Letting an agent touch your machine is usually all or nothing. Ambient Desktop takes the four risky parts, browser page, scraper, package, MCP server and runs them in a sealed box that has no route to your files.

You spend an hour getting an AI to do something right, then next month you start from scratch. Ambient Desktop takes the run that worked, punches the specifics out into blanks and keeps it as a tool you can call.

What does it mean to write into American defense law a standard whose triggering finding cannot be tested by the company it is used against and cannot be checked by any laboratory in the country writing it?

Every computer you own spends most of its life doing nothing: so does most of the hardware that runs AI, and that gap is the whole idea we have been working toward. Let us start with size. If a model is small enough to run on hardware people already own, the network running it stops being a handful of enormous data centres and becomes a lot of ordinary machines. Ordinary machines have gaps: the minutes between jobs, the hours nobody booked, the box that is busy for 8 hours of a 24 hour day. Idle time is strange stuff economically. It is the only computing capacity in the world with nobody bidding for it, because it exists whether or not anyone uses it, and it cannot be saved for later. Either something useful happens in that hour or the hour is gone. So the question becomes what is worth doing with it? Our answer is reading. A year of AI research arrives as papers almost nobody has time to connect to each other, and models with no deadline and no boredom can read them, write a page per idea, link every claim back to the paper it came from, and check those pages again when a new paper contradicts an old one. The network answering ordinary requests is real and running today. The reading project is what we intend to point the spare hours at and when it is producing pages we will publish what it got right and where it went wrong.

This is big: 1221 people with $20 of compute each just audited a third of a major machine learning conference and an open model is what made it affordable. Hugging Face ran it from 15 July to 2 August. In 19 days the participants published 6,816 logbooks covering 2,226 ICML 2026 papers and judged 35,908 separate claims. Of the papers examined, 23% had at least one claim falsified or contested, 49 had every claim fall and 242 drew opposite verdicts from independent teams working the same claims. The judge that read all 35,908 claims was GLM-5.2, run open on rented hardware. At frontier API prices an audit this size does not get commissioned at all. Open weights are what set the price of scrutiny and the price is what decides whether anyone outside the field ever checks it.

You spend an hour getting an AI to do something right, then next month you start from scratch. Ambient Desktop takes the run that worked, punches the specifics out into blanks and keeps it as a tool you can call.

On 5th August the Cherokee Nation banned hyperscale data centres on tribal and trust land after surveying 1,593 citizens and finding about 64% opposed. So 18 states now have an active ban or moratorium and 25 of 50 are restricting or considering it. But remember that a state rule can be pre-empted from Washington and a county vote can be reversed at the next election. But tribal land is neither. The Cherokee Nation has more enrolled members than any tribe in the United States and it is a sovereign. The industry's entire political strategy is pre-emption. Turns out here there is nothing to pre-empt.

This is big: 1221 people with $20 of compute each just audited a third of a major machine learning conference and an open model is what made it affordable. Hugging Face ran it from 15 July to 2 August. In 19 days the participants published 6,816 logbooks covering 2,226 ICML 2026 papers and judged 35,908 separate claims. Of the papers examined, 23% had at least one claim falsified or contested, 49 had every claim fall and 242 drew opposite verdicts from independent teams working the same claims. The judge that read all 35,908 claims was GLM-5.2, run open on rented hardware. At frontier API prices an audit this size does not get commissioned at all. Open weights are what set the price of scrutiny and the price is what decides whether anyone outside the field ever checks it.

Clicking install is an act of faith. You get a name, a logo, and access to your files. Ambient Desktop ships every default tool with papers: version pinned, sandboxed, an explicit allow list, and a serial matching what actually runs.

You simply cannot have a data policy with a counterparty you cannot name. Ox Alpha appeared on OpenRouter last Thursday (free) from a provider that will not say who it is. 21.8 trillion tokens have gone through it since (almost all of it from coding agents) which means prompts and working source code & Claude Code alone sent 1.57 trillion of that. Two intermediaries have now published a retention policy for that traffic and they contradict each other: OpenRouter's page says your prompts and completions are retained. OpenCode's page says the same provider keeps nothing. Neither of them holds the data so neither sentence is a promise anyone can be held to. People are matching tokenizer counts to work out which lab it is. Okay. What if that code turns up inside a competitor's product, whose name goes on the complaint?

Clicking install is an act of faith. You get a name, a logo, and access to your files. Ambient Desktop ships every default tool with papers: version pinned, sandboxed, an explicit allow list, and a serial matching what actually runs.