You Won’t Look At Apple’s Siri AI The Same Way After Watching This (Video)

Here's the story you probably heard: Apple fell behind on AI, panicked, and quietly handed Siri over to Google. So when a much smarter Siri showed up at WWDC, the case seemed closed. Siri was just Gemini wearing an Apple logo. It was a tidy narrative, repeated everywhere, and it turns out to be wrong.

During WWDC, I was invited to a closed-door media briefing with Craig Federighi and his team. Craig took the Google question head-on, first walking us through how a typical chatbot app like ChatGPT or Gemini works, then explaining what Apple built instead. His conclusion was blunt: "The amount of Google Assistant we use is none." No Gemini client code in iOS, none of the models Google deploys to its own customers, and not even Google Search as the foundation of Siri's world knowledge.

So what did Apple actually build? Five of its own foundation models, a family Apple calls AFM. Two run entirely on your device, two run on Apple's Private Cloud Compute servers on Apple Silicon, and a fifth heavy hitter handles the most demanding work. Google's role was in training, not at runtime. Apple used refined outputs from Google's frontier models to tutor its own models, the way a student learns from a great teacher and then has to take the exam alone. If that playbook sounds familiar, it should. It's what Apple did 25 years ago when it built Mac OS X on a licensed Unix foundation and then shaped it into something unrecognizable from its origins.

The piece that impressed me most is the System Orchestrator, a hyper-intelligent traffic cop buried deep in the operating system that decides where every request goes. Ask Siri what everyone is bringing to the neighborhood potluck, and the orchestrator flags that as personal, scours your messages, mail, and notes entirely on device, and answers without a single byte leaving your phone. Follow up by asking what drinks pair well with that menu, and it recognizes you've shifted to world knowledge, sends only the necessary context to the cloud, and never exposes who your friends are or what your texts said. The on-device model is a 20 billion parameter design that only activates the parts of its brain it needs for a given request, usually one to four billion parameters, which is the only reason it can run on an iPhone without obliterating your battery.

Now for the complicated part, because I want to be fair to the skeptics. The fifth model, AFM3 Cloud Pro, really does run on Google servers with NVIDIA GPUs. But Apple extended Private Cloud Compute directly into Google's data centers. Your request leaves your device sealed in an encrypted lockbox, gets processed inside a box that neither Apple nor Google can open, and is destroyed the instant your answer comes back. Nothing is stored, nothing is logged, and nothing feeds the next version of anyone's models. Apple even opened the underlying code to independent security researchers for continuous verification. That is a mountain of infrastructure nobody would bother building if the plan were simply to forward your queries to Gemini.

None of this means Apple has won. The new Siri lives in the iOS 27 developer beta, it's English only for now, the HomePod and Apple TV don't get it, and Apple still has to prove the whole architecture holds up under millions of real users. But "Siri is just Gemini in a trench coat" turned out to be a story that spread because it fit in a controversial tweet, not because it was true.