Michael Truell (co-founder and CEO)01 May 2025

The rise of Cursor: The $300M ARR AI tool that engineers can’t stop using

7Frameworks

15Insights

Episode overview

Lenny Rachitsky interviews Anysphere co-founder and CEO Michael Truell about Cursor’s origin, rapid growth, and ambition to create a higher-level form of programming beyond traditional code. Truell argues that engineers will increasingly become “logic designers,” specifying software intent while remaining in control, and explains why Cursor combines foundation models with fast custom models. He also discusses product focus, defensibility, hiring lessons, and his view that AI’s transformation of knowledge work will be consequential but unfold over decades.

Key ideas

Cursor aims to move software creation toward concise, human-readable specifications of intent while keeping people in the driver’s seat.
Engineering work may shift from implementing details toward logic design, product taste, and deciding exactly how software should behave.
Cursor began after GitHub Copilot and AI scaling research convinced the founders that useful AI products would improve rapidly; an initial mechanical-engineering idea was abandoned.
The team favors breaking AI coding tasks into small, reviewable steps, while using safe side projects to explore model limits.
Cursor trains specialized models for low-latency autocomplete, codebase retrieval, and converting foundation-model sketches into complete code changes.
Truell sees continuous product and research improvement—not permanent lock-in—as the main defense in a market with a high technical ceiling.
The company believes it initially hired too slowly and now relies heavily on targeted recruiting and two-day work-test projects.
Truell expects AI’s impact to exceed the internet’s but to arrive through a long, multi-decade transformation rather than an immediate leap to full automation.

Transcript available · source text is retained privately and is not published

Frameworks in this episode

Productivity4 steps

Chop-It-Up AI Collaboration Loop

Don't hand AI one giant spec — specify a little, review a little, repeat in tight loops.

Read →

Innovation4 steps

Complement-the-Frontier Model Strategy

Don't rebuild foundation models — train small specialty models that attack their weaknesses in speed, cost, and niche tasks.

Read →

Entrepreneurship4 steps

Dogfood-Driven Realism

Be the end user, use your product intensely daily, and never ship anything that isn't useful to you.

Read →

Self-Mastery4 steps

Fall-On-Your-Face Taste Calibration

Deliberately push AI past its limits in a safe environment to build a gut feel for what it can do.

Read →

Strategy4 steps

High-Ceiling Market Test

Attack markets with a high ceiling and easy switching; incumbents win only where there's little left to build.

Read →

Strategy4 steps

Knowledge-Work End-State Idea Generation

Pick an area of knowledge work and reason forward to its end state as AI matures — then build for that.

Read →

Leadership4 steps

Two-Day Work-Test Hire

Bring finalists on-site for two days to do a real end-to-end project — it screens skill and fit at once.

Read →

People & resources mentioned

Attributed to the moment in the episode. Timestamps are approximate.

People · 5

Michael TruellMentionsMichael is co-founder and CEO of Any Sphere, the company behind Cursor.
“Today my guest is Michael Tru. Michael is co-founder and CEO of Any Sphere, the company behind Cursor.”
▶ Watch at ~0:50approx.
Lex FridmanMentionsthis reminds me of a story I just listened to on Lex Friedman's podcast
“The only other podcast he's ever done is Lex Freedman.”
▶ Watch at ~1:30approx.
Lenny RachitskyMentionstoday we've got another very special compilation episode something I've been pulling on more and more with the podcast and the newsletter
“Michael, thank you so much for being here. Welcome to the podcast.”
▶ Watch at ~4:00approx.
Andrew LuoMentionsAndrew is CEO of One Schkeema, one of our longtime podcast sponsors.
“I'm excited to have Andrew Luo joining us today. Andrew is CEO of OneKema, one of our longtime podcast sponsors.”
▶ Watch at ~30:30approx.
Kevin WeilMentionsWe got connected through Kevin Wheel, who is former CPO at OpenAI, now head of science at OpenAI.
“I just had Kevin Wheel on the podcast, CPO of OpenAI, and he calls this the ensemble of models.”
▶ Watch at ~37:30approx.

Resources · 40

AnysphereMentionscompany
“Michael is co-founder and CEO of Any Sphere, the company behind Cursor.”
▶ Watch at ~0:50approx.
MITMentionscompany
“He studied computer science and math at MIT, did AI research at MIT and Google”
▶ Watch at ~1:00approx.
GoogleMentionscompany · Google
“He studied computer science and math at MIT, did AI research at MIT and Google”
▶ Watch at ~1:00approx.
YouTubeRecommendswebsite · Google
“don't forget to subscribe and follow it in your favorite podcasting app or YouTube.”
▶ Watch at ~1:30approx.
GranolaRecommendssoftware · Granola
“you get a year free of Perplexity, Linear, Superhuman, Notion, and Granola.”
▶ Watch at ~2:00approx.
SnowflakeMentionscompany · Snowflake Inc.
“EPO is a next generation AB testing and feature management platform built by alums of Airbnb and Snowflake”
▶ Watch at ~2:00approx.
Lenny's NewsletterRecommendsnewsletter · Lenny Rachitsky
“Check it out at lenny'snewsletter.com and click bundle.”
▶ Watch at ~2:00approx.
Lenny's NewsletterCoinednewsletter · Lenny Rachitsky
“if you become an annual subscriber of my newsletter”
▶ Watch at ~2:00approx.
PerplexityRecommendssoftware · Perplexity AI
“if you become an annual subscriber of my newsletter, you get a year free of Perplexity, Linear, Superhuman, Notion, and Granola.”
▶ Watch at ~2:00approx.
LinearRecommendssoftware · Linear
“you get a year free of Perplexity, Linear, Superhuman, Notion, and Granola.”
▶ Watch at ~2:00approx.
SuperhumanRecommendssoftware · Superhuman
“you get a year free of Perplexity, Linear, Superhuman, Notion, and Granola.”
▶ Watch at ~2:00approx.
NotionRecommendssoftware · Notion Labs
“you get a year free of Perplexity, Linear, Superhuman, Notion, and Granola.”
▶ Watch at ~2:00approx.
AirbnbUsescompany · Brian Chesky, Joe Gebbia, and Nathan Blecharczyk
“When I was at Airbnb, one of the things that I loved most was our experimentation platform”
▶ Watch at ~2:30approx.
EppoRecommendssoftware · Eppo
“Check out EPO at geto.com/lenny and 10x your experiment velocity. That's get epo.com/lenny.”
▶ Watch at ~3:00approx.
VantaRecommendssoftware · Christina Cacioppo
“Get $1,000 off Vanta when you go to vanta.com/lenny.”
▶ Watch at ~4:00approx.
CursorCoinedsoftware · Anysphere
“Our goal with Kerser is to invent sort of a a new type of programming.”
▶ Watch at ~4:30approx.
FigmaUsessoftware · Figma
“You can kind of gesture at it with Figma and you can gesture at it with writing down notes.”
▶ Watch at ~9:00approx.
GitHub CopilotUsessoftware · GitHub
“being really excited by using the the the first beta version of the Git Co-ilot.”
▶ Watch at ~13:00approx.
OpenAIMentionscompany
“the series of scaling on papers coming out of OpenAI and other places”
▶ Watch at ~13:30approx.
ChromiumMentionssoftware · Google
“We uh now we we use uh VS code kind of as a base like many browsers use chromium as a base”
▶ Watch at ~24:30approx.
Visual Studio CodeUsessoftware · Microsoft
“We uh now we we use uh VS code kind of as a base like many browsers use chromium as a base”
▶ Watch at ~24:30approx.
CursorUsessoftware · Anysphere
“it was after maybe five weeks that we were living on the editor full-time”
▶ Watch at ~25:30approx.
RampMentionscompany · Eric Glyman and Karim Atiyeh
“I know that you work with some of my favorite companies like RAMP and Vanza and Watershed.”
▶ Watch at ~30:30approx.
WatershedMentionscompany
“I know that you work with some of my favorite companies like RAMP and Vanza and Watershed.”
▶ Watch at ~30:30approx.
OneSchema FileFeeds 2.0Coinedsoftware · OneSchema
“So, we just launched the 2.0 of one schema file feeds. We've rebuilt it from the ground up with AI.”
▶ Watch at ~30:30approx.
Amazon S3Usessoftware · Amazon Web Services
“We support all of the trickiest file integrations, SFTP, S3, and even email.”
▶ Watch at ~31:00approx.
OneSchemaRecommendssoftware · Christina Gilbert
“If you want to learn more, head on over to onskeema.co.”
▶ Watch at ~31:30approx.
GPT-4Mentionssoftware · OpenAI
“we had done the calculation for what it took to to train GP4”
▶ Watch at ~32:00approx.
GmailMentionssoftware · Google
“lots of people are familiar with, you know, Gmail's autocomplete”
▶ Watch at ~35:00approx.
GeminiUsessoftware · Google
“another set of places where user models are to to help things like sonnet or gemini or gbt”
▶ Watch at ~36:30approx.
Claude SonnetUsessoftware · Anthropic
“another set of places where user models are to to help things like sonnet or gemini or gbt”
▶ Watch at ~36:30approx.
LlamaMentionssoftware · Meta
“are they based on like llama and things like that just open source models that you guys plug into and build on?”
▶ Watch at ~37:30approx.
SalesforceMentionscompany · Salesforce
“creating lock in and things like that where they're just for like Salesforce where it's just contracts with the entire company”
▶ Watch at ~41:00approx.
MicrosoftMentionscompany · Bill Gates and Paul Allen
“it's interesting that Microsoft was actually like right at this like at the center of this first with an amazing product.”
▶ Watch at ~44:00approx.
GPT-3Mentionssoftware · OpenAI
“this is GFD3 you know instruct doesn't exist”
▶ Watch at ~1:01:00approx.
DALL-EMentionssoftware · OpenAI
“instruct doesn't exist there's no doll there's no stable diffusion”
▶ Watch at ~1:01:00approx.
Stable DiffusionMentionssoftware · Stability AI
“instruct doesn't exist there's no doll there's no stable diffusion”
▶ Watch at ~1:01:00approx.
Apple PodcastsRecommendssoftware · Apple Inc.
“you can subscribe to the show on Apple Podcasts, Spotify or your favorite podcast app.”
▶ Watch at ~1:10:30approx.
SpotifyRecommendssoftware · Spotify
“you can subscribe to the show on Apple Podcasts, Spotify or your favorite podcast app.”
▶ Watch at ~1:10:30approx.
Lenny's PodcastRecommendspodcast · Lenny Rachitsky
“You can find all past episodes or learn more about the show at lennispodcast.com.”
▶ Watch at ~1:11:00approx.

Spot an error or want something removed? Request a correction or removal.

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 4

Hot Take08:30

Engineers become "logic designers" — taste replaces carefulness

Truell argues the most valuable future skill isn't visual design taste but having the right idea for what should be built and how the logic should work. As AI handles more of the 'how,' being an engineer will feel like being a logic designer focused on the 'what.' He predicts the field shifts away from the currently critical skill of carefulness toward taste.

Taste = having the right idea for what should be built, not just visuals
Engineering becomes specifying intent for how everything should work
The role moves from the 'how' toward the 'what'
Carefulness is critical today but will matter less as models improve

“I think that more and more being being an engineer will start to feel like being a logic designer.”
Michael Truell · 09:30

“we'll move a little bit from carefulness and a little bit more towards taste.”
Michael Truell · 10:30

#skills#taste#future-of-work#engineering

Hot Take10:30

Why vibe coding hits a wall you can't change

Asked about vibe coding, Truell calls it a controversial state of creation where you generate lots of code without understanding the details. The problem: by not understanding what's under the hood you quickly build something big enough that you can't change it, and the AI ends up making decisions you have no control over.

Vibe coding = generating lots of code without understanding the details
You get limited once the codebase grows too big to change
The AI makes decisions the maker can't control
Giving makers real control over the details is the open problem

“I think that v coding right now describes um exactly kind of this this state of creation that uh is pretty controversial”
Michael Truell · 10:30

#vibe-coding#ai-tools#control#limitations

Hot Take38:30

No moat but the best product: a market of many leapfrogs

On defensibility, Truell argues the ceiling is so high that no amount of entrenchment protects you — there are many leapfrogs still to come. He compares it to search engines circa 1999 and the rise of personal/mini computers, markets where an incremental smart-person-hour kept delivering value for a long time. The sad truth for incumbents but great news for the world: it's about consistently building the best thing, not lock-in.

The ceiling is so high that entrenchment can be leapfrogged
Analogous to search engines ~1999 and PC/mini-computer eras
Consumer-like dynamic: be the best product, not contract lock-in
R&D economies of scale in the right direction are the real defensibility

“maybe the sad sad truth for people like us but then like the amazing truth for the world is I think that there are many…”
Michael Truell · 40:30

#moats#defensibility#competition#ai-markets

Hot Take1:02:30

AI will be more consequential than the internet — over decades

On what people most misunderstand about where AI is heading, Truell rejects both extremes (all-at-once vs pure hype). He believes we're in a technology shift more consequential than the internet and than anything since the advent of computers — but that it will play out over multiple decades, with many independent problems on both the science and product sides needing to be solved.

Both 'happens instantly' and 'it's all hype' framings are wrong
More consequential than the internet, and than anything since computers
It's a multi-decade shift, not a sudden one
Progress requires knocking down many independent science and product problems

“I think it's going to be more consequential than the internet.”
Michael Truell · 1:03:00

#ai-future#predictions#technology-shift#hot-take

Explainer· 3

Explainer06:30

"After code": software logic you edit in English, not millions of lines

Truell lays out Cursor's core thesis that programming will move to a higher level of abstraction where you edit the logic of your software in something closer to English than TypeScript or Python. Instead of impenetrable millions of lines of code, you'd navigate a terser, human-readable representation of intent. He frames it as a gradual evolution of programming languages toward pseudo code, with humans still in the driver's seat.

Goal is a new, much higher-level way to build software distilled to specifying intent
Future code looks more like English / pseudo code than formal languages
Replaces impenetrable millions of lines with a terser, navigable representation
The human stays in control with a fast iteration loop, not slow background work

“it won't be kind of the the impenetrable millions of lines of code um it'll instead be something that's like much turser and easier to…”
Michael Truell · 07:00

“you can imagine in kind of an evolution of programming language towards pseudo code”
Michael Truell · 07:00

#future-of-coding#abstraction#cursor#ai-tools

Explainer34:30

Inside Cursor's model stack: a 300ms custom autocomplete

Truell details Cursor's 'ensemble of models' architecture. A custom autocomplete model no foundation model could serve on cost or speed predicts the next set of edits across multiple files and must return a completion within 300 milliseconds. Other custom models sit on the input side (a mini-Google search finding relevant codebase parts to show big models like Sonnet/Gemini/GPT) and the output side (fast models turning high-level changes into full code diffs).

Code is uniquely predictable — the next 5-30 minutes can be clear from context
The autocomplete model must complete within 300 milliseconds, run on every keystroke
It's trained to predict diffs (deletes and adds), not just next tokens
Input-side models act like a mini Google search over the codebase
Output-side fast models expand big models' sketches into full code diffs

“They need to give you a completion within 300 milliseconds.”
Michael Truell · 35:30

“you can kind of think about this as like a mini Google search that's specifically built for finding the, you know, relevant parts of a…”
Michael Truell · 36:30

#architecture#autocomplete#ensemble#custom-models

Explainer49:30

Junior vs senior engineers fall into opposite AI traps

Both cohorts benefit from Cursor, but they fail in opposite ways. Junior engineers go too wholesale, relying on AI for everything — which doesn't yet work end-to-end on professional tools in large codebases. Senior engineers, on average, underrate what AI can do for them and stick to their existing workflows. Truell likens it to Goldilocks: expect too much versus expect too little.

