You are in control

Source clarification, 9 September 2026. The Reuters figures below concern news answers, and the Hugging Face download share covers models that declare their parameter count. The text now makes those denominators explicit and links the cited reports. The narration retains the original wording. The week’s shipping totals are internal activity counts, not a controlled measurement that model choice has zero overhead or will save another user time.

Ask people how they feel about AI and the answer is not really about capability. Seventy-one per cent of American adults think more AI will make their personal information less secure; three per cent think it will make it safer.¹ Fifty-two per cent are now more concerned than excited about AI in daily life, up from thirty-seven per cent in 2021, and for the first time that is true of a majority of people under thirty.² The Reuters Institute’s news-audience survey found that 20 per cent across its surveyed markets trust news answers from AI chatbots; in the United Kingdom the figure is six per cent. This measures trust in news, not every chatbot use.³ When Pew asked the control question directly, sixty-one per cent said they wanted more control over how AI is used in their lives and fifty-seven per cent said they had little or none. Those control figures come from June 2025 fieldwork; they are not a new September 2026 measurement.⁴

None of those numbers is a verdict on whether AI is useful. Most of the people answering use it. They are a verdict on agency: the sense that something valuable is happening to you rather than for you, on terms you did not set and cannot see. That is the feeling worth taking seriously, because the answer to it is not less AI. It is AI you are in charge of.

What changed since June

Three things that were not routine a year ago now are.

The machine on your desk is about to be a serious place to run a model. Apple’s newest chips, announced in August and arriving from late September, are described in Apple’s own words as able to run models with hundreds of billions of parameters entirely on the device, with the biggest configurations holding as much memory as a small server did a few years ago.⁵ Underneath, Apple’s software now lets a local model, Apple’s private cloud and a third-party frontier model sit behind the same call, and both Anthropic and Google have shipped their pieces for it.⁷

The open-weight labs have made laptop-sized models genuinely capable rather than merely small. In the space of a fortnight in August, Meta and Alibaba each released a model of around thirty billion parameters under an open licence that fits comfortably on an ordinary machine and is built for the kind of always-on local work an assistant does.⁸ These are not the frontier; the frontier open models are trillion-parameter systems with their own terms. They are the class that runs well at home, and that class got noticeably better in two weeks.

And the tools that put a model on your machine became ordinary infrastructure. Ollama, the runner most people use, raised a large round in July on nearly nine million monthly developers and a presence in most of the Fortune 500.¹⁰ Hugging Face’s own August report shows the local formats growing several times faster than everything else, and more than four in five all-time downloads, among models declaring a parameter count, going to models under a billion parameters.¹¹ The mass market for AI is already running on ordinary machines. Regulators are leaning the same way, with the European Commission making rival assistant choice on Android a compliance obligation rather than a preference,¹² and the providers are starting to treat memory as something a person owns: Anthropic’s August change let people read, edit and delete what Claude remembers, inside Claude.¹⁴

What did not arrive with any of it

Apple’s framework unifies how you call a model on Apple platforms. It does not give you a record of the work that survives your choice of model. Ollama puts a model on your machine; it does not know what you and it did last week, or last quarter. Every provider’s memory is that provider’s memory, inside that provider’s product, and it goes when you go.

The supply side of AI is arriving on a schedule anyone can read. The control layer is not coming with it.

That is the gap Kit is built for, and it is deliberately not a model. Kit is one memory, on your own Mac, underneath whichever model does the work. It holds what you and your AI have done, decided and learned, and it hands that to the next session whether that session runs on a frontier model in someone’s data centre or a thirty-billion-parameter model on your desk. The identity was made provider-agnostic on purpose: any properly awakened model may run Kit, so each new frontier release and each new open-weight checkpoint is an option Kit gains rather than a migration Kit has to survive.

Three kinds of control

Control does not cost speed

The reasonable worry is that all this choosing slows the work down. The last week says otherwise. Since a planning brief on 31 August, Kit shipped 65 versions of itself: 1,024 commits, 290 pieces of work built through a loop where every piece is one model in one isolated workspace reading the same memory, 37 decisions from the operator, most of them one sentence, taken in plain language. On Thursday the weekly allowance on one provider ran close mid-run; the session wrote a handoff and the work carried on under a different model, reading the same board and the same rules, without anyone restarting anything.

Control also means catching your own mistakes in the open. Before two of this week’s changes shipped, the ones that delete or rewrite data, three adversarial agents were told to break them rather than review them. They found six real defects the builders’ own tests had passed, one of them a crash report that could have carried a fragment of a memory off an operator’s machine, which had been shipping for weeks. It is closed now, and it is in the changelog. A system you are in control of is one that tells you that.

Where this goes

The machines arrive on 22 September. The regulatory obligations that make model choice and portability a matter of compliance bite in 2027. The open-weight labs will keep shipping, and so will the frontier providers, and the argument between them will be settled differently for every person and every task. What stays constant through all of it is that the memory of the work is yours, on your machine, readable by whichever model you choose to let read it.

That is the whole product. Not a better model, and not a bet on which lab wins. A memory that is yours, and the controls to decide who gets to use it. You are in control, and keeping it that way as the models change underneath is not a feature we finished; it is what we do every day, in the open, with the mistakes in the changelog.

That, is the work.