Jun 11, 2025 · 37m · y-combinator
Cursor CEO: Going Beyond Code, Superintelligent AI Agents, And Why Taste Still Matters · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Y Combinator CEO Garry Tan interviews Michael Truell, Co-founder and CEO of Anysphere, exploring how Cursor achieved historic hypergrowth, the technical limits of autonomous AI agents, and why human taste remains essential in the future of software engineering.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The partners hold 18.3% of the talking time here. How this is scored →
The partners as informed peer 4.6 Guest teaching 4.7 Guest disagreement 1.3 The partners pushing back 1.2
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 2:15 Rejecting the viability of vibe coding
Michael firmly rejects the popular notion that vibe coding is suitable for production codebases, contrasting superficial prototyping with professional software development.
Hardest push from the partners ▶ 1:54 Challenging whether AI coding is already solvedGarry pushes Michael on claims that natural language coding is already a solved reality today rather than a future milestone.
Biggest teaching moment ▶ 6:08 Context windows vs continual learning realityMichael corrects the common assumption that infinite context windows solve code generation, detailing the lack of long-context data and organizational memory.
The partners hold their own ▶ 22:26 Connecting scaling conviction to Thielian contrarianismGarry synthesizes Michael's timeline with Sam Altman's doctrine on scaling laws and Peter Thiel's framework on contrarian secrets.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Reimagining Programming and the Core Mission of Cursor | 4 | 5 | 2 | 2 | Garry prompts Michael on whether prompt-to-code is already here today. Michael clarifies that while hobbyist 'vibe coding' works on small codebases, professional software engineering with millions of lines still requires reading code and handling n-th order effects. | |
| Analyzing Technical Bottlenecks and Context Window Limits | 5 | 4 | 1 | 2 | Garry asks if context window limitations and RAG are the primary technical bottlenecks when codebases reach millions of tokens. Michael explains the interplay between agent delegation, tab completion, and evolving the written logic UI. | |
| Overcoming Continual Learning and Computer Use Challenges | 5 | 6 | 2 | 1 | Michael educates Garry on why simply expanding context windows is insufficient, highlighting continual learning, time horizons, and computer-use modalities as true bottlenecks. Garry contributes insights on aesthetic reasoning in models. | |
| The Irreplaceable Value of Human Taste and Logic Design | 5 | 4 | 1 | 1 | Garry asks about the irreplaceable elements of engineering, referencing logic design. Michael explains how human taste separates high-level architecture from mundane human compilation steps. | |
| Origin Story: From MIT to Early CAD Experiments | 4 | 5 | 1 | 1 | Michael details Cursor's origin at MIT and their initial pivot away from 3D CAD modeling, explaining the data scarcity and CAD kernel hallucination challenges. | |
| Building High-Scale Model Infrastructure and Inference | 5 | 5 | 1 | 1 | Garry brings up custom model training and distributed cluster management. Michael describes their early experiences forking Megatron-LM and DeepSpeed to run tens of billions of parameters. | |
| The Decision to Pivot and Betting on Scaling Laws | 6 | 4 | 1 | 1 | Garry references Sam Altman and Peter Thiel on following scaling laws and non-consensus beliefs. Michael agrees and recounts calculating Codex training costs to pitch skeptical investors. | |
| Strategic Choice: Forking VS Code vs. Building an Extension | 5 | 6 | 1 | 1 | Michael explains the technical necessity of forking VS Code instead of building a simple extension, citing how even GitHub Copilot required deep editor modifications for ghost text. | |
| Scaling Operations, Product Dogfooding, and North Star Metrics | 4 | 4 | 1 | 1 | Garry queries Michael on North Star operational metrics. Michael clarifies they prioritized paid daily power users over vanity metrics to prevent optimizing for AI demo videos. | |
| Maintaining Talent Density and Evaluating Software Engineers | 4 | 5 | 2 | 1 | Garry asks if Cursor hires engineers based on AI tool proficiency. Michael surprises him by stating they ban AI tools during technical screening to evaluate raw reasoning and avoid bias. | |
| Preserving Hacker Culture and Structuring Bottom-Up Innovation | 4 | 4 | 1 | 1 | Garry asks about preserving hacker culture at a nine-billion-dollar valuation. Michael compares AI tooling moats to 90s web search distribution feedback loops rather than enterprise lock-in. |
Statements from this episode (26)
Tan: Cursor Reached $100M ARR in 20 Months at $9B Valuation
“They recently hit a nine billion dollar valuation and are one of the fastest growing startups of all time, reaching a hundred million dollars ARR just 20 months after launching.”
Truell: Cursor's ultimate goal is to replace traditional coding entirely
“Yeah, the goal with the company is to replace coding with something that's much better.”
Truell: Software creation will move to higher abstraction levels within a decade
“We think that over the next five to 10 years, it will be possible to invent a new way to build software that's higher level and more productive. That's just as still down to defining how you want the software to work and how you want the software to look.”
