The Liberation Window
Why Every Information Technology Starts as Freedom and Ends as Control
I. The Press (1450s)
In 1517, Martin Luther nailed ninety-five theses to a church door in Wittenberg. Or he didn't — historians debate the nailing. What is not debated is what happened next: within two months, his arguments had spread across Europe. Within a year, they had been translated from academic Latin into German and printed in editions of thousands.
The Church had managed the written word for a millennium. Monks copied manuscripts by hand, one page at a time. A single Bible took years. The cost was prohibitive, the access controlled, the interpretation mediated. You did not read the text. You received the text, interpreted, from someone authorized to interpret it.
Gutenberg's press broke that. Not gradually — shockingly fast. By 1500, fifty years after Gutenberg, an estimated ten million books had been printed. The number of books in Europe had increased by a factor of several hundred. Literacy followed. Vernacular languages gained authority. The idea that any literate person could read scripture directly — sola scriptura — became not just a theological position but a technological fact.
This is the liberation window. It is real. It is not metaphor. A technology arrives that the old gatekeepers cannot control, and for a period, information moves differently than it has ever moved before. New voices enter. Old hierarchies lose their chokehold. The world is genuinely more open than it was.
The window lasted perhaps sixty years at full openness.
By 1557, Queen Mary I of England had granted a royal monopoly on printing to the Stationers' Company of London. Only licensed members could print. The Catholic Church published its Index Librorum Prohibitorum — the Index of Forbidden Books — listing texts that could not be printed, owned, or read. Governments across Europe developed censorship regimes. The technology that had liberated text became the infrastructure through which text was controlled.
The mechanism was not complexity. It was not malice, necessarily. It was simpler: once you needed a press to reach an audience, and presses were trackable, the press became a chokepoint. Whoever controlled the chokepoint controlled the message.
The liberation window closed. The word did not stop spreading — but it spread through channels that the powerful had learned to own.
II. The Airwaves (1920s)
By 1927 there were 732 radio stations broadcasting in the United States. Labor unions had stations. Churches had stations. Universities had stations. Individual hobbyists built transmitters in their kitchens and broadcast into their neighborhoods. The airwaves were a commons in the most literal sense: anyone with equipment could speak, and the idea that you would need permission to do so hadn't fully formed yet.
The window lasted seven years.
The chaos moment came in January 1926, when Zenith Radio Corporation's Chicago station WJAZ began broadcasting on an unassigned frequency. The Commerce Department tried to stop them. Zenith took it to court and won — Judge James Wilkerson ruled the government had no legal authority to specify operating requirements. The Commerce Department lost control of the airwaves overnight. Stations could do whatever they wanted. The interference that followed was real, and immediate, and genuinely unmanageable.
The Radio Act of 1927 was the response. Signed by Coolidge on February 23rd. A new Federal Radio Commission. To receive a license, stations had to demonstrate they were operating "in the public interest, convenience, or necessity."
That phrase — public interest, convenience, or necessity — sounds reasonable. It was reasonable. It also became the chokepoint. Who decides what's in the public interest? The commissioners. Who gets licensed? Whoever the commissioners approve. The same Commission that solved a real coordination problem also created the mechanism through which commercial broadcasters — better resourced, better organized, better positioned to argue their case — steadily acquired the frequencies that hobbyists, unions, and universities had once occupied freely.
By 1934, the Communications Act had consolidated the structure. NBC and CBS dominated. The window had closed into the architecture we'd recognize for the next sixty years.
This is the structural insight the other examples point toward but radio makes explicit: capture does not require malice. The interference was real. The coordination problem was genuine. Licensing was a reasonable solution to a real emergency. And yet the result was identical to every other capture: the commons became a gatekept resource, and the gatekeepers were not the people who had been speaking freely into the air.
The mechanism is impersonal. That makes it harder to fight, and more important to name.
