The Internet Lives! — Page A6 — The Internet Tribune

Part I: The Factory

The Bespoke AI Model factory sits on four acres on Commerce Avenue, on what used to be the site of the Lakeland Tool & Die plant, which closed in 2009 and sat empty for fifteen years before the town’s founding and the beginning of what people here are already calling “the build.” The original building — the three-story brick structure that Lakeland workers called “the old girl” — has been preserved and restored, its façade cleaned and its windows replaced with the kind of glass that gives back a slightly amber light. The new fourth-floor addition, built in the same brick and with the same proportions, is almost indistinguishable from the original unless you know what to look for, which is: the windows are slightly larger, because the workers inside need more light.

The people who designed this building — a firm out of Milwaukee called Kramer-Pfeifer Architects, which specializes in what they call “industrial restoration with purpose” — said they were told by the company to make it look like a place that had been there for fifty years. “They wanted it to feel earned,” said lead architect Susan Kramer, who attended the ribbon cutting. “They didn’t want a tech campus. They wanted a factory. There’s a difference.”

There is. And inside, the difference is even clearer.

This reporter was given a brief tour of the first two floors on Wednesday afternoon, the day before the public opening, when the building was still being prepared. The lobby — a high-ceilinged room with polished concrete floors and the original Lakeland Tool & Die ironwork restored above the main entrance — smells of fresh paint and something else, something almost imperceptible: the faint, clean smell of electronics doing careful work.

The first floor houses what the company calls the Fitting Rooms: twelve suites, each the size of a small home office, with good chairs and better lighting and a workbench that looks, from a distance, like nothing so much as the counter at a tailor shop. Which is, the company says, precisely the point.

“What we’re doing here,” said Operations Manager Thomas Wierzbicki — no relation to this reporter, who checked — “is fitting. In the old sense. You come in. You tell us who you are. You tell us what you need. We ask questions. We listen. And then we build you something that fits.”

What they build you is a Bespoke AI Model. And that term — “bespoke” — is doing significant work in the company’s vocabulary. It is not, Wierzbicki was careful to say, just a marketing word.

“A bespoke suit is made for one body. It doesn’t work on anyone else. It’s not supposed to. That’s what makes it a suit instead of a shirt you bought at the Piggly Wiggly.” He paused, apparently reconsidering whether to name-check the town’s grocery store in a business context, then seemed to decide it was fine. “No offense to the Piggly Wiggly. But you understand what I mean.”

A Bespoke AI Model, in the company’s telling, is an artificial intelligence model custom-configured — “fitted” — to the personality, communication style, interests, and needs of a specific individual. The model is trained not on generic internet data but on a curated corpus developed collaboratively with the client over several sessions in the Fitting Rooms. The result, the company claims, is an AI that doesn’t feel like a search engine or a customer service bot but something more like a very well-read, very patient, very attentive companion who happens to know a great deal and who has been paying attention to you specifically.

“The difference between our model and the kind you get from one of those big tech companies,” said second-shift supervisor Darla Kaminski, a former administrative coordinator at the old Lakeland plant who joined the new company six months ago, “is the difference between getting your shoes from a machine and getting them from a cobbler. One knows your size. The other knows your feet.”

Part II: What Goes Into a Bespoke Model

Darla Kaminski has a way of explaining things that makes you feel, briefly, like you already understood them. She is fifty-one years old and has the confident, precise manner of someone who has spent her career moving between multiple kinds of work — administrative, technical, interpersonal — and has become fluent in all of them. She grew up twenty miles north of here. She raised her kids twenty miles north of here. When the Lakeland plant closed, she commuted to a call center in Green Bay for six years before she got the offer from the new company and drove down here on a Monday morning in March to see what it was all about.

“I walked in and I knew,” she said. “Not immediately. I need to be clear about that. The first thing I thought was that the lobby smelled too clean. The second thing I thought was that these are serious people who are serious about what they’re making. The third thing I thought was: I want to work here.”

What she works on, in the second-floor Calibration Suite, is the middle stage of what the company calls the “model lifecycle”: the period between initial fitting and final delivery when the bespoke model is trained, tuned, and tested against the client’s actual needs.

