John Gruber on Apple's WWDC moment of truth: vibe coding, Siri's Gemini secret, and software quality decline

John Gruber on Apple's WWDC moment of truth

Apple's WWDC keynote arrives at an awkward moment. Two years ago Apple announced a sweeping set of AI features under the Apple Intelligence banner. Most of the ambitious ones still haven't shipped. Last year's conference got a partial pass because Apple had already acknowledged the delay. This year, Gruber argues, there is no pass left. It's time to show cards.

The Gemini-as-Siri question

The deal Apple quietly struck with Google is the most consequential unreported story heading into the keynote. Gruber says Google's underlying AI technology will power Siri's foundation models — but Apple will never call it Gemini. Users who don't have a Google account and don't go looking will never encounter the word. The architecture keeps Apple's brand clean while outsourcing the hard model problem to Google.

Whether that works at all depends on one unresolved question: can you build a genuinely good AI system when you don't control the underlying models? Gruber says there's no clean precedent. The counterargument is Cursor — a product built on third-party models that became arguably the strongest example of an AI-native application — but Cursor is a developer tool with narrow scope. An ambient phone assistant touching mail, notes, calendar, and third-party apps is a different problem.

“It really is a big WWDC... Two years ago was the one where they announced the first crack at Apple Intelligence that did not pan out. And didn't pan out in a way that they had to announce in March last year. So that last year's WWDC said it's gonna be postponed a full year effectively... They've announced the deal with Google, very unusual to have a white label version of Gemini that won't be called Gemini.”

The extension model and what's at stake

Gruber expects the agent layer to look like a small app store, probably anchored by ChatGPT, Claude, and Gemini, with Apple deliberately keeping its own branded intelligence more conservative. The architecture serves Apple's brand constraints. Its own Siri will not answer things Apple considers inappropriate; the extensions can, and Apple washes its hands of it.

The commercial logic is also clear. Apple takes 30% (or 15%) of any subscription revenue paid through the extension system. But the bigger incentive, Gruber argues, is ecosystem lock-in. If third-party AI agents can access Apple Notes, Apple Mail, and Safari data natively, users have a reason to keep their digital lives inside Apple's apps. If Apple walls that off, it pushes people to consolidate everything in Google Calendar, Google Keep, and the Google ecosystem instead — or to hand the whole phone over to ChatGPT.

The practical ceiling on the current approach is that ChatGPT remains the only extension in iOS today. Until Claude and Gemini are added, the agent story is largely theoretical.

Vibe coding and the platform contradiction

Gruber puts Apple's response to mobile vibe coding in his top three or four most-watched items for the keynote — and says that if the keynote ends without addressing it at all, that silence is itself a red flag.

His frustration is structural. iOS is nearly 20 years old. An iPad Pro costs $2,000 and is as powerful as a MacBook. You still cannot use it to make software for itself. That was defensible when the hardware was limited and the App Store was new. It's harder to defend now.

He points to BitRig as the clearest example of what's being suppressed: a working, in-market tool for on-device vibe coding that Apple stopped accepting updates for without removing it. Not theoretical disruption — actual functionality, killed in place.

His read is that disruption theory applies here regardless of Apple's preferences. The company doesn't get to choose whether on-device software creation becomes normal. The question is whether Apple builds the best version of it or cedes that to someone else. A limited TestFlight-style model, capped at 10 devices, no App Store approval required, would be a plausible on-ramp — but right now there's no announced plan.

Software quality and the measurement problem

Gruber's diagnosis of Apple's software quality decline is precise. Apple has become very good at fixing things it can measure — app crashes, for instance, are nearly gone. What it has lost is the opinionated, top-down attention to things that don't produce a number on a dashboard.

The FaceTime notification problem raised in the conversation is his example. There is no metric that registers "user was confused about how to add someone to a call." No number lands in front of Craig Federighi. When Steve Jobs ran the product, he was the number one user; his personal frustration with a broken flow would have started a fix the next day. Tim Cook's operational strengths are different, and the organization has taken on his personality.

Gruber is careful to say Cook's tenure has produced real strengths. But on the software craft question, he says Apple has drifted — and points to John Ternus as the executive best positioned to make a course correction back toward opinionated, unmeasured quality. Not a full reversal, just a minor heading change: someone at the top who asks why adding a collaborator to a document looks the way it does, and whether it should.

iMessage search is his other example. It technically returns results. It does not behave the way anyone thinks of search behaving. That gap, invisible to any dashboard, has persisted for years.

The through-line across all of it: Apple's credibility on AI, on vibe coding, and on software craft all get tested at the same keynote. Two years of deferred promises means the 2026 WWDC is the conference where deferred stops being an option.