OpenAI Jupyter Agents: training LLMs with code execution

AI Impact Summary

OpenAI is introducing Jupyter Agents, a new approach to training LLMs by enabling them to execute code directly within Jupyter Notebooks. This allows the model to reason and solve data analysis tasks more effectively, mimicking a human data scientist's workflow. The initial experiments with Qwen-3 Coder demonstrate promising results, particularly when combined with a simplified scaffolding and a focused training dataset derived from Kaggle notebooks, highlighting the potential for smaller models to compete with larger ones in agentic data science scenarios.