Skip to content
Svenska
Menu
Screws, a felt wedge and a small paper fold set between the strings of a piano: a prepared piano. Black and white.
A prepared piano: the same instrument, its sound changed by what is set between its strings.

The short version

Many ways of knowing in work with language models

A transdisciplinary practice of priming language models to support people working together on the problems of our time

This is a way of working with language models for people who work together across fields. What a model writes depends on what it has in its context, the text it reads as it works, and to prime a model is to shape what it writes by shaping what it reads. Here the model reads a set of texts written in dialogue with several ways of knowing: not kinds of knowledge to be stored, but the ways in which people come to know something, attend to it and respond to it. The texts keep those ways in tension instead of blending them, and people maintain them and take responsibility for what is carried forward. I call it a polyphonic priming ecology.

What stands in front of the model

Most people meet a language model as one agreeable voice. It replies politely, summarizes evenly and settles on a sensible middle. That voice is not all the model can do. It is shaped by choices made in advance, and for everyone, by the company that runs the service: in how the model is trained, and in the instructions the company places at the start of every conversation. It leans toward the dominant ways of writing and knowing, often Western ones, and in my experience what comes out is too often the usual and the stereotyped, presented as balance.

A model writes one small piece at a time, a token (a word or part of a word), choosing among likely continuations, and what is in its context, the text it has before it, shapes which of them lead. The figure shows four things that can stand there, at the first answer and far into a long piece of work.

Choose what is added and when; the figure changes with your choice.

What is added in front of the model
When

1What stands there, and who put it there

Before the conversation begins

the company The company's instructionsplaced in front of every conversation, for everyone

The conversation

you, in the chat “How should we involve tenants in the renovation?”

The next word is chosen from all of this

2Which first words become likely

  • Here56 %
  • In11 %
  • The25 %
  • Before8 %
  • There0 %

Illustrative values, not measurements.

3How the answer begins

“Here are some key strategies.”

It is shaped by choices made in advance, and for everyone, by the company that runs the service.

A language model does not look up an answer: it continues the text in its context, so whatever is placed there changes which words come next. A careful prompt or a persona is a single addition, written once, and over a long piece of work it fades. A priming ecology is several texts that pull against one another, brought back to the model again and again by people who answer for them.

The figure shows only how an answer begins. Whether the work becomes more useful is something no bar can show; the four scenes further down show what it can look like.

What it is

A priming ecology works a little like the preparation of a piano: the instrument is not rebuilt, yet much of what it can sound like changes. It is the same model, with the same learned weights, the numbers fixed when the model was trained and left unchanged while it is used. What changes is what those weights are set to work on.

Choose a part to read about it.

the model, unchanged what the model reads at each answer outside: the people who write, read and answer

Maintained texts that someone answers for

Texts written to prime the model, each with its own task: what the work should care about, the concepts it uses, the direction carried from one session to the next. Some are given to the model each time it answers; others are entries in a knowledge base, a searchable collection whose relevant entries are found and supplied when the work calls for them. All are revised or retired as the work requires, and a named person, role or group answers for what is carried forward: takes responsibility for it, and can be asked about it.

Several ways of knowing acting on one another

Knowledge practices, approaches and kinds of experience, each keeping its own questions and terms and putting them to the others. No practice is treated as the authority over the others, and where they do not fit, the misfit is kept in view instead of being smoothed over.

Tension written into what the model reads

The texts hold pulls open against each other and trace tendencies instead of fixed categories. A rule tells the model what to conclude; a field of tensions gives it several pulls among which a situation has to be placed. The texts draw attention toward a difference without settling it. Which tensions are brought forward is chosen by someone, who answers for that choice. Keeping a difference open is a collective practice, carried by people in how they meet, read and decide; the model can support that work but cannot do it for them.

The practice around the model changes too: who writes the texts, who reads what comes back, and who decides what is carried forward. What counts is less what the model writes than what people go on to question, change, refuse or do. A priming ecology is built collectively, for each organization, company, project or collective, from its own ways of knowing and with the people who will carry it.

Where it can matter

A polyphonic priming ecology works at the level of shared practice, between regulation (laws that set what the companies must do, and rules on who may use their tools, and how) and the skill of the single user: in the groups, rooms and organizations where models are actually used. There it can matter in situations such as the four below. Each comes with a short scene. The organizations and people in the scenes are invented; the scenes draw on situations I have worked in. They show how the work can go, not that it does.

