Cursor’s Model Flywheel: From Fine-Tuning to Full Pre-Training and Recursive Improvement

Cursor is moving from fine-tuning an open-source base to full pre-training, controlling every layer. Agent usage data—now the majority of revenue—feeds an outer feedback loop of evals and an inner RL loop.

This lets them address specific behaviors, such as managing skill files or knowing when to push back on users. They harden evals against reward hacking by deleting git history and limiting network access.

Recursive model improvement: the best model generates derivative models for evaluation and reward, lifting all loops. Textual feedback nudges the model during RL by hinting at better tool use, improving credit assignment in long agent rollouts. They auto-generate coding tasks by stripping features from a greenfield app for the model to reimplement.

Organizational implications are significant. ML researchers orchestrate fleets of agents from Slack, automating experiment launches and monitoring. This human-agent coordination, where agents page researchers when infra fails, previews a near-future workflow. However, the approach demands massive compute, now supplied by SpaceX’s Colossus supercomputer and Terafab chips, putting it out of reach for most teams.