Agent Prediction
AI’s raw intelligence is something that is at least as vast as all digital human intelligence, but having the capability versus practically using that knowledge remains a wide gap in today’s models. AI is still currently most effective as something that is fundamentally a tool: a human wields it to leverage themselves. Creativity, tasks based on a vision rather than a product, and novel problems are some of the gaps that still exist, and proper learning environments are the bridge to help AI cross them.
We believe that countless numbers of these problems exist within alternative data sources and the frontiers of information exchange. Between large corpuses of messy data involving the communication between players of varied information levels and hundreds of different venues for this exchange, there are significant problems that require creative techniques to extract.
Frontier AI still struggles with this today, and we build the RL environments needed to improve those skills. The paradigm is perfect: rewards are fully verifiable, and there are countless techniques that are plausible, giving GRPO the rollouts that are needed.