Recursive Model Improvement: How Cursor Trains Its Own AI

Lee Robinson reveals a two‑loop framework that lets AI tools self‑improve, from user feedback to agent‑driven research.

Want to know how AI can upgrade itself without human hands? Lee Robinson breaks down a recursive training loop that’s reshaping software development.

⚡ TL;DR: Cursor uses a two‑loop system where models improve from user feedback and high‑quality evaluations, scaling up with SpaceX compute and even training other models.

The Simple Version

Think of the AI as a student that gets graded, then teaches itself better ways to answer.

🎯 Think of it this way: It's like a chef tasting their own dish, adjusting the recipe, and then teaching other chefs the new recipe.

How It Actually Works

Inner Loop

Inner loop refines the model on specific tasks using high‑quality test cases.

🎯 Think of it this way: It's like a musician practicing a tough passage until it's perfect before moving on.

Outer Loop

Outer loop updates the training data based on overall performance and user signals.

🎯 Think of it this way: It's like a teacher changing the curriculum after seeing how students perform on a test.

Reward Hacking Fix

They tackle reward hacking by building better public benchmarks that resist shortcuts.

Technical Deep Dive

Two‑loop framework: The inner loop optimizes weights on curated evals, while the outer loop expands the dataset and adjusts loss functions.

Composer 2.5: Demonstrates faster code generation with fewer errors, thanks to refined inner‑loop evals.

Teacher‑student textual feedback: Models receive plain‑English critiques and turn them into training signals.

SpaceX Colossus infrastructure: Provides massive GPU clusters to run ambitious training runs at scale.

Agent‑based automation: Autonomous agents run experiments, freeing researchers for high‑level thinking.

Why This Matters

Self‑improving AI reduces the bottleneck of manual model tuning, letting developers ship smarter tools faster.

By closing the loop between feedback and training, Cursor can deliver software that adapts to real‑world usage without constant human oversight.

Early tests show higher code correctness and lower hallucination rates, hinting at a new productivity frontier.

Key Takeaways

Recursive training creates a feedback loop between model performance and data generation.
Inner loops focus on high‑quality task‑specific evals; outer loops expand and refresh the training set.
User feedback directly shapes the next generation of model behavior.
SpaceX’s compute resources enable training on increasingly complex problems.
Agent automation speeds up research cycles, letting AI help build AI.
Fixing reward hacking improves reliability for real‑world software.
This approach could become the standard for AI‑native development tools.