Game Builders Seoul, Game meets AI
Games often begin with something small: a story someone wants to tell, a mechanic they want to try, or a moment from everyday life. Turning that idea into something playable takes experimentation, persistence, and a willingness to start again.
On August 31, OpenAI Game Builders Seoul brought together people exploring that process with Codex. Hosted by OpenAI and organized by Com2uS, the event was a chance to share games, compare approaches, and hear what builders had discovered along the way.
The day featured two tracks.
In Track 1: From Zero to One: Build with Codex, 40 shortlisted teams presented the games they had created during the Online Warm-up Challenge. Each team representative had up to three minutes for an elevator pitch and game demo, sharing the core idea, how Codex supported development, and what the team learned along the way.
In Track 2: Go Full Codex: 5 Hour Build, builders had five hours to create a playable browser game around a new theme revealed at the event. For builders coming from Track 1, it was a chance to start fresh and explore a different direction.
Game industry Legends and Special Guests also joined the event. They shared projects, creative processes, and perspectives on how AI is changing the way games are built and played.
Their experiences ranged from building an MMO alone to introducing AI across an entire studio. Others explored game balance, live operations, and AI as part of gameplay. Together, these stories offered a closer look at what becomes possible when more ideas can become playable.
Stories From Game Industry Legends
Building an MMO and a World AI Can Join
Song Jae-Kyung (Jake Song), the developer behind The Kingdom of the Winds, Lineage, and ArcheAge, shared his progress on an open-source MMO built with AI.
His work spans servers, shaders, physics, procedural terrain, and 3D assets. AI has helped him take on tasks that would usually involve several specialists, opening up areas that would have been difficult to explore alone.
He described writing around 30,000 lines of code over a year on earlier work. His current project reached approximately 160,000 lines in eight months. He estimated an eightfold increase in coding output, with an even greater change in overall content production.
The presentation also included practical discoveries from development. For the server, Song chose Rust. He found that AI worked effectively with the language, while compiler and type checks helped catch mistakes in generated code.
For assets, he connected tools suited to each stage. Generated character images became 3D models, which then went through rigging. Codex used Blender to adjust their scale and bone names to match the game’s requirements.
This let him describe the result he wanted and work toward it without first mastering every feature of every tool. The process still required choosing suitable tools, reviewing their outputs, and checking that everything worked in the game.
A Shared World for People and AI Agents : His project also explores how AI agents could participate alongside human players. Both connect through the same server protocol. People see the world through graphics. Agents receive a text representation of the same game state. Both follow the same rules.
Each part of the system has a clear responsibility. LLMs support conversation, interpret context, and set goals. Traditional algorithms handle pathfinding. The server enforces the game’s rules.
Early experiments exposed delays that interrupted conversations. He changed the design so agents activated in response to events, such as a player approaching or speaking. When no players were nearby, the system could skip unnecessary calls.
This suggested possibilities for a familiar challenge in long-running MMOs. Starter areas often become quieter as established players move on. AI companions could give newcomers someone to explore with. Characters that remember actions and respond to context could also make a world feel more personal. Song’s work explores where those capabilities might improve the experience of playing.
One Builder, One Game
Com2uS Chairman Song Byung-jun began by looking back at mobile game development 25 years ago. At the time, an entire game had to fit within 24KB. Games now occupy tens or hundreds of gigabytes. Yet the question at the heart of development remains familiar:
“How can the game be more fun?”
AI offers new ways to pursue an answer. One person can discuss an idea, turn it into code, play the result, and revise it. Song described the possibility of one builder creating an entire game.
As more people gain the ability to build, games can reflect a wider range of interests, perspectives, and ideas of fun. He also discussed the work that follows a playable build. Authentication, payments, notifications, user analytics, and server operations all help a game reach players.
Hive Axyl, introduced at the event, supports this stage. The game backend platform connects services through AI-assisted workflows and provides a console for configuration and live operations.
He described complementary roles for Codex and Hive: helping people expand their creativity, then helping those games reach players.
Creative Direction as Prototypes Get Faster
Kim Dae-hwon shared how AI is bringing more developers into a producer’s role.
That role involves setting a direction, reviewing work in progress, and deciding what needs more attention. It also means recognizing when the result has moved away from the intended experience.
He recalled action RPG prototypes that took 13 to 15 people around six months to build. Early versions often lacked the art, sound, and feel needed to communicate the finished game. Teams had to explain an experience that others could not yet play.
AI now allows smaller groups of experienced developers to create more representative prototypes earlier. Playing those builds gives teams a clearer basis for creative decisions. He reported efficiency gains of roughly three to five times on some tasks using coding agents. Faster implementation also creates more decisions about what to build and what to refine.
Which ideas belong in the game? What deserves another iteration? Where would more attention improve the experience? Those choices remain central to the creative process.
