Here is what we built, in plain words - and what it actually does for you.
Computers don't know where anything is. Before your phone can load a website, an app, or an advert, it has to ask for directions.
It asks a kind of phone book: "where do I find this?" The phone book answers with an address, and only then can your phone go there.
Normally that phone book belongs to your internet company, or to Google. They see every question your devices ask. You never do.
So we did something simple. We put the phone book in your house, on a computer you already own.
The moment the phone book moved in, every question became visible. And it turns out your devices talk constantly, even lying face-down on a table with nobody touching them.
That one was your phone quietly checking in with Samsung's artificial intelligence service. On a schedule. Nobody asked it to.
Once you can hear the questions, you can refuse to answer some of them. This is the whole trick, and it is gloriously stupid:
We're not hiding the advert after it arrives. We're stopping it from ever being fetched. It uses none of your data and none of your battery, because it never happened.
Here's the problem. Saying no works beautifully - if you know which names to say no to. And the list looks like this:
Over a thousand of those. Nobody is reading that, and nobody knows which ones are adverts.
So an AI reads them instead. It runs on the computer in the corner - not in the cloud, not anyone else's - and goes through the list one name at a time, saying what each thing actually is.
That's the bit that makes it work by itself. Instead of a list of names, the rule becomes: "block anything the AI says is advertising."
So when a brand-new advertising company shows up next week - one nobody has ever heard of - it gets recognised and blocked on its own. No one has to go looking for it.
It checks for new names once an hour, quietly, and adds them.
Before trusting it, we set a trap. We made up six companies that do not exist - names we invented ourselves - and mixed them in with real ones. There is only one correct answer for a made-up name: I don't know this.
The first AI we tried did that for all six fakes. Confidently, every time. If we had trusted it, our blocking list would have been full of invented nonsense and we'd never have been able to tell which parts were real.
So we changed two things. We switched to a different AI, and we changed how we ask - telling it plainly that "I don't know" is a perfectly good answer. Now it gets all six fakes right by refusing to guess.
And it still isn't fully trusted. The screen shows a little tag next to every answer: one colour when the fact is verified, another when the AI guessed it. About a fifth of the names it simply marks "don't recognise this" - and we leave that on the screen rather than filling it in.
That's the honest way round. An AI that admits what it doesn't know is far more useful than one that always has an answer.
Reading the list showed things nobody would guess. The loudest thing on the whole network wasn't a website - it was Adobe, checking in over and over while nothing was even open.
One news app refused to start at all once we began blocking. It turned out the app waits for its advert to load before it will open, and gives up on nothing else. Block the advert and the app simply sits there forever. We had to let that one advert through on purpose.
You cannot find that in any manual. It only shows up when you can finally hear what your own devices are saying.
Being straight about the edges, because everything has them.
None of these are faults to fix later. They're the honest shape of the tool - it is very good at one thing, and that one thing covers a lot.