Lab · Interactive piece
The Glass Ritual
A tiny neural network you run by hand, one station at a time. It is 335 numbers, and every one of them is on this page. Turn them until it names five sigils; then offer it your own, and watch it be fooled.
The artist Kieran Browne built The Other Side: a neural network run by hand, with a wooden apparatus and prayer beads, like a séance. It took about fifty minutes to “perceive” a handwritten 7. Writing about it, Ángel Azamar notes that running the machine as a ritual makes it less mysterious, not more.
This is a small glass version of that ritual. Every step you take is arithmetic you can check: multiply, add, squash, compare, nudge. Nothing is summoned and nothing wakes up. And yet something does happen. A heap of random numbers becomes a thing that tells a triangle from a moon, and you will have moved every number yourself.
The ritual
- I · Cast the circleFill 335 numbers at random from a seed. The same seed casts the same circle.
- II · OfferLay one of 25 offerings on the 7×7 grid. It is given only which squares are lit.
- III · SpeakThe forward pass: each eye adds the weights under the lit squares; each name adds up the eyes.
- IV · ListenCompare its answer with the truth. The gap is the error; its size is the loss.
- V · TurnNudge every number a little against its error (gradient descent).
- VI · RepeatGo round again. Nothing else happens. Only this, many times.
Keys: N next station · R let it run / stop · Esc stop. Speeds 1–3 show every station; 4 and 5 run whole rounds at a time.
It sees, in the only sense this page means: every offering is named correctly. Offer it a glyph of your own ↓
Show the whole sum for this offering
Offer your own glyph
Draw on the grid with a finger, pen or mouse (or focus it: arrow keys move, Space lights or darkens a square, Delete clears). The circle you trained above reads it at once.
Show the whole sum for this drawing
What this is, and is not
It is arithmetic. 49 squares, each 0 or 1. Six “eyes”, each with 49 weights and a bias: 300 numbers. Five names, each with six lines and a bias: 35 numbers. Speaking is adding and multiplying; tanh and softmax are two fixed formulas. Turning is the chain rule from school calculus, applied 335 times. The page checks its own gradient against a numerical one; they agree to within a ten-billionth.
“Sees” is in quotation marks. Here it means only that, for each of the 25 offerings, the right one of five totals comes out largest. It does not know what a moon is. It has met 25 patterns of lit squares, and the names are labels we attached. It does not understand, and it is not aware.
It can be fooled, and it cannot say “none”. Softmax always shares out the whole 100% among its five names, so a blank grid or a scribble still gets a confident answer. Confidence is a number it computes, not a sign that it is right.
Nothing here had to be hidden. Large models are built from the same operations (multiply, add, squash, nudge) at a vastly greater scale, with many more design choices, and scale really does make them harder to read. But Bram Hennekes argues that much of AI’s mystery repeats the old gatekeeping of Renaissance occult circles: methods kept private, specialists as keepers of interpretation. It is mathematically explainable, he says; its opacity is a business choice, not a law of nature. This page makes the other choice.
Why a ritual, then? Amina Inloes asks what older traditions would have called such made things: the talisman, the made servant, the question of whether an object can know. Those metaphors go beyond “demon or demigod”. Roberto Musa Giuliano shows that AI’s language of myth and religion comes from inside the field and shapes which futures people expect. The myths we use are active, so we choose ours openly: the ritual is the frame we built, and the machine inside it is only arithmetic.
Built on
- Ángel Azamar, Visualizing Occult Algorithms: Divination, Predictive Systems, and Artificial Intelligence in Contemporary Art (POST Arnhem, c. 2023). On Kieran Browne’s The Other Side.
- Bram Hennekes, From the Philosopher’s Stone to AI: Epistemologies of the Renaissance and the Digital Age (Philosophies, 2025).
- Amina Inloes, The Golem, the Djinni, and ChatGPT: Artificial Intelligence and the Islamicate Occult Sciences (Theology and Science, 2024).
- Roberto Musa Giuliano, Echoes of myth and magic in the language of Artificial Intelligence (AI & Society, 2020).
Our summaries of all four, and the roadmap this piece comes from, are on Resources.
How it was made: page and code written with AI assistance (Claude) and checked by hand. The network itself is not AI-generated content; it is the small learner on this page, trained by you. Nothing you draw leaves your browser.