Building AI that gets out of the way

For most of the past year, if you’d asked me whether Newsbin would ever have AI in it, I’d have said no. I’d written whole posts complaining about AI junk. And then I shipped an AI release.

So I owe a retrospective — not on what 2.0 added, but on how I decided to add it, and why I didn’t end up feeling like a hypocrite. The short version: I copied Apple’s homework.

One question, asked over and over

Every time I sketched a feature, I made myself answer the same question before writing a line of code: does this help you read, or does it get in the way of reading?

It sounds obvious. It wasn’t, because most of the AI features I could have built failed that test. A chatbot you have to go talk to gets in the way. A feed that quietly rearranges itself without telling you gets in the way. “Ask AI about this article” gets in the way — now there’s a button demanding a decision on every story. The easy, demo-friendly version of AI is almost always the intrusive one.

The features that passed were the boring-sounding ones. A summary you can glance at and move on. Clustering that de-duplicates a breaking story so you see it once. Tagging that organizes your feed while you’re looking the other way. A “For You” that learns without asking you to teach it. None of them announce themselves. That was the point.

Apple’s stance, borrowed wholesale

I didn’t arrive at this alone. I took Apple’s approach almost directly, because it matched what I already believed and Apple had thought about it far more than I had.

Apple’s whole framing of AI is that it should recede. It lives in the background, it’s personal to you, it’s private by default, and it’s useful without being a destination — you don’t “go use the AI,” it’s just quietly present in the things you were already doing. It’s not a product you visit; it’s a property of the software. Notably, Apple rarely even makes you look at the word “AI.” The feature just works, and the intelligence is implied.

That’s the exact opposite of the era we’re in, where every app has bolted a glowing “✨ AI” button onto its toolbar as if the badge itself were the value. I wanted Newsbin to feel like the intelligence had always been there — not like I’d stapled a language model to a feed reader and called it a version bump.

What the principle ruled out

Once “get out of the way” was the rule, a lot of decisions made themselves. No chatbot. Nothing on by default. Nothing routed through my servers. No AI badges sprinkled around the interface. No feature whose main job was to look impressive in a screenshot.

Saying no to those was harder than it sounds, because some of them are exactly what gets an app attention right now. A chat box would have been a great screenshot. It also would have been a lie about what Newsbin is.

Restraint turned out to be the feature

The hardest part of building the AI wasn’t building it. It was not building most of it — cutting features that worked fine but pulled focus, resisting the urge to surface the cleverness, and letting good work stay invisible.

There’s a strange discipline in shipping something you’re proud of and then hiding it. A summary that’s genuinely good still shouldn’t shout. The measure I kept coming back to wasn’t “is this impressive,” it was “would a longtime Newsbin user even notice a new thing is here, or would they just find themselves reading a little more easily?” The second one is the win. It’s also the one nobody claps for.

The feature that kept getting simpler

Here’s the clearest example. One of 2.0’s features is a plain-language summary of a list — you’re looking at a busy feed, and Newsbin can tell you, in a sentence or two, what’s in it. That’s the whole thing. It sounds almost too simple to have been hard.

It wasn’t simple to arrive at. It started ambitious and got humbler with every pass:

• First, it was a bullet list of the topics in the list. Tidy — but it just turned one list into another list.
• Then I added a thumbnail to each topic. Prettier. Also busier, and no more useful.
• Then I made the topics tappable: click one to filter the list down to the matching stories. Clever! And in practice, I never used it. Neither did the testers. It added a mechanic to learn in exchange for a job you could already do by scrolling.

Three rounds of adding, and each time the honest verdict was “this isn’t actually helping.” So I did the thing that’s hardest for a developer: I deleted all of it and shipped the boring version. A short summary. No chips, no thumbnails, no filtering. And that’s the one people actually use, because it answers the only question you had — what’s in here? — and then gets out of the way.

That whole arc is 2.0 in miniature. The best version of the feature was the one left over after I finally stopped adding to it.

On your device, or your own key

The privacy decision was the easiest of all, because it was already the whole brand. Newsbin’s intelligence runs on your device with Apple Intelligence, or through your own API key if you want higher-quality results. Nothing is routed through me. Nothing is used to build a profile. And all of it is opt-in — turn it off and Newsbin is exactly the app it was before.

That constraint made the features better, not worse. When you can’t phone home, you’re forced to build things that respect the person in front of the screen. It’s the same reason the app never had trackers: not because privacy is a marketing angle, but because building without surveillance makes you design more honestly.

What it quietly unlocks

There’s a second reason I’m glad I built the summarization layer the way I did: it’s a foundation, not a feature. Once Newsbin can reliably distill an article — or a whole list — into a few plain sentences, doors open that were closed before.

A summary is just text. But text can become speech. And that makes a few things I’ve wanted for a long time suddenly plausible: a CarPlay app that reads you the day’s news on your commute, or a TV app that plays spoken summaries and video segments while you’re making dinner. Newsbin on the screens where you can’t really read at all.

I want to be clear that none of that is close — these are far-off ideas, not a roadmap. But they were essentially impossible before 2.0, and now they’re merely hard. That’s the quiet value of getting the foundation right: the exciting part isn’t the summaries themselves, it’s everything they make possible next.

Did it work?

Honestly, I don’t fully know yet — that’s the nature of shipping. But I know I didn’t build the thing I was afraid of building. Newsbin 2.0 doesn’t have AI in your face. It has AI in the background, doing small useful things, staying out of your way, and asking your permission first.

If you’ve used it and the AI never really got your attention — if you just noticed your feed felt a little calmer and more organized — then it did exactly what I hoped. The best version of this stuff is the version you forget is there.

That still feels like the right answer. Build the tool. Then get out of the way.