Lenny's Podcast · Apr 20Building a magical AI code editor used by over 1m developers in 4 months: Inside Windsurf
Over a million developers tried it in four months, with hundreds of thousands active monthly. It has about 160 people, over 50 engineers and an 80-plus person go-to-market team, and its own IDE tripled autocomplete acceptance versus VS Code plugins. It trains retrieval and edit models in-house, uses Sonnet for planning, and sees 30 to 40% productivity gains.
Tips
- Build your own surface when host APIs cap the user experience▶ 15:21Autocomplete acceptance rate tripled after porting to Windsurf with the same ML models.
- Hire only when a function is underwater, not ahead of need▶ 25:44Windsurf has about 160 people, with over 50 in engineering and over 80 in go-to-market.
- Use understaffing to force ruthless prioritization▶ 27:05
- Reward engineers for big projects done with few people▶ 30:22
- Be upfront that the job demands hard work when hiring▶ 31:36Engineering pass rate is about 0.6%, and the take-home filters roughly another 10 to 15X.
- Invest in enterprise sales early if big companies come to you▶ 37:30The go-to-market team is now over 80 people.
- Differentiate on understanding very large enterprise code bases▶ 38:05A 100 million line code base is over 1.5 billion tokens, orders of magnitude beyond the largest context windows.
- Support the IDEs developers already use, including JetBrains▶ 39:40JetBrains IDEs are used by 70 to 80% of Java developers.
- Prompt AI coding tools with small, explicit changes first▶ 43:25Autocomplete is tolerable even when only 30% of suggestions are accepted because users know when to rely on it.
- Let non-technical staff build internal tools instead of buying SaaS▶ 50:05They saved over half a million dollars of SaaS products they were going to buy.
- Expect niche vertical SaaS to be undercut by custom-built tools▶ 51:55
- Show AI-driven code review and track user edits alongside agent edits▶ 47:00
- Combine frontier models for planning with in-house models for retrieval and edits▶ 54:25
- Train on proprietary behavior data that frontier models have not seen▶ 56:27Tens of millions of pieces of user feedback arrive every hour.
- Skip pure product managers when your team are the users▶ 1:01:10Three product strategy people versus over 50 engineers.
- Do not assume AI writing 90% of code means 10X productivity▶ 1:04:20Windsurf sees over 30, maybe close to 40% productivity improvements.
- Cannibalize your own product every six to 12 months▶ 1:07:25
- Pivot fast when your core assumption is proven wrong▶ 9:20They were free cash flow positive with a couple million in revenue and about 10,000 GPUs managed when they pivoted.
- Get hands-on with AI tools and ship changes across role boundaries▶ 1:11:23
