> We believe in training our models using diverse and high-quality data. This includes
data that we’ve licensed from publishers, curated from publicly available or open-
sourced datasets, and publicly available information crawled by our web-crawler,
Applebot.
> We do not use our users’ private personal data or user interactions when training
our foundation models. Additionally, we take steps to apply filters to remove certain
categories of personally identifiable information and to exclude profanity and unsafe
material.
> Further, we continue to follow best practices for ethical web crawling, including
following widely-adopted robots.txt protocols to allow web publishers to opt out
of their content being used to train Apple’s generative foundation models. Web
publishers have fine-grained controls over which pages Applebot can see and how
they are used while still appearing in search results within Siri and Spotlight.
Respect.
bitpush · 5h ago
When Apple inevitably partners with OpenAI or Anthropic, which by their definition isnt doing "ethical crawling", I wonder how I should be reading that.
I mean they also buy from companies with less ethical supply chain practices than their own. I don’t know that I need to feel anything about that beyond recognizing there’s a big difference between exercising good practices and refusing to deal with anyone who does less.
wmf · 4h ago
In theory Apple could provide their training data to be used by OpenAI/Anthropic.
bitpush · 4h ago
It isn't "apple proprietary" data to give it to OpenAI.
Also the bigger problem is, you can't train a good model with smaller data. The model would be subpar.
bigyabai · 4h ago
"Good artists copy; great artists steal"
- Famous Dead Person
fridder · 4h ago
Same way as the other parts of their supply chain I suppose.
simonw · 4h ago
One problem with Apple's approach here is that they were scraping the web for training data long before they published the details of their activities and told people how to exclude them using robots.txt
dijit · 4h ago
Uncharitable.
Robots.txt is already the understood mechanism for getting robots to avoid scraping a website.
simonw · 4h ago
People often use specific user agents in there, which is hard if you don't know what the user agents are in advance!
lxgr · 1h ago
That seems like a potentially very useful addition to the robots.txt "standard": Crawler categories.
Wanting to disallow LLM training (or optionally only that of closed-weight models), but encouraging search indexing or even LLM retrieval in response to user queries, seems popular enough.
wat10000 · 4h ago
If you're using a specific user agent, then you're saying "I want this specific user agent to follow this rule, and not any others." Don't be surprised when a new bot does what you say! If you don't want any bots reading something, use a wildcard.
lxgr · 1h ago
Yes, but given the lack of generic "robot types" (e.g. "allow algorithmic search crawlers, allow archival, deny LLM training crawlers"), neither opt-in nor opt-out seems like a particularly great option in an age where new crawlers are appearing rapidly (and often, such as here, are announced only after the fact).
simonw · 3h ago
Sure, but I still think it's OK to look at Apple with a raised eyebrow when they say "and our previously secret training data crawler obeys robots.txt so you can always opt out!"
astrange · 3h ago
> Using our web crawling strategy, we sourced pairs of images with corresponding alt-texts.
An issue for anti-AI people, as seen on Bluesky, is that they're often "insisting you write alt text for all images" people as well. But this is probably the main use for alt text at this point, so they're essentially doing annotation work for free.
barbazoo · 46m ago
> An issue for anti-AI people, as seen on Bluesky, is that they're often "insisting you write alt text for all images" people as well. But this is probably the main use for alt text at this point, so they're essentially doing annotation work for free.
How did you come to the conclusions that those two groups overlap so significantly?
simonw · 3h ago
I think it is entirely morally consistent to provide alt text for accessibility even if you personally dislike it being used to train AI models.
astrange · 2h ago
It's fine if you want to, but I think they should consider that basically nobody is reading it. If it was important for society, photo apps would prompt you to embed it in the image like EXIF.
Computer vision is getting good enough to generate it; it has to be, because real-world objects don't have alt text.
lxgr · 1h ago
Why would photo apps do what's "important for society"?
Annotating photos takes time/effort, and I could totally imagine photo apps being resistant to prompting their users for that, some of which would undoubtedly find it annoying, and many more confusing.
Yet I don't think that one can conclude from that that annotations aren't helpful/important to vision impaired users (at least until very recently, i.e. before the widespread availability of high quality automatic image annotations).
In other words, the primary user base of photo editors isn't the set of people that would most benefit from it, which is probably why we started seeing "alt text nudging" first appear on social media, which has both producer and consumer in mind (at least more than photo editors).
simonw · 2h ago
I actually use Claude to generate the first draft of most of my alt text, but I still do a manual review of it because LLMs usually don't have enough contents to fully understand the message I'm trying to convey with an image: https://simonwillison.net/2025/Mar/2/accessibility-and-gen-a...
epolanski · 58m ago
Respect actions, not words and PR.
bigyabai · 3h ago
Gotta polish that fig-leaf to hide Apple's real stance towards user privacy: arstechnica.com/tech-policy/2023/12/apple-admits-to-secretly-giving-governments-push-notification-data/
> Apple has since confirmed in a statement provided to Ars that the US federal government "prohibited" the company "from sharing any information,"
brookst · 3h ago
I mean if you throw out all contrary examples, I suppose you are left with the simple lack of nuance you want to believe
bigyabai · 2h ago
All examples contrary to what? Admitting to being muzzled by feds?
Take all the space you need to lay out your contrary case. Did the San Bernadino shooter predict this?
mittermayr · 5h ago
All I can say is, I asked Siri today (verbatim): What is 75 degrees fahrenheit in celsius, and what is 85 degrees in fahrenheit — and it offered a web search about fahrenheit. The "and" completely disabled its most basic ability to do metric conversions.
So, it's nice to see Apple is doing research and talking about it, but we're out here waiting, still waiting, for anything useful to make of it all on our thousand-dollar devices that literally connect us to the world and contain our entire life data. It's what I would've expected from one of the most valuable companies in the world.
al_borland · 3h ago
You asked 2 questions in a system made for 1 question at a time. Split these up and Siri answers them fine. You’re holding it wrong.
a_wild_dandan · 2h ago
"You haven't contorted your comically simple query enough to make the brittle tool work. Throw the chicken bones better next time."
al_borland · 2h ago
It’s been this way for over a decade. If someone hasn’t figured it out by now, that’s kind of on them.
