Apple's Private AI Will Run on Google's Servers

[montuori]

There is absolutely no way to process any data without it being unencrypted.

Apple is decrypting the requests; however, your assertion is not true more generally. It's possible to perform inference (albeit slowly) and train (really, really slowly) without decrypting either the model, prompt, or training data. See ConcreteML for some details. My point here is that one's intuition about these things is not always correct.

 

[nt5672]

If you have that expertise then you should give detail! I always enjoy reading well written posts from knowledgeable people. Don’t just dismiss me in an arrogant fashion!

So no experience, just blind trust. Got it.

 

[gummybear2026]

It's on their security blog: https://security.apple.com/blog/expanding-pcc/

"However, until today, those primitives have never been integrated into a comprehensive, end-to-end confidential inference pipeline capable of operating at global scale. That’s what we’ve done with PCC on Google Cloud, which incorporates PCC’s exceptional security and privacy properties at every stage, including the industry’s most comprehensive transparency guarantees that allow external security researchers to verify our privacy commitments."

A security blog is factual? That's hilarious. If the PCC was so secure why is AI in beta mode? Apple wants to cover themselves is the bottom line.

 

[Naraxus]

No, but Google will take it and Apple will be able to deny any wrongdoing.

Care to revise this statement? https://forums.macrumors.com/thread...ain-none-of-googles-gemini-assistant.2483636/

 

[Websnapx2]

Eh???? The overwhelming vast majority of Android devices are not Google hardware. (i.e. hardware they don't own). And Yet Gemini Nano runs just fine on them.

This who notion that Gemini can only run on specific hardware relatively baseless. Gemini runs on Google TPUs (hardware they own) , on Nvidia hardware (the configuration that Apple is using), and on a wide variety of hardware.

Apple licenese the Gemini models ( pragmatically software) so they can run it where they want to. This isn't the 'raw' Gemini models with a slick Apple facade slapped on top as a shiny wrapper.

You have to look at the context of the comment — I was talking about Google servers, not android devices. The context of the conversation was about PCC running on non-apple (specifically Google) server hardware.

 

[lowendlinux]

The cloud is just other peoples computers, if you trust Apple, Google, MS, and Amazon then this should be amazing it should make Siri actually smart and useful.

All of us will need to make a decision fairly soon, Apple was the last large AI holdout that time is now over so folks will have to make a decision before September.

 

[nt5672]

Care to revise this statement? https://forums.macrumors.com/thread...ain-none-of-googles-gemini-assistant.2483636/

Nope, lets wait till its cracked a few months after it is actually working.

 

[eifelbube]

These forums are just an Apple themed outrage factory a lot of the time. Apple can be as transparent as they want, and have many third-parties audit their security, but people's minds are made up. They're going to hate regardless.

I think there is truth in what you observe. Minds are made up, always. But there are also those who defend Apple regardless. Individuals from both extreme camps are easily identifiable.

 

[Luftkopf]

I think there is truth in what you observe. Minds are made up, always. But there are also those who defend Apple regardless. Individuals from both extreme camps are easily identifiable.

Sure, those people do exist, but off the top of my head, I couldn't name any.

 

[deconstruct60]

You have to look at the context of the comment — I was talking about Google servers, not android devices. The context of the conversation was about PCC running on non-apple (specifically Google) server hardware.

That context doesn't make much particular sense to this thread either because a substantial amount of Google Gemini in their data centers runs on Google TPUs/Tensor , not Nvidia nodes that Apple is latching on to. Gemini isn't 'restricted neither to TPUs nor Nvidia to run.

As Apple clarified in more talks around WWDC, they are not running 'pure' Google LLM. It is an customized model, that isn't purely Google's. So the underlying hardware is what Apple selects. Pretty likely Apple selected some general hardware security features ( Intel's and Nvidia for encryption) to minimize the 'lock in' their customized port is running on. ( The Titan key managers have a narrower role that would require some work to port the PCC stack, but shouldn't be huge.)

Apple is suppose to get a multiple billion dollar check from Google every year anyway. It is also likely just easier to take some of that in 'barter trade' for server hours. That is likely at least factor here than some magic pixie dust aspects to the Google server hardware being outlined.

 

[Websnapx2]

That context doesn't make much particular sense to this thread either because a substantial amount of Google Gemini in their data centers runs on Google TPUs/Tensor , not Nvidia nodes that Apple is latching on to. Gemini isn't 'restricted neither to TPUs nor Nvidia to run.

As Apple clarified in more talks around WWDC, they are not running 'pure' Google LLM. It is an customized model, that isn't purely Google's. So the underlying hardware is what Apple selects. Pretty likely Apple selected some general hardware security features ( Intel's and Nvidia for encryption) to minimize the 'lock in' their customized port is running on. ( The Titan key managers have a narrower role that would require some work to port the PCC stack, but shouldn't be huge.)

Apple is suppose to get a multiple billion dollar check from Google every year anyway. It is also likely just easier to take some of that in 'barter trade' for server hours. That is likely at least factor here than some magic pixie dust aspects to the Google server hardware being outlined.

The specific servers Apple is using for training their AFM are only the ones with NVIDIA chips. Google’s other servers do not factor in this conversation. Apple isn’t using it’s data, or the other hardware from it’s data centres so I’m not sure why you are bring them up other than they just exist. That’s why I mention context — the detail is very specific — you initially thought I was referring to android devices and I clarified I was referring to hardware on the server side, not the client side. I think we’re good here.