‘Siri’s 100 percent not into that.’ It’s a line Craig Federighi used at WWDC that does more than deflect romantic role‑play requests — it signals how Apple wants this next Siri to be understood: a utility stitched into your device, not a chatty companion vying for attention.
What changed at WWDC
Apple rolled out a third generation of Apple Foundation Models (AFM) and reintroduced Siri as a much more capable assistant. The suite now spans five models — from compact on‑device engines to heavy cloud systems — and promises features that feel genuinely integrated: quick cross‑app lookups, richer dictation, image understanding and even agent‑style tool use for complex tasks.
On stage and in follow‑ups, Apple framed the work as practical and privacy‑minded. Greg Joswiak said the company wasn’t doing "AI for AI’s sake"; the point, he emphasized, is making existing iPhone features actually better without forcing users to become prompt experts.
The hybrid under the hood
AFM 3 looks like a hybrid answer to competing approaches. Highlights that Apple disclosed:
- AFM 3 Core and AFM 3 Code Advanced run on devices. AFM 3 Core Advanced is notable: conceptually a 20‑billion‑parameter model that uses a sparse activation architecture, so only a few billion parameters light up depending on the request. That’s how Apple puts unusually big models on phones without burning battery or RAM.
- AFM 3 Cloud and ADM 3 Cloud (Image) power server‑side workloads like fast inference and image diffusion for the new Image Playground.
- AFM 3 Cloud Pro is the heavyweight server model that handles the trickiest reasoning and agentic features.
The practical outcome: smaller, private requests can stay local; heavier, multi‑step tasks can be escalated to cloud models when needed.
Why Google and NVIDIA entered the picture
Apple has long pushed to keep AI close to its silicon and data centers. But getting the performance and capabilities it wanted required partners. For AFM 3 Cloud Pro and the most demanding workloads, Apple extended its Private Cloud Compute (PCC) footprint onto Google Cloud and NVIDIA GPUs.
Apple is careful with how it describes that partnership. Federighi stressed Apple does not simply resell Google’s assistant or rely on Google Search as the knowledge base. Still, senior engineers acknowledged that outputs from Google’s Gemini Frontier models played a role in refining AFM during training — a pragmatic step to accelerate progress.
That compromise lets Apple access massive GPU fleets and sophisticated model tooling while trying to maintain control over model behavior and data flows.
Privacy architecture: PCC on third‑party clouds
To bridge Apple’s privacy promises with third‑party infrastructure, Apple expanded PCC beyond its own data centers. The security pitch is layered and technical: confidential compute primitives (NVIDIA Confidential Computing, Intel TDX, Google’s Titan chip), a cryptographic ledger of approved hardware, dual attestation roots, tight process isolation, short‑lived inference software, and an auditable supply chain. Apple says devices will only trust PCC software cryptographically approved by Apple.
Apple also pledged to publish binaries for inspection, offer research access to PCC nodes, and keep the same verifiable transparency that previously applied only to Apple silicon. Whether outside researchers will be fully satisfied remains to be seen — but Apple seems to be betting that engineering controls plus transparency can make cloud execution acceptable to privacy‑conscious customers.
A Siri that stays useful, not seductive
Federighi’s blunt refusal to let Siri become an AI girlfriend isn’t just PR theater. It’s a design choice shaping interaction: Siri is being positioned as an assistant that resists sycophancy and avoids engagement tricks that drive addictive behavior. Marcus Mendes’ technical readouts and Federico Viticci’s WWDC notes both underscore another point: Siri’s real advantage may be context — access to the nuts and bolts of your device and apps in a way third‑party chatbots can’t replicate.
That advantage depends on the app ecosystem. Siri’s usefulness will grow only as developers expose actions and as Apple lets the assistant act across apps. (If you’re curious about how Apple plans to expose Siri in new ways, it’s worth keeping an eye on the company’s plans for a standalone Siri app and business hub.)
Tensions and trade‑offs
There are trade‑offs baked in. Relying on Google Cloud and NVIDIA gives Apple scale and speed, but invites scrutiny over what ‘private cloud’ means when the metal and GPUs aren’t Apple’s. Apple’s answer: cryptographic attestations, code approvals, and an architecture designed to eliminate privileged access.
There’s also a product trade: by refusing to be an engagement engine, Siri may cede some of the conversational flair other assistants use to keep people returning. Apple’s bet is that people prefer a quiet helper that gets things done and respects privacy — if it’s actually useful enough to earn that trust.
Where this could lead
If Apple can deliver the promised on‑device fluency and make cloud escalations seamless and private, Siri could become an invisible but powerful layer across iPhone, iPad and Mac. Add richer inputs — cameras on accessories, for example — and the assistant’s capabilities widen; early reports suggest Apple is testing hardware that would give Siri new senses, like cameras on AirPods prototypes, which could change how the assistant interacts with the world (/news/airpods-cameras-late-testing-siri).
WWDC’s newsscape felt like a season shift: not a dramatic pivot to a single chatbot product, but an engineering effort to weave AI into software people already use. Whether that thread holds depends on execution — the model quality, app integrations, and how convincingly Apple defends the privacy claims it’s making.
Siri, reborn, is being pitched as practical and restrained. That’s deliberate. It’s also oddly refreshing.