I’m not even sure why those two things would be asked as a single question. It seems like a very unnatural way to pose those two questions. Most humans would trip on that, especially if it was asked verbally.
lxgr · 52m ago
> It seems like a very unnatural way to pose those two questions. Most humans would trip on that
I'd assume GP only gave an example. As a pretty frequent user, I can unfortunately only confirm that Siri trips over almost every multi-part question.
This would be forgivable if there weren't multiple voice-based AI consumer products available that can handle these kinds of requests perfectly.
sxg · 39m ago
OP isn't asking how to use Siri to do his contrived task. OP is saying that Siri in 2025 should be able to handle that relatively simple albeit contrived task.
CamperBob2 · 2h ago
Never mind that Infocom games running on my Apple ][+ could handle that sort of command in 1983.
(Well, with multiple direct objects, anyway.)
losvedir · 4h ago
> What is 75 degrees fahrenheit in celsius, and what is 85 degrees in fahrenheit
Err, what? As a native English speaker human that's a pretty confusing question to me, too!
hu3 · 3h ago
First, most of the English speaking world is not native.
"As of 2022, there were about 400 million native speakers of English. Including people who speak English as a second language, estimates of the total number of Anglophones vary from 1.5 billion to 2 billion."
Second, all popular models I tested did well with that query, including Gemini on Android (aka "ok Google"), except Apple's.
I am not sure why you go on the subject of English speaking world etc.
Anyway, the models you tested with that query, which I am not sure why we think is a good benchmark, are local models running on a wireless device or they use datacenter and only convey the text back and forth?
lxgr · 45m ago
I'm fairly sure Siri still sends user voice samples to a data center. At least for a while, it used to use multipath TCP to decrease latency over multiple available network connections if I'm not misremembering.
Some modern Apple devices support "local Siri", but it's a limited subset of both voice recognition performance and capabilities.
manoweb · 3h ago
Your usage of Siri today (probably on an old version of iOS) frankly has nothing to do with the article we are discussing. Sorry to say this but it is going to take time. Comparing the performance of a chatgpt running in a big data center with a model running locally on a phone device... give it a few years.
ninkendo · 2h ago
People have been giving Siri a few years for a decade now. Siri used to run in a data center (and still does for older hardware and things like HomePods) and it has never supported compound queries.
Siri needs to be taken out back and shot. The problem with “upgrading” it is the pull to maintain backwards compatibility for every little thing Siri did, which leads them to try and incorporate existing Siri functionality (and existing Siri engineers) to work alongside any LLM. Which leads to disaster, and none of it works and just made it all slower. They’ve been trying to do an LLM assisted Siri for years now and it’s the most public facing disaster the company has had in a while. Time to start over.
lxgr · 47m ago
As a user, I'd gladly opt into a slightly less deeply integrated Siri that understands what I want from it.
Build a crude router in front of it, if you must, or give it access to "the old Siri" as a tool it can call, and let the LLM decide whether to return its own or a Siri-generated response!
I bet even smaller LLMs would be able to figure out, given a user input and Siri response pair, whether the request was resonably answered or whether the model itself could do better or at least explain that the request is out of capabilities for now.
mrheosuper · 48m ago
Those little things have been broken for a while now, it's best to bite the bullet and integrate LLM to Siri now.
lxgr · 50m ago
> Your usage of Siri today (probably on an old version of iOS) frankly has nothing to do with the article we are discussing.
Yes, but isn't that infuriating? The technology exits! It even exists, as evidenced by this article, in the same company that provides Siri!
At least I feel that way every time I interact with it – or for that matter my Google Home speaker, ironically made and operated by the company that invented transformer networks.
basisword · 5h ago
>> What is 75 degrees fahrenheit in celsius, and what is 85 degrees in fahrenheit
Probably wouldn't have made a difference but the second half of that statement isn't exactly clear. 85 degrees what?
I also think when you're chaining these two separate calculations together you get a problem when it comes to displaying the results.
vosper · 4h ago
That exact phrase "What is 75 degrees fahrenheit in celsius, and what is 85 degrees in fahrenheit" given to ChatGPT produces the correct result (it infers that the second degrees must be Celsius) and ChatGPT gives me a nicely laid out formula for the math of the conversion.
So yeah, Apple is way behind on this stuff.
seydor · 4h ago
the fact is that gemini responds with this: 75 degrees Fahrenheit is 23.89 degrees Celsius, and 85 degrees Celsius is 185.00 degrees Fahrenheit.
bronco21016 · 4h ago
Meanwhile users have been conditioned to expect a system that understand the multiple queries and answers them appropriately.
JKCalhoun · 4h ago
True. But for most of us, only in the past year. I have a few friends/relatives who have still never conversed with an LLM.
ApolloVonZ · 4h ago
Despite all the “Apple is evil” or “Apple is behind” (because they don’t do evil). Well, what they made with the Foundation Model is great. The fact that they build a system within the Swift language that allows you to specify structured data models (structs) to be used like any other model in a modern programming language, and you actually get back generated data in that format is great. Unlike a lot of other AIs where you might get back a well formatted JSON after a carefully crafted request, but still you never can’t be sure and need to implement a bunch of safeguards. Obviously it’s still the beginning and other tools might do something similar. But as an iOS developer that makes the usage of AI so much simpler. Especially with the bridge to external AIs that still allows you to map back to the type safe structured Swift models. I try not to be a hater, every progress, even slow or underwhelming at first might lead to improvements everywhere else.
lxgr · 1h ago
How do you think their implementation works under the hood? I'm almost certain it's also just a variant of "structured outputs", which many inference providers or LLM libraries have long supported.
