Cursor.so: The AI-first Code Editor — with Aman Sanger of Anysphere

Episode Summary

Executive Summary: This episode traces Aman Sanger’s path from failed CAD-focused AI experiments to Cursor, arguing that coding is the first domain where general-purpose LLMs are already transformative. He explains why CAD lacked enough data and spatial reasoning, why Cursor needed its own IDE rather than a VS Code extension, and how product design, retrieval, and agent workflows must evolve for AI-assisted software development.

Main Topics: From CAD AI to Cursor (Priority: 5/5): Aman recounts the original plan to build AI for CAD/design software, why it seemed promising, and why the team pivoted after discovering the limits of data, transfer learning, and spatial reasoning. Why Cursor is an IDE, not a plugin (Priority: 5/5): He argues that owning the full editor is necessary to redesign software development UX, avoid platform risk from VS Code, and support future features like inline edits and agentic workflows. Prompting, system rules, and code UX (Priority: 4/5): The discussion covers how Cursor uses system prompts, language-specific guidance, and chained model calls to improve code generation, diffs, and editing behavior. Retrieval and context management (Priority: 4/5): Aman explains Cursor’s context-loading approach, current embedding-based retrieval, and future work on better search, reranking, and differentiable search over codebases. Inference economics and model behavior (Priority: 4/5): He breaks down why prompt tokens and completion tokens behave differently, why quantization helps only in some regimes, and why model choice depends on workload shape. Agents, sandboxes, and the future of coding (Priority: 5/5): The conversation explores code-interpreter-style agents, test-driven validation, tool use inside IDEs, and the need for safe execution environments to prevent prompt injection. Company strategy and AI outlook (Priority: 3/5): Aman reflects on AnySphere/Cursor’s team, hiring, internal tooling, and broader beliefs about generalist models, long-context limits, and the future of AI product companies.

Key Arguments: CAD failed as an AI target because there was too little training data, poor transfer from language/code models, and weak spatial reasoning in current LLMs. Cursor needed to own the IDE because extensions are constrained by incumbent platform APIs and cannot support deeper UX changes like inline edits and agentic workflows. Generalist models like GPT-4 changed the company’s direction because they were already strong enough at coding to make a full software-development product viable. Prompt engineering for code is often about controlling format, reducing ambiguity, and chaining models to repair diffs or improve outputs. Retrieval quality matters as much as model quality; better embeddings, rerankers, and symbol-level search can materially improve code assistance. Agentic coding will work best when tasks are broken into small, auditable steps with tests, tool access, and visible outputs rather than giant autonomous PRs. Long-context models are useful but expensive; future systems may need more elegant memory/recurrent approaches rather than simply scaling attention. Quantization is not a universal speedup: it helps low-batch, memory-bound inference, but can stop helping once workloads become compute-bound or KV-cache-bound.

Data Points: HumanEval score for GPT-4: 85% - Aman says early GPT-4 access in the playground achieved about 85% on HumanEval, which was a major signal for Cursor’s direction. Best open-source model on HumanEval at the time: 30% - He contrasts GPT-4’s performance with the best open-source model then available. Code DaVinci 2 HumanEval score: 47% - Used as a comparison point for earlier code models before GPT-4. Codex training data: 100 billion tokens of Python code - Referenced while explaining why the original Codex approach did not transfer well to CAD. Available CAD data: At most 10 billion tokens - Aman argues CAD data was far too scarce to train a useful large model. Cursor team size: 5 people - He says the company is currently a five-person team based in San Francisco. Public GitHub code size: 92 terabytes - Mentioned in a discussion of code data scale, though Aman later notes this estimate may be low. Google monorepo size: 86 terabytes - Used to illustrate that private codebases can be highly valuable training data. Prompt length in one real-world use case: 55,400 tokens - The host cites this as a recent long-context workload that still worked well. Completion length in that use case: 1,800 tokens - Shows the difference between prompt-heavy and generation-heavy workloads. GPT-3.5 vs Llama 2 workload claim: GPT-3.5 cheaper for completion-heavy workloads - Aman argues that for generation-heavy serving, GPT-3.5 can be more economical than Llama 2. AlphaCode benchmark result: 50th percentile - Used as an example of how test-generation and search can improve coding performance. AlphaCode base model score: ~28% on HumanEval - Referenced to show the jump from a base model to a search-and-test system.

Pivotal Quotes: "I think the best way here is probably redesigning the entire system." — Aman Sanger: Explaining why AI-assisted CAD likely requires a new product architecture rather than a plugin. "You need to own the full IDE and completely redesign the flow of producing software." — Aman Sanger: His core thesis for why Cursor is a standalone editor rather than a VS Code extension. "I really think it's this kind of long-term memory piece." — Aman Sanger: On the most important unsolved AI problem: lifelong learning and persistent memory.

Implications: The episode suggests AI coding tools will move from autocomplete toward full IDEs, agent loops, and safer execution environments. For builders, the winners may be those who control UX, retrieval, and workflow—not just model access.