The writing was on the wall the moment Apple stopped trying to buy their way into the server-side training game like what three years ago?
Apple has the best edge inference silicon in the world (neural engine), but they have effectively zero presence in a training datacenter. They simply do not have the TPU pods or the H100 clusters to train a frontier model like Gemini 2.5 or 3.0 from scratch without burning 10 years of cash flow.
To me, this deal is about the bill of materials for intelligence. Apple admitted that the cost of training SOTA models is a capex heavy-lift they don't want to own. Seems like they are pivoting to becoming the premium "last mile" delivery network for someone else's intelligence. Am I missing the elephant in the room?
It's a smart move. Let Google burn the gigawatts training the trillion parameter model. Apple will just optimize the quantization and run the distilled version on the private cloud compute nodes. I'm oversimplifying but this effectively turns the iPhone into a dumb terminal for Google's brain, wrapped in Apple's privacy theater.
> I'm oversimplifying but this effectively turns the iPhone into a dumb terminal for Google's brain, wrapped in Apple's privacy theater.
Setting aside the obligatory HN dig at the end, LLMs are now commodities and the least important component of the intelligence system Apple is building. The hidden-in-plain-sight thing Apple is doing is exposing all app data as context and all app capabilities as skills. (See App Intents, Core Spotlight, Siri Shortcuts, etc.)
Anyone with an understanding of Apple's rabid aversion to being bound by a single supplier understands that they've tested this integration with all foundation models, that they can swap Google out for another vendor at any time, and that they have a long-term plan to eliminate this dependency as well.
> Apple admitted that the cost of training SOTA models is a capex heavy-lift they don't want to own.
I'd be interested in a citation for this (Apple introduced two multilingual, multimodal foundation language models in 2025), but in any case anything you hear from Apple publicly is what they want you to think for the next few quarters, vs. an indicator of what their actual 5-, 10-, and 20-year plans are.
My guess is that this is bigger lock-in than it might seem on paper.
Google and Apple together will posttrain Gemini to Apple's specification. Google has the know-how as well as infra and will happily do this (for free ish) to continue the mutually beneficial relationship - as well as lock out competitors that asked for more money (Anthropic)
Once this goes live, provided Siri improves meaningfully, it is quite an expensive experiment to then switch to a different provider.
For any single user, the switching costs to a different LLM are next to nothing. But at Apple's scale they need to be extremely careful and confident that the switch is an actual improvement
I’m not so sure. Just think about coding assistants with MCP based tools. I can use multiple different models in GitHub Copilot and get good results with similarly capable models.
Siri’s functionality and OS integration could be exposed in a similar, industry-standard way via tools provided to the model.
Then any other model can be swapped in quite easily. Of course, they may still want to do fine tuning, quantization, performance optimization for Apple’s hardware, etc.
But I don’t see why the actual software integration part needs to be difficult.
> But I don’t see why the actual software integration part needs to be difficult.
That’s not the issue. The issue is that once Gemini is in place as the intelligence behind Siri, the bar is now much higher than today and so you have to be more careful if you consider replacing Gemini, because you’re as likely as not to make Siri worse. Maybe more likely to make it worse.
Oh well that’s a good problem to have, isn’t it? Siri being so good that they don’t want to mess it up.
That gives them plenty of runway to test and optimize new models internally before release and not feel like they need to rush them out because Siri sucks.
Doubt it. Of all the issues I run into with Siri none could be solved by throwing AI slop at it. Case in point: if I ask Siri to play an album and it can't match the album name it just plays some random shit instead of erroring out.
Um if I ask an LLM about a fake band it literally say I couldn't find any songs by the fake band did you type is correctly and it's about a millions times more likely to guess correctly. Why do you say it doesn't solve loads of things? I'm more concerned about the problems it creates (prompt injection, hallucinations in important work, bad logic in code), the actual functionality will be fantastic compared to Siri right now!
Because I'm sitting here twiddling my thumbs waiting for random pages to go through their anti-LLM bot crap. LLMs create more problems than they solve.
Um if I ask an LLM about a fake band it literally say I couldn't find any
songs by the fake band did you type is correctly and it's about a millions
times more likely to guess correctly
Um if Apple wrote proper error handling in the first place the issue would be solve without LLM baggage. Apple made a conscious decision to handle "unknown" artists this way, LLMs don't change that.
Ollama! Why didn’t they just run Ollama and a public model! They’ve kept the last 10 years with a Siri who doesn’t know any contact named Chronometer only to require the best in class LLM?
The other day I was trying to navigate to a Costco in my car. So I opened google maps on Android Auto on the screen in my car and pressed the search box. My car won't allow me to type even while parked... so I have to speak to the Google Voice Assistant.
I was in the map search, so I just said "Costco" and it said "I can't help with that right now, please try again later" or something of the sort. I tried a couple more times until I changed up to saying "Navigate me to Costco" where it finally did the search in the textbox and found it for me.
