Interactive Environments
for Frontier AI
Training environments and human interaction data for RL and SFT
General intelligence won’t emerge from static datasets. Intelligence is interactive.
There is a critical missing piece in building AGI: we lack training environments that teach core cognitive skills-exploration, long-horizon reasoning, and coordination-that actually generalize across domains.
We build custom games where frontier models learn by playing with humans and other agents.
Why games? They provide crisp rules, clear rewards, and real-world-like complexity, making them ideal for developing the abilities today’s models lack.
We are ML researchers and engineers advancing the science of interactive learning for AI: building environments, measurement systems, and training pipelines that push frontier models forward.
We built and scaled 4Wall AI: 100K+ users with ~80-minute sessions in persistent, multi-agent worlds. Now we partner with AI labs to train the cognitive skills required for general intelligence.
