A collaboration with Google is about as appealing as announcing a collab with Kanye West.
No thanks. Do not want.
No thanks. Do not want.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple Explains How Gemini-Powered Siri Will Work
- Thread starter MacRumors
- Start date
- Sort by reaction score
switz
Contributor
If Apple wanted the customer base to actually know what the heck is actually going on, they would have been all over the web and ads on their devices shouting the news. Saying nothing means there is no official wording that has to be retracted if this AI efforts fails dramatically or even a little bit.
Plausible deniability is the word of the day!
Tim tosses the word salad very carefully to cover his backside.
If there is no official wording, then the public will be unable to determine the scope of the failure, quite unlike those of us that acquired iPhone 16s on false promises.
So keep your sinus cavities clear, the better to sniff the bovine byproducts when presented.
Plausible deniability is the word of the day!
Tim tosses the word salad very carefully to cover his backside.
If there is no official wording, then the public will be unable to determine the scope of the failure, quite unlike those of us that acquired iPhone 16s on false promises.
So keep your sinus cavities clear, the better to sniff the bovine byproducts when presented.
Tagbert
macrumors 604
Hello, fellow Jayhawker,
This article and the announcement that it was based on were done to counter some misleading statements and assumptions (looking at your Gurman) that gave the impression that the new AI would run on Google’s AI servers. That prompted a lot of people to assume it meant that Google would have direct access to your prompts and data and would be able to train on that. We’ve seen a lot of people commenting on this site to that effect in the past couple of weeks.
This announcement, that Apple’s new, Gemini-based AI would still run on Apple’s Private Cloud Compute OS, should help show people that their privacy is still being maintained. We don’t know if that OS will run on Apple hardware or if it will run on Google’s TPU servers, but that hardware is less important than the OS that maintains the data integrity. If there does turn out to be a plan to run this on Google hardware, that could have been the cause of Gurman’s assertion.
This is only the most basic information about how this all will work. Rumors suggest that Apple will present more detailed information around the end of Feb and will release the next beta around that time, too.
Tagbert
macrumors 604
Start by reading this about Private Cloud Compute. We have also heard that Apple will reveal more about how this integration with Google’s AI will work around the end of Feb.
Private Cloud Compute: A new frontier for AI privacy in the cloud - Apple Security Research
Secure and private AI processing in the cloud poses a formidable new challenge. To support advanced features of Apple Intelligence with larger foundation models, we created Private Cloud Compute (PCC), a groundbreaking cloud intelligence system designed specifically for private AI processing...
jayhawk1941
macrumors member
Sure I do. You're changing the topic. I'm fully aware that we don't have 100% certainty with what happens to our data, regardless of the company. That said, I trust Apple 1000 times more than I trust Google. Back to the original point, I would like to see exactly how the handoff and retrieval of my data / queries between services is handled. And as I said earlier, there has been no transparency on what happens when a query is made. What route does that information take, how is it anonymized and encrypted, and how is it processed?
jayhawk1941
macrumors member
Rock Chalk! Thanks for the clarification. That was my understanding too (that Apple will use Google's Gemini AI on their own servers). I just wish they come out and state exactly how it works though rather than giving the most high-level overview that glosses over what I consider to be extremely important details in regards to privacy. Hopefully we hear more in the coming weeks.
switz
Contributor
There is a big difference between a dedicated connection between the processing companies and using the open internet to get there for processing. We must not forget the required NSA inter-connections for 100% monitoring as well.
Buying a new device for Apple Intelligence features is like buying a new car solely to upgrade your gear shift knob.
AgeOfSpiracles
macrumors 6502a
Black_Mage
macrumors 6502a
Good. That addresses my only concern with this partnership. I am surprised Google agreed to it.
lazyrighteye
Contributor
Am I crazy, or was what Tim said carefully worded to specifically NOT say that? The direct quote is:
What he said is that there will continue to be operations running on device and in PCC. Not that ALL operations, or that THE NEW operations will run on device or in PCC.
For example, if a hypothetical car manufacturer was currently building cars in the US and Canada, and had recently opened a factory in China, it would be entirely truthful for them to say, "We'll continue to build cars in the US and Canada and maintain our industry leading working conditions in doing so."
The quote I would have like to have seen is "These new unlocked experiences and innovations will continue to run on device and in Private Cloud Compute." But that's not what he said.
Maybe I'm paranoid, but until they discuss how it works I'm not convinced.
johnsawyercjs
macrumors 68000
I suspect we won't get a level of detail like that in any verbal statements from Apple, or even in written statements in response to questions, since those statements are intended for general-level consumption. The real details will be spelled out in papers they've written, as with their current papers on the inner workings of things like Private Cloud Compute, Apple Machine Learning Research, their Platform Security Guide, etc., all of which are publicly available but go into details that very few people understand.