0x457 · 4h ago
Guided generation is called "Structured Output" by other providers?
Well partially generated content streaming thing is great and I haven't seen it anywhere else.
ApolloVonZ · 3h ago
Sorry if I didn’t use the correct terms. Didn’t catch up on all the terminology coming from my native language. ;) But yes, I agree, the fact that parts, different parameters, of the model can be completed asynchronous by streaming the output of the model, is quite unique. Apple/swift was late with async/await, but putting it all together, it probably plays well with the ‘never’ (I know ) asynchronous and reactive coding.
astrange · 3h ago
An issue with this is that model quality can get a lot lower when you force it into a structured form, because it's out of distribution for the model.
(I'm pretty sure this is actually what drove Microsoft Sydney insane.)
Reasoning models can do better at this, because they can write out a good freeform output and then do another pass to transform it.
0x457 · 2h ago
I have this toy agent I'm writing, I always laugh that I, human, write a code that generates human-readable markdown, that I feed to llm where I ask it to produce a json, so I can parse (by code I, or it wrote) and output in a consistent human-readable form.
I'm thinking about let it output freeform and then use another model to use to force that into structured.
astrange · 1h ago
IIRC yaml is easier for models than json because you don't need as much recursive syntax.
t1amat · 1h ago
I doubt this is true anymore, if ever. Both require string escaping, which is the real hurdle. And they are heavily trained on JSON for tool calling.
a_wild_dandan · 2h ago
Huh? Grammar-based sampling has been commonplace for years. It's a basic feature with guaranteed adherence. There is no "carefully crafting" anything, including safeguards.
By the time Apple has an AI-native product ready, people will already associate it with dehumanization and fascism.
jonplackett · 6h ago
Every time I see a paper from Apple I just feel like, OK so why isn’t my iPhone actually doing any of this yet?
Why give this to developers if you haven’t been able to get Siri to use it yet? Does it not work or something? I guess we’ll find out when devs start trying to make stuff
imoverclocked · 6h ago
> why isn’t my iPhone actually doing any of this yet?
What exactly are you referring to? Models do run on iPhone and there are features that take advantage of it, today.
jonplackett · 4h ago
None of those features are in any way interesting though. Image playground is a joke, Siri is a joke, that generative emoji thing is a joke.
The AI stuff with photography sure, but that’s more like machine learning.
The photo touch up thing is… useable? Sometimes?
What is it you’ve been so impressed with?
astrange · 3h ago
The main features are text summarization, search, and writing tools.
lxgr · 44m ago
Yes, but why do I have to open a third-party app to do these things when Apple, the company that primarily popularized the entire genre of mobile voice assistants, could very feasibly bake all of that into theirs?
I mean, the thing even lets me ask ChatGPT things if I explicitly ask it to! But why do I need to ask in the first place?
bayindirh · 6h ago
> why isn’t my iPhone actually doing any of this yet?
Probably Apple is trying to distill the models so they can run on your phone locally. Remember, most, if not all, of Siri is running on your device. There's no round trip whatsoever for voice processing.
Also, for larger models, there will be throwaway VMs per request, so building that infra takes time.
lxgr · 35m ago
They just launched "Private Cloud Compute" with much fanfare to enable server-side LLM processing, so between that and the fact that Siri has been server-based for most of its existence (local processing is fairly new), I don't think that's their main constraint at this point.
That said, "Private Cloud Compute" does run on proprietary Apple hardware, so availability might be a concern (assuming they don't want to start charging for it).
jonplackett · 6h ago
It says there’s 2 models - one local. It’s already released to app developers to use locally I think (it was in the keynote for WWDC).
geoffpado · 5h ago
The model now available to developers (in beta, not in released versions of iOS) is the same model that powers stuff like the much-maligned notification summaries from iOS 18. So your phone does have features that are powered by this stuff… you may just not be particularly overwhelmed by those features.
jonplackett · 4h ago
That’s kinda my point though - is this only capable of things like this? If it ia capable of more, why isn’t there something more yet, it’s been a long time waiting…
totetsu · 3h ago
Apple silicone unified memory is amazing for running things like ollama. You don’t have to wait for them to release their own applications.
yalogin · 2h ago
I know apple is methodical and don’t show their hand but I cannot help but feel they are releasing all this research because they haven’t integrated any into the phone or provided a compelling AI functionality for their users. This is their only way to say “hey we are good with AI too”.
mensetmanusman · 4h ago
Apple can’t afford to run models, there are too many iPhones and not enough data centers.
Running on device is also risky because cycle limitations will make it seem dumb in comparison.
frankfrank13 · 5h ago
AFAICT this is the first commercial model trying to be marketed as responsibly-sourced. Love it, but it also seems like the noise around this issue has died down. Is this for legal cover? Or more apple-privacy marketing
Daedren · 5h ago
Stockholders are suing them over Apple Intelligence. Definitely legal cover.
woah · 4h ago
"Sorry we are hilariously far behind everyone else in the industry after having made a huge amount of fanfare about 'Apple Intelligence' for years. It's just that we have shot ourselves in the knee to satisfy Bluesky posters and the NY Time's lawyers"
msgodel · 5h ago
Do people have an issue with the smollm datasets? I guess it isn't really commercial.
chevman · 4h ago
Siri is literally a joke!
My son (he's 11 years old now and fairly skilled with all the main AI tools, eg chatgpt, gemini, etc) and I retry her every month or so, and this past time we just laughed. Can't handle basic questions - hears the question wrong, starts, stops, takes us to some random ass webpage, etc, etc.
"She's so jacked up!" he said.
Apple needs to get this under control and figured out, stat!
jiehong · 5h ago
Looks nice. I just wish they’d improve the models behind dictation on both iPhone and Mac to have better accuracy and on the fly multiple language transcription.
bitpush · 5h ago
The more I think about Apple, the more I realize that Apple is so far behind. While other companies are pushing the envelope (OpenAI, Anthropic, Google ..) Apple's ambitions seem much much smaller.
