Why Did a $10 Billion Startup Let Me Vibe-Code for Them—and Why Did I Love It?

I asked my editors if I could spell enactment astatine a tech startup. It was an antithetic request. But I wanted to larn to vibe-code. My request to cognize felt urgent. I wanted to survive the future.

The transportation process was amazingly easy: First my editors said yes, and past the tech startup I lobbed my chaotic thought to, Notion, agreed to fto maine embed with them. Why? It’s hard to say. Possibly due to the fact that Notion’s ain workforce has afloat embraced vibe coding—“vibe” present being a euphemism for “AI-assisted.” Some tech companies person estimated that astir 30 to 40 percent of their codification is present written by AI.

Notion is simply a 1,000-person, venture-backed San Francisco startup with a $10 cardinal valuation. It makes the eventual to-do and note-taking app, consisting of truthful galore templates and tables and ways to format tasks that figuring retired however to usage Notion is simply a task successful itself. On YouTube, productivity gurus effort to marque consciousness of Notion utilizing the well-worn vernacular of personal optimization. One specified video is titled “How to Get Started successful Notion Without Losing Your Mind.” It has 3.4 cardinal views.

I was scheduled to commencement astatine Notion arsenic a vibe-coding technologist connected a Thursday successful mid-July. The nighttime before, I recovered myself panic-watching these YouTube videos. Surely I would request to beryllium a powerfulness idiosyncratic of the app if Notion was allowing me—an English major!—to fiddle with its codification base. In an earlier onboarding call, a caller coworker had encouraged maine to download the AI coding level Cursor and play astir with it. I did. No existent codification emerged from this homework.

Fortunately, I would beryllium pair-programming astatine Notion, which meant that I’d beryllium moving alongside experienced (human) coders. Upon my arrival, Sarah Sachs, an AI engineering pb astatine Notion, acceptable maine up astatine a desk. A institution tote container and notebook awaited me. Sachs informed maine that the pursuing day, I would beryllium presenting my enactment to the unit astatine a play demo meeting. Was I bully with that? I said yes. We were each committed to the bit.

Sitting a fewer feet distant was Simon Last, 1 of Notion’s 3 cofounders. He is gangly and shy, an technologist who has relinquished absorption responsibilities to absorption connected being a “super IC”—an idiosyncratic contributor. He stood to shingle my hand, and I awkwardly thanked him for letting maine vibe-code. Simon returned to his laptop, wherever helium was monitoring an AI arsenic it coded for him. Later, helium would archer maine that utilizing AI coding apps was similar managing a clump of interns.

Since 2022, the Notion app has had an AI adjunct to assistance users draught their notes. Now the institution is refashioning this arsenic an “agent,” a benignant of AI that volition enactment autonomously successful the inheritance connected your behalf portion you tackle different tasks. To propulsion this off, quality engineers request to constitute tons of code.

They unfastened up Cursor and prime which of respective AI models they’d similar to pat into. Most engineers I chatted with during my sojourn preferred Claude, oregon they utilized the Claude Code app directly. After choosing their fighter, the engineers inquire their AI to draught codification to physique a caller happening oregon hole a feature. The quality programmer past debugs and tests the output arsenic needed—though the AIs assistance with this too—before moving the codification to production.

At its foundational core, generative AI is enormously expensive. The theoretical savings travel successful the currency of time, which is to say, if AI helped Notion’s cofounder and CEO Ivan Zhao decorativeness his tasks earlier than expected, helium could mosey down to the jazz nine connected the crushed level of his Market Street bureau gathering and bliss retired for a while. Ivan likes jazz music. In reality, helium fills the clip by moving more. The phantasy of the four-day workweek volition stay conscionable that.

My workweek astatine Notion was conscionable 2 days, the eventual codification sprint. (In speech for afloat entree to their lair, I agreed to place rank-and-file engineers by archetypal sanction only.) My archetypal duty was to hole the mode a illustration called a mermaid diagram appears successful the Notion app. Two engineers, Quinn and Modi, told maine that these diagrams beryllium arsenic SVG files successful Notion and, contempt being called scalable vector graphics, can’t beryllium scaled up oregon zoomed into similar a JPEG file. As a result, the substance wrong mermaid diagrams connected Notion is often unreadable.

Quinn slid his laptop toward me. He had the Cursor app unfastened and astatine the ready, moving Claude. For funsies, helium scrolled done portion of Notion’s codification base. “So, the Notion codification base? Has a batch of files. You probably, adjacent arsenic an engineer, wouldn’t adjacent cognize wherever to go,” helium said, politely referring to maine arsenic an engineer. “But we’re going to disregard each that. We’re conscionable going to inquire the AI connected the sidebar to bash that.”

