(Roughly) Daily

What happens when one combines practical carpentry skills with the the abundant theoretical knowledge of AI to build a chair? James Bridle recounts his experience…

In 1840, a surveyor took a chisel to one of the outer walls of St. John’s Church, on Old Haymarket in Liverpool, England. They carved a horizontal line into the stone, a couple of inches long, and below this, three radiating lines, getting wider at the base: a broad arrow. The broad arrow had been used in England since the fourteenth century, and later across Britain and its colonies, to mark government property, but the horizontal line was a relatively new addition. Into this crevice the surveyor would have placed an angle iron, or “bench,” used to support a leveling rod, from which elevations could be taken. Thus, the broad arrow and the line were known as a “benchmark.”

The benchmark on St. John’s Church, recorded in the records of the government’s Ordnance Survey, its official mapping agency, marked one hundred feet above sea level. Chiseling it into the stone meant that someone else could come along later and set up a leveling rod in exactly the same position and take their own reading, transforming it into a datum line: a reference point for further, and repeatable, surveying.

For a few years, the benchmark on St. John’s was the reference datum for the whole country, until it was realized that more accurate measurements should begin at the mean tide level. A pole was set up in the Victoria Docks, also in Liverpool, and the fluctuations of the tide observed for nine days. The resulting mark was used as the fundamental reference until the 1920s, when it was shifted again, to Newlyn, in Cornwall. Benchmarks continued to be carved into churches, civic buildings, milestones, and country walls across the country, all referencing the tidal datum at Newlyn.

“Sea level” is actually different everywhere, but in Britain, it remains a function, plus or minus, of ODN: Ordnance Datum Newlyn. This is the level measured within the town’s tidal observatory, constructed on the South Pier in 1915. The observatory takes the form of a hollow tower in which several floats are suspended, connected to a recording device. An inlet pipe allows the tower’s well to fill and empty with the Atlantic tides. Between 1915 and 2015 the mean sea level at Newlyn increased by 1.83 mm/year—and by 3.8 mm/year between 1993 and 2014. Each change in sea level is reflected in an incremental adjustment to elevations across the land.

The last official benchmark was carved on a milestone located outside a pub in Loughton, Essex, in 1993, before satellite-based systems took over the datum (today, a GPS antenna sprouts from the top of the tide tower in Newlyn). The Ordnance Survey still maintains a record of over half a million benchmarks, although many more have disappeared. Hiking clubs lead expeditions to find them, and hobbyist benchmarkers check them off their lists. When you start looking for them, you see them everywhere.

“Benchmarking” remains in use today as a term for establishing a baseline against which other data can be measured and judged: software performance, product quality, profitability, growth; quality, time, cost. But it’s worth remembering its origins not as a tool for establishing hierarchies, but for leveling; for connecting, rather than separating; and for understanding the whole earth, and everything that happens upon it, as a continuous, indivisible field, stretched across cities, forests, and mountains, and ever susceptible to the rhythm of the tides.

A couple of years ago, I found myself stuck in an apartment in Athens, and in desperate need of distraction. I’d been living with my family on a Greek island, where we had space, and time, and had started to build a house; but at the end of 2023 my partner, Navine, became severely ill, and we had to return to the city we’d lived in for five years previously. She spent nine months in hospital; I visited every day, while also looking after our three-year-old son between nursery school and sleep and school again. I had less space, and a lot more stress, but I also had time to fill, and was itching to use my nascent carpentry skills. I needed to make something, with the limited tools and materials I had to hand—anything would do, it just needed to keep my body, and my mind, busy.

It was around this time that the current wave of what we’re calling artificial intelligence started to excite, and demand, attention. I actually studied AI at university, twenty years ago: exactly the wrong time, it turned out, to be studying it. The history of AI is one of booms and busts: “AI winters,” they were called, when the field seemed to sputter out after a series of advancements, and twenty years ago it was decidedly chilly. We learned about expert systems and neural networks from aging professors, but all the sensible students were learning about classical logic and statistics as preparation for careers in digital finance. It was another decade before the field would regain its energy, and—this is a controversial take—it would have little to do with conceptual breakthroughs. Rather, the growth of the internet, with its accompanying boom in processing capacity and data storage, would allow those same ideas—primarily neural networks—to be supercharged by graphics cards, digitized text and artworks, surveillance and social media, and the wholesale transference of everyday life to the online world. The AI “revolution,” if that’s what it is, is not a technological breakthrough, but the overwhelming accumulation of power and data in the hands of a few companies: power and data which rightfully belong to, and have been stolen from, everyone. But I digress.