And this is after they made very big claims with Apple Intelligence last year, when they had everyone fooled.
This is like watching a train-wreck in slow motion.
halJordan · 4h ago
Apple's ambitions are actually bigger than openai or anthropopic. Only Google's ambition (surprise surprise) is similar. Apple fundamentally wants the llm to be a tool. It doesn't want the llm to be the product.
outworlder · 4h ago
Only if you think they _must_ compete with large models on the internet.
Uehreka · 4h ago
I wouldn’t go as far as GP, but yes, absolutely, they must compete with large models on the internet. Customers are now used to being able to ask a computer a question and get something better than “I just ran a web search for what you said, here are the uncurated, unsummarized results”.
Yes, this is in fact what people want. Apple is the biggest company in the world (don’t quibble this y’all, you know what I mean) and should be able to deliver this experience. And sure, if they could do it on device that would be aces, but that’s not an item on the menu, and customers seem fine with web-based things like ChatGPT for now. To act like Apple is doing anything other than fumbling right now is cope.
avianlyric · 3h ago
Erm, have you heard of these things called apps? It’s this magical concept where other companies can run code your iPhone, and deliver all the features you just talked about.
I don’t really understand why Apple has to provide a ChatGPT product, baked directly into their software. Why on earth would Apple want to get involved in the race to the bottom for the cheapest LLMs? Apple doesn’t produce commodity products, they package commodities into something much more unique that gives them a real competitive advantage, so people are willing to pay a premium for the Apple’s product, rather than just buying the cheapest commodity equivalent.
There is no point Apple just delivering an LLM. OpenAI, Anthropic, Google etc already do that, and Apple is never going to get into the pay-per-call API service they all offer. Delivering AI experiences using on-device only compute, that’s something OpenAI, Anthropic and Google can’t build, which means Apple can easily charge an premium for it, assuming they build it.
bitpush · 2h ago
> I don’t really understand why Apple has to provide a ChatGPT product
Control. It boils down to control. If you own a platform, you want to make your "suppliers" (apps in this case) as substitutable as possible.
If people start associating ChatGPT or Claude or Gemini as the main reasons to buy a phone, at some point in the future, they'll think - gee, most of what I'm doing on the phone is interacting with $app, and I can get the $app elsewhere.
const_cast · 3h ago
This usecase is run of the mill for someone like Google, who used to store and show you your location forever, but it's not in Apple style.
It's hard to be like "uhhh privacy" when you send all requests to a remote server where they're stored in clear text for god knows how long.
As of right now, there is no way to run big LLMs in a privacy preserving manner. It just doesn't exist. You can't E2EE encrypt these services, because the compute is done on the server, so it has to decrypt it.
There are some services which will randomize your instance and things like that, but that kind of defeats the a big part of what makes LLMs useful, context. Until we can run these models locally, there's no way to get around the privacy nightmare aspects of it.
Doesnt matter if it doesnt work. And by all accounts, Apple Intelligence has been a garbage fire.
Siri, even after decades of investment, is a joke. Apple does NOT have the talent or capability to deliver what people want.
GeekyBear · 4h ago
> I wouldn’t go as far as GP, but yes, absolutely, they must compete with large models on the internet
The people running large models want to charge a monthly fee for that.
I'm fine with having a free model that runs on device without slurping up my data.
specialist · 4h ago
I'm fine with Apple chilling on the sidelines for a bit.
slashnode · 2h ago
I think it's the right strategy for Apple.
They're not a model company. The risks of deploying something half-baked to their users is unacceptable. They're taking it slow and trying to do it in a way that doesn't damage/erode their brand.
Wait it out, let the best model(s) rise to the surface (and the hallucination problems to get sufficiently mitigated), and then either partner with a proprietary provider or deploy one of the open source models. Makes more sense than burning billions of dollars training a new foundation model
bitpush · 2h ago
This is a reasonable approach, but unfortunately misses what made Apple soooo successful. Apple is the master of controlling the brand. Apple DOES NOT like to highlight their suppliers. Nobody knows who makes iPhones displays, or sensors, or RAMs.
They love to "invent" brands that they control, so that they can commodotize the underlying supplier. Hey user, it is a retina display and dont worry whether it is LG or Samsung is making it.
Apple tried this with AI, calling it "Apple Intelligence". Unfortunately that faltered. Now Apple will have to come out and say "iPhone with ChatGPT" or "Siri with Claude". AND APPLE HATES THAT. HATES IT WITH PASSION.
People will start to associate smartness with ChatGPT or Claude, and Apple loses control and OpenAI/Anthropic's leverage goes up.
Apple has painted themselves into a corner. And as I said elsewhere, it is a train-wreck happening in slowmotion.
thebytefairy · 1h ago
They already deployed half-baked models (eg needing to disable news summaries because they were so bad), and haven't delivered on other aspects of apple intelligence. This is hard to call being cautious, this is them not being able to keep up.
steve-atx-7600 · 2h ago
Exactly. Another mobile.me moment that adversely impacts customers is worse than making something useful that works. Anyone that “needs” AI can use an app.
JKCalhoun · 4h ago
I see it as the opposite. Apple is absolutely positioned to own "chat". I am not worried they'll soon sort things out — and eventually we'll have an LLM integrated into the iPhone; call it Siri or otherwise.
With my history encrypted in the cloud, and the trust that Apple has built around privacy ... I think they're going to come out alright.
martinald · 4h ago
But they have de facto admitted failure of most of the strategy if the rumours are true that they are switching much harder to OpenAI/Anthropic for upcoming LLM products.
This is the first time in 10+ years I've seen Apple so far on the back foot. They usually launch category defining products that are so far ahead of the competition, even by the time they work through the 'drawbacks' in the first versions of them they are still far ahead. OS X, the iPhone and the iPad were all like that. They are still way ahead of the competition on Apple Silicon as well.
