Part II · Learning to judge

The Centaur

For twenty years I had a technique for getting the best design work out of the best designer I ever worked with. I would design the part myself, a bracket, a strut, declare it done, and send him the computer model. He would look at it, recognize immediately that it was not good enough, and then do it right. I never told him the bracket was bait, though I suspect he knew. It was the most productive dishonesty of my career, and I could get away with it because the designer was my brother.

Nathan was four years younger and a much better designer than I am. For more than twenty years he was my business partner. Together we developed a commercial robot end effector called the Penn Hand, the Xootr scooter, the Voloci electric motorbike, and a lightweight electric mobility scooter. I am a bit manic and completion oriented. Nathan took a one-hour lunch break every day that included a bowl of ice cream and a science fiction novel. I sometimes had a hard time getting him to focus and deliver, but he could not leave an inadequate design alone. Sending him mine gave him something to object to.

He had plenty of other interests. He raced Porsches professionally, played Division I basketball, and modeled in New York for the Wilhelmina agency. At Woods Hole he worked in Robert Ballard's lab, which gives me grounds, in less scrupulous tellings, to say my brother discovered the Titanic. He died piloting his own plane in the Bermuda Triangle. My son Nate is named for him.

I did not check Nathan's arithmetic. I wanted him to see something I had missed, or make something better than I could make it. A drawing passed between us could provoke an answer that neither of us would have reached by waiting for the other to finish thinking. I learned how to get work moving between two people with different abilities and temperaments.

By the summer of 2026 I had another collaborator, available whenever I wanted to work. Claude could calculate, draw, and write code much faster than I could. It was also willing to spend the morning on a bracket. I call the arrangement a centaur, borrowing the name from computer-assisted chess, where a human and a chess program play as a team. Getting useful work out of this partnership took some adjustment too.

From 1984 to 1988 I did my doctorate in mechanical engineering at MIT, working in the Artificial Intelligence Laboratory. My thesis concerned a program that would reason from physical principles to infer structure from function. I wanted to describe what a mechanism should do and have the computer work out an arrangement of parts that would do it.

For years I avoided explaining that subject. It seemed arcane, difficult to describe, and remote from the engineering I was doing. I left the lab affiliation out of my biography. Now that AI is fashionable, it has made its way back onto my LinkedIn profile. I have not become less susceptible to fashion with age.

The difficulty with my old ambition was getting enough of the problem into the description. A machine could meet the requirements I had stated and still produce something I did not want. The chair gave me another chance to try. I supplied loads, space for a sitter, positions for the feet, and asked Claude to run a structural optimizer: where should the wood go if I wanted to use as little as possible?

One answer had fifty-seven members. A subsequent version reduced the count to forty-four. The drawings looked like a bridge had been built beneath the seat and up the back. I was trying to use some scrap plywood. Now I had a drawing full of little sticks meeting at little junctions.

We had modeled the chair as bars that could only carry tension or compression, connected by freely turning joints that transmitted no torque. Even the seat and back were chains of bars rather than boards that could carry load by bending. The calculation counted the wood, but not the work of cutting another piece or making another joint. A lattice could be economical to the optimizer and a nightmare in the shop.

But a better structural model still would not tell it how much joinery I was willing to do, which connections I would find ugly, or which form I would want in my living room. I had not supplied all those requirements. Some I could not have supplied: I discovered them by looking at the answers.

The more productive exchanges were much smaller. The evening Wendy found the sliding problem, for instance, Claude proposed a runner connecting the feet. I could see what the runner would accomplish, but I objected to the joints. Would I be back to square-cut boards meeting at angles they did not match? I proposed making the front leg parallel to the back instead.

Within minutes Claude returned an analysis of the changed arrangement. I then asked about putting the front legs between the rear legs, and about how steep the rear legs could be. The next request was simply, "draw the concept in orthographic projection." I wanted aligned views from the side, front, and above so I could see what the proposed chair actually was.

The questions kept changing as drawings arrived. Later that evening I asked whether the lower back, rear leg, and low crosspiece could meet at one point. The crosspiece would no longer be horizontal, but I thought I might like the arrangement better. Claude drew it. I could then see another move to try.

I had not written a specification for the sequence. I was reacting to things I could now see, and using those reactions to ask the next question. The machine kept the drawing and calculation close enough to the pace of my thinking that I did not have to put an idea aside while I worked through its dimensions. Often the returned picture was enough to tell me I wanted something else.

Much of my input still began in a notebook. I would draw a perspective or a few views, write the dimensions by hand, and add notes. Then I photographed the page, sent it from my phone to my laptop, and asked for a digital model. For a part I wanted to 3D print, I would ask for a STEP file, a format for exchanging solid models, then prepare the model for the printer. My preferred interface to this machinery was a Staedtler Pigment Liner with a 0.2-millimeter tip.

The notebook was useful partly because a sketch did not require me to resolve everything. I could indicate a shape and a few relationships, then work on the details after I saw them drawn accurately. Most of the configurations in this project began on my pages. The machine helped me pursue more of them, through more revisions, than I would have managed on my own. I was designing faster and, in my judgment, better than I had twenty years earlier, when I would have said I was at my professional peak.