In organizational change

Where interests and ways of knowing conflict, power is unevenly shared, and a question has to be reframed before people can work out what to do together.

A fictional sceneA workshop that names the weatherRead the sceneClose the scene

A company that installs and services heating and ventilation systems asks for a day with its team leaders on how to lead change. Before the day, a facilitator drafts short workplace scenarios with a model whose texts keep three forces in view: what has been planned, what has settled over time, and what is forming now. On the day the model is not in the room, and in one group the complaint that nobody tries anything new becomes a question: what happened the last time someone did? Afterward the model's reading of her notes keeps rooms, formats and routines in view before anything is put down to people's attitudes, and she takes the question it leaves open to the manager who was there.

In research across disciplinary boundaries

Where each field's frames are opened to the questions of the others and can change in the meeting.

A fictional sceneA mapping that asks whose valuesRead the sceneClose the scene

A funder asks a team to map a network of neighborhood initiatives against the three values of a European program: beautiful, sustainable and together. The team first maps its own work with a gardener and a youth worker, and writes each value as a field of three pulls instead of sorting the initiatives by how well they live up to it. The model reads many reports quickly and drafts in the funder's language, noting what that language leaves out; the team makes every placement and decides which notes stay. Each placement carries the date it was made and is read by the initiative before the funder sees it, and when one initiative declines to be placed, the mapping says so.

In public and civic work

Where those affected are many, some of them are never in the room, and the way the work is done can change who comes to be in it.

A fictional sceneA housing question asked againRead the sceneClose the scene

A housing organization asks an analyst why only about one household in three has signed up for its renovation program. She works with a model whose texts let three practices question one another: the numbers, people's own accounts, and attention to whom a format was made for. Before any answer, she asks for a reading of the question itself, and two other phrasings come out of it; the staff decide where to begin, with two tenants at the table, and the tenants' wish to begin elsewhere is written into the record. The report carries both questions and who chose, and it is signed by the analyst instead of saying “the analysis shows”.

In artistic and collective practice

Where listening, improvising and composing together are the work itself, and where a model can be one more voice in the room, with its own limits.

A fictional sceneA listening score, and a session without the modelRead the sceneClose the scene

An artistic research group prepares gatherings where people listen together beside a stream, at the site of an old mill, and uses a model to draft the score. Even carefully primed, the first draft pulls toward the smooth arc of a well-run workshop; asked where it smooths something over, the model names three places, among them a closing circle in which each person speaks. In the second draft the circle is gone, the right to pass opens the invitation, and the silences are no longer timed: a change in the sound of the water can offer a turn. For the third gathering the group decides not to use the model at all, and no one is asked why.

It can also matter wherever models are already in use and the risk lies in how they are used: in what is handed over and to whom, in who answers for it, and in who has access to which models, at what cost.

What it does not do

A priming ecology does not take the place of the people who think, decide and answer for what is done. It cannot replace their experience, or the ways of attending, caring and taking part that live in their bodies and practices, or the presence in the room of people from other fields and backgrounds, and it cannot make sure that many ways of knowing are actually taken into account. It does not retrain the model, it is not an off-the-shelf product, and it comes with no guarantee: whether it helps can only be tried in the work, over time. Its own vocabulary has to stay open to challenge, since a priming ecology can grow blind to what its terms fail to notice. Anyone concerned can keep the model out of their own part of the work, and sometimes a smaller model, or none, is the better choice.

It also has costs. Long texts sent to a model again and again draw energy and water, the data centers that run the models weigh on the places around them, on local electricity and water supplies and sometimes the air, and the model is built on the work of many people who are seldom seen. To question extraction with a model is to draw on an extractive infrastructure, which is one reason why knowing when not to use a model belongs to the practice. A priming ecology built on a commercial model depends on a company that can change the model without notice, and anyone who can place text in a model's context can try to steer it. A priming ecology does not reach the larger questions about AI: who owns the models and the infrastructure, how they are governed and where the technology is heading. Today these are decided largely by engineers, executives, boards, investors and governments, who can be held to account for them.