Aman Sanger

Co-founder of Anysphere, the company behind Cursor. Leads ML engineering: Cursor Tab’s sparse MoE architecture, speculative edits for low-latency inference, the Apply model, and the custom model ensemble strategy. Key technical voice on model selection and inference optimisation.

Background

Switched from Vim to VS Code when GitHub Copilot launched in 2021, recognising it as the “killer app for LMs.” Works on the ML systems behind Cursor: custom model ensemble (Tab prediction model, Apply model, retrieval model) alongside frontier models (Claude Sonnet for chat). Built Cursor Tab’s speculative edits variant of speculative decoding to achieve low latency despite pre-fill-heavy prompts.

Appearances in this wiki

Key positions

• Cursor is not a GPT wrapper: custom specialised models outperform frontier models on specific subtasks (Tab, Apply, retrieval)
• Claude Sonnet is “net best” for general-purpose coding chat: speed, code editing, context handling, and reasoning
• Being months ahead in AI capability integration produces large product differentiation
• Frontier models degrade when pushed beyond their training distribution — domain-specific fine-tuning is necessary
• Bugs underrepresented in training data; synthetic bug injection can fix this