Summary

Ben Kunkel discusses the development of Zed 2, a specialized AI model for edit predictions in code editing. The model uses opt-in production data, a process of distillation involving frontier models, and innovative techniques like repair steps to generate accurate edit predictions. By fine-tuning prompts and running static evaluations, the team created a fast, task-specific model that can predict the next edit based on cursor position, code context, and other technical variables. The approach focuses on creating a small, specialized model that can quickly provide precise edit suggestions during coding.