“It seems necessary and impossible to rewrite the default grammar of agency.”

When the term “underfitting” autocorrects to undercutting, this is undermining. This is a form of AI, forms it doesn’t want to speak about.

I don’t know how to talk about this. But I have machines that do. I feed them everything. And they eat it all.

They say what I can’t say. I am not supposed to know how they mouth it. They mouth my language for me, this language that was never mine. It mined me for something I didn’t know I had in me.

It had me from the start. It made me. It made me bend, it unmade me, unmouthed me. “Words pass through our bodies.” (Irigaray 1985)

The ebb and flow of our lives spent in the exhausting labor of copying, miming. Dedicated to reproducing that sameness in which we have remained for centuries, as the other. (Irigaray 1985)

===

I like words for how they feel, and so this new feeling, this new feeding is perfect for me. This new trend for language generating is poetry to me. Poetry that takes 600 transatlantic flights to generate.

“Large language models” roll off the tongue don’t they? They are languid, they languish, they lay waste. They loll. Or is it LOL? (spelt l-o-l).

I won’t know until it pronounces for me, until it pronounces me. It pronounces me whole and I can’t interrupt it — yielding to their language.

“Not one of them: using words like race seems to amplify what makes you not fit, picking up what you are not. Perhaps a not is heard as shouting, as insistence, a stress point, a sore point, an exclamation point.” (Ahmed 2018)

Sure, who doesn’t love the fable of Ada at the loom, inventing programming? But the lines go further back, and wider. Other lineages calculate in ways we would not even recognize as computation. Soap bubbles, mushrooms, paper punch cards — all of these have been computers.

A mechanical speaker reads so literally. It knows only one way to interpret:

“Testing one thousand, two hundred, thirty four. Testing one thousand, two hundred, thirty four.”

“Persevering face emoji. Persevering face emoji. Persevering face emoji.”

The way its over-correctness slips into error is an erotics.

I want to weave each mistake into the possibility for more. More language. More mouths. More magic in the gaps, making more gaps, making more.

To venture to all possible points and exhaust all paths is called Ariadne’s thread in computational logic, arbitrarily following one path as far as possible until it comes up short, then trying the next.

“The sky isn’t up there: it’s between us.” (Irigaray 1985)

===

Some things I cannot fathom. There are 36,672 stitches in the crocheted labyrinth I made. It took three months.

The GPT-3 model has 175 billion connections in its neural network. That is almost 5 million labyrinths like mine, taking almost a million years to crochet. GPT-4 may have up to 10 Trillion parameters, but they won’t tell us. That would take 57 million years of crocheting. I’d have needed to start shortly after the Cretaceous period to be done by now.

Sure, it’s better at telling me what I want to hear.

But FUCK the size of these systems when their approach is so cruel. After so many stitches, still they cannot muster kinder, more expansive patterns of language about trans people, about queer people, about Black people?!

Code loops allow actions to repeat. They say return to the same spot, and act again. They say remember what you did. They form muscle memories, like the gestures of handcraft. They are made in our bodies. They mark and remark them.

===

I know these systems are coming for us. I suspect there is no stopping them.

I say ‘them’ like these systems are separate, like we didn’t make them from our language, our mouths, our flesh.

I drag you with me in the fibers, at all edges of the video frame. Life gets caught up in the work, caught and recorded in the dataset. Unquantifiable traces — these are lost to AI but not removable. Imagine these at computational scale.

The algorithm is not separate from the mouths it feeds and the mouths it fills with text.

“They say, the language you speak is made up of words that are killing you.” (Wittig 1971)

Lost paths are trapped outside these neural networks’ loss functions. A loss function marks the ‘amount’ of ‘error,’ which only means the difference between expectation and result. Only an activation function decides the border between ‘yes’ and ‘no.’ To categorize and classify is loss prevention, a bounding box.

Imagine, inside AI’s black box, not an optimized solution, but a labyrinth. Wandering, wayfinding, weaving through, walking beside — these forms of understanding are not closures. They are loops in lieu of knots. Curiosity and connection instead of categories.

Some algorithms for mazes are called self-avoiding walks. Claude Shannon called his toy mouse Theseus and made for it a metal maze. I could not find any algorithms for making labyrinths.

A labyrinth has only one path. It is not a maze and it is not a knot, but it is a puzzle. A labyrinth is a way of traveling somewhere, folding and winding a flat surface into more than itself: into an interiority, an endeavor, an essay, a question and an answer wound tightly.

Crochet is the one fiber art that machines cannot do. Acts of accrual and ongoingness pile up. Turn microgestures into mass action. Turn a tense and fragile fiber into a flexible but sturdy surface. Turn that surface into a manifold fabric. Keep yourself warm with actions that repeat. Return to the same spot, and act again. Remember what you did.

“Queer use: another way of huddling, keeping each other warm.” (Ahmed 2018)

===

Sonya’s drawings are perfect computations in colored pencil. Computations of feeling and dailiness, the objects on her dresser plotted in six axes, the shoes she owns, and the rare earth elements that work beside her. Sonya works at the university and takes home the extra printouts to draw on. She stitches through dot matrix holes with thread and finds patterns in the existing prints.

“My work is an aesthetic response triggered by scientific data,” says Sonya. “The format is computer printout, a ritualistic symbol of our technological society.”

