Recursive Model Improvement — Lee Robinson, Cursor, SpaceXAI

Summary

The main theme is training AI models at Cursor, focusing on how models learn to train themselves through recursive improvement and feedback loops. Key subjects include model behavior analysis, online metrics, A/B testing, and scaling training with compute and data. The practical takeaway is that Cursor is building increasingly capable AI models, exemplified by the popularity of their Composer 2.5 model, through a sophisticated iterative training process.