What I Worked On

What I Worked On. An essay by Paul Graham, February 2021.

Stories and punch cards

Before college, Graham worked on two things outside school: writing and programming. His short stories were almost all feeling and no plot.

His first programs ran on an IBM 1401 in a school basement, fed by punch cards. With no data to give it, there was little to do. One program never stopped.

Then came microcomputers, which answered as you typed. On a TRS-80 he wrote games, a rocket-height predictor, and a word processor his father used.

From philosophy to AI

In college he meant to study philosophy, the search for ultimate truths. But other fields had claimed almost all the ideas. Philosophy got the edge cases.

The classes stayed boring, so he switched to AI. A novel with a thinking computer, Mike, and a program called SHRDLU made it seem a few years away.

Cornell had no AI classes, so he taught himself Lisp, which blew open his idea of a program. His diploma printed his major in quote marks.

The AI that wasn't

In his first year of grad school at Harvard, he saw that the AI of the day was a hoax: programs turning sentences into formal facts in a list.

They handled only a small, formal slice of natural language. Teaching them more words would never reach real understanding. That road would never lead to Mike.

So he salvaged Lisp from the wreckage and started a book on it, though he knew little. Writing a book is a great way to learn. It became On Lisp.

Work that lasts

Computer science has theory, where people prove things, and systems, where they build. He wanted to build. But software goes obsolete.

Then, at a museum in 1988, looking at a painting, it hit him: paintings never go obsolete. And a painter answers to no boss.

So he took art classes, while doing a computer science PhD and writing On Lisp: many projects, none of them his thesis.

In April 1990 his advisor asked if he could graduate in June. With not a word written, he said yes, and wrote it in about five weeks.

Florence, and still lives

He got into RISD. Then a delayed letter from the Accademia in Florence invited him to its entrance exam. He crammed Italian, and passed.

A pattern was forming: a famous school, a prestigious subject, and disappointment. At the Accademia, teachers didn't teach and students didn't learn, by agreement.

So at night he painted small still lives in his room. A still life holds still, so you can copy what you see exactly, then stress cues like an object's edge.

Normally your brain hides detail: it tells you a bush, not where each leaf is. Painting makes you look closely, and it's full of surprises.

A year at Interleaf

Broke, he took a one-year job at Interleaf, whose software had a thin layer of Lisp over a mountain of C. He admits he was a bad employee.

But the pay! In Florence his budget had been $7 a day. Now he made over four times that per hour. He saved for RISD and paid off his loans.

The big lesson: the low end eats the high end. Be the cheap, entry-level option. If you don't, a rival will, and pin you to the ceiling.

Which means prestige is a warning sign. He would use this idea at Viaweb, and again at Y Combinator.

Art school, then New York

Back at RISD in 1992, painting was much like the Accademia, only pricier. Students were to express themselves, often by inventing a signature style.

A signature style is a schtick: it marks a painting as yours at a glance, and it's mostly why buyers pay so much. He felt closer to the kids who came to draw.

In 1993 he dropped out and moved to a cheap apartment in New York. To pay for painting, he wrote another Lisp book, hoping to live on the royalties.

One day, helping the painter Idelle Weber, he heard about a young, very rich fund manager. A thought: why not get rich? Then he could work on anything.

The web app

The web was coming, and Graham wanted on the next train out. His first idea, putting art galleries online, flopped: they wouldn't take it free.

But online stores looked like their gallery sites plus order buttons. Store software then meant a copy installed on every computer, which he hated writing.

Then, lying on a mattress one morning, an idea: put the software on the server, and let people drive it by clicking links. All they'd need is a browser.

By August 12 it worked: a web app, before they were common. No versions, no installs. They called it Viaweb, and took $10,000 for 10%.

Running Viaweb

They meant to be the Word, not the Interleaf: simple and cheap. At $100 to $300 a month they were cheaper than they knew; $300 sounded like a lot to them.

Desperate for users, they built stores for them. It taught them retail: if a shirt's picture must be tiny, show the collar, not the whole shirt.

Stores went from about 70 to 500 in 1997, about seven times in a year. Growth rate is the real test of a startup, but he fretted about the totals.

He hired too many people, so Viaweb only broke even around when Yahoo bought it, in 1998. The sale was a huge relief.

Rich, and adrift

After the sale he was worn out, and Yahoo's cube farm felt like Interleaf. Once his first options vested in 1999, he left, to paint.

His boss assumed nobody walks away from that much money to paint; surely he was starting a rival startup. He wasn't. He just needed a vacation.

In California he had no energy to paint, so he went back to his old New York apartment. Same life, but now rich: taxis, and lunch out.

In 2000, a new idea: a web app for building web apps. He started a company, then saw he didn't want to run one, and kept a piece: a new Lisp, Arc.

Anyone can publish

In 2001 he put a talk about Lisp on his site. Someone posted it to Slashdot, and it got 30,000 views in a day. There was an audience.

In print, editors guarded a narrow channel to readers. Online, anyone could publish. A whole new generation of essays was waiting to be written.

That was the turning point. From then on, whatever else he did, he'd always write essays.

Online essays looked low-status. But unprestigious work keeps serving him well: still lifes, Lisp, Viaweb, YC. Being drawn to it anyway suggests the right motives.

Starting Y Combinator

In 2003, at a party at his house, he met Jessica Livingston. Later, as she waited to hear about a VC job, he listed what VCs got wrong.

At a Harvard talk he said the best seed money comes from successful founders. Every eye turned to him. He said not him, then wondered why not.

On March 11, 2005, walking home with Jessica, three threads met: his VC ideas, angel investing, and Robert and Trevor. They'd start their own firm.

To learn investing, they'd fund a batch at once, over a summer. The Summer Founders Program drew 225 applications. They funded 8, at $6,000 per founder.

Batches turned out better for founders too. They cured isolation, and alumni helped each other and even became each other's customers.

YC takes over

The plan was a third each: hacking, essays, YC. To test a new Arc, he wrote Hacker News, which became his biggest source of stress.

YC grew to fill his days. It ate Arc, then started on his essays. With a new batch every six months, the problems came to him.

In 2010, Robert Morris, who almost never gave advice, gave him some: “You should make sure Y Combinator isn't the last cool thing you do.”

Slowly Graham saw it: on this path, YC would be the last thing. Either it was his life's work, or he'd have to leave. It wasn't. So he would.

Handing over

In 2012 his mother had a stroke, caused by colon cancer. Flying to Oregon to visit her, he had time to think, and saw he was ready to hand YC over.

They asked Sam Altman to be president, and in October 2013 he agreed. It was a full handover, so YC could outlast its founders.

His mother died in January 2014. He stayed until March to see that batch through, then stepped away.

He painted most of 2014, and got better. Then in November, mid-painting, he ran out of steam. Attention is zero-sum: the wrong project crowds out the right one.

Bel

In 2015 he went back to Lisp. At its core, Lisp is a language defined by an interpreter written in itself: John McCarthy's idea, from 1960.

McCarthy could only test it by hand. Now computers could run it, so you could keep extending his Lisp, just as carefully, into a full language: Bel.

He banned himself from essays, so it looked like he was doing nothing. But solving a problem while his boys played in tide pools, it felt like doing life right.

Bel was done in 2019. Asking what to do next, he looked at how he'd chosen before. The answer was long and messy, so he wrote it up for others.

In short

Even for him, choosing what to work on was long and messy.

Unprestigious work kept paying off: prestige is a danger sign.

Be the low end, and judge a startup by its growth rate.

Attention is zero-sum: drop work that isn't right for you.

The original is long, funny and moving. Read it: paulgraham.com/worked.html