Cursor's development process is heavily reliant on internal use. New features, such as their AI agent, undergo multiple internal prototypes and are only shipped when the team finds them indispensable for their own daily coding tasks.
The conversation explores the future of coding as a collaborative process between humans and AI. The vision is to move beyond simple code completion to a higher-level interface where developers edit pseudocode or architectural plans, while the AI handles the low-level implementation details.
Cursor employs a sophisticated multi-model strategy, carefully selecting different LLMs for specific tasks based on their unique strengths. For instance, they find Anthropic's Sonnet ideal for fast, interactive agentic tasks due to its speed and coherence, while also running custom, cost-effective models like DeepSeek V2 at massive scale for code completion.
To power its AI features, Cursor has built a robust, large-scale infrastructure. This includes a custom code completion model handling over 100 million daily requests, an indexing system processing billions of files, and a disaggregated vector database architecture using Turbopuffer and S3.
A key challenge for AI coding assistants is understanding the high-level architecture and implicit rules of a codebase. Cursor is exploring novel solutions, such as a `readme.ai.md` file to provide explicit context, and systems that automatically extract and prune rules from developer activity.
Keep pulling the thread on Sualeh Asif.