AI summary by TwinMind

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

Y Combinator

37m 29s

Company Vision and Goal

  • End goal: replace coding with something much better
  • Decade ahead will magnify ability to build for everyone
  • Philosophy: building company is hard, work on what excites you most
  • Vision: invent new type of programming where you describe what you want and it gets built

Cursor's Current State and Growth

  • $9 billion valuation, fastest growing startups of all time
  • 100 million ARR reached in just 20 months after launching
  • Over half billion model calls per day on own inference
  • AI writes 40-50% of lines of code produced within Cursor
  • Users are professional programmers who use AI 4-5 days per week

Evolution of Programming with AI

  • Current phase: AI as helper - delegating tasks and looking over shoulder (tab form factor)
  • Next 6 months to year: make both form factors order of magnitude more useful
  • Tab form factor can predict 10-15-20 minutes of work ahead
  • Agent form factor for delegating end-to-end tasks
  • Goal: 25-30% of professional development can lean entirely on these without looking at code

Technical Challenges and Bottlenecks

  • Context window limitations: 10 million lines of code = ~100 million tokens
  • Need models that can ingest, be cost effective, and pay attention to full context
  • Continual learning problem: knowing organizational context, past attempts, coworkers
  • Long time horizon tasks: progress has gone from seconds to ~1 hour recently
  • Computer use capabilities needed: running code, checking logs, interfacing with tools
  • Modality challenges: need to see and interact with actual output

Product Philosophy and Interface Design

  • Even with superhuman coding AI, text box interface is imprecise for humans
  • Need evolution of programming languages to higher level
  • Direct manipulation of UI: pointing at screen elements, adjusting values
  • Always need control over finest details (moving pixels, editing specific logic)
  • Higher level abstraction of written logic will be necessary

Irreplaceable Human Elements

  • Taste: defining what you actually want to build
  • Taste applies to visual aspects and non-visual logic
  • Current programming bundles product definition with implementation details
  • Human compilation step (translating ideas to for loops, if statements) will go away
  • Computers will fill in gaps and details
  • Taste for what's useful and what to build will never go away

Early Days and Founding Story

  • Founded 2022 by Michael Tru, Swale, Arvid, and Aman (all MIT)
  • Started with ambitious idea exercise, all found programming young
  • Early AI projects: robotic reinforcement learning, Google competitor using neural networks
  • Two key 2021 moments: GitHub Copilot (first viscerally useful AI product), OpenAI scaling laws research
  • Initial focus: pick knowledge work area, build product that evolves as AI matures

CAD Experiment and Pivot

  • Initially worked on mechanical engineering co-pilot for CAD (Solid Works, Fusion 360)
  • Trained 3D autocomplete models, predicted next geometry changes
  • Challenges: not excited about mechanical engineering, science wasn't ready for 3D
  • Orders of magnitude less CAD data than code on internet
  • Learned: should have worked at mechanical engineering company for 3 weeks instead of hundreds of user interviews

Technical Infrastructure Learning

  • Trained tens of billions of parameter models when few people were doing it
  • Forked Megatron LM, Microsoft Deep Speed, modified internals
  • Experience with large-scale training and inference proved valuable
  • Back then calculated Codex cost only ~$90-100k to train

Decision to Focus on Coding

  • Always excited about coding, had trepidation due to competition
  • Confidence came from: seeing slower progress than expected, believing all coding would flow through models in 5 years
  • Existing players didn't seem to aim for completely different type of coding
  • Phrase: "follow the line" - always plan for where scaling laws were heading

Key Product Decisions

  • Built editor instead of extension (non-obvious choice at time)
  • Came from belief all programming would flow through models
  • GitHub Copilot needed editor-level changes even for simple ghost text
  • Initially built editor from scratch, later switched to VS Code base
  • Got flack for editor decision but proved necessary

Growth and Metrics

  • Year in wilderness working on CAD precursor
  • 3 months from first line of code to public beta
  • Another year of iterating in public at small scale
  • 9 months to year of dialing in product and custom models before growth clicked
  • Key metric: paid power users using AI 4-5 days per week
  • Focus on professionals who build software for living

Hiring and Scaling Philosophy

  • Just co-founders for long time, then co-founders plus few people
  • Agonized over first 10 hires - going slow initially to go fast later
  • First hires act as immune system, maintain talent density
  • Looked for generalist polymaths: product-minded, commercially-minded, but trained models at scale
  • Need to break "don't grow headcount more than 50% year-over-year" rule
  • Two-day on-site as final hiring step: work on project, demo, get energy/passion sense

Technical Interviews and AI Tools

  • Still interview without AI (except autocomplete) for first technical screens
  • Programming without AI remains good test of skill and intelligence
  • Hire great programmers with no AI experience, teach them on job
  • Mine product insights from beginner's mind using tools first time

Maintaining Culture at Scale

  • Hire passionate people through rigorous process
  • Encourage bottom-up experimentation alongside top-down alignment
  • Section off engineering teams for independent experimentation
  • Founders still write code

Market Strategy and Moats

  • Market resembles search in late 90s, not enterprise software
  • High product ceiling, long runway for improvement
  • Distribution helps make product better (like search clickthrough data)
  • See user acceptance/rejection patterns, corrections after acceptance
  • Inspiration from consumer electronics (iPod/iPhone moments)
  • Focus on racing toward breakthrough moments fastest

Future Implications

  • Professional developers will become much more productive
  • Thousand-person software projects currently move very slowly due to existing logic weight
  • Next distributed training frameworks, databases, visual design tools will be faster to build
  • AI model labs bottlenecked on engineering capacity will improve
  • Many more pieces of niche software will exist
  • Example: biotech companies needing internal software won't need full engineering teams

Personal Excitement

  • Decade where ability to build will be magnified for everyone
  • Will become accessible to many more people beyond professional developers
  • Transformative potential for democratizing building capabilities

TwinMind

Get summaries like this for your own meetings

TwinMind takes notes in the background during meetings, lectures, and calls, then writes the summary for you.

Free plan to get started

No bot joins your calls

Phone, desktop & Chrome apps

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome