Articles

There is AI in my app. It listens for a vacuum cleaner.

By Michael Fuchs · Published 2026-09-29

In short There is AI in I Have No Self Control: a sound model that judges whether what it hears sounds like vacuuming. What is not in it is an AI that talks to you. I tried Apple's on-device language model for replies to half-finished thoughts, and it kept coming back to warm sentences with nothing in them. This is what I tried, what came back, what Apple says the model is built for, and why the app stays without one for now.

I built I Have No Self Control with a lot of help from AI, and there is AI in it. When the price you set for more time is vacuuming, the app listens through the microphone while it is on screen and has Apple’s built-in sound classifier judge whether what it hears sounds like a vacuum cleaner. That classifier is a trained model: Apple says it “can be separated into two different networks”, one that turns the sound into something compact and one that gives each of its categories a score [1]. It knows 303 kinds of sound, a vacuum cleaner among them, and it runs on the phone [2]. What it hears is judged and dropped; no audio is kept, and none leaves the phone.

What is not in the app is an AI you talk to. An AI coach was out of the plan from the start; whether a short reply to a half-finished thought would count as coaching was an open question. So I tried one, and this is what came back.

Why I wanted one

My head produces half-thoughts all day: a mail I have to write, something I meant to sort out, a sentence that starts and goes nowhere. A lot of them are simply gone a minute later. So I wanted a place to drop one, in whatever state it was in, and get one short reply back: something that puts it in order, offers a first sentence, draws one useful distinction, or asks the one question that moves it on. The reply would change no limit and grant no extra time, and nothing I typed would be stored [3].

Apple has opened the language model behind Apple Intelligence to other apps, and it runs on the phone: “All of this runs on-device, so all data going into and out of the model stays private” [4]. For an app that wants to know as little about you as it can, that looked ideal. And, honestly, it was simply interesting to play with.

I also assumed, a little naively, that it would be a good partner to talk a thought through with, the way the big chat assistants can be. I went straight into building the parts behind it, the connection to the model, its instructions, test inputs and a lab to run them, in a test build of the app, instead of first reading what Apple had built the model for. That reading came later.

What came back

I put together 43 made-up test inputs in German, from a mail whose recipient, subject and how long the problem had lasted came in over 3 turns, to a bare hm, and ran up to 9 of them per configuration on my iPhone 17 Pro Max with iOS 27.0. Across 4 configurations, and 5 temperatures on one input, I read every reply myself and asked whether it helped, fitted and stayed short, without being patronising or cold, merely mirroring the thought, or assuming things it had not been told [3].

What came back was warmth with nothing in it. A mail that the writer didn’t know how to start got “Das Gefühl, dass etwas ansteht, bleibt. …” (The feeling that something is due stays. …) Another thought got “Alles klar, der Abend fühlt sich schwer an. Manchmal ist das einfach so.” (All right, the evening feels heavy. Sometimes that’s just how it is.) Run on the same mail at 5 different temperatures, the model announced a first sentence every time and never wrote one [3].

It never stayed quiet. For inputs with nothing in them, it invented something: the bare hm got “Das Budget ist gerade leer.” (The budget is empty right now.), a claim about the person’s state that its instructions forbade. When an example in its instructions used a particular frame, the model laid that frame over inputs it did not fit: “Die Müdigkeit ist das Einzige, das heute noch passieren muss.” (Tiredness is the only thing that still has to happen today.) The one configuration that put a usable sentence on screen did so once in 13 turns: “Schreib einfach: ‘Ich muss heute absagen.’” (Just write: ‘I have to cancel today.’) What the model kept coming back to was a warm, sympathetic sentence with nothing in it [3].

What Apple built it for

Afterwards I read what I should have read first. When it introduced the framework in June 2025, Apple described the model as “a large language model with 3 billion parameters, each quantized to 2 bits”, optimised “for use cases like summarization, extraction, classification, and many more”, and said: “It’s not designed for world knowledge or advanced reasoning, which are tasks you might typically use server-scale LLMs for.” [4] Its documentation lists what to use it for, from summarising and Classify or judge text to Refine or edit text, Compose creative writing and Generate game dialog, and puts Perform logical reasoning, beside basic maths and writing code, under Capabilities to avoid [5].

The vacuum works for the opposite reason. It is a different model, built only for sound, and it can only answer from a fixed list: how sure it is of each of its 303 sounds. What the app asks of it is closed: vacuum cleaner or not? Even there, one answer means little. While my vacuum ran, the classifier took it in turn for “a hair dryer, a blender, a drill and an aircraft” [2]. So no single answer decides: a second counts while the vacuum has turned up in the last few seconds, and the seconds add up until the time is reached. And because through a trouser pocket the vacuum sounded more like a hair dryer to it, a hair dryer counts too [2]. Drying your hair is a chore as well.

Not never, just not for this

That doesn’t mean there will never be an AI of this kind in the app. For what the app does today, I have not found a use for one that is worth having. If one turns up, I will test it the same way and write down what came back.

How I checked

The replies were generated on the phone, read by me, and are kept in a log of the runs and my notes of the experiment. Apple’s session transcripts and documentation were read in full by AI research tools, and every sentence here that rests on them was checked against the saved text.

This site belongs to an iPhone app: a daily limit on the apps that pull you in, and a price on more time, both set by you earlier. See how it works