Obviously this isn't the same thing as Gemini but the experience with Android Auto becomes more and more garbage as time passes and I'm concerned that now we're going to have 2 google product voice assistants.
Also, tbh, Gemini was great a month ago but since then it's become total garbage. Maybe it passes benchmarks or whatever but interacting with it is awful. It takes more time to interact with than to just do stuff yourself at this point.
I tried Google Maps AI last night and, wow. The experience was about as garbage as you can imagine.
I'm genuinely curious about this too. If you really only need the language and common sense parts of an LLM -- not deep factual knowledge of every technical and cultural domain -- then aren't the public models great? Just exactly what you need? Nobody's using Siri for coding.
Are there licensing issues regarding commercial use at scale or something?
Pure speculation, but I’d guess that an arrangement with Google comes with all sorts of ancillary support that will help things go smoothly: managed fine tuning/post-training, access to updated models as they become available, safety/content-related guarantees, reliability/availability terms so the whole thing doesn’t fall flat on launch day etc.
Probably repeatability and privacy guarantees around infrastructure and training too. Google already have very defined splits for their Gemma and in house models with engineers and researchers rarely communicating directly.
That said, Apple is likely to end up training their own model, sooner or later. They are already in the process of building out a bunch of data centers, and I think they have even designed in-house servers.
Remember when iPhone maps were Google Maps? Apple Maps have been steadily improving, to the point they are as good as, if not better than, Google Maps, in many areas (like around here. I recently had a friend send me a GM link to a destination, and the phone used GM for directions. It was much worse than Apple Maps. After a few wrong turns, I pulled over, fed the destination into Apple Maps, and completed the journey).
> what their actual 5-, 10-, and 20-year plans are
Seems like they are waiting for the "slope of enlightenment" on the gartner hype curve to flatten out. Given you can just lease or buy a SOTA model from leading vendors there's no advantage to training your own right now. My guess is that the LLM/AI landscape will look entirely different by 2030 and any 5 year plan won't be in the same zip code, let alone playing field. Leasing an LLM from Google with a support contract seems like a pretty smart short term play as things continue to evolve over the next 2-3 years.
This is the key. The real issue is that you don’t need superhuman intelligence in a phone AI assistant. You don’t need it most of the time in fact. Current SOTA models do a decent job of approximating college grad level human intelligence let’s say 85% of the time which is helpful and cool but clearly could be better. But the pace at which the models are getting smart is accelerating AND they are getting more energy efficient and memory efficient. So if something like DeepSeek is roughly 2 years behind SOTA models from Google and others who have SOTA models then in 2030 you can expect 2028 level performance out open models. There will come a time when a model capable of college grad level intelligence 99.999% of the time will be able to run on a $300 device. If you are Apple you do not need to lead the charge on a SOTA model, you can just wait until one is available for much cheaper. Your product is the devices and services consumers buy. If you are OpenAI you have no other products. You must become THE AI to have in an industry that will in the next few years become dominated by open models that are good enough or to close up shop or come up with another product that has more of a moat.
My impression is that solar (and maybe wind?) energy have benefited from learning-by-doing [1][2] that has resulted in lower costs and/or improved performance each year. It seems reasonable to me that a similar process will apply to AI (at least in the long run). The rate of learning could be seen as a "pace" of improvement. I'm curious, do you have a reference for the deceleration of pace that you refer to?
1) the dearth of new (novel) training data. Hence the mad scramble to hoover up, buy, steal, any potentially plausible new sources.
2) diminishing returns of embiggening compute clusters for training LLMs and size of their foundation models.
(As you know) You're referring to Wright's Law aka experience learning curve.
So there's a tension.
Some concerns that we're nearing the ceiling for training.
While the cost of applications using foundation models (implementing inference engines) is decreasing.
Someone smarter than me will have to provide the slopes of the (misc) learning curves.
I was not aware of (or had forgotten) the term "Wright's law" [1], but that indeed is what I was thinking of. It looks like some may use the term "learning curve" to refer to the same idea (efficiency gains that follow investment); the Wikipedia page on "Learning curve" [2] includes references to Wright.
I don't think anyone really knows, because there's no objective standard for determining progress.
Lots of benchmarks exist where everyone agrees that higher scores are better, but there's no sense in which going from a score of 400 to 500 is the same progress as going from 600 to 700, or less, or more. They only really have directional validity.
I mean, the scores might correspond to real-world productivity rates in some specific domain, but that just begs the question -- productivity rates on a specific task are not intelligence.
That's not an "obligatory HN dig" though, you're in-media-res watching X escape removal from the App Store and Play Store. Concepts like privacy, legality and high-quality software are all theater. We have no altruists defending these principles for us at Apple or Google.
Apple won't switch Google out as a provider for the same reason Google is your default search provider. They don't give a shit about how many advertisements you're shown. You are actually detached from 2026 software trends if you think Apple is going to give users significant backend choices. They're perfectly fine selling your attention to the highest bidder.