I am not very confident on their on device strategy at least in the short to medium term. Nearly all their devices do not have enough RAM and even if they did SLMs are very far behind what users "know" as AI - even the free ChatGPT plan is leap years ahead of the best 3B param on device model. Maybe there will be huge efficiency gains.
Private cloud is used AFIAK for virtually 0 use cases so far. Perhaps it will be more interesting longer term but not very useful at the moment given the lack of a suitable (ie: non Chinese), large (>500b param) model. They would also struggle to scale it if they roll it out to billions of iOS devices especially if they put features that use a lot of tokens.
Then they've got OpenAI/Gemini/Anthropic via API. But this completely goes against all their private cloud messaging and gives those providers enormous potential control over Apple, which is not a position Apple usually finds itself in. It will also be extremely expensive to pay someone per token for OS level features for billions of iOS/Mac devices and unless they can recoup this via some sort of subscription will hit services margins badly.
To me its clear the future of "OS" is going to involve a lot of agentic tool calling. These require good models, with large context windows and a lot of tokens - this will definitely not work on device. Indeed this is exactly what the Siri vapourware demo was.
I'm sure they can potentially get to a great UX (though these missteps are making me question this). But having such a core feature outsourced does not leave them in a good position.
JKCalhoun · 1h ago
You're right about the RAM, of course. Apple will no doubt have to run that up. At the same time it's an obvious "top tier" feature for the "Apple aiPhone 17 Max". And it will cost dearly.
alwillis · 3h ago
> Private cloud is used AFIAK for virtually 0 use cases so far.
Applications using Apple's foundation models can seamlessly switch from on-device models to Private Compute Cloud.
Research is already showing the use of LLMs for people's most intimate relationship and medical issues. The usual suspects will try to monetize that, which why Private Cloud Compute is a thing from the jump.
> Then they've got OpenAI/Gemini/Anthropic via API. But this completely goes against all their private cloud messaging
Using ChatGPT via Siri today, no personally identifying information is shared with OpenAI and those prompts aren't used for training. I suspect Apple would want something similar for Google, Anthropic, etc.
At some point, there will be the inevitable enshitification of AI platforms to recoup the billions VCs have invested, which means ads, which won't happen to Apple users using foundation model-based apps.
> Nearly all their devices do not have enough RAM and
Every Apple Silicon Mac (going back to the M1 in 2020) can run Apple Intelligence. 8 GB RAM is all they need. Every iPhone 15 Pro, Pro Max and the entire 16 line can all run Apple Intelligence.
Flagship iPhone 17 models are expected to come with 12 GB of RAM and all current Mac models come with at least 16 GB.
Apple sells over 200 million iPhones in a given year.
There's no doubt Apple stumbled out of the gate regarding AI; these are early days. They can't be counted out.
dialup_sounds · 3h ago
Apple is only "behind" if you think they're in the same race. They haven't shown any interest in developing frontier models or taking on the enormous costs of doing so.
bitpush · 2h ago
Did you even watch Apple Intelligence ads? They were very much in the race, just that they got ahead of themselves a bit.
They were touting the same features that other companies are now delivering. Point the phone at something, and it'll tell you what you're looking at. Or summarize news articles etc. Instead we got .. emojithingy
visarga · 4h ago
The paper was a very nice read, and they did many creative things. It's a pity this model won't be directly accessible, only integrated in some apps.
alwillis · 4h ago
> It's a pity this model won't be directly accessible, only integrated in some apps.
It's already accessible using Shortcuts, even to non-developers "iOS 26 Shortcuts + Apple Intelligence is POWERFUL " (Youtube) [1].
When the Blackberry ruled the Earth, people asked 'Why doesn't Apple do a smartphone?'.
JacobJack · 6h ago
I'd really like to be able to use this 3B model on my little 4GB GPU card!
It looks very capable for a reasonable weight.
Maybe one day on HhuggingFace
sneilan1 · 5h ago
I feel like this is the most exciting news today about AI on hn. I really hope apple shows that small models can be just as capable as the bigger ones. Maybe they have the people on perplexity working on these small models.
JKCalhoun · 7h ago
I wonder if we'll see these models running on the phone (aiPhone) hardware in the future.
alwillis · 6h ago
As someone mentioned, this model is available in the beta version of iOS 26; it's also part of macOS 26, iPadOS 26 and visionOS 26. Anyone with a free developer account can install the developer betas; the public beta is expected next week.
There's a WWDC video "Meet the Foundation Models Framework" [1].
It does. You can use it directly on iOS 26 beta - without writing a line of code I can toy with the on-device model through Shortcuts on my 16 Pro. It’s not meant to be a general purpose chatbot… but it can work as a general purpose chatbot in airplane mode which is a novel experience.
The above took like 3 seconds to generate. That little box that says On-device can be flipped between On-device, Private Cloud Compute, and ChatGPT.
Their LLM uses the ANE sipping battery and leaves the GPU available.
JKCalhoun · 6h ago
Wild to see what improvements might come if there is additional hardware support in future Apple Silicon chips.
ivape · 5h ago
What’s the cost of pointing it to Private Cloud Compute? It can’t be free, can it?
floam · 2h ago
It’s “free”, as in it doesn’t charge you anything or require a subscription: it’s a part of Apple Intelligence which is basically something bought with the device. It’s in the cloud so theoretically one shouldn’t need a quite new iPhone or Mac but - one does.