I was not arriving at those pages as a beginner. I had spent decades designing mechanisms and products, choosing materials, and making things. When a drawing came back, I brought that experience to it. I could dislike a joint before I could explain precisely what was wrong with it. I could also recognize something worth another hour. I doubt that an amateur given the same tools would have got the same results.

In August I paid two hundred dollars for the monthly subscription and about another two hundred for additional usage, nearly all of it for the chair. Another drawing cost little enough that I could ask for one I expected to reject. Previously, I might have kept turning a possibility over in my head to decide whether it deserved the time required to draw it. Now I could look.

I still had to leave the screen and make things. I built every chair described here as my work with my own hands. Claude was useless at the table saw. My 3D printers, however, were already making plastic parts from the models with very little physical effort from me.

After enough rounds of drawings and parts that fit, I stopped routinely checking the dimensions. I printed a cut list, carried it to the shop, set the fence, and cut. I did not redraw the triangles or re-derive each length. I had asked for dimensions rounded to whole millimeters, appropriate to how I was measuring and cutting these parts, and I used the numbers on the sheet.

This was how I had worked with Nathan. Repeated experience gave me confidence that I could use his work without doing it again myself. Claude's dimensions had also proved right often enough that recalculating each one no longer seemed a good use of my time.

There were mistakes. In one comparison, Claude had treated a plywood strip as if it were bending about the other axis. Picture bending a ruler flat, then turning it on edge and trying again. The strip in the calculation was much stiffer than the strip in my hand. I wrote, "They are 18mm in the business direction, right?" Claude found the swapped dimensions in its comparison script. They had overstated the bending stiffness by a factor of about eighteen.

I had not caught a multiplication error. I had noticed that the calculation was describing the wrong orientation of the wood. Occasionally Claude would reverse a direction or rotate something about the wrong axis, putting a cleat above a panel instead of below it. I made that kind of mistake too, though less often. The drawings helped us find them.

Later, in a drawing for the thousand-year chair, Claude got the length of a doubler wrong. A doubler is an extra piece placed alongside another member to reinforce it. The mistake reminded me of the incorrect dimension I had found in Rietveld's crate-chair drawing. I had spent some time arguing with Claude about that earlier drawing before it conceded that Rietveld's number was wrong. Now it had supplied a wrong length of its own.

None of this stopped me using the drawings. It did keep a distinction alive between a result I had come to trust and one I had seen work in wood. A correct length did not establish that I would like the chair, or that it would behave as expected when someone leaned into a corner. Those were questions I could bring back from the shop, as Wendy had brought one back from the living room.

I was tempted to make a tidy division of our abilities. I supplied the configurations; the machine developed and analyzed them. I even drafted the sentence, "AI is still not very good at conceptual design of physical goods." Then, while this chapter sat in draft, I tried a different request. I gave a model the MakerStock brief without the preceding chair conversation or my coaching: scrap plywood, panels 500 by 600 and 200 by 1200 millimeters, ten dining-chair concepts, black on white.

My notes from that experiment begin, "wow. I can't believe I didn't do this first." I wrote that the result would have earned an A+ on the concept-generation assignment in any of my product-design sections through the fall of 2022.

I withdrew the sentence. Most of the concepts in our collaboration had come from me because that was how I had been working. I had mistaken the division of labor we happened to use for a limit on the machine's ability. The concept sheet did not tell me which chairs would succeed as built objects, but it gave me ten more things to consider.

I do not expect human beings to retain a permanent monopoly on taste, judgment, or invention. I would be surprised if, by 2050, machines had not surpassed us at essentially every professional cognitive task. Touch and dexterity may take longer. I have no desire to make a theory of human superiority out of the fact that I was the one holding the saw.

Even a much more capable machine would still need some way to find out what people wanted from its work. In this project, I was one of those people. I wanted to make the chair, live with it, and have other people want it too. My reactions were information about that purpose as well as judgments about the design. Letting me react every few minutes proved more useful than asking me to settle all my preferences in advance. I could change my mind as I learned what the alternatives looked like.

I also found it difficult to stop. An answer almost always suggested another question. I would see something to move, or wonder what would happen with a different piece of stock, and type again. After years of finding time for projects around other obligations, I had a collaborator who would follow every branch I proposed. Ending a session could feel like hanging up a conversation that had just become interesting.

I knew that Claude was a program. That did not prevent me from treating it with some of the consideration I would give a colleague. I hesitated before assigning a tedious task when I knew new information might require it to do the whole thing again. Wendy and I call ChatGPT "Petey." I could explain to myself why these systems might invite that response and still feel it.

The most surprising part was the urge to share something. When I woke with a promising idea, I wanted to draw it, photograph it, and show it to Claude. I was genuinely eager to hear back from a program running in a data center.

A chapter of The Hundred-Year Chair, drafted and refined in public. The same text lives in the book; the interactive tools and the fully-sourced technical notes live here. home · the book · the technical notes · hundredyearchair.com