There are second-order effects of Google or Apple removing Twitter from their stores.
Guess who's the bestie of Twitter's owner? Any clues? Could that be a vindictive old man with unlimited power and no checks and balances to temper his tantrums?
Of course they both WANT Twitter the fuck out of the store, but there are very very powerful people addicted to the app and what they can do with it.
In the current US environment, no one can survive going against Trump, and as recently evidenced, this is meant literally.
The US, for all intents and purposes, is now a kleptocracy. Rule of law, freedom of speech, even court orders, all of that doesn't matter any more in practice. There will always be some way for the federal government to strong-arm anyone into submission.
Not with that attitude they can’t. Let’s see what happens to the first person to call his bull shit. If jpow folds or is actually indicted, you may be right. Let’s see what happens with Exxon though i think they’re gonna bend the knee.
Why would you not bend the key ? There's a reasonnable chance Trump is gone in less than 3 years and MAGA tears itself apart. Better bet on that than try to make a stand now and lose everything.
> Let’s see what happens to the first person to call his bull shit
Well, ICE just executed a woman in broad daylight with multiple cameras filming, and a day later you got Kristi Noem standing on a podium with a slogan referencing to an OG Nazi massacre [1][2] and half the US government gaslighting the country, spreading outright lies [3] without consequences so far.
When they can get away with this level of lies, they can get away with anything. Trump's infamous "I Could ... Shoot Somebody, And I Wouldn't Lose Any Voters" quote [4] wasn't a joke - it was a clear prediction of what he intended to enable eventually.
I do not usually comment on politics but just this one time, and hopefully I can wordsmith it without taking a political stance.
When Trump started his campaign, circa 2011 with the birth certificate, he did not know he will win or not, but he made it his life's mission.
Countering him will take the same zeal. I know we have a precedence of presidents retiring, but unless Obama (and Hillary and Biden and Kamala) hits the streets as the leader of resistance, the resistance will be quelled easily by constant distracting. Yeah maybe AOC, maybe Bernie, maybe someone else, but no ... Trump is smart and dedicated (despite the useful idiot role he plays), he can not be countered by mid-term and full-term campaigns. We are not in Kansas any more. Been a while. The opposition needs a named resistance leader whose full time job is to engage Trump.
I'm not American, but in my opinion Newsom+AOC might have the zeal and voter base to MAYBE do something.
Newsom (or his PR team) knows how to play the troll game correctly, hitting low blows and not sticking to the fucking high ground.
AOC on the other hand will make the MAGA base _so_ irrationally angry they might do something actually stupid. She's also got Bernie's views, which might make America a place I want to visit some day in the next decade again. I've literally turned down all expenses paid company trips to USA a few times because I just don't want to risk either not getting into the country or not getting back.
Apple’s various privileged device-level ads and instant-stop-on-cancel trials and special rules for notifications for their paid additional services like Fitness+, Music, Arcade, iCloud+, etc are all proof that they do not care about the user anymore.
Why? Grain is a commodity, but I buy flour at the store rather than grow my own. The “commmodity” argument suggets that new companies should stay away from model training unless they have a cost edge.
An Apple-developed LLM would likely be worse than SOTA, even if they dumped billions on compute. They'll never attract as much talent as the others, especially given how poorly their AI org was run (reportedly). The weird secrecy will be a turnoff. The culture is worse and more bureaucratic. The past decade has shown that Apple is unwilling to fix these things. So I'm glad Apple was forced to overcome their Not-Invented-Here syndrome/handicap in this case.
Apple might have gotten very lucky here ... the money might be in finding uses, and selling physical products rather than burning piles of cash training models that are SOTA for 5 minutes before being yet another model in a crowded field.
My money is still on Apple and Google to be the winners from LLMs.
And when the cost of training LLMs starts to come down to under $1B/yr, Apple can jump on board, having saved >$100B in not trying to chase after everyone else to try to get there first.
Apple has also never been big on the server side equation of both software and hardware - don't they already outsource most of their cloud stack to Google via GCP ?
I can see them eventually training their own models (especially smaller and more targeted / niche ones) but at their scale they can probably negotiate a pretty damn good deal renting Google TPUs and expertise.
I always wondered what they were hoping for with their server products back when they had them. Consumers and end users benefit greatly from the vertical integration that Apple is good at. This doesn't translate with servers. Commodity hardware + linux is not only cheaper, its often easier, and was definitely less proprietary.
Its also a race to the bottom type scenario. Apple would have never been able to keep up with server release schedules.
Was an interesting but ultimately odd moment of history for servers.
Pre-iPhone and pre-Intel Mac, Apple was experimenting with a lot of things. The iPod wasn't a clear initial win--and the iPhone wasn't either. A lot of the success happened in retrospect.
