Recent advances in AI allow us to turn our attention from how to what: what capabilities are the most important to build out first? What is critical on the path to strong AI? Our view is that there is exactly one right answer. We must first teach the models to perform ML research.

Existing frontier models are brittle when applied to real-world ML tasks. The present bottleneck is the lack of high-quality RL training environments. We build RL environments that reflect real-world complexity, with diverse tasks and robust reward functions.