Traditional model improvement pipelines often suffer from a disconnect between training objectives and real-world usage, leading to reward hacking and slow iteration. Lee Robinson from Cursor introduces a two-loop training framework that directly tackles this tension: an inner loop focused on high-quality evaluations to prevent reward hacking, and an outer loop that captures genuine user interactions from the product as training signals. This structure closes the feedback loop between model behavior and user needs, making improvement recursive rather than manual.

The concrete technical path involves using Composer 2.5 as a case study, where the outer loop feeds user feedback from the product into model training, while the inner loop relies on evals designed for real-world scenarios instead of public benchmarks. A novel teacher-student textual feedback method enables the model to learn from natural language critiques generated by stronger models, accelerating inner-loop learning. On the compute side, Cursor scales infrastructure through a partnership with SpaceX‘s Colossus cluster, allowing flexible allocation of training compute without sacrificing iteration speed.

For builders, the key takeaway is that recursive model improvement is a sustainable strategy for AI-native products: it requires investing in both high-quality eval data that matches deployment context and infrastructure that can scale training dynamically. Cursor‘s approach demonstrates how an AI product can continuously improve by treating every user interaction as a potential training signal, aligning model behavior with real human needs without relying on static datasets or manual pipelines.