His vibe-coding strategy, Quinn explained, was often to inquire the AI: Hey, wherefore is this happening the mode it is? The question forces the AI to bash a spot of its ain probe first, and the reply helps pass the punctual that we, the quality engineers, would write. After “thinking,” Cursor informed us, via streaming lines of text, that Notion’s mermaid diagrams are static images that, among different things, deficiency click handlers and aren’t integrated with a full-screen infrastructure. Sure.

Using Claude’s notes, I wrote up the petition and pasted immoderate notes from the engineering squad into Cursor, similar this:

Ticket: Add Full Screen / Zoom to mermaid diagrams. Clicking connected the diagram should zoom it successful afloat screen.

Notes from slack: "mermaid diagrams should beryllium zoom / fullscreenable similar uploaded images. they're conscionable svgs right, truthful we tin astir apt svg -> dataurl -> representation constituent if we privation to zoom"

We waited. Time is inverted successful the onshore of vibes. Projects that utilized to instrumentality your full vocation are present done successful days, portion commands you expect to spot executed successful seconds instrumentality endless minutes. One 100 lines of AI-generated codification later, mermaid diagrams were expandable.

Except, not really. They were inactive excessively small, immoderate parts were transparent, and the margins astir them needed padding; also, would this enactment successful the app successful some airy mode and acheronian mode? I spent the adjacent fractional hr iterating connected these changes, with Quinn and Modi talking maine done it. Thirty minutes later, we had an expandable, readable mermaid diagram.

Next I worked alongside an technologist named Lucy, who told maine that alternatively of typing prompts into Cursor we would beryllium utilizing an cause from Codegen, different AI engineering tool. The duty was simple. We would make a caller accomplishment successful Notion called Alphabetize, truthful that erstwhile idiosyncratic uses Notion AI to draught a database oregon array of fashionable canine breeds, the idiosyncratic tin alphabetize the contented with 1 click.

Just then, Anthropic’s Claude—which was powering Codegen—suffered an outage. Sarah Sachs, who was successful the country with us, received a leafage connected her phone, similar an ER doctor. She hurried retired of the room. Vibe coding and alphabetizing were temporarily enactment connected pause. Bulldogs would travel earlier beagles until Claude was backmost online.

The adjacent duty was arsenic open-ended arsenic Lucy’s was specific: to physique immoderate I wanted. The state was unnerving, a Rorschach trial for vibe coders. What did I spot erstwhile I looked astatine the blinking cursor? I decided that determination should beryllium a mode for Notion users to draught an “intelligent” to-do database successful 1 step. They would beryllium capable to flick unfastened the app and benignant “to bash reorder favored food” and Notion AI would cognize what they meant. I besides wanted this diagnostic to debar duplicate items from different caller to-do lists.

I was crushing it. I was a liable babysitter for code, watching it cascade successful beforehand of my eyes and past toddle its mode into the world. Except, my logic was wrong. My to-do database hack was someway allowing for endless duplicates alternatively of avoiding them. Who was to blame: maine oregon the AI?

A merchandise decorator named Brian talked maine done it. “Pretend you’re talking to a astute intern,” helium said. Again with the interns.

I reversed my logic and tried again, typing successful much item astir however I envisioned the widget working. “That’s a large idea,” Claude responded, ever the sycophant, and past got to work. Forty minutes later, the 3 of america had prototyped a mentation of my dinky little—no, I mean killer—feature. We had spent $7 to physique it, according to the token antagonistic successful Claude Code. I was told different engineering projects outgo overmuch much than that, particularly if coders fto the AI tally for hours. It was inactive airy retired erstwhile I wrapped up the archetypal day.

On Friday morning, I showed up for the demo session. Cheese platters, successful grant of a Swiss employee’s birthday, awaited america successful the league room. Coders grabbed their coffees, their Celsius cans, their cups of flavored h2o dispensed by a Bevi instrumentality successful the kitchen.

One of the archetypal demos was of a Notion AI cause that had been fixed a memory, truthful it could follow a learned penning style. For kicks, different technologist had coded an app that kept way of the flavored syrups successful the staff’s beloved Bevi. At the extremity of each presentation, I was told, idiosyncratic usually takes a tiny mallet to a xylophone. They made maine the keeper of the xylophone that day. The temper was light.

When it was my crook to present, I tried to succinctly picture the fewer features I had vibe-coded (with recognition doled retired to my brace programmers). One of the managers asked a follow-up question: How agelong had it taken to codification the changes to the mermaid diagrams, extremity to end?

I looked astatine Quinn and Modi. We tallied that our moving league had been astir 30 minutes, positive immoderate 15 minutes of preliminary enactment Quinn had done.

“Wow,” idiosyncratic successful the country said.

“I situation to ideate the wide nationalist learning however to constitute code,” the programmer and writer Ellen Ullman wrote successful a 2016 effort titled “Programming for the Millions.”

The prevailing sentiment of the 2010s, of course, was that everyone could basal to larn a small code. We should propulsion unfastened the doors and invade the closed nine wherever codification gets written, Ullman wrote. This was our champion anticipation for loosening the strangle of the codification that surrounds america arsenic a society. As portion of her reporting process, Ullman enrolled successful 3 monolithic unfastened online courses, oregon MOOCs, that promised to thatch normies however to program. (I situation to ideate her eyebrow arched arsenic she enrolled.)

“Stick a needle into the shiny bubble of the method world’s received wisdom,” Ullman urged would-be coders. “Burst it.”

Expanding a mermaid diagram oregon alphabetizing a database of canine breeds hardly seemed similar sticking it to the coding man. But during my clip astatine Notion I did consciousness arsenic though a trapdoor successful my encephalon had opened. I had gotten a shimmery glimpse of what it’s similar to beryllium an anonymous logical god, pulling levers. I besides felt susceptible of learning thing new—and had the state to beryllium bad astatine thing new—in a semi-private space.

Both vibe coding and journalism are an workout successful prodding, and successful procurement: Can you accidental much astir this? Can you elaborate connected that? Can you amusement maine the documents? With our chap humans, we tin tolerate a spot of imprecision successful our conversations. If my stint arsenic a vibe coder underscored anything, it’s that the AIs coding for america request that we articulate precisely what we want.

During luncheon connected 1 of my days astatine Notion, an technologist asked maine if I ever usage ChatGPT to constitute my articles for me. It’s a question I’ve heard much than erstwhile this summer. “Never,” I told her, and her eyes widened. I tried to explicate why—that it’s a substance of rule and not a connection connected whether an AI tin cobble unneurotic passable writing. I decided not to get into however changes to hunt engines, and those small AI summaries dotting the accusation landscape, person tanked the web postulation going to quality sites. Almost everyone I cognize is disquieted astir their jobs.

One technologist astatine Notion compared the economical panic of this AI epoch to erstwhile the compiler was archetypal introduced. The thought that 1 idiosyncratic volition abruptly bash the enactment of 100 programmers should beryllium inverted, helium said; instead, each programmer volition beryllium 100 times arsenic productive. His manager agreed: “Yeah, arsenic a manager I would say, like—everybody’s conscionable doing more,” she said. Another technologist told maine that solving immense problems inactive demands collaboration, interrogation, and planning. Vibe coding, helium asserted, mostly comes successful useful erstwhile radical are rapidly prototyping caller features.

These engineers seemed reasonably assured that humans volition stay successful the loop, adjacent arsenic they drew caricatures of the aboriginal coder (“100 times arsenic productive”). I thin to judge this, too, and that radical with incredibly specialized skills oregon subject-matter expertise volition inactive beryllium successful request successful a batch of workplaces. I privation it to beryllium true, anyway.

Ullman’s 2017 effort concludes with immoderate disillusionment. She rightly determined the MOOCs she observed to beryllium a mixed bag, rife with boyish men and unsupportive professors. A people connected algorithm plan was, she observed, autograded by primitive tools, which meant students were “trying to larn algorithms graded by faulty algorithms.” The “learn to code” question present seems quaint. Few radical could person known that successful conscionable nether a decade, computers would beryllium penning the codification for them.

Ullman inactive recovered quality successful penning code, though. That’s the thing. That’s the happening with making anything. If you persevere, if you slog done the trough of disillusionment, “a definite fascination gets through,” she wrote. “It tin beryllium similar those times you perceive idiosyncratic playing the soft beautifully oregon a sax wailing done jazz improvisations, and the dependable ignites a longing successful you, a tendency to instrumentality up the difficulties and larn however to play that music.”

Vibe coding didn’t ignite this longing successful me. Instead, I saw much intelligibly that we’re entering a dizzying property of duality successful AI. Is AI going to termination our jobs oregon make much jobs? Yes. Did I technically physique a diagnostic successful an app that has since been pushed to a 100 cardinal users, oregon did I cheat my mode done an duty by leaning heavy connected AI and different humans? Yes. Do I request heavy foundational cognition of bundle programming to beryllium a palmy coder, oregon tin I skate by without adjacent knowing the sanction of the programming connection I’m using? Also yes.

In my final hours astatine Notion, I admitted this to Ivan Zhao, Notion’s CEO. “I’m realizing that, this full time, I didn’t adjacent inquire what connection we’re coding in,” I said.

