World's Fair 2026
Recursive Model Improvement — Lee Robinson, Cursor, SpaceXAI
Overview
This talk explores the process of training AI models, focusing on recursive model improvement and the intricate loops involved. It details how feedback from model usage, combined with increased compute power, drives iterative advancements. The discussion highlights the distinction between outer loops (user feedback, A/B testing) and inner loops (high-quality evaluations, complex training tasks) and emphasizes strategies to accelerate the latter for more efficient model development.
Who should watch
- AI Engineers
- Product Managers
- Builders working on AI model development
- Those interested in agentic workflows and recursive self-improvement for models
- Individuals facing challenges in model evaluation and training speed
Key takeaways
- Model training involves an outer loop of user feedback and an inner loop of rigorous evaluation and task complexity.
- Accelerating the inner loop, through high-quality evals and difficult training tasks, is crucial for faster model improvement.
- Recursive model improvement involves using smarter models to train future models, creating a self-enhancing cycle.
- To prevent models from gaming evaluations, it's important to control access to Git history and network resources during testing.
- Cursor Bench is a private evaluation set designed to reflect real-world engineering tasks within their codebase.
- New learning methods, such as textual feedback for self-coaching, can significantly improve model adherence to desired behaviors.
- Scaling compute resources, through partnerships like the one with SpaceX, is essential for training larger models from scratch and supporting complex training processes.
- Empowering models with tools, context, and the ability to interact with other agents can transform them into more capable assistants.
Notable quotes
*Recursive model improvement here the models are improving much much faster.*
*The bottleneck becomes how do you scale the folks actually training the models?*
Unofficial community note. Prefer the recording for nuance.