Fireside Chat, Interview, Conference Presentation, Keynote
Michael Truell: Building Cursor At 23, Taking On GitHub Copilot & Advice To Engineering Students
- Origin of Cursor's Vision: Founders concluded that while existing players were incrementally improving AI tools, no one was aggressively pursuing a future where "all of coding as we know it today gets automated" and software development "flows through models" entirely.
- Founder Backgrounds:
- Michael Truel (24 at time of founding) learned to code in middle school via Objective-C after a failed attempt to build a mobile game with his brother.
- High school projects included creating a viral high-score spoofing app and building robotic arms using reinforcement learning (RL) to play ping pong with minimal data points.
- The founding team (graduates of MIT) included collaborators with backgrounds in computer vision, security research (end-to-end encrypted messaging), and recommendation systems.
- Early Pivot History (2021–2022):
- Idea 1 (CAD/Mechanical Engineering): Attempted to build a Copilot for mechanical engineers using CAD files (SolidWorks/Fusion 360); abandoned due to market fragmentation, lack of exportable data, and rudimentary training infrastructure.
- Idea 2 (Encrypted Messaging): Built an end-to-end encrypted messaging system to solve metadata leakage issues (similar to Signal/WhatsApp); shelved due to scalability trade-offs and lack of B2B traction.
- Idea 3 (Security/Quants): Prototyped niche tools for security CVE detection and quantitative researchers; pivoted away after realizing a broader "best way to code with AI" opportunity.
- Strategic Decision to Build an Editor:
- Despite GitHub Copilot generating ~$100M in revenue in 2022, the team pivoted to code completion in late 2022 due to a lack of excitement for previous sectors and a belief that "nobody working on the space was really taking [a fully automated future] seriously."
- Product Launch Timeline: Roughly three months from first line of code to General Availability (GA).
- Technical Approach: Initially built a custom editor from scratch (using CodeMirror, custom SSH, and custom PINs); pivoted to base the product on VS Code after realizing the complexity of replicating a "feature-complete" editor.
- Model Strategy Evolution:
- Initially attempted proprietary model training (citing the original Codex paper cost of ~$100k) but pivoted to using API models for pragmatism in 2023.
- Later reintroduced custom models for specific product levers, such as "tab" next-edit predictions, to leverage product data for improvement.
- Growth Metrics & Adoption:
- 2023: Reached 1 million users through product-led growth and founder evangelism on Twitter (specifically regarding open-source models like Flan T5); maintained a team of under 10 people.
- 2024: Grew from 1 million to 100 million users; adoption spread through YC batches, where usage jumped from single-digit percentages in 2023 to ~80% of the batch in 2024.
- Growth Drivers: Rapid iteration on product features (codebase awareness, next-action prediction, speed, and action automation) created immediate compounding effects via word-of-mouth among professional developers.
- Future Outlook on Coding:
- Near-Term Vision: Code remains essential for professional engineers; AI will act as an "advanced colleague" or "advanced compiler" rather than a total replacement.
- Skill Requirements: Fundamentals of programming and computer science (like math) will remain critical for reading, reviewing, and editing AI-generated logic; the specific tech stacks studied in school are less crucial than the ability to learn.
- Long-Term Trend: AI is expected to be a transformative industry-wide effort over the next two decades, requiring independent capabilities to coalesce before software building is fully transformed.