Dispatch — AI as material
For more than a decade, various thinkers I respect have treated technology as a material with a grain. Recent pieces bring the idea to AI: Nick Foster on distortion, Eric Rodenbeck on model output as ink, Simone Rebaudengo on testing intelligence like wood, and Robin Sloan on software you ask for.
For more than a decade, a strand of design writing has treated technology as a material, something that has a grain you can learn to work with, instead of a tool that executes instructions. Tom Armitage put it that way in 2011, Russell Davies’ post-digital asked for digital and analogue tools to be judged on equal footing, and I wrote about both in 2012 in a piece on eversion for the first The Alpine Review. In 2020, a Dispatch on publications looked at the grain of the web and of print. Two recent essays, an interview and a slightly older talk brought this idea to mind, this time as it relates to AI. Nick Foster finds the grain in distortion, the flaw engineers spent a century removing and musicians turned into rock and roll. Eric Rodenbeck has his students treat model output as ink. Simone Rebaudengo bends and tests intelligence the way you would test wood. Robin Sloan builds his own software by asking for it, then insists the result stand on its own. Collectively they make a case for approaching AI the way a craftsperson approaches a new material: with curiosity, with attention to what it does under pressure, and with a clear idea of which skills you want to keep using.
On distortion
In a 2024 conference talk, futures designer Nick Foster, most recently head of design at Google X and author of Could Should Might Don’t, uses the history of the guitar amplifier to argue that AI’s flaws may be where its most interesting uses lie. Turn an amplifier up too high and the top and bottom of the waveform get clipped, producing a snarl that engineers spent the following century trying to eliminate in pursuit of high fidelity. Musicians heard something completely different and went the other way. Jackie Brenston’s band recorded “Rocket 88” through an overdriven amp in 1951, Link Wray’s instrumental “Rumble” was banned from American radio, and the line runs on through metal, punk and grunge. The Moog didn’t sound like the instruments it tried to imitate, and the Roland 808, built with cheap transistors because memory chips cost too much, gave us “Sexual Healing” and “Planet Rock.” Foster opens with bentwood chair makers, who have to find the direction of the wood or it splinters; the talk is about finding that grain in a technology. He and Simone Rebaudengo made the same point in a 2020 essay (now removed from Medium), also called On Distortion:
We prefer the term “Machine Intelligence” as it allows these technologies space to breathe, to be unencumbered from comparison, and allows the technology to show its own grain, its own vagaries and find its own direction.
—Nick Foster and Simone Rebaudengo
With AI, the engineers’ reflex is the same: iron out the errors until the system meets a benchmark, which assumes we already know what it’s for. For Foster, the dominant framing, set by technologists, investors and LinkedIn, is AI doing what we already do, faster and cheaper. He’d rather look at the oddities. Ask Midjourney for a tall yellow crane over a city and you get a bird, and the back-and-forth that follows, which he calls “negotiation computing,” is a new way of dealing with computers. His examples include Ross Goodwin’s AI-written short film Sunspring, James Bridle’s lopsided chair built from an AI’s instructions, and Simone Rebaudengo and Matteo Loglio’s spoons, generated by a small model and made by one of Italy’s best cutlers. Early AI video, like the Will Smith spaghetti clip, is to him the visual equivalent of an amp turned up too high. OpenAI’s Sora is technically better and duller, “a regurgitation of our existing image and video culture.” The people who push a medium into new territory, he concludes, are rarely the ones writing strategy documents.
AI as a design medium
Eric Rodenbeck, founder of map wizards studio Stamen Design, has spent two years teaching a course at the Harvard Graduate School of Design where students work inside the archives of MoMA, the Harvard Art Museums, the Institute of Black Imagination and the American Museum of Natural History. In an essay on AI as a design medium, he describes what happens when model output is something to work on instead of something to accept. His first rule for students is to interrogate what comes back and stay with it longer than feels efficient. Roy Zhang asked ChatGPT for twenty keywords describing a single image from Houghton Library, twenty times with the same prompt. The early lists were wide and a little strange; later ones narrowed toward conventional descriptions, as if the system had settled. Useful if you want efficiency, says Rodenbeck, and both a problem and an opening if you treat AI as a medium. Kevin Tang and Yuanqing Xie split Dario Calmese’s podcast archive into sentences and rebuilt it around themes, so listeners follow ideas across episodes. Cindy Emefa Coffee and Ieva Lygnugaryte took MoMA photographs whose authors and subjects are unknown and had AI animate the subjects walking out of the frame, using the tool to widen a gap in the archive instead of filling it.
It introduces noise. It reflects back assumptions you did not know you had. It collapses difference. It invents things. It stabilizes too quickly, then surprises you somewhere else. Those are not bugs to eliminate. They are properties to work with.
—Eric Rodenbeck
His other rule is to show your prompts as part of the work: what you asked, what you got, what you thought of it, what you asked next. It slows things down, makes the chain of decisions visible, and gives back some authorship to output that tends to feel anonymous. It also keeps in view a risk he sees in students who can now build whole projects through prompts: fluency without understanding. When Google Maps launched in 2005, a few years into Stamen, he assumed map design was finished; instead, maps became so easy to deploy that everyone wanted their own, and web mapping had a golden age. He expects something similar for AI in design, provided designers ask more of it than speed. Like Foster, he treats AI as a material with properties of its own, including a measurable pull toward the familiar.
