How the frieze was made

This page goes with the frieze Machinic Vision: Machine intelligence as infrastructure, shown at ICIS 2026 in Lisbon. It describes each layer of the frieze: where its material came from, which models processed it, and which rules placed it. Every layer is rendered from a recipe, so the frieze can be rebuilt at any length or resolution.

The frieze

The frieze is 15 m long and 1.07 m high. It reads from left to right: a prelude, five stations and a coda. A margin of 2.5 inches at the top and bottom is kept free of text for the grommets. The frieze is composed in eight layers, in this order: the photographs and their parts; the arcs between the stations; the timeline of ideas along the top; the phrases; the prelude; the coda; the art and design row; and the keys.

Photographs

The artist photographed five places between 2022 and 2026. The frieze uses 120 of these photographs.

Twenty of the photographs are printed faintly under their stations. The rest contribute only their parts.

Parts

Two models found and cut out the equipment in each photograph. Grounding DINO (the tiny model, from IDEA Research) finds objects named in a list of words. Each station has its own list. For the port, the list includes crane, ship, hull and container. For the grids, it includes transformer, insulator, lattice tower and cooling tower. SAM 2.1 (the large Hiera model, from Meta) then cuts out each object found. A detection is kept if its score is at least 0.22. A box covering more than 85% of the photograph is rejected.

Sometimes the named parts cover less than 15% of a photograph. In that case the photograph is kept as an unnamed field: the whole frame minus the sky, water, trees and roads. Twenty-five photographs are shown this way, most of them from the grids.

The parts of each station are gathered within that station's stretch of the frieze, in the order of the photographs. They keep their true relative scale, and larger parts sit lower.

Phrases

The words over the photographs come from three collections of text.

1. IS journals. The collection has 3,865 articles published since 1 September 2021 in eleven journals, retrieved from Crossref. The journals are MIS Quarterly, Information Systems Research, Journal of the AIS, Journal of MIS, Journal of Information Technology, European Journal of Information Systems, Information Systems Journal, Journal of Strategic Information Systems, Decision Support Systems, Information & Management, and Information and Organization. Of these, 705 articles mention AI, algorithms, automation, machine learning, generative models, robots or agents in their title or abstract, and the phrases come from those.
2. AI research. The collection has 3,000 arXiv abstracts, the 1,500 most recent in each of cs.AI and cs.CY, submitted between June and September 2026.
3. The technical public. The collection has 5,614 Hacker News story titles about AI, from September 2021 to September 2026, each with more than 20 points.

In each collection, two- and three-word phrases were ranked by TF-IDF: a phrase ranks high when it is frequent in a few documents but not spread across most of them. Common English words and generic phrases were removed. A phrase contained in a higher-ranked phrase was dropped. A phrase had to occur exactly as written in at least two documents. The frieze sets about 25 phrases from each collection. A phrase that appears in more than one collection is set only once, in the first collection that has it.

Each phrase is placed at the photograph that CLIP matches it to. CLIP is an image and text embedding model (ViT-L/14, trained on LAION-2B). The phrase is written as "a photograph of …", and CLIP compares it with each photograph. The phrase goes to the photograph with the highest similarity. If two photographs score almost the same, it goes between them. Phrases from the IS journals are set level and in black. Research phrases rise and are in blue. Phrases from Hacker News fall and are in red.

The arcs between the stations

A line along the bottom of the frieze names each station. Arcs join a later station to an earlier one when the later station depends on the earlier one or recalls it. The sentences on the arcs were drafted by a language model (Claude) and edited by the artist. There are six arcs.

The colours come from a small graph. It encodes each station's main idea, the five places, and the causal links between the ideas. A program reads the graph and colours each arc by a fixed rule:

1. If an idea at the earlier station acts on the later station, the arc is red (builds on).
2. If the later station acts back on the earlier one, the arc is blue (explains).
3. If both are true, the arc is black (the stations cause each other).
4. If there is no causal link, the arc is grey and dashed (refers back).

Improvisers call this kind of link reincorporation: a later scene brings back something from an earlier one (Johnstone, K., 1979, Impro: Improvisation and the theatre, Faber and Faber).

The timeline of ideas

The timeline along the top edge shows 93 works from 1911 to 2026, in four rows: neural networks; symbolic AI, search and evolution; learning from reward; and organizations and systems. It is encoded as a graph, which holds 97 ideas and 162 links between them. Sixty-six of the links join a later work to an earlier work whose idea it reused, and all 66 are drawn as arcs. They are coloured by the same rule as the arcs between the stations. Two further colours appear only in the timeline. A green arc marks a later work that answers an objection to an earlier one. A violet arc marks a later work that reads an earlier idea as the same thing, as when attention was read as a modern Hopfield network. Every work and its full citation are on the history page.

Below the timeline, a row in ochre places ten artists and designers on the same time scale, from Ladislav Sutnar (1944) to Wolfgang Tillmans (2000).

Diagrams, type and the prelude

The diagrams and the arcs are drawn in TikZ. The type is Helvetica Neue, set by XeLaTeX. Behind the prelude, set very faint in Menlo, is part of the program that builds the frieze: the function that gathers each station's parts.

The crystal of agents

The crystal in the coda is grown by a program, not drawn. Agents sit on a diamond lattice 30 cells long, 7 wide and 8 high, divided into seven grains, one for each domain of work. Two of the grains are refined three times, which makes their structure finer. A few agents are missing (4%), the lattice is bent a little, and a few links are rewired across grains. Forty larger white spheres, each linked by thin rust-coloured lines to the agents around it, stand for people. They sit in four overlapping ellipsoids outlined in fine dots, which stand for institutions: a firm, a profession, a public body and a university. Small antennae on the upper surface are sensors. Larger brown spheres on the lower face are actuators, and the pale scaffold below them is what they have built. The result has 11,603 agents and 21,589 links.

The crystal is rendered in Blender (Cycles). The film circles the structure, flies through it and pulls back. It lasts 90 seconds, fades to black and loops. It has 2,160 frames at 1920 × 1080 and 24 frames per second.

The station films

Each station has a three-minute film. In the first minute, the photographs dissolve into one another. In the second, the parts stand out as the scene fades to white. In the third, the station's diagram is drawn over the parts. The sound is a drone rebuilt from the beating tones of a phone recording, with a different centre frequency for each station: 172 Hz for Busan, 124 Hz for Copenhagen, 98 Hz for the grids, 148 Hz for Providence and 212 Hz for the data centre.

The send page and the table

A visitor who sends a photograph gets it back read by the same machines as the frieze. The steps are these:

1. Grounding DINO and SAM 2.1 find and count the parts.
2. CLIP retrieves the nearest phrases from the three collections.
3. A language model (Claude) writes one sentence about the photograph.

The reading page shows five views of the photograph: its parts, their labels, the nearest station of the frieze, a depth map (Depth Anything V2) and a line sketch. Visitors can choose photographs from the gallery and ask in their own words for something to be made from them. Each result records what it was made from, so the table builds a tree of remixes.

At the table, a NextDraw pen plotter draws cards from the photographs. Faces, name badges and number plates are pixelated in the public gallery. Everything is deleted on 31 December 2026 unless the sender asks for it to be kept.

Use of generative AI

The models named on this page are the medium of the work. Claude also wrote, under the artist's direction, the code that composes the frieze and the program that grows the crystal. The artist chose the places and the photographs, their order and the art and design references, and edited the text.