Sam Altman, talking to a room of interns:
in the next 6 months, we’re close to a world where a descendant of ChatGPT can watch your screen, record every meeting and call, and have perfect context of your whole life.
We are, he added, “only one model generation away from this being incredibly useful”.
We have built this. Not something adjacent to it. This. So it is worth saying plainly what the hard part actually is, because it is not the model, and no model generation touches it.
Your life is not on your screen
“Perfect context of your whole life” is a claim about data, not about intelligence. And your life is not in the window you happen to have open.
It is in ten years of Mail. It is in a Messages database going back to your first iPhone. It is in WhatsApp, which has no supported export at all. It is in your calendar, your contacts, your photos, and a browsing history that quietly records what you were actually curious about at 1am.
On a Mac, every one of those sits behind a lock. Mail, Messages, Safari history and Time Machine are gated behind Full Disk Access. Contacts, Calendar, Reminders and Photos each carry their own separate permission. Apple put those locks there deliberately, and they do not open for a cleverer model.
What ChatGPT can actually reach today
Worth being specific rather than rhetorical, so here is the shipping product. ChatGPT’s “Work with Apps” on macOS reads content through the macOS Accessibility API, and the supported list is published:
- Can read. Xcode, VS Code, JetBrains, Cursor, Terminal, iTerm, Warp, Prompt, Notes, TextEdit, Notion, Quip, BBEdit, Nova, TextMate, Raycast, Alfred.
- Not on the list. Mail, Messages, Calendar, Contacts, Photos, browser history, WhatsApp, anything archived.
Read that list again. Code editors, terminals and notepads. And because it works through the accessibility layer it reads text and only text: Apple’s screen reader cannot describe an image or a layout to it.
So the accurate description of what ChatGPT ships today is this. It can read the text in a window you have open, in an allowlist of apps that is mostly developer tools.
That is a genuinely useful feature. It is also a very specific one, and the specificity is the tell.
An allowlist is a statement about who a product is for, so count that one. Eleven of the seventeen apps are developer tools. Two more are power-user launchers. The remaining four are notepads. Nobody’s mother is on that list, because nothing on that list is where anybody’s mother keeps her life.
Which is fine, and genuinely useful, and exactly what you would expect a company of engineers to ship first: the tool they wanted themselves. But it does mean the promise and the product are pitched at different people. “Perfect context of your whole life” and “perfect context of your working directory” are not the same sentence, and only one of them has shipped.
And the gap between those two sentences is not a model generation. It is a native app, a pile of operating-system permissions, a signed and notarised installer, and a great deal of unglamorous parsing of formats that were never meant to be read by anyone else.
We know, because that is where our time goes. Almost none of our engineering effort is spent on the model. It is spent getting at the data at all, and on the part nobody writes threads about: doing it without the data ever leaving the machine.
Watching and knowing are different products
Watching. Observes you continuously. Sees what is in front of you as it happens. Needs nothing from you except your permanent presence. Comparatively cheap to build, works the same for everybody, and carries one unavoidable property: the only way for it to have context is to never stop looking.
Knowing. You grant access once, to the archive you already own. It reads a decade in an afternoon. And then it does not need to watch you at all.
A product that lives in a browser tab, or in a sandboxed app talking to a data centre, can realistically only build the first one. Not for want of ambition, but because the second requires being trusted with the whole archive, and that is a very different conversation when the archive is going to leave the building.
Ostler asks for more access, not less
Here is the part it would be dishonest to leave out.
Ostler asks for Full Disk Access. That is a larger ask than ChatGPT’s Mac app makes today, and we are not the more restrained option.
But look carefully at who is doing the reading. The app on your Mac reads your Mail, your Messages, your calendar and your browsing history. It builds a graph out of them, on your own disk, and that graph knows a very great deal about you.
We never see any of it. Not a message, not a contact, not one line. There is no server with a copy, no account of yours we can open, no analytics pipe, no support engineer who can look. Nothing for us to sell, leak, or be compelled to hand over, because none of it was ever ours to hold.
The permission you grant is to your own computer. It is not a permission you grant to us, because there is no us in the loop.
The difference has never been how much it sees. It is that nobody else ever does.
Which is exactly where the law lands
This is not a marketing posture. It is the structure of the regulation.
GDPR carries a carve-out at Article 2(2)(c), the household exemption: the regulation does not apply to processing carried out by an individual in the course of a purely personal or household activity. You keeping a record of your own family, friends and correspondence, for your own purposes, is not what data protection law was written to police.
But read the next part carefully. The exemption covers you. It explicitly does not cover a provider: controllers and processors that supply the means for those personal and household activities remain fully within scope.
So the same activity has two completely different legal shapes depending on a single architectural decision. Your assistant builds a picture of your life on your own machine, and you sit inside a carve-out written precisely for that. Your assistant sends the same material to a provider’s servers, and that provider becomes a controller processing personal data at scale.
And it is not only your data. This is the part easiest to miss, and a commenter on Altman’s announcement put it better than we have: your daily life is not composed of you alone.
An assistant that remembers your life remembers your family. Your doctor. The friend who told you something in confidence over Messages six years ago. Somebody you stopped speaking to in 2019. None of those people signed up for anything. They are not your users. When that stays on your machine, they are part of your household record, which is exactly what the exemption exists to protect. When it goes to a provider, that provider holds personal data about people it has no relationship with, who never consented, and who in most cases will never know.
One honest caveat, and it is a real limit rather than a footnote: the household exemption is a shield for personal life, not a loophole for business use. Lean on it to justify processing your client list and it falls away. Anyone telling you otherwise is selling something.
So…?
Altman is not wrong that this is coming. He is wrong about what stands between here and there, and he is wrong about how far away it is.
It is not six months and a model generation. It is not the model at all, and it has not been the model for some time. It is permissions, plumbing, formats, and the question of whose building the data ends up in.
We answered that last question first and built backwards from it. Everything else has been parsing, and the parsing turned out to be the product.
Which is why we are not waiting for the next model generation to arrive. We did not need it. The thing is built, what is left is polish, and polish is measured in weeks.
The architecture is documented here. Questions, corrections, disagreements – [email protected].