III. The Network (1990s)
In 1998, if you built a website, anyone who typed the right URL would find it. No intermediary. No algorithm deciding whether your page deserved to exist in a feed. No platform whose terms of service you had agreed to and whose recommendation engine you had no visibility into. You made a page, you put it on the open web, and the open web was between you and whoever was looking.
GeoCities had fifteen million of those pages — organized into neighborhoods, connected by webrings where you clicked "next" and discovered whatever was next, laterally, without curation. SiliconValley for tech. SunsetStrip for music. WestHollywood for community. Fifteen million people with something to say and a place to say it that they had built themselves.
Yahoo bought GeoCities in 1999 for $3.57 billion. Shut it down in 2009. Fifteen million pages — gone. The personal web didn't die because it failed. It died because the audience moved somewhere else, and the somewhere else had owners.
The platform consolidation timeline ran fast. Google became the dominant surface of discovery by around 2005 — which meant that being findable now depended on satisfying an algorithm you couldn't inspect. Facebook opened to the public in 2006. The iPhone launched in 2007, accelerating the shift from open web to app ecosystems controlled by two companies. By 2012, the majority of web traffic flowed through a handful of platforms. The open web was still technically there. You could still build a personal website. But the audience had moved behind walls, and walls have owners.
The breakout attempts were real. RSS let you subscribe to anything without a platform mediating your feed — Google killed Google Reader in 2013, effectively ending mainstream RSS adoption. ActivityPub and Mastodon built federated social infrastructure — technically sound, never reached escape velocity. Open-source software was the one partial success: the infrastructure captured, but the tools stayed free.
The window lasted roughly fifteen years. Less than the printing press. More than radio.
The pattern was becoming visible, for those paying attention.
IV. The Window We're In Now (2020s)
In 2020, running a capable AI assistant required access to a data center. The compute was prohibitive for individuals, the models proprietary, and the only way to use them was through an API owned by a company that logged every query, associated it with your account, and used it to improve the model — or the targeting, or the profile. You got the capability. They got everything you asked.
By 2025, the same inference that required racks of GPUs in 2020 runs on a consumer laptop. The NPUs shipping in mobile chips and desktop processors are powerful enough to run models that, five years ago, existed only in research labs. The bottleneck — compute — has democratized faster than almost anyone predicted.
This is the liberation window. Right now, today.
For the first time in the history of AI, powerful capability does not require centralized infrastructure. You can run a model that reasons, drafts, analyzes, and remembers entirely on hardware you own. The query does not leave. The answer does not travel through someone else's server. The loop closes on your device.
The capture is already being built alongside it.
Cloud AI is the default. The business model is the same one the internet settled into: give the capability free, take the data. Build the product at a loss until the infrastructure is established, then install yourself at the chokepoint. The platform consolidation playbook, being executed in real time, in a market where the players are larger and the infrastructure investment deeper than anything the internet era produced.
There is a pattern in the timeline that is worth naming directly.
The printing press: sixty years before licensing locked in. Radio: seven years. The internet: fifteen. Each window is a different length, but none of them stayed open, and the mechanisms of closure are recognizable across five centuries: infrastructure creates chokepoints, chokepoints attract incumbents, incumbents become gatekeepers. The coordination problem is usually real. The solution is usually reasonable. The capture is usually permanent.
We are, at most, a few years into the liberation window of on-device AI.
The fork is visible right now, if you know what you're looking at. On one side: AI that requires your data to function, that runs on someone else's servers, that you access through a subscription or an API whose terms can change when the business model shifts. On the other: AI that runs on hardware you own, with no network dependency, no account, no server with your name on it.
The difference is not capability. The models are comparably capable. The difference is architecture — and architecture is destiny.
The window is open now, barely, because the compute happened to democratize before the consolidation locked in. It will not stay open. The infrastructure investment required to close it is being assembled.
The question is not whether the window closes.
The question is what gets built before it does.
We built Private Assistant to close the loop on your hardware — because we watched what happened when the loop stayed open.
Private Assistant runs entirely on your device. No cloud. No subscription. No data collection.
Join the Waitlist →