I’d been noodling with this idea for some time: What if we used AI for something actually useful? It feels like we’re applying this new tool to all the wrong things: beating us at the games we enjoy, supercharging predatory capitalism, extracting and burning fossil fuels, depleting the planet’s resources, conflating images with art, encouraging extremism, suppressing rights, hallucinating, and so on and so forth. Wouldn’t it be infinitely more interesting if AI helped us do useful things in new and better ways, ways which we could not have imagined alone? Intelligences which we engaged with more like colleagues and collaborators than slaves and masters.

The one thing computers have always been really good at, one of the few qualities which genuinely impress me, is optimization. They can take a human-designed thing, an object or an algorithm, and ruthlessly strip it to the bare essentials while maintaining its functionality. Give a clever machine some widget, like a structural bracket or a complex workflow, and it will produce some alien-looking thing, a third of the weight and twice as efficient, which a human mind would not have conceived. That’s cool, that’s interesting. Of course, applied to culture, society, and the wider environment, such an idea of efficiency is a disaster, but it’s pretty neat for objects, as long as a human is keeping an eye on things.

I thought I might be able to make use of this particular skill to create something with what I had to hand: By combining my rudimentary carpentry abilities with the machine’s knack for efficiency and strangeness, something interesting might happen. Specifically, I didn’t want to design a process which would ultimately replace me, or take over the rewarding bits of work, but something more like a smart carpenter’s mate, that could do some of the sums and make useful suggestions and leave me to do the fiddling and the sanding. A collaborator, not a boss…

[Bridle recounts the first iteration of his collaboration of his collaboration with ChatGPT…]

… There are two qualities of the chair that ChatGPT and I built that stand out. The first is that it looks flimsier than it is, but it works. The AI told me to use a sheet of very thin ply for the seat, so thin that I considered laminating two sheets of it together. It gives a little when you sit on it—but it holds. That’s exactly the kind of optimization I was looking for: a decision I wouldn’t have had the confidence to make, but, again, it works. It might even be slightly comfier that way—although I have no way of knowing if the machine actually intended it, structurally or comfort-wise; it might have been going on dimensions alone. Still, the chair stands.

The second quality of the chair is that it is unnervingly big. Not too big—it’s actually pretty comfortable for a six-footer like me—but it’s definitely bigger than you’d expect, verging on the uncanny. This I find to be absolutely fascinating.

What we have here is, quite literally, a chair designed by someone who has read a lot about chairs but never actually sat in one. They know what a chair should look like, what it should do, the mechanics of how one is put together, the physics of holding you up. But they have no embodied experience of the world, no history with chairs or any other object, no physical memory. They don’t know how it feels, not merely to sit in a chair, but to encounter one in the world, to place it among other things in a home, to live with it, to be around it every day as one goes about the mundane business of life. This is the idea of a chair brought to life; but things—objects, beings, relationships—are not mere ideas, and you certainly can’t get every aspect of them from books. The chair is low stakes, but, once again, you wouldn’t want to run a society, or a culture, or a planet this way.

Since building that first chair, others have been created. This is where the benchmarking bit comes in. A few weeks after the ChatGPT chair, I made another one, this time with an AI model called Mistral. Mistral is a French company which made a lot of claims to be more ethical than ChatGPT (whatever that really means), and I’d been told it was more accurate and precise when it came to technical matters. Once again, I described my request, and we went back and forth a bit, and then I tried to do what it told me.

Mistral’s chair didn’t stand up at all without some serious modifications: I had to add a bunch of cleats and bracing around the legs to make it stable enough to sit on, which eliminated any purported gains in efficiency. It was definitely more interesting than ChatGPT’s doll’s-house effort, with a short back and slanted armrests. It was also waaaayyy too high. My feet didn’t even touch the floor. So much for Mistral’s accuracy and precision—but points for making something a daring European chair designer might actually have come up with.