Both junior and senior engineers get big benefits from the tools
Juniors over-rely on AI, which doesn't work end-to-end professionally yet
Seniors on average underrate AI and stick to old workflows
Dev-experience teams (often very senior) are among the most boundary-pushing

“on average as a group the senior engineers underrate what AI can do for them and stick to their existing workflows.”
Michael Truell · 50:30

#engineers#adoption#juniors-vs-seniors#antipatterns

Story· 3

Story12:30

Cursor spent 4 months building for mechanical engineers first

Cursor began as a solution in search of a problem, sparked by the first GitHub Copilot beta and OpenAI's scaling-laws papers. The founders ran an idea-generation exercise to pick a 'boring, uncompetitive' area of knowledge work and spent about four months building tools for mechanical engineers before realizing they weren't passionate about it and pivoted back to programming.

Two triggers: the first Copilot beta and OpenAI scaling-law papers
They deliberately sought a sleepy, uncompetitive area of knowledge work
First bet was automating/augmenting mechanical engineering for ~4 months
Founders weren't mechanical engineers — a 'blind man and the elephant' problem
Came to their senses and returned to coding, seeing rivals lacked ambition

“Carer kind of started as a solution search of a problem.”
Michael Truell · 13:00

“there was a period of you know four months to begin with where we were actually working on a very different idea which was helping…”
Michael Truell · 15:30

#origin-story#cursor#pivot#startups

Story32:00

The counterintuitive lesson: Cursor builds its own models

The most counterintuitive thing Truell learned is that they never expected to do their own model development — they'd calculated that training a GPT-4 was out of reach. Yet Cursor now does a ton of model development, and every magic moment in the product involves a custom model. The trick was picking spots that complement the big foundation models' weaknesses rather than reinventing pre-training.

They assumed they'd never train their own models
Replicating GPT-4-scale pre-training was infeasible and the wrong focus
Every magic moment in Cursor now involves a custom model
Strategy: target foundation models' weaknesses, don't reinvent the wheel

“we we definitely didn't expect to be doing any of our own model development when we started.”
Michael Truell · 32:00

“at this point every magic moment in Kerserv involves a custom model in some way”
Michael Truell · 33:00

#custom-models#cursor#counterintuitive#ai

Story51:30

Cursor's hiring lesson: they hired too slow, and the 2-day work test

Counter to the common 'we hired too fast' regret, Truell says Cursor actually hired too slow early on, fussing over getting a world-class team more than almost anything but the product. They eventually recruited world-class people over years. Their signature filter is a two-day onsite work-test project in their codebase — something they thought wouldn't scale but which had surprising staying power and doubles as a 'do you want to be around this person' test.

Contrarian take: they hired too slow, not too fast
Recruiting world-class people sometimes took many years
They over-weighted young, high-credential, well-known-school profiles at first
The two-day onsite work-test project became a load-bearing hiring signal
It reveals real work product and whether you want to be around the person

“many people you hear hired too fast. I think we actually hired too slow to begin with.”
Michael Truell · 53:00

“this two-day work test, which we thought would not scale past a few people, has been has had surprising staying power.”
Michael Truell · 55:30

#hiring#team-building#interviewing#startups

Q&A· 2

Q&A18:30

Why an IDE, not a pure model or agent: humans in the driver's seat

Truell explains Cursor chose to build a full IDE rather than just a coding model or an end-to-end agent because they want humans to control every decision in what they build. Being 'intense realists' about today's tech — models have massive issues — pushed them to keep humans in control. Owning the whole application is necessary because the form factor of programming will change so much.

Model-only and agent-only players assume AI makes all the decisions
Cursor wants humans controlling every decision in the end tool
Daily dogfooding instills realism about where the tech actually is
Owning the whole app is required if the UI/form factor will change a lot

“we care about giving humans control over all the decisions um in kind of the end tool that they're building”
Michael Truell · 19:00

“our our product development process was really about dog fooding and using the tool intensely every day and we we never wanted to ship anything…”
Michael Truell · 20:00

#product-strategy#ide#human-in-the-loop#cursor

Q&A44:00

Why Microsoft's Copilot fell behind despite the lead

Asked why Microsoft's Copilot — an early inspiration for Cursor — seems to be falling behind, Truell cites both structural and historical reasons. Structurally, this market isn't friendly to incumbents: switching is easy and the ROI between products is large, so people buy the innovative tool rather than the bundled one. Historically, the original Copilot team has largely moved on and coordinating across Microsoft's departments is hard.

