Apple announced Apple Intelligence at WWDC 2024. The rollout started with iOS 18.1 in October 2024, picked up more features through 18.2 and beyond, and by mid-2026 we have about two years of Apple’s AI story playing out in real devices used by hundreds of millions of people. That’s a more interesting thing to write about than any single conference keynote.
I want to focus on what Apple has actually built, because the real architecture is more thoughtful than the marketing around it tends to communicate.
Private Cloud Compute
This is the piece of Apple’s AI story that I find most technically interesting and that gets the least attention in general coverage.
Apple’s approach to AI on device is well-understood: the Neural Engine in their chips handles on-device model inference, keeps data local, and doesn’t require a network connection. Great for privacy, limited by hardware constraints on model size and capability.
But not all requests can be handled by a small on-device model. For more complex tasks, Apple built Private Cloud Compute — dedicated Apple Silicon servers that run inference in a privacy-preserving way. The key claims: data sent to PCC is not retained after the request, not accessible to Apple employees, and the server-side code is verifiable by security researchers who can inspect what’s running. Apple published the PCC software for external review.
Whether you trust those claims depends on your threat model and your level of trust in Apple’s architecture and auditing. But the design intent is meaningfully different from sending your request to a conventional cloud API. The infrastructure is designed so that even Apple can’t read your data. That’s an architectural commitment, not a privacy policy commitment, and the distinction matters.
What Apple Intelligence Features Have Actually Shipped
Writing tools: summarization, rewriting, proofreading integrated across apps. This works. It’s not groundbreaking — similar functionality is available from other sources — but the integration is tight and it’s available wherever you’re writing across the system. The quality is consistent with mid-tier model capability: useful for drafts and cleanup, not a replacement for careful editing.
Notification summaries. Also works. Also occasionally produces summaries that miss nuance or compress information in misleading ways. Apple has had to iterate on this. The failure mode — getting the gist wrong — is more annoying than not having the feature, so calibrating when to trust the summary is an ongoing thing.
Image Playground and Genmoji — image generation features. These are clearly running in Apple’s infrastructure given the model sizes required. The results are in the “useful for casual creative stuff, not professional output” category. About what you’d expect.
Priority notifications and smart summaries in Mail and Messages. The prioritization is genuinely helpful for high-volume inboxes. Not perfect, but a net positive.
ChatGPT integration: Apple shipped integration with ChatGPT (with user opt-in) for requests that exceed what Apple Intelligence handles locally or in PCC. This was a pragmatic decision — Apple’s first-party models aren’t at GPT-4 capability levels — and it reflects Apple being more realistic about the state of their models than the marketing language sometimes suggests.
What’s Still Not Great
Siri. The honest answer is that Siri is much better than it was two years ago and still behind what you’d hope for from Apple’s AI investment. It understands more, handles more complex requests, and integrates with more apps. It still fails in frustrating and somewhat random ways — particularly on anything requiring real-world knowledge or multi-step reasoning. The gap between Siri and the best conversational AI products remains noticeable.
Apple is working on it. The pace of improvement suggests they’re treating it seriously. But two years into the Apple Intelligence era, Siri is the most visible place where the marketing and the experience diverge.
The Developer Tools Side
Apple has been building out the on-device ML tooling for developers steadily. Core ML, Create ML, and the APIs for working with the Neural Engine give iOS and macOS developers access to on-device inference for custom models. The friction has come down significantly — shipping an app with an on-device model used to require significant expertise, and Apple has worked to make it more accessible.
The tradeoff developers deal with: on-device models that are small enough to ship in an app are limited in capability. If you want your app to do something that requires a frontier model, you’re making an API call — and you’re back to the same privacy and reliability questions that apply to any cloud service. The on-device story is strongest for tasks where a well-trained small model can handle the domain.
Apple’s Actual Advantage
It’s not the most capable models. It’s distribution and integration. Billions of devices with Neural Engines, a unified OS experience, and deep integration with the system mean Apple can deploy AI features to a scale that no other company matches for consumer hardware.
Whether Apple translates that distribution advantage into model capability advantage over the next few years is the interesting question. Right now, the capability is conservative by frontier standards. The integration and privacy architecture are the genuine differentiators.
Two years in, Apple Intelligence is a solid foundation that hasn’t yet delivered on the most ambitious parts of the pitch. That’s about where I expected it to be.