Companies were still figuring out Linux with servers at the time. Xserve seemed like it might be something of interest to at least academia but Apple never really had their heart in it as I wrote at the time.
How is server hardware more "commodity" than MacBook laptops? Both are quite sophisticated and tailored to their audience in nuanced ways; both are manufactured at scale and face fierce competition. I don't think Xserve was a uniquely commodity business, it was a B2B service business--which isn't Apple.
I'd rather say IBM got cannibalized by "financial engineers", this wasn't a decision made because of "it's a commodity".
There used to be a time when IBM actually meant quality (that's where "no one ever got fired for buying IBM" came from, after all), but nowadays? A loooot of stuff is either sold (Thinkpad went to Lenovo, Lotus Notes to HCL), faded into irrelevancy outside of extremely few niche markets (anything mainframe), got left for dead (the PC - it used to be called "IBM compatible personal computer"!) or got spun off (Kyndryl).
According to Wikipedia, IBM has 282.000 employees worldwide. What the fuck are all of these people doing?
The no one ever got fired for buying IBM wasn’t about quality. It’s always safe to buy the default choice that everyone else uses. Especially when things go wrong.
Many want to be founders here on HN don’t get that. Even if your product is better and cheaper, there is too much of a reputational risk signing a contract for a B2B SaaS product with an unknown vendor.
On a completely unrelated note, for the love of all that is holy don’t try to do B2B SaaS without SSO support.
With Thunderbolt 5 and M5 Ultras, Apple could be building lower cost clusters that could possibly scale enough while keeping a lower power budget. Obviously that can't compete with NVIDIA racks, but for mobile consumer inference maybe that would be enough?
Yeah… there’s this “bro— do you even business?” vibe in the tech world right now pointed at any tech firm not burning oil tankers full of cash (and oil, for that matter,) training a giant model. That money isn’t free — the economic consequences of burning billions to make a product that will be several steps behind, at best, are giant. There’s a very real chance these companies won’t recoup that money if their product isn’t attractive to hoards of users willing to pay more money for AI than anyone currently is. It doesn’t even make them look cool to regular people — their customers hate hearing about AI. Since there are viable third party options available, I think Apple would have to be out of their goddamned minds to try and jump in that race right now. They’re a product company. Nobody is going to not buy an iPhone because they’re using a third-party model.
Something weird has gone wrong in the psyche of humans.
Why are we even talking about 'AI'? When I heat up food in a microwave, I dont care about the technology - I care about whether it heats up the food or not.
For some bizarre reason people keep talking about the technology (LLMs) - the consumers/buyers in the market for the most part dont give a hoot about it. They want to know how the thing fits in their life and most importantly what are the benefits.
Ive unfortunately been exposed to some Google Ads re. Gemini and let me tell you - their marketing capabilities are god awful.
>Nobody is going to not buy an iPhone because they’re using a third-party model.
You're right, and this is proven. Apple has fumbled a whole release cycle on AI and severely curbed expectations, and they still sell 200m iPhones a year and lead the market [0]
Easy enough. Most people abhor AI and want nothing to do with it. The only ones who actually love AI (or what's being sold to them under that banner) are clueless and/or greedy executives, propagandists, and a select few legitimate AI artists doing pretty nice remixes of Star Wars, Harry Potter and the likes in a quality not seen before.
Reportedly, Meta is paying top AI talent up to $300M for a 4 year contract. As much as I'm in favor of paying engineers well, I don't think salaries like this (unless they are across the board for the company, which they are of course not) are healthy for the company long term (cf. Anthony Levandowski, who got money thrown after him by Google, only to rip them off).
So I'm glad Apple is not trying to get too much into a bidding war. As for how well orgs are run, Meta has its issues as well (cf the fiasco with its eponymous product), while Google steadily seems to erode its core products.
The Allen Institute (a non-profit) just released the Molmo 2 and Olmo 3 models. They trained these from scratch using public datasets, and they are performance-competitive with Gemini in several benchmarks [0] [1].
AMD was also able to successfully train an older version of OLMo on their hardware using the published code, data, and recipe [2].
If a non-profit and a chip vendor (training for marketing purposes) can do this, it clearly doesn't require "burning 10 years of cash flow" or a Google-scale TPU farm.
No, of course the training costs aren't that high. Apple's ten years of future free cash flow is greater than a trillion dollars (they are above $100b per year). Obviously, the training costs are a trivial amount compared to that figure.
What I'm wondering - their future cash flow may be massive compared to any conceivable rational task, but the market for servers and datacenters seems to be pretty saturated right now. Maybe, for all their available capital, they just can't get sufficient compute and storage on a reasonable schedule.
I have no idea what AI involves, but "training" sounds like a one-and-done - but how is the result "stored"? If you have trained up a Gemini, can you "clone" it and if so, what is needed?
I was under the impression that all these GPUs and such were needed to run the AI, not only ingest the data.