AI is unfortunately very hot right now. At a time when we could be attending to decarbonizing our societies and building more robust communities to help us through the turbulent years ahead, we seem to be giving most of our attention to a technology which not only burns through energy and other resources, but actively works to make our relationships less secure and our information environment less coherent. The art world is not immune to this tendency, and so when I received requests for AI-themed work for events and exhibitions, I turned the chair-making process into a set of instructions which anyone could follow, and sent it over.

The first iteration took place in Tokyo at an exhibition called “AI is Not Magic” in June of 2025. The curators followed my instructions, inviting Anthropic’s Claude AI to make them a chair from a selection of materials in the studio. It turned out a pretty nice chair, but the quality of the finish made me suspect that an experienced carpenter was involved at their end. Conversely, when a conference called ThingsCon in Amsterdam asked Elon Musk’s Grok AI to have a go, the system told them to build a lounge chair and gave very precise cutting instructions, but none of it fit together at all. As the ThingsCon team noted, “The system could produce—confidently, fluently, at length. But it couldn’t understand what it had made.” The chair they eventually ended up with, after much modification, was “grammatically coherent and practically useless. […] It’s a strange object—part furniture, part evidence.”

Each chair is indeed a piece of evidence—a benchmark, in the truest sense. Each iteration of the work produces a different outcome, which we can judge and compare. Each AI model reveals its own tendencies, biases, and limits, in physical form, easily appraised by a layperson. We are benchmarking; we are benchmaking; we are literally making a bench (fine, a chair, whatever). I appreciate processes where you learn to do a thing, a thing that you didn’t know how to do before, and at the end of that process you have not only done it, but you can see the results. And, in this case, sit on them.

This benchmarking has a deeper meaning as well. Carving a broad arrow into a church wall doesn’t reduce or isolate anything. Rather, it reminds us that every action is made in context; each benchmark refers to every other one, in a continuous, connected line through space and time. When we build, when we think, when we act, we do so in honor of every object, thought, and action which has preceded this one, and we are changed by the process. AI wants us to forget this, by making every output its own miraculous creation. But we owe the world more than that…

[Bridle reflects on the experience…]

… Building the AI chairs taught me something about AI, but it didn’t teach me much about making chairs. Ultimately, I was the teacher, taking what I’d learned elsewhere to make the “practically useless” suggestions into something stable and self-supporting. The AI undoubtedly learned more than me from these interactions: how to speak, what interests (some) humans, how to flatter, how to slant and tailor its responses to my demands.

The anarchist anthropologist David Graeber, may he rest in peace, had a good bit on authority. He said that anarchists don’t oppose authority per se, but they study and critique it, and appraise its legitimacy. What he liked, he said, was self-subverting authority, authority which undermines its own basis. He gave the example of a teacher: A teacher knows something useful to their student, and so they have a natural authority; but if they do their job well, at the end of the process the student knows what the teacher knows, and their authority is spent. The hierarchy is flattened, and no further unequal power can accrue. The student is changed, and so is the world.

Of course it’s possible to learn things from AI, but, for me, real intelligence resides not in an ability to extract and dominate, but in the capacity to share, co-exist, and thrive together. There are many ways of being intelligent, as many as there are living beings, each one unique and valuable. The real challenge of AI is not in building ever more powerful systems—this is difficult, but inevitable—the challenge is to find what relationships between different intelligences benefit us all. The challenge is to choose the kind of world we all want to inhabit. Benchmarking—the practical, embodied work of exploration and mediation—is an attempt to find our way towards such a world…

On finding a healthy relationship with AI: “Benchma(r)king” from @jamesbridle.bsky.social in @emergencemagazine.bsky.social.

* Ludwig Mies van der Rohe

###

As we measure twice before cutting, we might recall that it was on this date in 1985 that the first discovery of a fullerene was published in the journal Nature. In September of 1985, the American chemists Robert F. Curl, Jr. and Richard E. Smalley, colleagues at Rice, met with Sir Harold W. Kroto of the University of Sussex, England. In 11 days of research, they discovered the first fullerene, C60, a spherical cluster of carbon atoms. The discovery of this new structure (AKA “buckminsterfullerene” or “buckyball”) opened a new branch of chemistry, and all three men were awarded the 1996 Nobel Prize for Chemistry for their work.

source