Startup decelerator

Why slow, now

The faster the situation moves, the slower we think you need to be about choosing what deserves years. Everybody is running after a train that has already left. The useful move is to stop for a second, work out where it is going, and meet it at the next station.

We did want to build a startup. Every idea we discussed came with the same doubt: would AI be doing this on its own in six months? We could have aimed for something that made money for three or four years and then sold it, but we wanted to make something that lasts, and we couldn't explain to ourselves why any one idea deserved that. A plan made in a hurry today is a bet on where the train was this morning.

Slow is about the big commitment, and only that one. Taking money, hiring and promising a product to strangers can all be walked back, at a price paid mostly by the people who trusted us, so we slow down for those, the way Why Castor puts it. Everything that tells us something and can be undone, we do fast: a piece of open-source software we can write in a week, an essay, a conversation. Two people who live together can decide that kind of thing over breakfast, which is the advantage a small team always has over a giant.

The Romans had a phrase for this, festina lente, make haste slowly. Suetonius says Augustus used it often, and fifteen centuries later the Venetian printer Aldus Manutius made it his mark, a dolphin wrapped around an anchor.³

Moving without a destination

Say you are walking somewhere four blocks north and three blocks east, in a city laid out as a grid. Walk the four north, turn once, walk the three east, and it is seven blocks. Mix the turns up however you like and it is still seven. What changes is the traffic lights. One of us takes whichever crossing is green, as long as it leads the right way. Walk all four blocks north first and you have spent your choice: every crossing left has to take you east.

That is how we run Castor while we decelerate. We go in several directions at once: this blog, and open-source software that circles one theme, from a server that lets an agent work in the browser tabs you already have open to a router that spreads coding work across the AI subscriptions you already pay for. Each project is small enough to finish or drop in a few weeks, so a turn costs little.

The walker knows where they are going. We don't have a destination yet, but we have a direction. Our values are written down on this blog, dated, and they already rule out some green lights. An agent that acts in your name without asking would be quicker to build and easier to demo; Gaddi, our browser server, holds actions like publishing until you approve them yourself, with Touch ID, on your own Mac.

Watching for what

From the outside, sitting back and watching looks exactly like procrastination, and on some days that is what it is. The difference we try to hold on to is that procrastination waits for a feeling, the feeling of being ready, which never comes. A decelerator waits for information, and should be able to say which.

We watch where a person still makes the result better and where keeping one in the loop is ceremony, which is the question behind Castor. We watch what agents can do this month that they could not do last month. And we write, because writing is how we catch ourselves: put "we're exploring" on a page and the reader will ask what we have learned. OmniRead, our web reader for agents, came out of something we kept running into: an agent cannot tell a page it read from a page it half read, so it reasons over both with the same confidence. Now every read comes back with evidence of whether it was complete. When we can't say what a project taught us, maybe we enjoyed making it and haven't learned much yet.

The weak spot is the stopping rule. Sometimes the street ahead is the one you need, and its light is red. A large project is mostly red lights: long stretches where the interesting choices have been made and the rest still has to be done. If every dull stretch sends us looking for a new direction, we will end up with a fine collection of beginnings. So we are looking for a question that keeps coming back after the excitement of a new project wears off, because several projects have left it unanswered, or because the answer matters enough to sit through the boring parts. We can say where to look. We cannot yet say what we will see on the day it is time to commit.

The case for hurrying

Everything above argues with the person who did most to define what a startup is. "A startup is a company designed to grow fast," Paul Graham wrote in 2012. "The only essential thing is growth. Everything else we associate with startups follows from growth."⁴ At Y Combinator, which he co-founded, a good week means growing 5 to 7 percent, and "if you can only manage 1%, it's a sign you haven't yet figured out what you're doing."⁴

He answered this essay, too, in 2006. "Startups make all kinds of excuses for delaying their launch. Most are equivalent to the ones people use for procrastinating in everyday life."⁵ Among the causes of delay he lists "working on too many different things", which is a fair description of a walk across a grid. He thinks the risk runs almost entirely one way: "Launching too slowly has probably killed a hundred times more startups than launching too fast". And the evening at the event was a small case of another of his lines: "it's only by bouncing your idea off users that you fully understand it."⁵

We are playing devil's advocate here, and we know it. A startup's nature is to grow fast and to move fast to get there; that is what the word means, and Graham has watched far more of them than we have. We don't know which of us is right, and the opposite of this essay may well be true. We often argue the other side on purpose, to see whether the strong case holds up.

Even he allows that a startup's growth begins with "an initial period of slow or no growth while the startup tries to figure out what it's doing."⁴ Our bet is only about that first period: that while what machines can do changes every few months, it should be longer than it used to be, and spent shipping small things within a week and saying them out loud to strangers, which is his advice too. If we are wrong, a year from now we will have shipped plenty and committed to nothing, and the last paragraph of this essay asks you to call that out.

The pitch problem

A decelerator pitches badly. The pitch asks for the conclusion we have chosen not to draw yet, so the honest answer loses most rooms to someone with a slide and a number.

The evening gave us something for the price. Saying a project out loud to a stranger, with a little pressure on, found in a few seconds a gap that weeks of building had hidden. Nothing rides on that project yet, so the gap cost us one awkward sentence and a better question to work on. Next time we will say what we are watching for and show one thing that works.

If a year from now these essays still say "we're exploring" and none of them says what we chose, it will have been procrastination, and you should tell us so.