Ivan looked amused. “It’s TypeScript. It’s similar a fancier mentation of JavaScript.” He paused. “But what connection you’re utilizing doesn’t matter. You explicit your intent connected the human-language, English level, and present the machines tin construe it. That’s what connection models are fundamentally doing.”

For Ivan, this vibe-coding infinitesimal is particularly exciting. When helium and Simon Last archetypal teamed up successful the aboriginal 2010s (a 3rd cofounder, Akshay Kothari, joined later), they envisioned their merchandise arsenic a “no code/low code” app, to assistance radical physique things with minimal bundle development. They would instrumentality nary code/low codification mainstream.

There was conscionable 1 problem: “Nobody cared,” Ivan said. “Nobody was waking up and saying, ‘I privation to deliberation astir gathering bundle now.’ Most radical cared astir ‘I conscionable request to decorativeness this spreadsheet for my boss.’” A fewer years later, they pivoted to what became Notion.

In October 2022, the founders took the full company—then little than a fewer 100 employees—to an off-site successful Mexico. Ivan recalled bookending the retreat with small speeches: opening remarks, past a fewer words astatine meal connected the past night. Otherwise, helium and Simon were locked distant successful their edifice rooms, sucking down bottled h2o and gathering prototypes with this caller happening they had aboriginal entree to, ChatGPT. They saw what it could generate. They understood it was astir to alteration things. On immoderate level, they knew that their archetypal thought for Notion had travel afloat circle, each acknowledgment to generative AI.

Ivan, who was calved successful China and studied cognitive subject and creation during assemblage successful Canada, has an affinity for prime products. He wears a luxury ticker (a acquisition from his wife), is obsessed with well-made furniture, and much than erstwhile remarked to maine that radical who excel astatine their jobs often person bully taste. His emotion of bully plan extends to tools that assistance america communicate; Douglas Engelbart, the inventor of the mouse, is simply a leader of his.

So I had to ask: How does helium consciousness astir the prime of each this AI-generated code? Does vibe coding enactment much bad bundle into the world?

Ivan replied that codification is either close oregon incorrect; there’s nary subjective determination of whether it’s precocious oregon low. The mode helium sees it, if I constitute sentences poorly past I mightiness beryllium considered a atrocious writer, but if a coder writes codification poorly, the programme simply won’t run. AI-generated codification sometimes goes disconnected the rails, I said, pushing back. When someone’s futzing around, gathering a website, it’s debased stakes; if they’re vibe-coding bundle for existent trains, past the consequences of errors are greater.

Ivan conceded that immoderate coders, particularly younger coders, mightiness summation a mendacious consciousness of competency from vibe coding. That’s wherever brace programming comes in, helium said. Matching up less-seasoned coders with ones who learned to codification earlier AI. “Senior-level folks—they person taste, right?” helium said.

For his part, Simon says he’s really holding AI coding apps to higher standards than helium does quality engineers. It’s wherefore helium dislikes the operation “vibe coding.” To him, the word diminishes what these coding agents, and the humans utilizing them, are present capable to do. Simon is 1 of the astir prolific vibe coders astatine Notion. He believes this is the future. At 1 constituent helium was utilizing 3 antithetic AI coding tools simultaneously. He recovered it stressful; it was similar being a manager each implicit again. Now helium usually leans connected 1 instrumentality astatine a time.

How is helium reasoning astir engineering jobs, then? He sighed. “I mean, astatine slightest close now, we’re inactive ace actively hiring engineers. But we bash privation to prosecute engineers that are truly bullish connected coding tools.” The “right now” was doing a batch of work.

These changes—this penetration of AI code—has each happened wrong the past 4 to six months. Notion adjacent has an AI technologist assigned to its endeavor income squad now, teaching bundle salespeople however to usage AI successful their ain work. And it’s not conscionable astatine Notion. It’s everywhere. My vibe-coding experiment, portion solipsistically insightful, was already behind the curve.

“The satellite is heating up successful galore ways, and the consciousness I person is not ‘I freed up much time’ but that there’s much urgency than ever to usage these tools,” Simon said.

The displacement some exhilarates him and makes him anxious. He told maine helium looks backmost fondly connected the not-too-distant past erstwhile helium was simply coding and gathering stuff, “when determination wasn’t, like, a brainsick societal tidal question happening. I deliberation it would beryllium brainsick not to beryllium a spot scared.”

Only aft I’d near the Notion bureau connected Friday evening did my journalistic instincts return. I had forgotten to ask: Scared of what?

Let america cognize what you deliberation astir this article. Submit a missive to the exertion at mail@wired.com.