Tagbert
macrumors 604
That is what Apple described when they first announced it. I was just speculating, based on Gurman suggesting that the AI might run on Google’s servers. Google has some very capable TPU-based AI servers. It is not inconceivable for Apple to port PCC to run on a TPU server. If so, that would explain Gurman. It would still be privacy protecting because it’s the PCC OS that maintains that privacy. Apple already uses Google and AWS and probably Azure servers to host some of their stuff. It’s still secure because the communication and data storage are encrypted.
This was just speculation based on a rumor, not a serious assertion that this will happen.
Tagbert
macrumors 604
i hope that we get some more details in February when they start talking about this. I don’t expect a white-paper level treatise, just something with a little more spelled out would help a lot of the people who are nervous about Apple partnering with Google.
johnsawyercjs
macrumors 68000
True, it shouldn't take a white paper to describe just the details relevant to the general public.
There already is a white paper
.
“ …
Introducing Apple’s On-Device and Server Foundation Models …”
Introducing Apple’s On-Device and Server Foundation Models
At the 2024 Worldwide Developers Conference, we introduced Apple Intelligence, a personal intelligence system integrated deeply into iOS 18…
In the charts there, substitute Gemini for Gemma and very likely works the exact same way. Apple already probably ran open source models on PCC ( to do ‘apples-to-apples’ comparison to their own models .not sure why would throw hardware variations into the mix if trying to equalize to compare just software differences ).
Why use a set of open source models if not trying to port them to the same hardware ?? ( if wanted to bury them in hardware difference could just pull the none Apple ones from web if data is anonymized ). There may be licensing restrictions on benchmarking on paid for versions ,but porting is a clear option here .
Gemini is just replacing Gemma with a related code code base.
But that likely is not enough for the tinfoil hat folks . Multiple models running on PcC means nothing … Google is evil Google and slid a back door in that pumps all the data out of Apple data center without Apple noticing at all. …
Last edited:
The service isn’t even out yet so why would Apple explain how it works?
Apple regularly updates their platform security guide which explains in detail how they keep devices/data secure. I’m sure Apple will be adding a section to this document for their upcoming AI services.
There is little need to port to port the whole OS. PCC is not the only Private AI compute model in the universe.
Private AI Compute: our next step in building private and helpful AI
Introducing Private AI Compute, our new way to bring you helpful AI with the power of the cloud, while keeping your data private to you.
PCC is a stripped down iOS . The iOS boot process is deep tied to Apple’s hardware. ( Apple chucked UEFI for Macs for the Apple Silicon transition).
There is also a chance to spill-over to TPU hardware is at least to get chance to run a model that doesn’t need a hardware port at least as much to run on ‘bigger hardware’ .
What is missing is how long of is taking Google+Apple to port these models to PCC hardware and tune them up.
But PCC is also a system management design . I doubt Google is handing over raw hardware to be run. There is custom Google cloud infrastructure ( both hardware and software ) to deal with.
Similarly any utilities ( compilers , etc ) and Google specific drivers would neeed to be ported to this iOS instance .
Right, but they don’t have to ‘nuke’ the host operating system image to get that. Just initialize the keystore enclave on the server and enable the encrypted memory features.
I doubt all of the PCC was going to be outsourced to. Google. Apple is also treating Apple Foundational model as an OS library that is free to use ( not a direct revenue model ) . So very expansive use of Google servers is being paid by who ???
We will see what Apple provides for folks who want to jump out of Apple Foundation model and go into the for pay mode.
Glad that it is running on Apple's Private Cloud compute. This should maintain the privacy levels that are existing today.
You are a likely a little crazy. You are inserting lots of stuff that isn't there. The 'We' in cooks comments are not about him and the other humans in the room. It is slighly like a 'royal' we in that it is speaking as the 'kingdom' not a group of people. We == Apple there. So basically the sentence says without the pronoun usage.
"Apple believes that Apple can unloock a lot of experiences and innovate in a key way due to the collaboration. Apple will continue to run on the (Apple) device and run in Apple Private Cloud Compute and maintain Apple's industry leading privacy standard in doing so. "
The ported Gemini model is pragmatically be Apple software since it is inside of Apple Foundation Models.
When Apple used gcc to build MacOS they used the GNU std c libaries. It wasn't the GNU Apple OS. When Apple shifted to LLVM compiler they used that compiler stack std c libraries. Gemini is being treated like a rather large library here. That is it. Apple didn't write 100% of all of the code in gcc or LLVM. Apple collaborated on those open source projects. There are lots of bits and parts of macOS, iOS , etc that Apple produces that are licensed from other folks that Apple bundles into their OS offerings.
