Forty-One Years at Sea
A game begins as a thought I can barely make out. How does developing games change when I can pursue many of those thoughts rather than a single treasured one? I want to find out whether a game can evolve that way, through attention and selection: many ideas tried, few kept. Making games with AI is how I’m trying that out.
Before AI, I could spend months building a game before finding out whether it was any good. And I’d only see what players made of it once I’d found the money to launch and advertise it. With AI, I’ve developed three games in the past two weeks. In total, maybe a couple of hours of prompting per game. Making each one by hand would have taken me maybe three long months at the computer. I’d have needed at least a year just to learn to make the art. Another year for sound. Now all I have to do is make my intention clear enough, and AI gives my colleague work to approve.
The AI writes and revises the code. I barely look at the output. I send it to my colleague, who sits next to me. They have their own AI review it and barely look at it either. After a few back and forths, one of us sends the word OK, and the other presses the green button.
Told like that, it sounds efficient. It only starts to look strange when you slow it down and follow a single key.
Painted plastic over a metal spring over wired circuitry. A human finger. My bones and flesh, shaped by evolution into a finger that bends at the knuckle. Fifteen years of falling down, getting up, pulling towels off the table, cups breaking. One day I can move that finger on its own towards the O key.
The finger strikes a keycap made from petroleum that spent millions of years under pressure. The keystroke lights up an O. The O forms upside down in my eye. The optic nerve carries the signal to a brain that has spent years learning to tell one shape from another by getting them wrong. Oh, it’s a circle. In the 220 or so milliseconds it takes me to recognise the circle, I’ve already forgotten pressing the key.
Another age passes, and the letter K appears on the light box. O-K.
The finger finds the bigger piece of plastic, the one with the arrow on it, and the key goes click. The two letters go into packets with an address. All of this while the muscle is still releasing its tension. The signal goes by radio to the box that connects me to the internet. From there the message travels by cable under the Atlantic to a headless server deep underground. The server stores copies of OK in 56 places and sends the message back across the seabed. The two letters arrive at the computer next to mine, at the person sitting next to me. A notification appears on their light box. My hand is finally retracting from the plastic.
Their eyes pick up a small blue circle beside the outline of a church bell. They move a clump of plastic on a tail, named after a rodent that lives in our walls. The arrow on the light box turns into the semblance of a hand. After fifteen years of human development, my colleague knows the five humps mean they can click the bell. They press down. The click crosses the ocean and comes back as rectangles within rectangles within rectangles. Inside the last one there are two glyphs beside each other, one saying O, one saying K. They find the green box marked merge.
The two letters are the only part of it either of us wrote. They went under an ocean twice to travel about a metre.
The work was already sitting behind the green button before either of us said OK. Deciding to say it is what feels like an eternity.
Put the wait for an OK and the trip it takes on the same clock. Make the signal cross the ocean at the speed of a sailboat from the 17th century. At three knots, a ship takes about forty-five days to cross six thousand kilometres of Atlantic. The Mayflower took sixty-six. Weather, errors and sickness could stretch a voyage like hers or end it. Through a fibre cable, the signal makes the same crossing in about thirty milliseconds. Now slow every other step by the same amount, so that each crossing of the signal is one of those voyages. Even slowed down this far, the processor still ticks more than twenty times a second.
On that clock, the finger holds the key down for almost six months. Eleven months for the eye to react to the light on the screen. By the time my colleague has noticed the small blue circle, the signal could have crossed seven times. The two crossings that brought them the letters took three months at sea. Typing them and pressing enter takes about four years.
The AI writes about one word every month. It writes five pages in 164 years. Reading them takes more than fifteen hundred.
My colleague doesn’t read them. They run the review, walk away and take a shower. The AI reviews for the length of the shower: about two and a half thousand years. They come back, see that the AI says OK, and look for maybe ten seconds. That is forty-one years at sea, a working life. Then they press the button.
And the work itself? Each of my games took a couple of hours of prompting: about thirty thousand years at sea. One of them, hand coded the old way over three long months, would take thirty-two million years. Make all three by hand and add the year I would have spent learning the art and the year learning the sound. More than three hundred and fifty million years.
On any clock, ten seconds of looking is a little over a millionth of three months of making. Even against the couple of hours I spent prompting, it is one part in seven hundred. I still have to ask what we could have understood together in those ten seconds.
A disagreement about how something should work used to give us a reason to explain it to each other. We kept at it because we thought we could follow each other’s reasoning all the way to a decision. Now I don’t understand. I lose the thread of what the AI changed. How much of the work can I ask my colleague to make sense of when I can barely make sense of it myself?
So we both approve a change neither of us could explain to the other.
The two letters look the same as they did when we could explain the change to each other. The ships still make the crossing in about thirty milliseconds. But they are no longer carrying understanding in their cargo holds. The OK is not empty acknowledgement either. We trust each other. We know the other person tried.
It’s camaraderie in ignorance.
Even reading all five pages before we say OK wouldn’t end that ignorance: we’ve only seen the version we got. Behind it are all the other versions the AI could have produced. In the three months it would take me to hand code one game, I could prompt about a thousand of them. But each of those thousand means writing a prompt and making sense of what comes back. I send a message of 200 words to an AI and get five pages back. On our clock, from typing the prompt to finishing the five pages, nearly three thousand years pass. Almost all of it is a person typing and a person reading.
Why spend all that time typing and reading when we could talk? Sometimes we should. My colleague sits next to me, and for a small decision, leaning over is faster than any of this. Talking to another human was often the best tool we had for deciding not to do something. Now it’s one of the only tools we have for understanding what we did do. We now have a game to talk about, rather than an idea to abandon. To get there, we push everything the AI makes back through the slowest parts of us: five pages for the eyes, two letters for the fingers.
Why are we still using our fingers? Suppose I have a thousand ideas for games and can tell the AI aloud what I want to play. I reject all the game ideas except one. AI turns that one into a game I can play. If the game does what I had in mind, I pass it on for someone else to try before a real ship from 1620 would have made it halfway across.
Could we do it that quickly and still tell each other what we’re saying OK to?
I used Claude, ChatGPT and ProseMap to shape my dictated observations, interspersed with manual editing.
— Van