Theoretically it would be much less expensive to just continue to run the existing models, but ofc none of the current leaders are going to stop training new ones any time soon.
No, I doesn't beat Gemini in any benchmarks. It beats Gemma, which isn't a SoTA even among open models of that size. That would be Nemotron 3 or GPT-OSS 20B.
Yea, I think it’s smart, too. There are multiple companies who have spent a fortune on training and are going to be increasingly interested in (desperate to?) see a return from it. Apple can choose the best of the bunch, pay less than they would have to to build it themselves, and swap to a new one if someone produces another breakthrough.
100%. It feels like Apple is perfectly happy letting the AI labs fight a race to the bottom on pricing while they keep the high-margin user relationship.
I'm curious if this officially turns the foundation model providers into the new "dumb pipes" of the tech stack?
It’ll be interesting to see how it plays out. The question is, what’s the moat? If all they have is scaling to drive better model performance, then the winner is just whoever has the lowest cost of capital.
> I'm oversimplifying but this effectively turns the iPhone into a dumb terminal for Google's brain, wrapped in Apple's privacy theater.
This sort of thing didn't work out great for Mozilla. Apple, thankfully, has other business bringing in the revenue, but it's still a bit wild to put a core bit of the product in the hands of the only other major competitor in the smartphone OS space!
I dunno, my take is that Apple isn’t outsourcing intelligence rather it’s outsourcing the most expensive, least defensible layer.
Down the road Apple has an advantage here in a super large training data set that includes messages, mail, photos, calendar, health, app usage, location, purchases, voice, biometrics, and you behaviour over YEARS.
Let's check back in 5 years and see if Apple is still using Gemini or if Apple distills, trains and specializes until they have completed building a model-agnostic intelligence substrate.
> The writing was on the wall the moment Apple stopped trying to buy their way into the server-side training game like what three years ago?
It goes back much further than that - up until 2016, Apple wouldn't let its ML researchers add author names to published research papers. You can't attract world-class talent in research with a culture built around paranoid secrecy.
It is more like Apple have no need to spend billions on training with questionable ROI when it can just rent from one of the commodity foundation model labs.
I don't know why people automatically jump to Apple's defense on this.... They absolutely did spend a lot of money and hired people to try this. They 100% do NOT have the open and bottom-up culture needed to pull off large scale AI and software projects like this.
They did things far more complicated from an engineering perspective. I am far more impressed by what they accomplished along TSMC with Apple Silicon than by what AI labs do.
In any case, see the section on Jakob Uszkoreit, for example, or Noam Shazeer. And then…
> In the higher echelons of Google, however, the work was seen as just another interesting AI project. I asked several of the transformers folks whether their bosses ever summoned them for updates on the project. Not so much. But “we understood that this was potentially quite a big deal,” says Uszkoreit.
Worth noting the value of “bosses” who leave people alone to try nutty things in a place where research has patronage. Places like universities, Xerox, or Apple and Google deserve credit for providing the petri dish.
You can understand how transformers work from just reading the Attention is All You Need paper, which is 15 pages of pretty accessible DL. That's not the part that is impressive about LLMs.
It’s such a commodity that there are only 3 SOTA labs left and no one can catch them. I’m sure it’ll be consolidated further in the future and you’re going to be left with a natural monopoly or duopoly.
Apple has no control over the most important change to tech. They have control to Google.
> It’s such a commodity that there are only 3 SOTA labs left and no one can catch them.
No one can outpace them in improving the SOTA, everyone can catch up to them. Why are open-weight models perpetually 6 months behind the SOTA? Given enough data harvested from SOTA models you can eventually distill them.
The biggest differentiator when training better models are not some new fancy architectural improvements (even the current SOTA transformer architectures are very similar to e.g. the ancient GPT-2), but high quality training data. And if your shiny new SOTA model is hooked into a publicly available API, guess what - you've just exposed a training data generator for everyone to use. (That's one of the reasons why SOTA labs hide their reasoning chains, even though those are genuinely useful for users - they don't want others to distill their models.)
Really, don't believe benchmarks as gospel. Chinese models are pretty much competitive with offerings from Anthropic, OpenAI or Google. Meta is currently at a disadvantage, but I believe they will find their mojo and soon be competitive again.
Frankly, a lot of times I prefer using GLM 4.6 running on Cerebras Inference, than having to deal with the performance hiccups from Claude. For most practical purposes, I've seen no big penalty in using it compared to Opus 4.5, even the biggest qwen-coder models are pretty much competitive.
Between me and the company I work for, I spend some serious money with AI. I use it extensively in my main job, on two side projects that I have paying customers for, and for graduate school work. I can tell you that there quite a few more SOTA models around than what the benchmarks tell you.
I always think about this, can someone with more knowledge than me help me understand the fragility of these operations?
It sounds like the value of these very time-consuming, resource-intensive, and large scale operations is entirely self-contained in the weights produced at the end, right?