Truell: Developers on small codebases already let AI agents make all changes
“Already there, we see people kind of stepping up above the code to a higher level of abstraction and just asking essentially agents and AIs to make all the changes for them.”
Truell: 'Vibe coding' doesn't work for large professional codebases
“I think that the whole idea of kind of vibe coding or coding without really looking at the code and understanding it doesn't really work. There are lots of end order effects. You know, if you're dealing with millions of lines of code and dozens or hundreds of …”
Truell: AI writes 40% to 50% of all code produced inside Cursor
“You know, on average, we see about people using, you know, having AI write 40%, 50% of the lines of code produced within Cursor, but it's still a process of, you know, reading everything that comes out of the AI.”
Truell: AI coding tools will become 10x more useful within a year
“And I think that the game in the next six months to a year is to make both of those, you know, an order of magnitude more useful.”
Truell: AI labs face a severe shortage of high-quality long-context training data
“I think that there's a dearth of really good long context data available to the institutions that are training these models.”
Truell: Computer use and log inspection are essential for future AI coding
“But so computer using is kind of, is going to be important for the future of code. Being able to run the code, being able to look at data dog logs and interface with those tools that humans use.”
Truell: Chat text boxes are fundamentally too imprecise for controlling coding AI
“And then, you know, one thing I will note kind of harkening back to a last response is that even if you had something you could talk to that was human level at coding, Or faster and better, better than a human at coding, you know, sort of the skill of an entir…”
Truell: Human taste in product logic will remain irreplaceable despite AI advances
“We think that one thing that will be irreplaceable is taste. So just defining what do you actually want to build? People usually think about this when they're thinking about the visual aspects of software. I think it's also, there's a taste component to the no…”
Truell: AI will eliminate the manual 'human compilation' step of programming
“And so I think that more and more of that, like, human compilation step will go away, and computers will be able to kind of fill in the gaps, fill in the details.”
Truell: Major AI labs are currently bottlenecked by software engineering capacity
“The next AI model, which, you know, if you talk to the labs, largely they're bottlenecked on engineering capacity.”
Truell: Automated AI coding will cause an explosion of hyper-niche software
“I think that, you know, one second order effect too will be many more pieces of niche software will exist.”
Truell: Working on-site beats hundreds of user interviews for vertical AI startups
“And looking back on the time we spent on that, I think it would have been better up front to actually just go work at a company that was employing
Mechanical engineers for three weeks.”
Truell: Cursor processes over 500 million daily model calls on custom inference
“Now in Cursor, we do over half a billion model calls per day on our own inference.”
Truell: OpenAI's original Codex model cost roughly $100,000 to train
“Is we did the back of the envelope math for what Codex, the first coding model, costed to train. From my memory, it only cost about 90 K or a hundred K by our calculations.”
Truell: ChatGPT supposedly cost only 1% more to train than GPT-3
“But it rumored that to go from GPT-III, which, you know, had existed for a while and didn't You know, impressed some people, but was certainly not the breakout moment ChatGPT was. To ChatGPT, it was like a one percent increase in the training costs.”
Truell: AI models will mediate all programming, requiring complete UI control
“All of programming is going to flow through these models. It's going to look very different in the future. You're going to need to control the UI.”
Truell: Cursor originally built a custom editor before forking VS Code
“We actually initially started by building our own editor from scratch, obviously using lots of open source technology, but not, you know, basing it off of VS code, kind of like how browsers are based off of Chromium. It was a little bit more akin to building, …”
Truell: Cursor launched in three months, then iterated for a year pre-PMF
“I think that it was from lines of code to first public beta release, it was three months. But then there was this year of iterating in public at very small scale where we had, did not have lightning in the model.”
Truell: Fast-growing AI startups must break rules against 50% annual headcount growth
“And I think rules of thumb around don't grow headcount more than 50% or year over year. Yeah. Have to be broken, I think.”
Truell: Cursor's primary metric was paid users utilizing AI 4+ days weekly
“For us, the main activity metric we looked at, or the, yeah, the main top line metric we looked at revenue. We looked at paid power users measured by, are you using the AI four or five days a week out of seven days a week? And that was the number we were tryin…”
Truell: Optimizing for video demos is a dangerous trap for AI startups
“One of the siren songs involved in building AI products is optimizing for the demo. We were really nervous about optimizing for the demo because with AI, it's easy to kind of take a couple of examples and put together a video where, you know, it looks like you…”
Truell: Cursor bans AI tools during its initial software engineering technical screens
“So for interviewing, we actually still interview people without allowing them to use AI other than autocomplete for our first technical screens.”
Truell: AI coding tools mirror web search's product-improving distribution flywheel
“One of the things that characterize search, and I think also characterize our market, is distribution is really helpful for making the product better. And so if you have lots of people using your thing, you have an at scale business, you get a sense of where t…”