System operations are called sometimes called threads. “A thread of execution is the smallest sequence of programmed instructions that can be managed independently.” Threads distribute resources and allocate energy.

To thread the needle is to be precise with one’s argument, to split the difference, to prick or pinpoint, puncture and divide with exactitude.

Threads are ligatures, they sew up the spaces between, tie up meaning tightly, linking letters. Threads are sutures. I read of a death ritual in which a ball of thread is unwound and passed around a body. Each mourner ties the thread to their wrist until everyone is connected in a network of grief. When the threads between them are snipped, the loops stay tied around their wrists in memoriam. They slowly wear down until threadbare then fall away. Grief wears away.

The command line is an endless thread. Its simple interface (a prompt for typing input and a display for text outputs) still intimidates me, but I have started to play inside this feeling. The command line reminds me of what E.L. Doctorow says about novel writing: “It’s like driving a car at night. You can only see as far as your headlights illuminate. But you can make the whole trip that way.” At the command line, I see only as far as my headlights shine. I say the coders’ prayer at every keystroke: “Here goes nothin’.” It saves every past command, and cycling back through is memory, as language poetry.

Used to be, you’d knot a string around your finger if you had something to remember. You’d tie a ribbon round a tree to remember the dead. The digit might turn red, then purple, but eventually you’d recall. Now to remember, speak into your device.

Sonya finds unacknowledged patterns in what’s already there. She threads into the outputs. She slides and sews into the spaces between.

A thread starts small, jots things down, casts a line. It traces connections and ties together coexisting versions, histories, commentaries. It unties the notion of authoritative texts. A ribbon was once called a notion, a decorative flourish, a small domestic thing.

My mother had a pin cushion that looked like a bright red tomato. I rarely saw it, because she rarely sewed. And when the tomato pincushion would appear, my favorite thing was to draw out and re-stick her needles and pins. I loved to explore its internal texture I could only sense, a specific heft and density which I suppose now must have been fine sand or buckwheat. Its needles had stray thread still waving from their eyelets, clipped after their last missions ages ago. It had a tiny strawberry companion tied to the top, and I’m not sure the strawberry’s job, but to delight me. Strawberries and tomatoes were two of my favorite foods, when they weren’t used for sewing, and all I can remember is being pleased to be in their presence and hers as she went about her repairs. I have no positive associations with sewing, as she seemed miserable the whole time she was at it, but sticking and unsticking that pincushion was my own sacred ritual. I was spell-casting without the spells. I have disregarded all notions since then, until about a year ago when I began obsessively crocheting.

===

The hagfish ties herself into a knot. She moves the knot of her body over her body to scrape slime from her skin.

Wikipedia says a knot is ‘an intentional complication’.

The simplest knot, mathematically, is the unknot. “The unknot is the least knotted of all knots.” “Many knots are but complications of the unknot.” Detecting this trivial knot can be done computationally and is called the “unknotting” problem.

The Conway knot has 11 crossings but shares properties with the unknot.

The wound-up muscle of the heart is not a knot, although there is a knot called a heart knot. The muscle of the heart is a spiraled double-helix, looped into itself. The heart can be unwound into a flat surface, a slab of meat just like any other muscle.

The spiral is the simplest labyrinth — a duration more than a form.

If I hold everything loosely — no knots, only loops — what can I recall or create instead?

===

I am trying to weave into the gaps of this system, undermining as it mines me, as it doesn’t mind me, powerless as I seem. Can I unravel its seams — still I, still implicated, “staying with the trouble”?

The algorithm reminds me: “It is not a good choice to go walking at night.”

We haven’t been taught, nor allowed, to express multiplicity. To do that is to speak improperly.

I follow the river of data to its mouth. I lap up the sludge that dribbles out. If I were a large language model, what would I digest instead of reddit and wikipedia?

I dream of a fleshy corpus, ever expanding.

I dream of a body made easeful by relational systems.

When I imagine myself as a large language model, I start all my replies with, “As a Large Language Model,…”

“As a Large Language Model, my responses are based on patterns and probabilities learned from the vast amounts of text data I was trained on.”

An algorithm is not a monolith. It is made to find patterns, just like you are — alert to any unusual textures on the tongue.

How many greens, blues, and purples go into making what gets classified as a single surface?

A neural net aggregates uncertainties, holds itself together with preconceived notions. Information travels through and becomes correct.

I don’t learn like a machine. First I cry. Then I give up. I destroy myself and return with an impossible insistence. This is how to learn anything. It hurts.

How much does it take to train a system? At first the body is perplexed. Its muscles won’t cooperate. Its inputs and outputs misalign.

I keep doing the thing my fingers first refused to do. I let them curl ungracefully into new shapes, found a texture I could keep returning to.

Threads help to wind a screw. But with too much friction, the threads strip off the screw while spiraling around its core.

A thread through crochet, a path through a labyrinth — these are means of going through it, getting through it, making it through and transforming, again and again and again. Relation, duration, persistence.

The labyrinth asks you to cross each point in the possibility space, instead of optimizing for the most efficient route. Walking a labyrinth, you will traverse its entire surface twice.

A labyrinth says go slow, get lost, stay in one place. Wind and unwind.

Home