No comments yet
bigyabai · 6h ago
It would be interesting to see the tok/s comparison between the ANE and GPU for inference. I bet these small models are a lot friendlier than the 7B/12B models that technically fit on a phone but won't accelerate well without a GPU.
mrheosuper · 37m ago
fitting 7B model on phone with 8gb ram for the whole system is impressive.
gleenn · 6h ago
I thought the big difference between the GPU and ANE was that you couldn't use the ANE to train. Does the GPU actually perform faster during inference as well? Is that because the ANE are designed more for efficiency or is there another bigger reason?
wmf · 6h ago
GPUs are usually faster for inference simply because they have more ALUs/FPUs but they are also less efficient.
kingnothing · 7h ago
> The new Foundation Models framework gives access to developers to start creating
their own reliable, production-quality generative AI features with the approximately
3B parameter on-device language model. The ∼3B language foundation model at the
core of Apple Intelligence excels at a diverse range of text tasks like summarization,
entity extraction, text understanding, refinement, short dialog, generating creative
content, and more. While we have specialized our on-device model for these tasks,
it is not designed to be a chatbot for general world knowledge. We encourage app
developers to use this framework to design helpful features tailored to their apps
Zee2 · 7h ago
> a ∼3B-parameter on-device model
ThomasBb · 6h ago
There are even already some local AFM to Open AI API bridge project on GitHub - that lets you point basically any Open AI compatible client at the local models. Super nice for basic summarisation and completions.
JKCalhoun · 6h ago
I was worried "device" was a Mac mini, not an iPhone. (I already have been running models on my MacBook Pro.)
leot · 7h ago
The dozens of "contributors" being presented in random order is, one would suppose, an anti-poaching tactic?
zamadatix · 7h ago
It's hard to know what it isn't for certain but there are many other reasons papers list contributors in a flat structure (be it random or alphabetical order). Particularly with large numbers of collaborators.
JKCalhoun · 7h ago
"References" section sort of narrows the field anyway.
browningstreet · 6h ago
As someone whose last name is near the end of the alphabet, that's not the first presumption I had seeing that page.
44520297 · 3h ago
Considering a large portion of the contributors have names originating in a script and language that has no relationship whatsoever to English’s already arbitrary letter ordering, this list configuration is as good as any.
ml-anon · 6h ago
Well meta already got Ruoming so he can obviously give them a ranked list of who to grab.
Most of his team are former Google brain so GDM knows who is good.
rafram · 7h ago
Not very hard to look people up on LinkedIn and figure out who the core researchers are. I think this is just a very surface-level overview paper that encompasses a bunch of different research projects conducted by different teams, and it would be difficult to order the contributors in any meaningful way.
EGreg · 4h ago
Here is my question…
This is the first time that millions of people will actually download and run a model on their own devices.
The question is… will Apple be constantly tweaking these models, or only during OS upgrades?
I for one really like local software. Call me old-fashioned, but I enjoy when a company doesn’t switch up software anytime on the server, or phone the results home all the time in order to extract more profits from their collective users.
alwillis · 3h ago
> The question is… will Apple be constantly tweaking these models, or only during OS upgrades?
Certainly when new updates are released--going from macOS 26 to 26.1).
They can probably push model updates between releases if necessary.
floam · 2h ago
Per the PDF in this post:
> “Adapters produced by the toolkit are fully compatible with the Foundation Models framework. However, each adapter is compatible with a single specific model version, meaning that a new adapter must be trained for each new version of the base model.”
Any changes should require retraining any LoRA adapters that has been built & distributed by third party developers, so they wouldn’t update the models outside OS updates at the drop of a hat I don’t think.
LoRA adapters can be distributed via Background Assets, but the base model itself should be version-locked to the OS build (e.g. iOS 26.0 → 26.1) and updates only when Apple ships a new OS image.
alwillis · 1h ago
Makes sense; thanks for the clarification.
wmf · 4h ago
The model is gigabytes so I doubt they will push updates frequently.
robotresearcher · 3h ago
Educate me: is there any work on modifying models in a way that changes relatively few parameters, so an update is a smaller payload?
wmf · 3h ago
Yeah, LoRAs. Apple uses them to specialize a single model for different uses.
poszlem · 5h ago
In the meantime, when I ask Siri to set a timer for 15 minutes, about 10–15% of the time it just says, “Here’s what I found about setting a timer for 15 minutes,” instead of actually setting the timer"
> We do not use our users’ private personal data or user interactions when training our foundation models. Additionally, we take steps to apply filters to remove certain categories of personally identifiable information and to exclude profanity and unsafe material.
> Further, we continue to follow best practices for ethical web crawling, including following widely-adopted robots.txt protocols to allow web publishers to opt out of their content being used to train Apple’s generative foundation models. Web publishers have fine-grained controls over which pages Applebot can see and how they are used while still appearing in search results within Siri and Spotlight.
Respect.
https://www.bloomberg.com/news/articles/2024-06-12/apple-to-...
Also the bigger problem is, you can't train a good model with smaller data. The model would be subpar.
- Famous Dead Person
Robots.txt is already the understood mechanism for getting robots to avoid scraping a website.
Wanting to disallow LLM training (or optionally only that of closed-weight models), but encouraging search indexing or even LLM retrieval in response to user queries, seems popular enough.
An issue for anti-AI people, as seen on Bluesky, is that they're often "insisting you write alt text for all images" people as well. But this is probably the main use for alt text at this point, so they're essentially doing annotation work for free.
How did you come to the conclusions that those two groups overlap so significantly?
Computer vision is getting good enough to generate it; it has to be, because real-world objects don't have alt text.
Annotating photos takes time/effort, and I could totally imagine photo apps being resistant to prompting their users for that, some of which would undoubtedly find it annoying, and many more confusing.
Yet I don't think that one can conclude from that that annotations aren't helpful/important to vision impaired users (at least until very recently, i.e. before the widespread availability of high quality automatic image annotations).
In other words, the primary user base of photo editors isn't the set of people that would most benefit from it, which is probably why we started seeing "alt text nudging" first appear on social media, which has both producer and consumer in mind (at least more than photo editors).
> Apple has since confirmed in a statement provided to Ars that the US federal government "prohibited" the company "from sharing any information,"
Take all the space you need to lay out your contrary case. Did the San Bernadino shooter predict this?