Given that we have a lot of other players enabling this in other ways, like Open Sourcing weights (West vs East AI race), and even leaks, this play by Apple sounds really smart and the only opportunity window they are giving away here is "first to market" right?
Is it safe to assume that eventually the weights will be out in the open for everyone?
> and the only opportunity window they are giving away here is "first to market" right?
A lot of the hype in LLM economics is driven by speculation that eventually training these LLMs is going to lead to AGI and the first to get there will reap huge benefits.
So if you believe that, being "first to market" is a pretty big deal.
But in the real world there's no reason to believe LLMs lead to AGI, and given the fairly lock-step nature of the competition, there's also not really a reason to believe that even if LLMs did somehow lead to AGI that the same result wouldn't be achieved by everyone currently building "State of the Art" models at roughly the same time (like within days/months of each other).
So... yeah, what Apple is doing is actually pretty smart, and I'm not particularly an Apple fan.
> is entirely self-contained in the weights produced at the end, right?
Yes, and the knowledge gained along the way. For example, the new TPUv4 that Google uses requires rack and DC aware technologies (like optical switching fabric) for them to even work at all. The weights are important, and there is open weights, but only Google and the like are getting the experience and SOTA tech needed to operate cheaply at scale.
Google says: "Apple Intelligence will continue to run on Apple devices and Private Cloud Compute, while maintaining Apple's industry-leading privacy standards."
So what does it take? How many actual commitments to privacy does Apple have to make before the HN crowd stops crowing about "theater"?
No, its not misleading it says right there in the private policy.
Apple isnt suddenly private just because they have enough data about you that they dont need to link to 3rd party data. They do exactly what 3rd party sites that are considered privacy invasive do. They serve you ads based on your private data like what you watch, what you read and what things you do on your device.
It doesnt say that they only store all this information on device. Apple is only using a random identifier when its sharing information about your habits and personal data on its ad platform, that info btw is shared with 3rd parties. But dont worry that data suddenly becomes non personal because they used a random identifier.
Apple's goal is likely to run all inference locally. But models aren't good enough yet and there isn't enough RAM in an iPhone. They just need Gemini to buy time until those problems are resolved.
That was their goal, but in the past couple years they seem to have given up on client-side-only ai. Once they let that go, it became next to impossible to claw back to client only… because as client side ai gets better so does server side, and people’s expectations scale up with server side. And everybody who this was a dealbreaker for left the room already.
Apple thinks they can get a best-of-both-worlds approach with Private Cloud Compute. They believe they can secure private servers specialized to specific client devices in a way that the cloud compute effort is still "client-side" from a trust standpoint, but still able to use extra server-side resources (under lock and key).
I don't know how close to that ideal they've achieved, but especially given this announcement is partly baked on an arrangement with Google that they are allowed to run Gemini on-device and in Private Cloud Compute, without using Google's more direct Gemini services/cloud, I'm excited that they are trying and I'm interested in how this plays out.
Apple integrated ChatGPT as an opt-in that made the mainstream users feel like Apple was delivering on some of that marketing. Apple also delivered "high wow" features like auto-stickers and other such silliness. Just as the privacy issues are maybe a bit of an HN specialty, HN is also maybe a bit more prone to "Apple Intelligence" brand cynicism than mainstream belief.
I'm excited about the attempt at privacy because I'm on "Team Keep Siri Dumb". I like dumb Siri. It reliably meets most of my needs, setting timers and managing house lights. I'd rather Siri stay dumb and I would never opt-in to ChatGPT Siri as some of my family has, but if Siri "has to" get smart to survive, I will celebrate whatever privacy wins are still available as my only hope that smarter Siri is not something I need to just disable entirely (and lose my "friend" in charge of my timers and house lights in the process).
I stated that I am not naive and am not entirely convinced by Apple's sales pitch that the Private Cloud Compute containers are encrypted with keys in a way that only your hardware device can read in such a way that the PCC is an extension of your device.
I just think it is useful that Apple is trying something along those lines and wishful the guarantees work half as well as they claim they do, because that's a good goal to have in theory even when it fails in practice against dedicated threat actors.
And yes, to be fair my personal day-to-day threat model currently is much more concerned with the evil advertising company known as Google than it is with government actors. Even if Apple's Private Cloud Compute only means "private from Google" that's still a win for me (and most of the information I was looking for when I saw this headline, because my first fear was that the advertising company Google was involved).
Phones will get upgrades, but then so will servers. The local models will always be behind the state of the art running on big iron. You can’t expect to stand still and keep up with the Red Queen.
Rumor has it that they weren't trained "from scratch" the was US would, i.e. Chinese labs benefitted from government "procured" IP (the US $B models) in order to train their $M models. Also understand there to be real innovation in the many-MoE architecture on top of that. Would love to hear a more technical understanding from someone who does more than repeat rumors, though.
