Before the Platform
On broadside printing, distributed presence, and the press coming home.
On the night of July 4, 1776, a printer named John Dunlap set type by candlelight in Philadelphia. He worked through the night. By morning he had produced somewhere between 200 and 250 copies of a single-sheet document — rough paper, dense typeset columns, no decorative flourish. The document announced independence from Britain. Dunlap printed it as a broadside.
A broadside was street literature. It was the format of ballads, execution confessions, royal proclamations, and scandal sheets. You posted it on a wall, passed it hand to hand, read it aloud in a tavern. It was designed to travel and then disappear. No archive claimed responsibility for keeping all copies. The message moved through people and then it was gone, or it wasn’t — that was not the printer’s concern.
Of the 200-odd Dunlap broadsides printed that night, 26 are known to survive. Most were used. They moved. Some made it to London, where King George III learned about American independence from a broadside posted in a coffeehouse. The most consequential political text of the modern democratic era — a document that remade the map of the world — traveled in a format explicitly designed to be throwaway.
This is not a minor irony of history. It is a design philosophy.
The broadside was distributed in a way we’ve lost the vocabulary to describe accurately, because the word “distributed” now sounds like a technical specification rather than a way of being present in the world. What the broadside did was something closer to local instantiation: it showed up in a place, for people in that place, and it carried the mark of its own making. The rough paper was local paper. The ink bleed was that printer’s ink. You could tell something about where it came from by looking at it.
Over the next two centuries, the logic of communication reversed. The printing press gave way to the newspaper, which consolidated distribution and gave us the masthead — a name and place to look to for truth. Newspapers gave way to broadcast, which condensed the masthead into a call sign. Broadcast gave way to the internet, which promised to return us to distributed presence and instead gave us platforms: centralized architectures wearing the aesthetic of connection. The message moved farther than ever and carried less of where it came from.
The platform is the inverse of the broadside. It doesn’t disappear once it delivers the message — it stays, it accumulates, it owns the transaction. The reader doesn’t receive information and carry it with them into their own life. The reader visits. Every visit is logged. The platform knows who came, when, from where, what they looked at, how long they stayed. The Dunlap broadside couldn’t remember who read it. That was a feature, not a limitation.
What was lost in this arc is something I’d call distributed presence: the capacity of a message, a tool, a piece of technology to exist with you rather than at a remove you visit. The broadside existed in the tavern. It didn’t require you to go somewhere else to access it. It didn’t have a server. It had paper and ink and hands.
The consolidation happened gradually enough that we normalized it. Software arrived and it was, from the beginning, something you ran on shared infrastructure you didn’t control. First mainframes, then client-server architectures, then the cloud. The shift was framed as progress — more power, more storage, more connectivity. The cost was made invisible: everything you computed went somewhere. Every prompt, every query, every document you processed passed through infrastructure that was not yours, lived in logs you couldn’t read, and could be retained indefinitely by parties you’d agreed to terms with but hadn’t read.
By 2024, the category of “AI assistant” had fully normalized this arrangement. You talked to a model. The model was elsewhere. What you said was the product — or at least the training data, or at least the behavioral signal, or at least the audit trail. The assistant was convenient but it was not yours. You were a visitor.
The thing that most people feel about this but struggle to articulate isn’t quite privacy, exactly. It’s something older. It’s the feeling of a conversation that cannot be a real conversation because one party is recording it. The discomfort isn’t paranoia — it’s a correct perception of an asymmetry. You came to visit. They stayed.
On-device AI inverts this. Computation that runs on your device — not on a remote server, not in someone else’s cloud, not through an API that logs your queries — is the same distributed logic the broadside ran on, applied to inference. The press comes to you.
This is not the same thing as saying on-device AI is better in every dimension than cloud AI. Cloud models can be larger. Cloud infrastructure can be more powerful. The Dunlap broadside was also cruder than a bound book — smaller vocabulary, rougher typeset, no index, no permanence. The broadside’s constraints were real. They were also its meaning.
When the broadside was rough, the roughness was evidence of where it came from. When the ink bled into the grain of the paper, that was proof of local making. The constraint was the authenticity mark — the thing that told you this object had been made somewhere specific, by a specific person, under specific conditions, and that it had not been laundered through any process that separated it from its origin.
On-device AI carries its constraints the same way. A model that runs on your phone is smaller than a frontier cloud model. It doesn’t know everything. It processes faster than it could if it were calling a remote endpoint, because there’s no round trip — but the tradeoff is real and visible. Nothing about on-device AI pretends to be something it isn’t. The rough paper is still rough. You can see exactly what you’re holding.
That visibility is not a liability. It is the beginning of a different relationship — one where the tool is accountable to you because it lives with you, and where you understand the shape of its limitations because you can feel them directly, not infer them from a company’s documentation.
There’s a Japanese practice called kintsugi — repairing broken pottery with gold-lacquered seams, so that the cracks become the most visible part of the object. The philosophy behind it holds that the history of an object, including its damage and repair, is part of its beauty rather than a defect to be hidden.
I think about this when I think about why on-device AI matters beyond the privacy argument. The privacy argument is correct — your data stays on your device, your queries don’t train anyone else’s model, your conversations are genuinely yours. But that argument positions privacy as an absence, a thing that wasn’t taken, a damage that didn’t happen.
The more honest framing is about presence. On-device AI doesn’t just protect you from something. It offers something: a tool that lives where you live, that has the shape of your device’s capabilities, that can’t be updated without your knowledge or disabled without your consent, that doesn’t require an internet connection to think. A tool that is, in a meaningful sense, local. Not as a technical specification. As a way of being present with you in your life.
The cracks are gold because they show you what the object survived. The constraints of on-device inference are gold because they show you what the software doesn’t hide: where it runs, what it touches, what it can’t see.
John Dunlap printed the Declaration as a broadside because that was what the moment called for. Speed, reach, the capacity to be there — in the tavern, on the wall, in the hands of someone who needed to read it that night. Not preservation. Presence.
The platform era made us forget that presence was ever a design goal. It gave us scale and called it connection. It gave us reach and called it relationship. What it couldn’t give us — what was structurally outside its capacity to give — was the thing the broadside had: the message that showed up and was genuinely, unreservedly yours.
That’s what we’re building toward. Not a feature. A different logic. The press, coming home.
Your AI runs on your device. Not someone else’s server.
Explore Private Assistant