Paul Graham

The pattern they actually represent

In the summer of 1995 Paul Graham and Robert Morris started Viaweb on $10,000 from a friend named Julian Weber. Weber took ten per cent, and for it he set them up as a company, taught them what business was about, and stayed calm in a crisis. Graham's own verdict on that trade is the entry in one line: that is why he knew there was room for something like Y Combinator. Yahoo acquired Viaweb effective 10 June 1998, for about $48.6 million by the Form 8-K/A, and the software shipped on as Yahoo Store.

Seven years later he re-issued the deal he had been given. On 11 March 2005, walking home from dinner in Harvard Square, he and Jessica Livingston decided to do it. He put in $100k, Robert Morris and Trevor Blackwell $50k each, and Y Combinator started with $200k. The first thing it announced was a Summer Founders Program, described at the time as an experimental replacement for the conventional summer job.

Now the part worth stealing, which is not the money. Graham writes that they did not initially have what turned out to be the most important idea — funding startups synchronously, instead of asynchronously as it had always been done before. They funded a group at once because that was the fastest way to learn how to be angel investors. A shortcut became the product. YC still runs three-month batches four times a year, in groups, ending in a day when founders present to an outside audience, on a shape Graham says is almost identical to that first summer.

Read what he did not do. He never told founders working alone to network harder. He changed what was around them: a fixed group, one clock, a date in public.

The blockage it speaks to

Read this if the only person who has looked at your work this month is you.

From the inside it does not feel like isolation. It feels like focus. No meetings, no opinions to manage, nobody slowing you down. What is actually happening is that the loop has closed. You generate the idea and you also grade the idea, with the same instrument, on the same assumptions.

Two things follow, and neither is a discipline problem. The first is that you keep studying. Another course, another teardown, another architecture thread at one in the morning — which is the rational move, because with nobody to say "that is solved, move on", reading is the only new input available to you. The second is that decisions swell. A choice a colleague would settle in four minutes takes you three weeks, because the only test you can run is putting it to yourself again.

Graham is the right study for one reason. His answer was never find a mentor, which is a person you have to be lucky enough to meet. He built a container instead, and the container has a parts list anyone can copy.

Three moves you can steal

Recruit two colleagues, not a network. Graham's line is that it is better to have one or two great ones than a building full of pretty good ones. Two is a number you can act on this month; a network is a wish. The test is not agreement. Good colleagues offer surprising insights and see things you cannot, so if everything they say is something you had already thought, the loop is still closed.

Give yourself a batch. Copy the structure rather than the institution: three or four other people building something, a start date and an end date three months apart, and a booked event where you show the work to people who are not your friends. The end date is the load-bearing part. Without one, a group is a chat. With one, a group is a deadline other people can watch you miss.

Recruit your first users by hand this week. Graham's position is that you cannot wait for users to come to you, and that the feedback from engaging directly with your earliest users is the best you will ever get. This is also the release valve for the reading. Ten people using a rough version generate more correction in a week than a quarter of courses will.

Where the pattern breaks

In September 2024 Graham published an essay that opens on a talk Brian Chesky gave at a Y Combinator event, then generalises from that one talk about one company into a claim about two ways of running a company. He names the gap himself. As far as he knows there are no books specifically on the subject, asked whether the idea is good or bad he answers that we still don't know, and a footnote predicts that founders unable to delegate will use it as an excuse. In December 2024 Charity Majors published a detailed objection: the case generalises from outliers and presents context-dependent choices as universal. That is a limit of applicability rather than a fault. Advice minted inside a dense network arrives compressed from that network's own sample, and the compression is invisible from inside it.

What to do this week

Pick two names. Not a network — two specific people who are building something and whose opinion has surprised you at least once. Send both the same message: thirty minutes on a call, same time every week for the next eight weeks, where each of you shows the actual work and the others say what is wrong with it. Put all eight in the calendar today, not just the first one.

Then set the end date. Eight weeks out, name a day and one thing you will show on it, and tell somebody outside the group who will be watching. A demo call, a post, a room above a bar.

That is an hour of scheduling. What it buys is the input you have been trying to manufacture out of your own head, and a date that arrives whether or not the work is ready.