Gemini happens to be 'closed source' but that just means have to pay to look at the code. ( and there is a open source variant of Gemini called Gemma. So it not as though Google has shared nothing previously. )
The "all new" is mostly misdirection. The next OS will be the next OS would a new set of Apple libraries bundled. That doesn't change MacOS/iOS security model because the library have new features.
Gemini ( and Gemma) are simply pieces of software. If ported to the underlying OS and hardware it will run anywhere that can run a decent sized program ( some 80's vintage of MS-DOS probably won't work. But most modern apps wouldn't work either. There is nothing particular novel in that. )
Private Cloud compute is just a iOS variant plus hardware combo, it is essentially just as decoupled from the OS/hardware as another other app that runs on iOS. The 'Privacy' of PCC is largey done by the OS, not the model. If Apple substitutes another model implementation run on PCC-OS that is essentially just like running another app on iOS. The 'new' iOS app would be just as sandboxed as all the other principal iOS apps because they all run sandboxed.
This is deeply flawed because Cook mentions nothing about opening a compute platform in "another country at all".
Gemini doesn't only run on Google TPU.
This is more like hypothetical cvar manufacturer saying they are going to buy and import German steel for the structure frame instead of US steel using the same factories they already had. It really doesn't functionally change the car at all. It is still steel.
If Cook has said something along the lines that they were tossing PCC into the trash can you might get some traction there. But there is nothing there. Just like there are no for forward looking product statements. He didn't promise say they were going do a new iPhone this year either. It is just standard Apple corporate policy.
Actually, it pretty much is what he said. As the Apple foundational model improves and is more widely deployed that will create a better value for their ecosystem.
Apple has already issued serveral technical outlines of how private cloud compute works and how they they are looking to split workload between the two.
A model falling out of the Gemini/Gemma traing bucket is just an app library ( a piece of software). It isn't the entirety of the finished Apple Foundation models.
JSRinUK
macrumors 6502a
AI’s been around long enough by now that we should all have formed our own views of it. For me, I hope there’s an “off” switch for it.
I'm in the UK too and, moving closer to the topic of this thread (Siri), as a smart speaker enthusiast one thing I can't resist mentioning is just how bad Alexa is in terms of knowing what is around it even in a relatively big (in terms of GDP) country like the UK.
I currently have both Google and Amazon smart speakers in quite a few rooms in my home. I use Alexa because it hears me better when speaking softly or when there is background noise and I use Alexa because for most stuff beyond turning devices on and off Google gives noticeably better answers.
As an example of Alexa's failings a new branch of Lidl opened about 100 metres from where I live. Google detected the new branch within about 24 hours and when I asked "What time does Lidl close today?" it would answer "Lidl at <the correct address> closes at <whenever>" whereas for 6 months after the new branch opened Alexa would continue to say "Lidl at <a branch about 3 km away> closes at <whenever>". It took Alexa a full 6 months before it finally realised that my closest Lidl was about 100 metres away.
I use my smart speakers quite a lot for distances, walking times, public transport timetables and opening/closing hours. That's why I'm somewhat obsessed with the quality of the ground truth in the various mapping apps because for my use cases it's quite a big part of how satisfactory my experiences are with my smart speakers.
I would dearly love to consolidate everything to just HomePods but I need to be convinced that (a) they will hear me and respond (and not false-respond) to the wake word as well as Echo devices do and also that Siri, presumably with the aid of Apple Maps ground truth, is at least as good as Google Assistant when it comes to local UK knowledge.
Since I was whittering on about Amazon and Google's responses to shop opening hours, ignoring the issues with Alexa's local knowledge there's another little thing that, when added to lots of other little things, really makes a difference to how good the responses from a smart assistant feel.
If you ask Alexa "When does Lidl close?" it's actual response is "Lidl at <address> is open from <opening-time> to <closing-time> today" whereas if you ask Google Assistant (in a Nest smart speaker) that exact same question it answers "Lidl at <address> closes at <closing-time> today" or, if that closing time is within an hour of the current time, it replies "Lidl at <address> closes in <x> minutes". It's such a nice extra touch that makes me feel that Google Assistant actually took the time to fully understand what I was asking and give me the direct answer vs Alexa doing something closer to a keyword search and quick lookup without actually trying to fully tailor the answer to the question. That query is one of the first things that I'll be giving to the new hopefully improved Siri to see how it does.
That behaviour has nothing to do with LLMs by the way. I've been impressed with Google Assistant doing stuff like that since I bought my first Google smart speaker in 2018.