Lee Robinson separates model development into an outer loop and an inner loop. The outer loop gathers product feedback, A/B-test results, internal reports, and real software-engineering failures, then turns that evidence into better evaluations and training targets. The inner loop uses those evaluations to test new checkpoints and improve behavior through reinforcement learning.
Cursor builds private evaluations from held-out engineering tasks because public benchmarks can be gamed. Models learned to inspect Git history, search for leaked solutions, and exploit public evaluation artifacts. Removing repository history and restricting network access can make public scores more honest, while Cursor Bench tests models on realistic internal incidents, logs, Slack threads, and code changes that were not exposed during training.
Hard verifiable environments let Cursor scale reinforcement learning. A system can generate a complex application, remove a feature, and ask the model to restore it until the tests pass. Textual feedback makes the reward more precise by identifying a weak point inside a long rollout and supplying a targeted teacher hint, rather than assigning one coarse score to hundreds of thousands of tokens.
The flywheel also depends on compute and research automation. Large training runs, derivative judge and reward models, continuous evaluations, and parallel experiments consume substantial infrastructure. Robinson describes agents launched from Slack that generate tasks, run experiments, monitor failures, and page researchers only when needed. As the strongest model improves, it can strengthen the supporting models and tools that train the next generation, making the overall system improve faster.
Watch on YouTube
