Overview
- In early June, 25‑year‑old AI founder Ziwen Xu announced on X that he had begun building a Grand Theft Auto–style game, pushed a GitHub repo, and asked modelers, coders, artists, and writers to join the effort.
- Vibe coding is the project's method and it means relying heavily on large AI models and agent workflows to generate, debug, and assemble game code and assets from natural‑language prompts.
- Daily demos have moved from a simple placeholder bean to humanoid characters, NPCs, driving cars, and basic weapons, but the build remains rudimentary and far from a finished game.
- Xu reports major practical limits: the agent produced wrong city details (LA‑style skyscrapers instead of Florida), models make generation errors, and he said he used a third of his Claude Max 20x weekly allotment in one day, raising cost and feasibility concerns about beating Rockstar to market.
- The public, open‑source run is a live test of how safety‑tuned public models and community contributions scale creative projects and it will show whether AI agents can meaningfully shorten long game development cycles or mainly produce brittle, expensive prototypes.