: advancing Meow Island’s development, and pushing Meowa from a development tool toward a platform.

Li Chi uses the relationship between Valve, Steam, and Half‑Life as an example; in the future, he wants to build Miao Jituo’s own game platform. He envisions that if large‑scale, high‑quality game generation becomes feasible, the distribution logic of the platform might also change.

Unlike the “shelf‑style” display of traditional game platforms, players might prefer to discover and experience games through recommendation feeds, forming a model similar to “game version of Douyin”. This is an area many game creators are interested in, and investors are more willing to pay for it.

“Everyone is very curious about what the next ‘Douyin‑style’ game content platform will look like.”

However, all of this depends on whether these AI games themselves are high‑quality and fun enough to truly attract and retain players.

Li Chi says Meow Island wants to explore more cutting‑edge possibilities of combining AI and games. It may still need some polishing and optimization time, “but we will definitely do our best to make it good and serve everyone a truly delicious ‘dish’.”