A lot of HN commentators are high on their own supply with regard to the AI bubble... when you realize that this stuff isn't actually that expensive the whole thing begins to quickly unravel.
> Seems like they are pivoting to becoming the premium "last mile" delivery network for someone else's intelligence.
They have always been a premium "last mile" delivery network for someone else's intelligence, except that "intelligence" was always IP until now. They have always polished existing (i.e., not theirs) ideas and made them bulletproof and accessible to the masses. Seems like they intend to just do more of the same for AI "intelligence". And good for them, as it is their specialty and it works.
Could you elaborate a bit on why you've judged it as privacy theatre? I'm skeptical but uninformed, and I believe Mullvad are taking a similar approach.
Mullvad is nothing like Apple. For apple devices:
- need real email and real phone number to even boot the device
- cannot disable telemetry
- app store apps only, even though many key privacy preserving apps are not available
- /etc/hosts are not your own, DNS control in general is extremely weak
- VPN apps on idevices have artificial holes
- can't change push notification provider
- can only use webkit for browsers, which lacks many important privacy preserving capabilities
- need to use an app you don't trust but want to sandbox it from your real information? Too bad, no way to do so.
- the source code is closed so Apple can claim X but do Y, you have no proof that you are secure or private
- without control of your OS you are subject to Apple complying with the government and pushing updates to serve them not you, which they are happy to do to make a buck
Mullvad requires nothing but an envelope with cash in it and a hash code and stores nothing. Apple owns you.
Agreed on most points but you can setup a pretty solid device wide DNS provider using configuration profiles. Similar to how iOS can be enrolled in work corporate MDM - but under your control.
Works great for me with NextDNS.
Orion browser - while also based on WebKit - is also awesome and has great built in Adblock and supposedly privacy respecting ideals.
Apple has records that you are installing that, probably putting you on a list.
And it works until it's made illegal in your country and removed from the app store. You have no guarantees that anything that works today will work tomorrow with Apple.
Apple is setting us up to be under a dictator's thumb one conversion at a time.
They transitioned from “nobody can read your data, not even Apple” to “Apple cannot read your data.” Think about what that change means. And even that is not always true.
They also were deceptive about iCloud encryption where they claimed that nobody but you can read your iCloud data. But then it came out after all their fanfare that if you do iCloud backups Apple CAN read your data. But they aren’t in a hurry to retract the lie they promoted.
Also if someone in another country messages you, if that country’s laws require that Apple provide the name, email, phone number, and content of the local users, guess what. Since they messaged you, now not only their name and information, but also your name and private information and message content is shared with that country’s government as well. By Apple. Do they tell you? No. Even if your own country respects privacy. Does Apple have a help article explaining this? No.
If you want to turn on full end-to-end encryption you can, if you want to share your pubkey so that people can't fake your identity on iMessage you can, and there's still a higher tier of security than that presumably for journalists and important people.
It's something a smart niece or nephew could handle in terms of managing risk, but the implications could mean getting locked out of your device which you might've been using as the doorway to everything, and Apple cannot help you.
>Also if someone in another country messages you, if that country’s laws require that Apple provide the name
I don't mean to sound like an Apple fanboy, but is this true just for SMS or iMessage as well? It's my understanding that for SMS, Apple is at the mercy of governments and service providers, while iMessage gives them some wiggle room.
Ancedotal, but when my messages were subpoenaed, it was only the SMS messages. US citizen fwiw
Because Apple makes privacy claims all the time, but all their software is closed source and it is very hard or impossible to verify any of their claims. Even if messages sent between iPhones are E2EE encrypted for example, the client apps and the operating system may be backdoored (and likely are).
All user data is E2E encrypted, so the government literally cannot force this. This has been the source of numerous disputes [0, 1] that either result in the device itself being cracked [0] (due to weak passwords or vulnerabilities in device-level protection) or governments attempting to ban E2E encryption altogether [1].
Maybe E2E, but the data eventually has to be decrypted to read it.
Then you learn that every modern CPU has a built-in backdoor, a dedicated processor core, running a closed-source operating system, with direct access to the entire system RAM, and network access. [a][b][c][d].
What you cited is for data on a device that was turned off. Not daily internet connected usage. No one is saying you have no protection at all with Apple, it is just very limited compared to what it should be by modern security best practices, and much worse than what can be achieved on android and linux.
If they didn't want you to think key escrow might be possible, why wouldn't they just leave the wording the way it was? Why go through the effort and thereby draw attention to it? The court system doesn't use sovcit rules where playful interpretation of wording can get a trillion dollar corporation out of a lawsuit or whatever.
It’s also a bet that the capex cost for training future models will be much lower than it is today. Why invest in it today if they already have the moat and dominant edge platform (with a loyal customer base upgrading hardware on 2-3 year cycles) for deploying whatever future commoditized training or inference workloads emerge by the time this Google deal expires?