So, it's nice to see Apple is doing research and talking about it, but we're out here waiting, still waiting, for anything useful to make of it all on our thousand-dollar devices that literally connect us to the world and contain our entire life data. It's what I would've expected from one of the most valuable companies in the world.
I’m not even sure why those two things would be asked as a single question. It seems like a very unnatural way to pose those two questions. Most humans would trip on that, especially if it was asked verbally.
I'd assume GP only gave an example. As a pretty frequent user, I can unfortunately only confirm that Siri trips over almost every multi-part question.
This would be forgivable if there weren't multiple voice-based AI consumer products available that can handle these kinds of requests perfectly.
(Well, with multiple direct objects, anyway.)
Err, what? As a native English speaker human that's a pretty confusing question to me, too!
"As of 2022, there were about 400 million native speakers of English. Including people who speak English as a second language, estimates of the total number of Anglophones vary from 1.5 billion to 2 billion."
Second, all popular models I tested did well with that query, including Gemini on Android (aka "ok Google"), except Apple's.
https://en.m.wikipedia.org/wiki/English-speaking_world
Some modern Apple devices support "local Siri", but it's a limited subset of both voice recognition performance and capabilities.
Siri needs to be taken out back and shot. The problem with “upgrading” it is the pull to maintain backwards compatibility for every little thing Siri did, which leads them to try and incorporate existing Siri functionality (and existing Siri engineers) to work alongside any LLM. Which leads to disaster, and none of it works and just made it all slower. They’ve been trying to do an LLM assisted Siri for years now and it’s the most public facing disaster the company has had in a while. Time to start over.
Build a crude router in front of it, if you must, or give it access to "the old Siri" as a tool it can call, and let the LLM decide whether to return its own or a Siri-generated response!
I bet even smaller LLMs would be able to figure out, given a user input and Siri response pair, whether the request was resonably answered or whether the model itself could do better or at least explain that the request is out of capabilities for now.
Yes, but isn't that infuriating? The technology exits! It even exists, as evidenced by this article, in the same company that provides Siri!
At least I feel that way every time I interact with it – or for that matter my Google Home speaker, ironically made and operated by the company that invented transformer networks.
Probably wouldn't have made a difference but the second half of that statement isn't exactly clear. 85 degrees what?
I also think when you're chaining these two separate calculations together you get a problem when it comes to displaying the results.
So yeah, Apple is way behind on this stuff.
Well partially generated content streaming thing is great and I haven't seen it anywhere else.
(I'm pretty sure this is actually what drove Microsoft Sydney insane.)
Reasoning models can do better at this, because they can write out a good freeform output and then do another pass to transform it.
I'm thinking about let it output freeform and then use another model to use to force that into structured.
By the time Apple has an AI-native product ready, people will already associate it with dehumanization and fascism.
Why give this to developers if you haven’t been able to get Siri to use it yet? Does it not work or something? I guess we’ll find out when devs start trying to make stuff
What exactly are you referring to? Models do run on iPhone and there are features that take advantage of it, today.
The AI stuff with photography sure, but that’s more like machine learning.
The photo touch up thing is… useable? Sometimes?
What is it you’ve been so impressed with?
I mean, the thing even lets me ask ChatGPT things if I explicitly ask it to! But why do I need to ask in the first place?
Probably Apple is trying to distill the models so they can run on your phone locally. Remember, most, if not all, of Siri is running on your device. There's no round trip whatsoever for voice processing.
Also, for larger models, there will be throwaway VMs per request, so building that infra takes time.
That said, "Private Cloud Compute" does run on proprietary Apple hardware, so availability might be a concern (assuming they don't want to start charging for it).
Running on device is also risky because cycle limitations will make it seem dumb in comparison.
My son (he's 11 years old now and fairly skilled with all the main AI tools, eg chatgpt, gemini, etc) and I retry her every month or so, and this past time we just laughed. Can't handle basic questions - hears the question wrong, starts, stops, takes us to some random ass webpage, etc, etc.
"She's so jacked up!" he said.
Apple needs to get this under control and figured out, stat!
And this is after they made very big claims with Apple Intelligence last year, when they had everyone fooled.
This is like watching a train-wreck in slow motion.
Yes, this is in fact what people want. Apple is the biggest company in the world (don’t quibble this y’all, you know what I mean) and should be able to deliver this experience. And sure, if they could do it on device that would be aces, but that’s not an item on the menu, and customers seem fine with web-based things like ChatGPT for now. To act like Apple is doing anything other than fumbling right now is cope.
I don’t really understand why Apple has to provide a ChatGPT product, baked directly into their software. Why on earth would Apple want to get involved in the race to the bottom for the cheapest LLMs? Apple doesn’t produce commodity products, they package commodities into something much more unique that gives them a real competitive advantage, so people are willing to pay a premium for the Apple’s product, rather than just buying the cheapest commodity equivalent.
There is no point Apple just delivering an LLM. OpenAI, Anthropic, Google etc already do that, and Apple is never going to get into the pay-per-call API service they all offer. Delivering AI experiences using on-device only compute, that’s something OpenAI, Anthropic and Google can’t build, which means Apple can easily charge an premium for it, assuming they build it.
Control. It boils down to control. If you own a platform, you want to make your "suppliers" (apps in this case) as substitutable as possible.
If people start associating ChatGPT or Claude or Gemini as the main reasons to buy a phone, at some point in the future, they'll think - gee, most of what I'm doing on the phone is interacting with $app, and I can get the $app elsewhere.
It's hard to be like "uhhh privacy" when you send all requests to a remote server where they're stored in clear text for god knows how long.
As of right now, there is no way to run big LLMs in a privacy preserving manner. It just doesn't exist. You can't E2EE encrypt these services, because the compute is done on the server, so it has to decrypt it.
There are some services which will randomize your instance and things like that, but that kind of defeats the a big part of what makes LLMs useful, context. Until we can run these models locally, there's no way to get around the privacy nightmare aspects of it.
Siri, even after decades of investment, is a joke. Apple does NOT have the talent or capability to deliver what people want.
