We know that Siri AI and other Apple Intelligence features are powered by Google’s Gemini models, but Apple has been at pains to point out that this is not the same as running Gemini on iPhone.
While there are still some unknowns, a far clearer picture is emerging about exactly what all of this means …
Google hasn’t exactly helped matters by using the term Gemini to describe quite distinct things.
Gemini is the name given to a series of Google AI models, while Gemini Assistant is the name of Android’s equivalent to Siri. However, Google has often omitted the “Assistant” part, simply using Gemini to refer to the intelligent assistant.
However, this much is clear: while both assistants use Gemini models, they are entirely separate. Siri AI is not simply a rebadged version of Gemini Assistant.
Apple talks about its own models.
At the heart of this architecture is our third generation of Apple Foundation Models (AFM), a family of five foundation models
However, that sentence continues:
… custom-built in collaboration with Google.
Macworld’s Jason Cross unpacks what Apple has and hasn’t said to conclude that four of the five models are custom versions of Gemini running on Apple Silicon, while the fifth – and most sophisticated – is essentially Google’s standard model running on Google servers but likely using a different set of training data.
Siri AI doesn’t pull info from Google’s web search or knowledge graph; it uses its own. However, Federighi is not claiming that Apple’s models themselves are not based on Gemini code. In fact, he explicitly says the four models made to run on Apple Silicon are “trained using proprietary data with reinforcement learning and refined using outputs from Gemini frontier models.” It’s likely that the biggest model is trained using both Google and Apple’s proprietary data.
So what does that mean? It seems like Apple started with Gemini’s foundation models, optimized and rebuilt them for Apple Silicon and the model sizes it needs, and retrained them with its own data, weights, and guardrails.
Two of the four models run on-device. This means that none of the data ever leaves your device, and so absolute privacy is assured.
The next two run on Apple Silicon chips on the company’s own Private Cloud Compute (PCC) servers. Apple has designed the architecture in such a way that absolutely no data is retained or exposed to either Apple or Google – and this fact is independently verifiable by security researchers. In other words, you don’t have to trust what Apple says: any qualified expert can check for themselves.
With the most powerful model, this is running on Google servers. However, these are servers dedicated to Apple use, and although NVidia GPUs are used in place of Apple Silicon, the company says that the PCC architecture still applies. An Apple security blog says that exactly the same protections apply.
Our core PCC requirements remain exactly the same: stateless computation, enforceable guarantees, no privileged runtime access, non-targetability, and verifiable transparency.
PCC on Google servers is not the same as PCC on Apple servers, but the company appears confident that it is just as safe.
In my view, this is the one area where we effectively have to take on trust that the implementation of PCC principles works as well as it does on Apple’s own servers. By this, I don’t mean to imply that Apple would mislead us, but it is uncharted territory, so there may be vulnerabilities the company hasn’t yet discovered.
FTC: We use income earning auto affiliate links. More.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。