A new startup called Moonlake AI just came out of stealth with a massive $28 million seed round to tackle one of the biggest bottlenecks in artificial intelligence: creating interactive worlds. The funding round, co-led by Threshold Ventures, AIX Ventures, and Nvidia’s venture arm NVentures, is packed with a who’s who of AI and tech luminaries, including YouTube co-founder Steve Chen, Google AI chief Jeff Dean, and Ian Goodfellow, the inventor of generative adversarial networks (GANs).
Founded by former Stanford and Nvidia researchers Fan-Yun Sun and Sharon Lee, Moonlake isn’t just another text-to-image generator. The company is building what it calls “reasoning models” designed to generate entire interactive simulations and games from natural language prompts. The goal is to move beyond the static images and videos produced by models like Midjourney and Sora and into creating dynamic, physics-based environments that users and AI agents can actually interact with.
In a LinkedIn post, co-founder Fan-Yun Sun argued that current AI models are “hitting a wall” because they lack the practical simulations needed to learn and improve through reinforcement learning. You can’t teach an AI to drive a car or navigate a complex environment with a library of JPEGs. It needs a world with rules, consequences, and interactive logic. Moonlake aims to provide exactly that, compressing what is traditionally a months-long prototyping process for games or robotics simulations into just a few minutes.
From text prompts to playable worlds
Moonlake’s platform reportedly uses a multi-layered approach. It combines multi-modal reasoning for spatial layout, program synthesis to generate the underlying interactive logic and code, and a real-time diffusion model for rendering the 2D or 3D visuals. This means a user could theoretically ask the AI to “create a low-gravity platformer game with bouncy floors” and get a playable prototype almost instantly, complete with the ability to make real-time edits to physics and game mechanics.
The high-profile backing underscores the significance of this challenge. Ian Goodfellow noted that generating simulations was one of his original motivations for developing GANs back in 2014. With this Moonlake AI funding, his early vision is getting a serious injection of capital and talent.
For developers and creators, this could radically lower the barrier to entry for building games and interactive experiences. But the bigger implication is for the field of AI itself. By creating a tool that can instantly spin up complex training grounds, Moonlake could unlock new capabilities for reinforcement learning agents in everything from robotics to autonomous systems.
The company is currently in a private preview, but its ambition is clear: to give anyone the power to build and share a world, not just a picture of one.