Knowing the Recipe, Judging the Taste
Kim Yong-ha, executive producer of Blue Archive, offered a cooking analogy: AI may know the recipes, but it cannot taste the food.
“AI may know the recipes,
but it cannot taste the food.”
AI can implement familiar mechanics, solve well-defined problems, and offer alternatives. People still need to judge whether a character is appealing, movement feels satisfying, or a game makes them want to play again.
He shared an example from an internal game jam at Nexon Games. Around 50 teams used AI to create 50 playable games in two days. The results varied with how participants guided the tools. Knowing what to request, and when to redirect the work, helped builders bring their ideas into focus.
A game built around one compelling mechanic can leave a lasting impression. As implementation becomes more accessible, experience and taste help determine which ideas deserve further development.
The Possibility of a One-Person Studio
Park Sang-jun, director of Com2uS AX HUB, explored how AI could change who gets to create.
YouTube and social platforms gave more people a way to make and share content. AI could extend that opportunity to projects that once required larger organizations.
One person or a team of two or three can increasingly build a game around a particular interest or personal vision. Park described the potential for a new generation of one-person studios.
Lower development costs make it easier to test distinctive ideas. Concepts that once stayed in a design document can become playable prototypes. Small teams gain more room to pursue ambitious work and discover what makes their games unique.
Learning Together Across a Studio
Aquatree CEO Park Beom-jin shared his experience introducing AI across a studio of around 100 people.
The team had been focused on launching a game, leaving limited time to explore new tools. Seeing programmers begin to use AI prompted him to investigate further.
After a month of studying the tools himself, he organized an internal AI Day. Programmers, designers, artists, and colleagues across the studio shared examples from their work.
The change he emphasized was an expansion of what people could take on. Team members continued to apply their expertise while exploring tasks beyond their usual responsibilities. AI gave individuals new ways to contribute, learn from one another, and work toward larger goals.
Making More Time to Playtest
Second Dive CEO Ban Seung-cheol shared an AI workflow for balancing MMORPGs.
Combat, progression, rewards, the economy, and monetization affect one another. Individual data checks can pass even when the overall game feels unbalanced.
Second Dive encountered this during large-scale testing ahead of a global launch. Playing the game revealed problems that internal checks had missed.
The team built an AI workspace that could interpret game data and code. It combined knowledge of the game’s rules with task-specific tools and safeguards against incorrect changes. Peer review and human approval remained part of the process.
Ban wanted to change how designers spent their time. Data entry had taken around 70%, leaving about 30% for testing. His goal was to reduce data work to roughly 10% and devote more time to repeated play.
He reported a substantial reduction in data work and a 3.5- to fivefold increase in throughput measured by commits. Designers could also review combat, rewards, drops, and progression together. The team used the time saved to play more, examine the results, and improve the game.
Giving AI Direction and Room to Explore
Ko Jeong-hwan, co-CEO of Supervillain Labs, discussed how to guide AI toward a better play experience.
In his experience, AI can favor solutions that are easier to produce. Clear principles and quality expectations help align the work with the developer’s intentions. Reviewing a result from several perspectives can uncover useful improvements. Even a small observation may reveal something worth investigating.
He also emphasized leaving room for exploration. Instructions that define every check too narrowly can limit the problems AI is able to identify. A useful workflow provides clear standards while allowing different approaches to emerge.
Tools People Can Rely On
Hong Seong-jin, technical advisor at Devsisters, described three requirements for useful AI tools.
They should solve a concrete problem, meet the expectations of domain experts, and support ongoing quality measurement and recovery. A promising demonstration gives people a reason to try a tool. Reliability in everyday work gives them a reason to keep using it.
When AI Becomes Part of Gameplay
ReLU Games developer Kim Dong-gyu shared lessons from experimenting with AI-based games.
He suggested asking what part of the experience would disappear if AI were removed. That question helps clarify its contribution. AI might support production, or it might enable a mechanic that gives the game its character. Understanding that role can help a team decide where to take the idea next.
Stories From the Builders
40 Shortlisted Teams, Many Ways to Begin
The shortlisted teams presented ecosystem simulations, cooperative puzzles, strategy games, rhythm games, text adventures, and multiplayer experiences.
Some ideas began with exercise, healthcare, public transit, or a small frustration from everyday life.
The builders included designers, artists, students, a nurse, experienced game developers, and first-time builders. Each brought a different starting point and a different story about turning an idea into something fun to play with.
A Nurse Brings a Story to Life
The builder participating as 코딩하는 간호사, meaning “a nurse who codes,” had worked at a university hospital for eight years. Team explained that writing even simple code would have been difficult without AI. Their submission was 피안화 ㅡ 꺼지지 않는 등불, a 3D horror adventure.
“The project began with a story
I wanted to tell.”