AI is a material, not a tool or a magic button
Simone Rebaudengo, who with Matteo Loglio, made the spoons in Foster’s talk, has designed intelligent objects for over a decade, first at Frog and in the speculative collective Automato, now at OIO, the studio the two founded. In an October 2026 interview with Matthaios Tsimitakis for Creatives Unite, he describes how the studio treats intelligence as a material. Thinking about intelligence instead of AI widens the range of options: for an IKEA project on sustainable furniture, OIO took intelligence to mean survival and evolution, and made pieces that adapt to the people who use them. A material, for him, is something you use, bend and twist until you find its breaking point, the way you would with wood, instead of taking it as it is. He traces the idea to the raw text models of 2017 and 2018, and to Robin Sloan, who trained his own autocomplete and found that writing with it was harder. A system that makes the work different, or harder, produces other results and involves the person more than one that promises a perfect result at the click of a button.
Rebaudengo left gallery-bound art because it was “preaching to the choir,” and now puts conceptual ideas into objects people can own: a chair that evolves, Murmur, a speaker with a small generative world inside, the spoon above, made in collaboration with an AI. He places the useful work between the promise of a perfectly automated world and outright refusal, in the choices designers and companies make about transparency, involvement and friction. “With friction comes creativity, pleasure, and learning,” he says. He also points out that today’s language models share one angle, which is to keep people using them, and that hardly anyone tweaks them toward a different point of view, even an adversarial one. Of the four, he is the most direct about why people should build alternatives themselves: if every example comes from companies chasing profit, that becomes the reality, so he would rather show another path than predict where this one leads.
Ask for what you want
In 2020, Robin Sloan wrote about apps as home-cooked meals, software made for one household. Now that AI agents write the code, he revisits the idea and proposes an exploratory program for creative people who are wary of AI: pay for one month of a top model and have it build a desktop app you actually want. On a drive this summer, he decided to stop putting up with his notes app. Before opening the agent, he spent an hour writing a document describing what he wanted, with references to Notational Velocity and a long list of what the app should not do, then asked for a plan and let the agent work. By bedtime he had a native Mac app he still uses every day, later an iOS capture companion, and eventually an email client that shows him only part of his inbox. He doesn’t look at the code. When designer and user are the same person, conventional ideas of good design stop applying, since the logic is always clear to you. The work becomes continuous co-building: nitpicking screenshots over coffee, and ripping out features that didn’t pan out.
What he values in the agents is patience and diligence more than intelligence. They will speed up an app or tidy messy code as many times as you ask, and they keep records of every configuration he never kept himself. He is also wary of depending on them: he wants finished, standalone programs that will keep working if the API fails, gets more expensive or changes, and he can feel his own programming skills fading, something he intends to keep confined to this one domain.
The moment of asking for an app and getting it is dangerous, because you feel like you accomplished something. You didn’t — not really.
—Robin Sloan
His rules for personal AI-built software: use it for two months before posting about it, ideally never post about it, name it yourself, and don’t distribute it. Watching an agent grind through a bug also left him with a visceral unease, the relentless effort of something that, for now, stays inside the world of apps and files.
As the 14-year-old piece shows, my interest in technology as a material—in something closer to craft than pushing buttons—goes back a long way. I’ve never been able to say quite why. Nostalgia doesn’t fit, nor a taste for authenticity, since I’m more drawn to technology than missing another period or looking for authentic things. The closest I can get is that the grain of a material only shows when you push on it. You learn what a technology actually does, as opposed to what it’s sold as, and that’s where the edges and cracks are. The “solidity” I associate with older crafts may really be about limits: you can feel the limits of wood, print or an overdriven amp, while digital products work hard to hide theirs—which also reminds me of a debate from the same period, seamless vs using seams in design†.
Treating something as a material brings those limits back into view, and with them the ability to judge what it’s good for. It also keeps you in the maker’s seat, which all four pieces come back to: Rodenbeck’s students being fluent without understanding, Rebaudengo’s language models tuned to keep us using them, Sloan feeling his own skills fade. The post-digital stance, judging digital and analogue each on its own merits, would then follow from the rest, as the way those reasons turn into practice. AI is the first material that can pass itself off as the maker, which makes it even harder to keep it at some distance, or at least isolated in how we use it. ◼
Further reading
- What I think about when I think about Claude Code. Matt Webb on imagining the software in detail before narrating it to the agent, and the calmer rhythm of coding that way.
- Your AI is not a tool. L. M. Sacasas argues AI is an environment we live in rather than a tool we pick up.
—
† It’s still insane to me that within seconds a memory from years ago, quickly explained to an LLM, can return such a precise and useful answer resurfaced from the exact same scenius I had in mind.
Also, I had a great chat with a reader today, where we both fondly remembered (without having been there) the BERG/RiG/Silicon Roundabout London scene. It’s a bit weird how the thinking from the place and period has, to this day, influenced mine.