DEV Community

Jiahui Miao
Jiahui Miao

Posted on

Your AI Has No Character Because You Keep Changing the Room

TL;DR: In 1948 Irving Penn wedged the most famous people in the world into a corner narrower than ninety degrees — two studio flats and an old carpet — and photographed how each one dealt with it. Some leaned. Some pushed back. Some refused to be diminished at all. The corner never changed; that is why the pictures could be believed. We do the opposite with AI. Every demo gets a new prompt, new tools, new scenery, and then we wonder why the agent has no readable character and no two evaluations can be compared. Hold the room still. Character is a time series, not a screenshot.

Penn's corner was, first of all, a productivity device, and he said so plainly. Confinement, he wrote, seemed to comfort people: the walls were a surface to lean on or push against, and limiting his subjects' movement relieved him of part of the problem of holding on to them. In an open studio, a sitter has to invent how to stand — and what they invent is their public repertoire, the poses they have rehearsed for years. The corner deleted the invention. It left one physical question — how do you inhabit a space that is slightly too small for you? — and the answer could not be rehearsed for long. Marcel Duchamp met it with unperturbed elegance. Truman Capote curled into a chair, small inside an oversized coat. Georgia O'Keeffe, arms straight, simply simmered. Same angle, same carpet, three different human beings, and the difference is entirely theirs.

Critics have never fully agreed with Penn that the corner comforted anyone; it has also been read as a ruse, a coercive little trap that made people quickly available to the camera. Both readings are on the record, and the honest lesson includes the second one: a constraint is a measurement instrument, and like any instrument it needs an ethics. Used on yourself, or on a willing collaborator who can stop at any time, it reveals. Used on someone who cannot refuse, it just extracts.

Same curtain, same light, 252 trades

Penn then did the thing almost nobody does: he kept the format and changed only the person. For his Small Trades series in 1950–51 he brought 252 workers — in Paris, London and New York — into the studio exactly as they were, in work clothes, carrying the tools of their trade, and photographed them full-length against the same grey backdrop, in the same north light, with the same precision he gave Picasso. The backdrop itself was an old theatre curtain he had found in Paris in 1950; it followed him from studio to studio for sixty years. Dignity, in that series, is not supplied by luxurious scenery. It is supplied by equality of format: the chimney sweep and the celebrity get the same cloth, the same light, the same care — and the famous have to stand out, if they can, on their own.

That is a controlled experiment wearing a photographer's clothes. Background, light and framing are the control variables. The person is the only variable left, so differences can finally be attributed to the person. Most of what we call an AI's "personality" fails this test completely. We change the system prompt on Monday, add three tools on Wednesday, shoot the demo in a hand-built scenario on Friday, and declare the agent charming, or unreliable, or improved. We changed the instrument three times and then tried to read a trend. There is no trend to read. There is only a pile of unrelated screenshots.

What I changed in my own operation

I run a personal AI built for exactly one person, and I now hold its room still on purpose. The recurring checks run in a fixed format, against fixed limits, and I read the reactions, not the poses: does it lean on a rule when the rule is wrong, push back with evidence, curl up into safe vagueness, or stand its ground? A single answer tells me almost nothing. The same constraint, met the same way for the twentieth time, tells me what the system actually is — and a change in that reaction is the earliest warning that something upstream has drifted. My portrait work follows the same contract, for the same reason: one standard grey backdrop, one side light, one small named constraint, shot after shot, so the series becomes a time series of a person instead of a stack of unrelated good pictures.

Penn's second lesson is about the other end of the process. From 1964 he taught himself the platinum-palladium process and spent some twenty years reprinting work he had shot in 1950 — because for him the photograph was not finished when the shutter fired. It was finished at the printing table. Our version is less romantic and just as neglected: the export, the crop, the compression, the version a reader actually receives. If nobody specifies the delivered version, every platform re-authors your work for you. Delivery is not storage. It is the second act of authorship, and it has to be specified and checked as such.

So stop re-shooting your agent in a new room every day. Give it one corner — a fixed format, a fixed limit, a test it cannot charm its way around — and photograph what it does there, over and over. Freedom produces performances. A constant constraint, honestly used, produces evidence.

— Jiahui Miao, founder of vertciti, building a personal AI operating system for one.

Top comments (0)