This market isn't friendly to incumbents — bundling advantages don't apply
Users can easily try tools and switch to whichever is best
Incumbent-friendly markets are commoditized or have high switching costs — this isn't one
The original Copilot team largely left; cross-department coordination is hard

“I think that this is a market that's not super friendly to incumbents”
Michael Truell · 44:30

#microsoft#copilot#incumbents#competition

Takeaway· 3

Takeaway16:30

Attack the hottest space if incumbents lack ambition

Contrary to the usual 'go after a boring industry' advice, Cursor went into the crowded AI-coding space because the existing players weren't being sufficiently ambitious about where things were heading. Truell's lesson: even if it looks too late and a GitHub Copilot exists, a high enough ceiling plus a flaw in incumbents' approach means a big opportunity remains.

Existing coding tools felt disconnected and under-ambitious about the future
A high ceiling means the best tool in a field still has far more to build
Leapfrogs are possible where the ceiling is high — unique to AI right now
'Too late' can be wrong if rivals under-reach

“it felt like they weren't being sufficiently ambitious about um where everything was going to go in the future and how kind of all of…”
Michael Truell · 16:30

#strategy#competition#ambition#startups

Takeaway27:00

The secret to 0-to-$100M ARR: sustained paranoia, not the launch

Truell downplays the initial three-month launch — that first version wasn't very good. The real driver of Cursor's historic growth was a sustained paranoia that the tool could always be better, and continuously improving it rather than over-rotating on the initial push. Early on they let normal fires (sales, marketing) burn and just made the product great.

The three-month first version wasn't very good
Growth came from continuous improvement, not the initial launch
They deliberately let sales and marketing fires burn early on
Focus and picking the right things to build was the hard part

“the three-month version wasn't very good”
Michael Truell · 27:00

“a sustained paranoia about you know there are all of these ways in which this this thing could get better.”
Michael Truell · 27:00

#growth#product#focus#startups

Takeaway46:00

Two tips to master Cursor: chop tasks up and go for broke safely

Truell's advice for new users: don't hand the model one giant task and hope — that's a recipe for disaster. Instead, chop work into small bits, specifying and reviewing incrementally. Second, on a safe side project, explicitly try to fall on your face and discover the model's limits, because most people underestimate what the models can do.

One-shot giant prompts are a recipe for disaster right now
Chop tasks up: specify a little, get a little, review, repeat
On a side project, deliberately push to find where the model breaks
Most people give AI too little credit and underrate its abilities

“I think that will be a little bit of a recipe for disaster right now.”
Michael Truell · 47:30

“I would encourage people to explicitly try to fall on their face and try to discover the limits of uh what these models can do”
Michael Truell · 48:00

#how-to#cursor#workflow#prompting

Key ideas

Frameworks in this episode

Chop-It-Up AI Collaboration Loop

Complement-the-Frontier Model Strategy

Dogfood-Driven Realism

Fall-On-Your-Face Taste Calibration

High-Ceiling Market Test

Knowledge-Work End-State Idea Generation

Two-Day Work-Test Hire

People & resources mentioned

People · 5

Resources · 40

Insights & moments

◆Hot Take· 4

Engineers become "logic designers" — taste replaces carefulness

Why vibe coding hits a wall you can't change

No moat but the best product: a market of many leapfrogs

AI will be more consequential than the internet — over decades

✶Explainer· 3

"After code": software logic you edit in English, not millions of lines

Inside Cursor's model stack: a 300ms custom autocomplete

Junior vs senior engineers fall into opposite AI traps

❝Story· 3

Cursor spent 4 months building for mechanical engineers first

The counterintuitive lesson: Cursor builds its own models

Cursor's hiring lesson: they hired too slow, and the 2-day work test

?Q&A· 2

Why an IDE, not a pure model or agent: humans in the driver's seat

Why Microsoft's Copilot fell behind despite the lead

▲Takeaway· 3

Attack the hottest space if incumbents lack ambition

The secret to 0-to-$100M ARR: sustained paranoia, not the launch

Two tips to master Cursor: chop tasks up and go for broke safely

Hot Take· 4

Explainer· 3

Story· 3

Q&A· 2

Takeaway· 3