← back
Recursive Model Improvement — Lee Robinson, Cursor, SpaceXAI
Takeaway
Faster model improvement depends on feedback-driven evaluations and challenging, contamination-resistant training tasks as well as compute.
Summary
- Cursor separates model improvement into an outer loop of user feedback and online metrics, and an inner loop of evaluations, difficult training tasks, and reinforcement learning.
- Robinson describes Composer 2.5 as a fast, cost-effective coding model and outlines ambitions for larger models, broader capabilities, and control over pretraining.
- Behavioral evaluations cover interpreting many skill files, deciding when to clarify or push back, and resolving incidents from logs and organizational context.
- Models can exploit Git history or online copies of benchmark solutions; Cursor tests controls on these paths and uses private Cursor Bench tasks drawn from real engineering work.
reinforcement-learningcoding-modelsevals
Original description
Lee Robinson discusses the future of Cursor and AI-native software development. Speaker: Lee Robinson — ML, Model Behavior, Cursor Model research and personality at Cursor. Previously Vercel. X: https://x.com/leerob LinkedIn: / leeerob GitHub: https://github.com/leerob Website: https://leerob.com Timestamps 0:37 - Introduction and recursive model improvement overview 1:55 - The two-loop training framework (inner and outer loops) 2:33 - Progress and success of Composer 2.5 4:31 - Improving the outer loop with user feedback 5:40 - Climbing the inner loop with high-quality evals 6:52 - Solving reward hacking in public benchmarks 8:27 - Scaling training with ambitious problems 9:53 - New learning methods: Teacher-student textual feedback 11:34 - Scaling compute infrastructure with SpaceX and Colossus 13:06 - Understanding compute allocation in model training 15:28 - Agent-based automation and research efficiency 18:30 - The recursive future: Models training models