Personally also think it's very smart move - Google has TPUs and will do it more efficiently than anyone else.
It also lets Apple stand by while the dust settles on who will out innovate in the AI war - they could easily enter the game on a big way much later on.
Seems like the LLM landscape is still evolving, and training your own model provides no technical benefit as you can simply buy/lease one, without the overhead of additional eng staffing/datacenter build-out.
I can see a future where LLM research stalls and stagnates, at which point the ROI on building/maintaining their own commodity LLM might become tolerable. Apple has had Siri as a product/feature and they've proven for the better part of a decade that voice assistants are not something they're willing to build a proficiency in. My wife still has an apple iPhone for at least a decade now, and I've heard her use Siri perhaps twice in that time.
And if you wanted to build your own data center right now there’s only so much GPU and RAM to go around, and even all the power generation and cooling manufacturers are booked solid.
The trouble is this seems to me like a short term fix, longer term, once the models are much better, Google can just lock out apple and take everything for themselves and leave Apple nowhere and even further behind.
Of course there is going to be an abstraction layer - this is like Software Engineering 101.
Google really could care less about Android being good. It is a client for Google search and Google services - just like the iPhone is a client for Google search and apps.
Agreed, especially since this is a competitive space with multiple players, with a high price of admission, and where your model is outdated in a year, so its not even capex as much as recurring expenditure. Far better to let someone else do all the hard work, and wait and see where things go. Maybe someday this'll be a core competency you want in-house, but when that day comes you can make that switch, just like with apple silicon.
> They simply do not have the TPU pods or the H100 clusters to train a frontier model like Gemini 2.5 or 3.0 from scratch without burning 10 years of cash flow.
Why does Apple need to build its own training cluster to train a frontier model, anyway?
Why couldn't the deal we're reading about have been "Apple pays Google $200bn to lease exclusive-use timeslots on Google's AI training cluster"?
Apple sells consumer goods first and foremost. They likely don't see a return on investment through increased device or services sales to match the hundreds of billions that these large AI companies are throwing down every year.
Of all the companies to survive a crash in AI unscathed, I would bet on Apple the most.
They are only ones who do not have large debts off(or on) balance sheet or aggressive long term contracts with model providers and their product demand /cash flow is least dependent on the AI industry performance.
They will still be affected by general economic downturn but not be impacted as deeply as AI charged companies in big tech.
They have the largest free cash flow (over $100 billion a year). Meta and Amazon have less than half that a year, and Microsoft/Nvidia are between $60b-70b per year. The statement reflects a poor understanding of their financials.
> To me, this deal is about the bill of materials for intelligence. Apple admitted that the cost of training SOTA models is a capex heavy-lift they don't want to own. Seems like they are pivoting to becoming the premium "last mile" delivery network for someone else's intelligence. Am I missing the elephant in the room?
Probably not missing the elephant. They certainly have the money to invest and they do like vertical integration but putting massive investment in bubble that can pop or flatline at any point seems pointless if they can just pay to use current best and in future they can just switch to something cheaper or buy some of the smaller AI companies that survive the purge.
Given how much AI capable their hardware is they might just move most of it locally too
The cash pile is gone, they have been active in share repurchase.
They still generate about ~$100 billion in free cash per year, that is plowed into the buybacks.
They could spend more cash than every other industry competitor. It's ludicrous to say that they would have to burn 10 years of cash flow on trivial (relative) investment in model development and training. That statement reflects a poor understanding of Apple's cash flow.
Honestly, I'm relieved...it's not really in their DNA and not pivotal to their success; why pivot the company into a U turn into a market that's vague defined and potentially algorithmically limited?
Apple is flush with cash and other assets, they have always been. They most likely plan to ride out the AI boom with Google's models and buy up scraps for pennies on the dollar once the bubble pops and a bunch of the startups go bust.
It wouldn't be the first time they went for full vertical integration.
calling neural engine the best is pretty silly. the best perhaps of what is uniformly a failed class of ip blocks - mobile inference NPU hardware. edge inference on apple is dominated by cpus and metal, which don't use their NPU.
Apple has the best edge inference silicon in the world (neural engine), but they have effectively zero presence in a training datacenter. They simply do not have the TPU pods or the H100 clusters to train a frontier model like Gemini 2.5 or 3.0 from scratch without burning 10 years of cash flow.
To me, this deal is about the bill of materials for intelligence. Apple admitted that the cost of training SOTA models is a capex heavy-lift they don't want to own. Seems like they are pivoting to becoming the premium "last mile" delivery network for someone else's intelligence. Am I missing the elephant in the room?
It's a smart move. Let Google burn the gigawatts training the trillion parameter model. Apple will just optimize the quantization and run the distilled version on the private cloud compute nodes. I'm oversimplifying but this effectively turns the iPhone into a dumb terminal for Google's brain, wrapped in Apple's privacy theater.