The flaw is the feature.
“It doesn’t look right.”
That’s usually how it starts. A homeowner, halfway through a kitchen renovation, gets a call from her contractor. The second tile shipment has arrived. She goes to look, holds one up against the wall, and frowns. She ordered the same thing she ordered three months ago. Same product name, same catalog number. But something is off. The color isn’t quite matching. The surface catches the light differently. The grout lines, once she steps back, reveal a seam that shouldn’t be there.
I spent two years as a brand manager for a Japanese tile manufacturer. Sales handled those calls. My job was something harder to name: bridging the world of the customer who sees only the finished surface, and the world of the people actually firing clay in a kiln. Because the gap between those two worlds is not a quality control problem. It is, at its core, a philosophical one.
The complaint isn’t wrong. That’s the first thing to understand.
Ceramic tiles, particularly those made in the yōhen (窯変) style, aren’t cookie-cutter products. The color in the catalog is a representative sample, not a promise. What actually comes out of the kiln depends on the iron, copper, and cobalt content in the glaze; the temperature curve during firing; where each piece sits relative to the heat source; even the season, because humidity inside the kiln shifts with the weather outside. These variables work together — or against each other — to produce the surface texture, the pattern, the particular quality of light that bounces off a finished tile. Two pieces from different production runs won’t be identical. Sometimes two pieces from the same run won’t be. Each one is, in a meaningful sense, its own thing.
What made my job difficult, and what I found fascinating about it, is that the thing customers sometimes didn’t like was also the thing that made the product worth making. Bridging that understanding felt critical. Because “I don’t like how it looks” and “this is a defect” are not the same sentence, even when they feel that way to the person saying them.
What the Kiln Keeps
There’s a word I want you to know: keshiki (景色).
It means “scenery,” but let me explain what it actually carries. In ceramics, keshiki refers to the unplanned variations that fire leaves behind — flux marks, color gradients shifting from rust to ash to deep indigo within a single piece, subtle warping at the rim where heat moved unevenly. These aren’t defects. They aren’t even considered imperfections. They’re read as evidence of a specific encounter: this clay, this glaze, this kiln, this fire, this moment. A landscape compressed into an object. You’re not looking at a flaw. You’re looking at a record.
The yōhen (窯変) aesthetic takes this further. Yōhen means something close to “kiln transformation.” The potter sets the conditions. The kiln does the rest. What comes out is genuinely unknown until the door opens. This might sound like romantic imprecision, but it’s closer to a design philosophy — one that builds the unknown into the process on purpose, because the results that emerge from real heat, real chemistry, and real time cannot be replicated by any other means.
What the homeowner noticed in the kitchen backsplash was the seam between two separate kiln firings. She wasn’t imagining it. The tiles were different because they are different. The question underneath her frustration isn’t whether someone made an error. It’s what we expect from objects that carry the visible record of how they were made. Whether we’re willing to look past the surface and find what’s in there.
The Model That Came Out Different
I’ll be honest: when I started writing this piece, I wasn’t sure the comparison would hold.
The keshiki on a fired bowl is beautiful. Then I found something that genuinely surprised me: if you train the same AI model twice with identical data and settings, the outputs aren’t identical. Close, but not identical. The process looks identical from the outside. What comes out doesn’t. Something in there resists exact repetition, and no one can fully explain why.
I sat with that for a while.
The reason, it turns out, is that randomness is structural in how these systems learn. And here’s the part that stopped me: the field discovered early on that removing that randomness makes the output worse. A technique called dropout randomly switches off parts of the system during training. On purpose. The result is something that handles the unexpected better, precisely because it was never allowed to rely on everything going perfectly.
The instability, introduced deliberately, produces robustness. The kiln, it turns out, understood this before we did.
景色 and What Nobody Designed
There is a phenomenon in large AI systems that researchers still don’t fully understand: at a certain scale, after enough repetition, models begin doing things that weren’t explicitly taught. Reasoning through problems. Making connections across unrelated domains. Things that look, uncomfortably, like inference.
No one designed this in. It appeared.
The closest thing I know to this in the physical world is keshiki — the scenery that emerges when a kiln does its work at scale, over time, under conditions no one fully controlled. The potter didn’t plan the particular gradient of indigo fading into ash at the rim. The gradient happened, given the right conditions and enough heat.
