Apple shipped the Siri it promised in 2024. The delay tells the real story.

What happened

Apple's Apple Intelligence landing page hit 527 points on Hacker News this week — not because the page changed, but because the feature the page has been promising since June 2024 finally works on a shipping iPhone. The 'more personal Siri' demoed at WWDC 2024, quietly pulled from iOS 18.1, 18.2, 18.4, and eventually punted out of the iOS 18 cycle entirely, landed in the iOS 26 developer beta seeded after this year's keynote.

The build that shipped does roughly what the original demo promised: cross-app actions ('send the photos from yesterday's hike to mom'), on-screen awareness, and personal context retrieval that pulls from Mail, Messages, Calendar, and Photos without any of it leaving the device or Apple's Private Cloud Compute (PCC) cluster. The on-device model is now confirmed as a ~3B parameter foundation model running on the Neural Engine; the server model runs on Apple silicon servers inside PCC with attestation receipts that third parties can verify. The Foundation Models framework — the developer-facing API that lets third-party apps call the on-device model — is out of beta and stable for the iOS 26 SDK.

The HN thread isn't celebrating; it's autopsying. Top comments are forensic: why did this take two years, what changed architecturally, and is the result actually competitive with what GPT-5 and Gemini 3 can do in 2026. The answers are more interesting than the launch itself.

Why it matters

The 2024 Siri demo was a fake. Not in the bad sense — Apple was clear it was a 'preview' — but in the sense that no working prototype existed end-to-end. Reporting from Mark Gurman and later Apple insiders confirmed what was already obvious by mid-2025: the team had a planning model that could decompose user intent, a retrieval layer that could pull personal context, and an action layer that could execute App Intents — but stitching them together with acceptable latency and acceptable hallucination rates was a genuinely unsolved problem at Apple's quality bar.

Apple's quality bar is the story. A consumer-grade LLM that confidently sends the wrong photo to the wrong contact is a product Apple cannot ship. OpenAI can A/B test ChatGPT behavior on 500M users and roll back in an afternoon. Apple ships a binary to a billion devices and lives with it for a year. The asymmetry forces a different architecture: smaller, more constrained, more deterministic, with hard guardrails on what the model is allowed to do without explicit user confirmation. That architecture is what shipped — and it's also why the feature took two years longer than the keynote implied.

Compare the approaches as they stand in June 2026. OpenAI is shipping GPT-5-class agents that operate a virtual browser and can hallucinate themselves into a refund loop. Google is shipping Gemini-on-Android with Pixel-exclusive 'magic' features that work great in demos and inconsistently in the field. Apple spent two extra years building a system where the model cannot execute an irreversible action — sending a message, deleting a file, moving money — without an OS-level confirmation surface that no third-party app can spoof. That's not behind. That's a different bet.

The community reaction is split along predictable lines. The 'Apple is behind on AI' camp points out that ChatGPT integration is still the escape hatch for any query the on-device model can't handle, and that the on-device model's benchmarks (Apple's own published numbers from the 2024 ML research blog, never independently reproduced) are well behind frontier models. The 'Apple is playing a different game' camp points out that no other vendor has shipped a verifiable end-to-end attestation chain from on-device prompt to server-side inference, and that Private Cloud Compute is the only major-vendor architecture where a regulator can audit what the model was allowed to see.

What this means for your stack

If you ship an iOS or macOS app, the Foundation Models framework is now real and stable. The honest assessment: it's a 3B-parameter model with App Intents integration, not a GPT-class reasoning engine, and you should architect accordingly. Use it for structured extraction, summarization of content the user already has on-device, and intent-routing to your existing logic. Don't use it for open-ended generation, don't use it for anything where wrong answers have cost, and budget for sub-second response latency for short prompts and 2-5s for longer context windows. The on-device model has no internet access by design; if your feature needs current information, you're routing through PCC or your own backend.

For agentic workflows, the App Intents surface is now the primary integration point. If your app doesn't expose its capabilities as App Intents with proper parameter schemas and undo support, Siri will not call your app — it'll call a competitor that did the wiring work. This is the equivalent of the 2009 'is your app on the App Store' moment for AI-mediated discovery on Apple platforms. The apps that ship intent schemas in the next six months will be the ones Siri actually reaches for.

For everyone else — backend, web, Android — the relevant takeaway is the Private Cloud Compute architecture as a design pattern. Attestable inference, where the client can cryptographically verify which model binary processed its request and that no operator could see the plaintext, is a primitive that didn't exist at scale before Apple shipped it. Expect to see this pattern adopted by regulated industries (healthcare, finance, government) over the next 18 months, and expect the major cloud vendors to start offering 'attestable inference' as a product line.

Looking ahead

The two-year Siri delay will be studied as either Apple's biggest AI mis-step or its most disciplined product call, depending on what happens to the agent-driven AI market in 2027-2028. If consumers turn out to tolerate frontier-model unreliability for the upside of frontier-model capability, Apple's caution will look like a strategic blunder. If the first major class-action over an LLM-initiated wrong action lands in 2027 — and the architecture that ships is the one that can prove the model was constrained — Apple will look like the only adult in the room. Either way, the WWDC 2026 keynote was the first one in three years where the AI section ended with a feature you can actually use, not a feature you'll see next year.