The game follows a girl returning to an abandoned school where her older sister disappeared ten years earlier. Exploring the school reveals lost memories and clues to what happened. The builder worked with Codex on code, story development, scene planning, and the 3D asset workflow. It also supported the preparation of presentation materials. AI helped them work beyond their technical starting point and give that story a playable form.
Finding the Idea Worth Playing Again
공룡달리기 · GongRyong Run, by team 동그라미, emerged after several attempts to find the right concept.
The builder developed 11 ideas and made seven prototypes. Playing them revealed that none delivered the experience they wanted.
While working on the seventh, their stomach rumbled. Their wife joked, “Are you raising a dinosaur in there?” The remark inspired a dinosaur running game. A new prototype followed that day.
This time, the builder found themselves starting another run. That desire to play again helped them recognize the idea worth submitting.
AI made it easier to try several possibilities and move on when they fell short. Lower prototyping costs gave the builder more freedom to keep looking.
Finding a Way of Working
TEAM PANGU created 아르카나의 기록, an interactive narrative combining tarot and personality tests. The team had begun exploring vibe coding about a month earlier.
They built a workspace where people could edit the story and character values. Codex then applied those decisions to the game. AI also ran the game and captured specific scenes. The team checked screenshots for cropped images and misplaced elements, then repeated the process of fixing and testing.
‘Unknown Cake’ built Metro Villain, a social deduction game inspired by difficult passengers on public transit. None of its three members had prior game development experience. They used AI to implement the game and run hundreds to thousands of simulations, refining its rules and systems along the way.
For 10초만 왕국, people defined the rules and intended experience. Codex supported implementation and verification. When playtesting uncovered a bug, AI reproduced the situation and worked on a fix. Regression tests helped prevent the same issue from returning.
Other teams created developer modes for editing objects, spaces for testing sound and features, and sites for organizing assets.
These tools grew out of each project’s needs. A repeated instruction could become a control that a person adjusted directly. A shared workspace could make the next change easier to understand and review.
Letting Play Guide the Next Step
Across the presentations, builders returned to the importance of playing their games.
They tested projects themselves, shared them with friends, and watched how people responded. Those experiences helped them decide what to change and what to develop further. Codex supported implementation. Playing, sharing, and refining helped each team bring its idea into focus.
Five Hours to Try Something New
Track 2 began with the theme “Go Limitless - Surprise me.”
Builders had five hours to choose an idea, define its scope, and create a playable browser game with Codex.
Some teams tested several concepts before choosing one. Others built the core mechanic first, then added art and sound. The time limit brought priorities into focus. Each team had to decide which elements mattered most and where another iteration would make a difference. The games offered different interpretations of the theme. Some stood out for their technical ambition. Others found humor or an unexpected perspective in a familiar situation. Each reflected the choices its builders made throughout the afternoon.
Final Recap
The event brought together different scales of work and different stories.
Song Jae-Kyung explored solo development and AI agents in a shared world. Kim Dae-hwon and Kim Yong-ha examined creative direction and taste. Park Beom-jin described learning across an organization.
Ban Seung-cheol and Hong Seong-jin focused on domain knowledge and reliable workflows. Kim Dong-gyu and Ko Jeong-hwan considered how AI could contribute to gameplay.
Lee Jung-woong, Lim Hyeon-soo, and Lee Jun-hee also shared approaches to trying more ideas, moving on quickly, and discussing a concept before implementation.
The 40 shortlisted teams showed what these possibilities looked like in practice. Time saved in development became time to test another idea, share an early build, or refine a promising mechanic. Across the projects, Codex supported different creative processes. The builders’ interests, decisions, and experiences gave each game its character.
What’s Next?
Every submission to the Online Warm-up Challenge helped shape OpenAI Game Builders Seoul. Builders brought their ideas, worked through problems with Codex, and made something another person could play.
For many, that playable build is the beginning of another stage. Reaching players requires authentication, notifications, in-game mail, and other backend services. For individuals and small teams, building and maintaining that infrastructure can take considerable time.
Helping Games Reach Players With Hive Axyl
Hive Axyl is a game backend platform designed for AI-native development workflows. Builders can describe the services they need in natural language and use tools such as Codex to generate integration code. This extends a familiar way of working into launch preparation and live operations.
The platform draws on Hive’s experience supporting games worldwide. According to Com2uS, Hive has served more than 980 million cumulative users across approximately 250 games. Hive Axyl provides SDKs for authentication, payments, push notifications, in-game mail, and coupons. These services reduce the infrastructure developers need to build themselves.
Build with Codex, Play with Hive Axyl
After launch, its console supports user analytics through events such as logins and payments. Teams can also manage targeted campaigns and service settings. Reducing backend work can give developers more time to understand their players and improve the game. For builders who have created a playable project with Codex, Hive Axyl offers tools to help take that next step.
Thank you to all the builders, game industry Legends, Special Guests, and everyone who helped bring OpenAI Game Builders Seoul to life.
Build with Codex, Play with Hive.