Four of six · Making stories by machine
It is possible to build a machine that invents stories: you write the rules rather than
the words, set it running, and read what comes out.
The hard part is not making it produce a great deal. Anything can produce a great deal. The
hard part is making the things it produces feel meaningfully different from one another.
The other five explain something and then show you the evidence. This one goes the other way round: it asks for your opinion first, and opens the workings at the end. It seemed only decent to say so before starting rather than afterwards.
It will make a good deal more sense at the end than it does at the beginning, which is the arrangement.
Below are four objects recovered from a single archaeological site. Do what anybody does when handed a set of things: notice a pattern and form a view about what it means.
Nobody is available to ask. The people who made these are gone, they left no writing, and everything anybody will ever know about them has to come out of the objects themselves.
A curved pin in worked bone, carrying the spiral mark. Showing almost no wear. Recovered in a pit outside the wall.
A shallow bowl in bronze, carrying the spiral mark. Showing almost no wear. Recovered under a collapsed threshold.
A socketed hook in polished stone, with no spiral mark. Showing almost no wear. Recovered in a river channel.
A pierced disc in bronze, carrying the spiral mark. Heavily worn. Recovered in a burial with two others.
Some of them carry a spiral mark and some do not. What is your best guess about what the mark means?
Four more sections follow. Choose one of the four answers above and the next one opens.
The obvious way to judge a machine like this is by how much it can produce. By that measure the generator on the left is far better than the one on the right. It has more sounds to draw on, longer words, more possible combinations, and every output it has ever made is unique.
Read them and see whether that is how it feels. This is a problem the game designer Kate Compton gave a name to: you can serve ten thousand bowls of porridge, each one mathematically distinct from all the others, and every person you serve will say they have been given the same thing twice.
The numbers underneath are worth taking literally. Think of a generator as a bag of tiles rather than as a machine. You are not writing the words. You are deciding which tiles go in the bag, and then reaching in. The number is how many tiles are in there.
The second generator can produce roughly a hundred million times fewer things than the first, and it is the one that reads as a language. Its words could belong to the same people. You could probably invent a plausible new one yourself, which is a sign that you have picked up its rules without being told them.
The first has more variety and no character at all, so its outputs blur into one another. Variety and distinctiveness are not the same quantity, and past a certain point they pull in opposite directions.
Between them they define the two ways anybody builds a story machine, and both are easier to see as objects on a table than as software.
The first is a deck of cards. Each card has a scene written on the front and a condition on the back saying when it may be played, and the machine deals whichever cards are currently legal. Every word the reader sees was written by a person. It fails by running out: there is only ever as much as somebody found time to write.
The second is a board with pieces on it that move according to rules, where the story is whatever you find yourself narrating while you watch. Nobody wrote any of it. It fails by porridge: it will produce events for ever, and almost all of them are indistinguishable from one another.
Neither failure is a fault to be fixed later. Each is the standing cost of the approach, and choosing between them is mostly a decision about which problem you would rather spend the next several years on.
Here is the same generator with nothing switched on. It can produce an enormous number of words and none of them belong together.
Underneath are four prohibitions. Each one removes possibilities and not one of them adds anything at all. Switch them on and watch both numbers move: the size of the bag goes down, and the sense that these words come from somewhere goes up.
Tap any rule to switch it on or off. The list of words changes immediately.
Back to the bag. The thing you are actually designing is not any of the words that came out of it. It is what is in the bag, and every one of those switches takes tiles out. None of them puts any in.
Which is the part that takes some getting used to. Adding a rule feels like adding something, and it is the opposite. A generator gets better by being allowed to do less.
This section has so far made only one side of the case, so here is the other. Every prohibition deletes outputs, and some of them were good. A rule forbidding a consonant from following itself has just removed every name that would have been memorable for precisely that reason, and it removed them silently, because a bag of tiles gives you no way of inspecting what is no longer in it.
Push it far enough and coherence turns into monotony: a bag small enough that a reader sees the whole of it and stops looking. There is no formula for where that point sits. It is a judgement, made by ear, and it is the part of the job that stays a craft.
The spiral meant something because it was never once cut by the wrong workshop.
A plot is a sequence of events that somebody arranged so that it would land properly. A past is a sequence of events that merely happened, most of which nobody bothered to write down, and which reaches you through people who were not trying to help.
Generating a plot is very difficult. Generating a past is not, so long as you build three things rather than one.
The first is consequence. Events have to change the state of the world, so that later events can only happen because earlier ones did. A war costs both sides strength; weakness invites revolt; a kingdom that runs out of strength stops existing. Without that you have a list, and a list is porridge with dates on it.
The second and third are the two kinds of loss. Things that were never written down, and things that were written down and did not survive. A record that captures everything is not a record. It is a transcript, and no history has ever been one.
Below is a century of an invented country, in three versions. Start with what survives, then look at what was written, then at what actually happened.
In this particular run, 8 events were never written down at all, 2 records were written and then lost, and 1 of what remains is either under the wrong year or has two separate events fused into a single episode.
Notice one rule in particular. Every collapse goes unrecorded, and it does so by design, because the kingdom that would have written it down is the thing that stopped existing.
Nothing in the first column is false, and the first column is not what happened. A reader given only that view will build a perfectly coherent account of this century, and it will be wrong in ways that nothing available to them could possibly expose.
Which is where this joins up with the rest. The first three sections put you in front of a record and let you draw conclusions from it. This one is the same machine seen from the other side: not interpreting a past, but manufacturing one with the gaps already built in, so that somebody else can come along afterwards and get it wrong in the ordinary way.