The people running large models want to charge a monthly fee for that.
I'm fine with having a free model that runs on device without slurping up my data.
They're not a model company. The risks of deploying something half-baked to their users is unacceptable. They're taking it slow and trying to do it in a way that doesn't damage/erode their brand.
Wait it out, let the best model(s) rise to the surface (and the hallucination problems to get sufficiently mitigated), and then either partner with a proprietary provider or deploy one of the open source models. Makes more sense than burning billions of dollars training a new foundation model
They love to "invent" brands that they control, so that they can commodotize the underlying supplier. Hey user, it is a retina display and dont worry whether it is LG or Samsung is making it.
Apple tried this with AI, calling it "Apple Intelligence". Unfortunately that faltered. Now Apple will have to come out and say "iPhone with ChatGPT" or "Siri with Claude". AND APPLE HATES THAT. HATES IT WITH PASSION.
People will start to associate smartness with ChatGPT or Claude, and Apple loses control and OpenAI/Anthropic's leverage goes up.
Apple has painted themselves into a corner. And as I said elsewhere, it is a train-wreck happening in slowmotion.
With my history encrypted in the cloud, and the trust that Apple has built around privacy ... I think they're going to come out alright.
This is the first time in 10+ years I've seen Apple so far on the back foot. They usually launch category defining products that are so far ahead of the competition, even by the time they work through the 'drawbacks' in the first versions of them they are still far ahead. OS X, the iPhone and the iPad were all like that. They are still way ahead of the competition on Apple Silicon as well.
I am not very confident on their on device strategy at least in the short to medium term. Nearly all their devices do not have enough RAM and even if they did SLMs are very far behind what users "know" as AI - even the free ChatGPT plan is leap years ahead of the best 3B param on device model. Maybe there will be huge efficiency gains.
Private cloud is used AFIAK for virtually 0 use cases so far. Perhaps it will be more interesting longer term but not very useful at the moment given the lack of a suitable (ie: non Chinese), large (>500b param) model. They would also struggle to scale it if they roll it out to billions of iOS devices especially if they put features that use a lot of tokens.
Then they've got OpenAI/Gemini/Anthropic via API. But this completely goes against all their private cloud messaging and gives those providers enormous potential control over Apple, which is not a position Apple usually finds itself in. It will also be extremely expensive to pay someone per token for OS level features for billions of iOS/Mac devices and unless they can recoup this via some sort of subscription will hit services margins badly.
To me its clear the future of "OS" is going to involve a lot of agentic tool calling. These require good models, with large context windows and a lot of tokens - this will definitely not work on device. Indeed this is exactly what the Siri vapourware demo was.
I'm sure they can potentially get to a great UX (though these missteps are making me question this). But having such a core feature outsourced does not leave them in a good position.
Applications using Apple's foundation models can seamlessly switch from on-device models to Private Compute Cloud.
Research is already showing the use of LLMs for people's most intimate relationship and medical issues. The usual suspects will try to monetize that, which why Private Cloud Compute is a thing from the jump.
> Then they've got OpenAI/Gemini/Anthropic via API. But this completely goes against all their private cloud messaging
Using ChatGPT via Siri today, no personally identifying information is shared with OpenAI and those prompts aren't used for training. I suspect Apple would want something similar for Google, Anthropic, etc.
At some point, there will be the inevitable enshitification of AI platforms to recoup the billions VCs have invested, which means ads, which won't happen to Apple users using foundation model-based apps.
> Nearly all their devices do not have enough RAM and
Every Apple Silicon Mac (going back to the M1 in 2020) can run Apple Intelligence. 8 GB RAM is all they need. Every iPhone 15 Pro, Pro Max and the entire 16 line can all run Apple Intelligence.
Flagship iPhone 17 models are expected to come with 12 GB of RAM and all current Mac models come with at least 16 GB.
Apple sells over 200 million iPhones in a given year.
There's no doubt Apple stumbled out of the gate regarding AI; these are early days. They can't be counted out.
They were touting the same features that other companies are now delivering. Point the phone at something, and it'll tell you what you're looking at. Or summarize news articles etc. Instead we got .. emojithingy
It's already accessible using Shortcuts, even to non-developers "iOS 26 Shortcuts + Apple Intelligence is POWERFUL " (Youtube) [1].
[1]: https://youtu.be/Msde-lZwOxg?si=KJqTgtWjpdNDxneh
There's a WWDC video "Meet the Foundation Models Framework" [1].
[1]: https://developer.apple.com/videos/play/wwdc2025/286
https://share.icloud.com/photos/018AYAPEm06ALXciiJAsLGyuA
https://share.icloud.com/photos/0f9IzuYQwmhLIcUIhIuDiudFw
The above took like 3 seconds to generate. That little box that says On-device can be flipped between On-device, Private Cloud Compute, and ChatGPT.
Their LLM uses the ANE sipping battery and leaves the GPU available.
No comments yet
Most of his team are former Google brain so GDM knows who is good.
This is the first time that millions of people will actually download and run a model on their own devices.
The question is… will Apple be constantly tweaking these models, or only during OS upgrades?
I for one really like local software. Call me old-fashioned, but I enjoy when a company doesn’t switch up software anytime on the server, or phone the results home all the time in order to extract more profits from their collective users.
Certainly when new updates are released--going from macOS 26 to 26.1).
They can probably push model updates between releases if necessary.
> “Adapters produced by the toolkit are fully compatible with the Foundation Models framework. However, each adapter is compatible with a single specific model version, meaning that a new adapter must be trained for each new version of the base model.”
Any changes should require retraining any LoRA adapters that has been built & distributed by third party developers, so they wouldn’t update the models outside OS updates at the drop of a hat I don’t think.
LoRA adapters can be distributed via Background Assets, but the base model itself should be version-locked to the OS build (e.g. iOS 26.0 → 26.1) and updates only when Apple ships a new OS image.