Recursive Model Improvement — Lee Robinson, Cursor, SpaceXAI

Model Training Strategy & Infrastructure - Cursor employs a recursive model improvement loop, scaling compute and data feedback to iteratively enhance model performance [2][3][4] - The training process is divided into an outer loop (user feedback and online metrics) and an inner loop (high-quality evaluations and difficult task generation) [5][6][15] - Cursor leverages massive compute resources, including a partnership with SpaceX to utilize supercomputing clusters like Colossus, which added 100,000 GPUs in 92 days, following an initial 100,000 GPU build-out in 122 days [32][33] Product Performance & Market Positioning - The "Composer" model, released in May, has become the most popular model within the Cursor platform, balancing speed, intelligence, and cost-effectiveness [7][9] - Agent usage currently accounts for the vast majority of Cursor's revenue, providing a critical data source for training future model iterations [13] - Internal evaluations utilize private datasets derived from real-world engineering tasks to prevent "reward hacking" and ensure models are tested against genuine software development challenges [21][22] Research & Operational Efficiency - Cursor implements "textual feedback" in Reinforcement Learning (RL) to provide precise guidance to models during training, improving adherence to tool-calling and stylistic requirements [29][31] - The engineering team utilizes a fleet of agents to automate research workflows, allowing researchers to launch and monitor experiments directly via Slack to eliminate human bottlenecks [48][49] - Recursive self-improvement is achieved by using the most intelligent model checkpoints to create derivative models, which then optimize the evaluation and reward generation processes for subsequent training cycles [51][52]