What both share: the interesting thing didn’t come from the instructions. It came from what happened when the process ran long enough, with enough variables that the maker couldn’t control.
We tend to describe this in AI as mysterious, even slightly alarming. I wonder if the ceramicist would find it mysterious at all.
The Explanation That Runs Out
Ask a ceramicist why a particular yōhen piece turned out the way it did and they can tell you the contributing factors. Temperature. Glaze chemistry. Placement in the kiln. But the explanation runs out before the phenomenon does.
AI researchers face the same wall. Why did the model weight this feature the way it did? The field has an entire research area dedicated to that question, and it remains largely unsolved.
Both the kiln and the model produce outputs that exceed the explanation of their inputs. The maker sets the conditions. Something else does the rest.
I find this less alarming than most people seem to. Growing up in Asakusa, objects were understood to hold more than their materials. Craftspeople have lived with this for centuries. We’re only now catching up.
What Are We Trying to Make?
There is a version of this story where the tile complaint gets resolved cleanly. The manufacturer improves quality control. The variation between lots narrows. The homeowner gets a backsplash that matches the catalog exactly.
And something is lost in that resolution. Not dramatically. Just quietly, the way things always get lost when a process gets fully optimized.
I’ve been thinking about what it would mean to apply that same logic to AI. To train out the randomness. To eliminate the variation. To produce a system that gives the same output every time, reliably, cleanly, without the seam showing.
We already know how to do this, technically. And we keep choosing not to, because the models that retain their structured randomness perform better in the real world. The variation isn’t a problem to be solved. It’s doing something.
The yōhen potter and the ML researcher are, in this sense, working with the same material: a process that produces more than it was told to. Both have learned, through different routes and different centuries, that trying to remove that excess tends to remove the thing that made it interesting.
So here is the question, and I don’t have a clean answer for it.
We are building systems of extraordinary capability. Systems that learn from more human output than any single human could read in a lifetime. Systems that, at scale, begin doing things nobody designed them to do.
What are we trying to make?
If the answer is a tool that produces correct outputs reliably, that’s one kind of project. If the answer is something that carries the record of what went into it, something whose surface, looked at carefully, tells you something about the conditions of its making, that’s a different project entirely.
The kiln doesn’t know which one we want. Neither, I suspect, does the model.
But here’s the thing: the potter always knew. Before the firing, before the door opened, before the keshiki revealed itself, there was an intention. Not a guarantee. But a direction.
We keep talking about what AI will do. We talk less about what we’re trying to make, and why, and for whom. That question doesn’t live inside the model. It never did.
It lives in the room where someone decides to open the kiln.
Further Reading & Resources
engawa: In Praise of Friction — On why the smoothest path isn’t always the most meaningful — and what craft can teach product thinking when variation isn’t a bug.
engawa: Wabi-Sabi in the Age of Perfect Algorithms — Imperfection, intention, and what we ask of systems that are trained toward an ideal.
FAQ
What is yōhen in Japanese ceramics? Yōhen (窯変) means “kiln transformation.” It refers to the unpredictable changes that occur during firing — shifts in color, texture, and surface quality that result from heat, glaze chemistry, and atmospheric conditions inside the kiln. No two yōhen pieces are identical. The term describes both the phenomenon and the aesthetic that embraces it: that what the kiln does on its own is often more interesting than what the potter planned.
What is keshiki and why does it matter? Keshiki (景色) literally means “scenery.” In ceramics, it refers to the unplanned variations that fire leaves on a surface — flux marks, color gradients, subtle warping at the rim. These aren’t considered flaws. They’re read as a record of a specific encounter between material, heat, and time. A landscape compressed into an object. The concept matters because it reframes imperfection not as failure, but as evidence that something real happened.
What does a kiln have to do with AI? More than it might seem. Both kilns and machine learning models produce outputs that exceed the explanation of their inputs. Both perform better when a degree of randomness is preserved rather than eliminated. And in both cases, the most interesting results tend to emerge not from perfect control, but from what happens when a process runs long enough under conditions the maker couldn’t fully predict. The parallel isn’t decorative. It’s structural.
Taishi Okano writes about the intersection of technology, craft, and culture from New York and Tokyo. engawa is where he works things out.