The process, simplified considerably for the purpose of this article and any resulting legal liability, works like this:

A client comes in — or, eventually, connects remotely through the town’s internet — and spends several sessions in the Fitting Rooms with a Model Fitter. The fitter’s job is not to take measurements in the traditional sense, but to understand the client: how they communicate, what they care about, what frustrates them, what delights them, what they need in a companion who has read more than any human being ever could and who will, unlike most human beings, always pick up on the first ring.

“You’d be surprised what comes up in those sessions,” said Model Fitter Bobby Stankiewicz, twenty-seven, who previously worked the counter at his family’s auto parts store. “People tell you things. Not secrets, necessarily. Just… truths. Things they haven’t thought about or said out loud in a long time. Like ‘I need someone who doesn’t interrupt me when I’m thinking.’ Or ‘I need someone who pushes back sometimes. My last AI just agreed with everything I said and it drove me crazy.’” He laughed. “We get a lot of that one, actually.”

From those sessions, the fitter generates what the company calls a Client Profile: a document that is part personality sketch, part communication preference map, part list of hard rules (“always cite sources,” “never suggest meal ideas before 11 a.m.,” “don’t mention my sister’s podcast unless I bring it up first”). The Profile then goes upstairs to Calibration, where Darla and her team use it to tune the base model.

“Think of the base model as the bolt of fabric,” she said. “It’s good fabric. It’s quality material. But it’s not a suit yet. Our job is to cut it right.”

The tuning process, which takes between two and four weeks depending on the complexity of the client profile, involves a combination of fine-tuning on curated data sets, Reinforcement Learning from Human Feedback conducted by the factory’s RLHF team (three people on each shift, chosen, Operations Manager Wierzbicki says, primarily for their “patience and ability to disagree politely”), and a final review session in which the model is tested against a battery of scenarios developed specifically from the client’s stated preferences and observed communication patterns.

The result is, by almost any technical definition, a version of the large language models that have existed for several years now — refined, adjusted, configured. A sophisticated user from the broader AI world might question whether this process produces something meaningfully different from what is available through commercial AI services. That sophisticated user was not at the ribbon cutting this morning. Three hundred other people were.

“I know what the tech people say,” Bobby Stankiewicz said, when this reporter raised the comparison. “I’ve seen the forums. I’ve read the Reddit posts. ‘It’s just fine-tuning. It’s just RLHF. It’s nothing you can’t do yourself.’” He shrugged. “You know what you can’t do yourself? Get on a plane to Italy and have someone measure you for a suit that fits like a second skin. The tech is there if you want to hack it. But most people don’t want to hack it. Most people want it to work, and they want to feel like somebody cared that it worked.”

There is a philosophical case to be made here, and the company’s founders are not shy about making it. The before-times internet, they argue — the internet of 1998, of 2002, of the era that gets referenced in their marketing materials with the kind of warm-toned nostalgia usually reserved for Polaroids and drive-in movies — was an internet in which the relationship between user and technology was still human-scaled. In which a website was made by a person, for a reason. In which the algorithm had not yet learned to predict your desires before you knew you had them and redirect them toward purchases.

“The internet used to be a tool,” said the Mayor of the Internet, in a conversation Thursday afternoon while his staff cleaned up the ceremonial ribbon scraps. “A really good tool. Like a library and a telephone and a newspaper and a record store and a post office all in one. And then it became something else. It became a machine that consumes attention. That optimizes engagement. That turns people into products.” He paused. “We’re trying to go back to when it was a tool. A really good tool. Only now the tool is smarter than it was.”

The irony of using a sophisticated artificial intelligence to recapture the simplicity of the pre-AI internet is not lost on the Mayor. He is, it turns out, a man with a good sense of irony.

“Yes, I know,” he said, and smiled. “We are aware of the irony. We’ve been aware of the irony since Day One. We think the irony is actually the point.”

“The internet used to be a tool. A really good tool. Like a library and a telephone and a newspaper all in one. We’re trying to go back to that. Only now the tool is smarter.”

The Mayor of the Internet • Post-ceremony interview

The question of what the town’s new internet is, exactly — beyond the factory, beyond the ribbon, beyond the feeling — is something that will take more than a single morning to answer. What it is right now is a promise. A very specific promise, made in brick and glass and carefully printed signage on Commerce Avenue, that the technology can be made to serve the people rather than the other way around.

Whether it keeps that promise is the story this newspaper was founded to cover. Today, it’s enough to say: the factory is open. The mayor is smiling. The ribbon is on the ground. And something new has begun.