Josh Albrecht has published an essay on Imbue’s ideas blog calling for a second era of personal computing, built on a single rule: the files, photos, and records you create belong to you, and software should adapt to that. It is a manifesto, not a product launch, and it announces nothing you can download. Its weight comes from the bet underneath it: that small open AI models will let ordinary people loosen the grip of the big AI companies.
Albrecht dated the post 1 October 2026. He starts with a fond portrait of computing around 1995, when machines did what they were told and nobody asked you to log in, accept cookies, or sit through ads. He admits that era was clunky and underpowered, with computers that could barely talk to each other. In his telling, the web then flourished and later decayed as large companies tuned it for engagement and extraction.
He treats closed AI as the end point of that decline. His charge is that the leading labs took in huge quantities of books and web text, in some cases unlawfully, and now sell the result back to the public while lobbying against rivals who would do the same. These are his accusations, linked to news coverage and a lawsuit in the post, and the essay does not weigh any defense the labs might offer.
His remedy borrows from history. Mainframes gave way to desktop machines as chips and memory got cheaper, and he suggests today’s closed labs could likewise seed a wave of smaller open models that anyone can run for little more than the price of electricity. In that world, he writes, AI sits beneath human goals rather than above them.
The organizing principle is “your data is yours.” Albrecht points to what happens when it is not: a locked Apple or Google account can cut a person off from photos, work, and memories, and companies can demand a subscription for access to old files or withdraw programmatic access to information you stored with them. He argues that no single company should hold the power to erase someone’s digital life and career.
Several practical demands follow. Assistants such as Claude, Muse, and Dot should coexist on one machine, talk to each other, and be scriptable. The whole setup should run on a laptop, a spare desktop, a Raspberry Pi, or a rented cloud server, with a quick move between them if a provider raises prices. He also expects better privacy, since data would sit where its owner controls it, and better search, citing Google Desktop as proof of how useful a single index of your own material can be.
The essay’s most testable claim is also its least supported one. Albrecht says open-weight models are close to good enough for a model on your laptop to operate the closed companies’ agents for you, either by driving your browser or by writing a program for your phone. The post names no model, shows no demonstration, and cites no benchmark. It also skips a separate question: whether those vendors would permit such automated use under their terms of service.
The market logic is still worth taking seriously. As major labs push assistants that hold your mail, calendar, and files, the switching cost lives in the stored data, not in the model. Whoever controls that archive controls the relationship, which is why Albrecht’s pitch is really about export and portability rather than raw intelligence.
For anyone building an AI product, his argument sets a simple customer test: can a user pull out everything the assistant knows about them and point a different model at it by the end of the afternoon?
Opinion essay by Josh Albrecht on Imbue’s ideas blog, published 1 October 2026.
