Playgrounds for AGI

There is a critical missing piece in developing AGI: there are no training environments available for building core cognitive skills like exploration, long-horizon reasoning, and coordination, so that they generalize to diverse domains. This is why today's models still fail at real-world tasks like doing your taxes, or autonomously running an organization.

To solve this, we create custom games where frontier LLMs learn from playing with humans and other agents.

Why games? Games provide structured rules, clear rewards, and mimic real-world complexity, making them ideal for developing the abilities current models struggle with.

We built and scaled 4Wall AI, a platform with over 100K users and active members spending over 80 mins daily creating and interacting with persistent, multi-agent worlds powered by AI characters. We are now expanding from consumer entertainment to working with AI labs to train the cognitive skills required for general intelligence.