Cursor CEO: Going Beyond Code, Superintelligent AI Agents, And Why Taste Still Matters

Cursor CEO Michael Truell argues taste is the last irreplaceable engineering skill as AI crosses 40–50% of lines written inside Cursor today.

• Cursor hit $100M ARR 20 months after launch and a $9B valuation; now runs over 500M model calls per day on own inference.
• AI currently writes 40–50% of lines of code produced inside Cursor, but professional devs still read every line — full delegation is still years away.
• Truell and co-founders initially built a CAD autocomplete startup; Codex cost roughly $90–100k to train, which they used to convince early investors.
• Going from GPT-3 to ChatGPT required only ~1% more training cost via RLHF — most of the leap was fine-tuning, not compute scale.
• GitHub Copilot’s origin: a tiger team wandered for nearly a year on wild product ideas before landing on simple autocomplete; even that required mainline VS Code changes.
• Cursor was built as a standalone editor, not an extension — a contrarian bet that all coding would eventually flow through AI and require a control UI.
• Key north-star metric was paid power users (AI used 4–5 days out of 7), not DAUs or MAUs — sustainability demanded the paid tier.
• Long-context reliability and continual learning remain the hard blockers to superhuman agents; max forward-progress time has grown from seconds to roughly one hour across recent models.

2025-06-11 · Watch on YouTube