Moonlake AI's Fan-Yun Sun is building interactive AI-generated worlds using code as the world model

Summary

Fan-Yun Sun founded Moonlake AI to build interactive worlds where code serves as the underlying world model rather than pixels. The distinction matters because most foundation models fail at long-horizon tasks. They lack genuine world state prediction. Genie, DeepMind's pixel-first model, produces impressive visuals but loses coherence after roughly a minute. Moonlake splits the problem into two components. Code and symbolic representations handle the deterministic, interactive logic. A pixel model handles appearance. The engine is a customized fork of Godot, post-trained to let Moonlake's model access tools that off-the-shelf models cannot. A recent demo showed an AI-generated environment where players can bowl, interact with an arcade cabinet running Space Invaders, and hear ambient music, all built by the model. Sun describes this as a code-generation model trained to construct worlds, not just render them.

Consumer game creation

Moonlake's near-term commercial target is consumer game creation. The pitch is that creative leverage shifts from technical skill to taste. The platform lets anyone with a good idea build and monetize games without domain expertise. It already supports one-click deployment of multiplayer experiences, with the model automatically configuring databases and multiplayer infrastructure on request. Sun points to a Roblox-style network effect as part of growth strategy. Users get a persistent Moonlake identity that travels across games built on the platform. Moonlake handles distribution and monetization.

Longer vision

Sun wants to train models that can plan across long horizons in both virtual and physical environments. The ambition extends beyond gaming into multimodal reasoning.

Funding

Moonlake has raised $28 million in total. Investors include Jeff Dean, whose participation was cited as a notable endorsement.