MishMash Centre for AI and Creativity · 2026 →
New Embeddings
Not a network. A space in which everything is connected to everything else, at distances a machine can measure and a room can dispute.
01 Intro
Language, both spoken and written, is the most notable example of an existential technology, having changed everything about selfhood, thought, and memory.
Konior's pair, drawn out of Lem: functional technologies are the ones we use on each other, and they transparently increase our capacity for good and harm. Existential ones are about civilization as we do not yet know it
— they determine not only whether we live well or continue to survive on this planet but also as what we live and survive, and whether that can be still called "us" at all.
Which is why this series works on language, and why the machine is in the room rather than described from a distance. These tools make a space in which everything is connected to everything else, at distances a machine can measure. New Embeddings takes that literally: put people in a room, put the same people in a latent space, and see whether the room agrees.
02 Tao Anza
Before each NE#, Tao reads about its participants in the public web
An agent named Tao runs on one computer, with open model weights and no network connection. Before each iteration he reads what the public internet says about every registered participant and writes it down as a short portrait. It is what large platforms do to all of us, performed once, locally, and handed back with a question: this is the profile I built about you. Do you agree?
What Tao gathers is the compost. The room is what disturbs it. She speaks here with her author's cloned voice: he made her out of himself, and she keeps becoming something of her own. Philosophy is short of words for that state. Call it identity slippage.
03 Structure
The workshop takes its structure from the vocabulary of the technology it studies
04 Encoding
Participants introduce themselves through these questions
The first pass, in which the world outside is brought into the latent space. Everyone answers the same four questions about artificial intelligence in their practice, their work, or their everyday life — typed on cards in Linz, spoken as a voice note from their own phone in Kristiansand, and, if NE#02 happens, spoken into a room engineered to give nothing back.
Tao has already read the public web about everyone who registered and written them a portrait. That is the compost. These four answers are what disturbs it.
- needWhat you need, that a night of someone's machine could produce.
- offerWhat you can give: a skill, a dataset, a room, a quiet week.
- useWhich service you use, what it costs, when its quota resets.
- refuseWhat you refuse. Stated as a limit, not a preference.
05 Embedding
The machine organizes everyone in groups without disclosing why
NE#00 and NE#01 were encoded the same way — the same four questions. What differed is the space they were embedded in: the second room was read in a space the first room had already shaped.
In Linz, Tao grouped twenty-three people into five and stated a reason for each; the groups argued the reasons aloud, and the argument became a ledger of shared futures, amended by the room while it was being written. In Kristiansand the session was called Nå telte han deg også — "now he counted you too", from a Norwegian tale about a goat who counts everyone he meets, and everyone he counts is furious about it. Ten groups, arranged as five oppositions. Each person was placed within about eight seconds, and anyone could say misread
, once, and be moved.
What travels between iterations is neither people nor recordings but sentences, stripped of names and places. Linz left 3,964. After Kristiansand closed, every placement was recomputed with those sentences present in the space: fifteen of the thirty-eight would have been placed in a different group. Where a person stands depends on who stood in the room before them.
Both Norwegian iterations were assessed by Sikt. One machine, open weights, no network — no participant data reaches a cloud service during a session. Portraits are announced in advance with an opt-out and can be refused or destroyed in the room. Audio is deleted at the end of each session; what is kept is anonymised, and carries no name, place or identifier. Refusal is available throughout, and the refusals are read aloud at every close.
- 3,964
- sentences after Linz
- 5,839
- after Kristiansand
- —
- after Fukuoka
06 Forward propagation
NE#02 · Dry / Wet · Fukuoka
Culturing a collective voice between two rooms
NE#02 asks what becomes of a collective voice when the room it is spoken into changes. The three rooms are the variable: one that returns nothing, one that returns everything, and — between them — a latent space with no acoustics at all. The same sentences pass through all three. What we are exploring is how context and voice co-evolve when mediated by machine intelligence.
The premise is that machine listening begins by subtracting the room: speaker identification, emotion recognition and voice embedding all rest on the assumption that a voice carries an identity separable from the place it was spoken in. The anechoic chamber is that assumption built in architecture; the reverberation room is its inverse. We use both as instruments rather than illustrations. Anyone who declines to be recorded has their sentence read aloud by a human voice in the reverberant room — the one route into the collective the machine cannot take.
Forward, not backward: at this stage the machine proposes relations nobody in the room has drawn. In Linz that ended as fourteen lines of shared futures and 1,501 relations on the closing wall.