Apple says the new personalized Siri upgrade is on track in its development and release. Apple never claimed it was coming in iOS 26.4, nor that there would be split releases of it over time, nor did they say they would use multiple different models for Siri.
Apple is pissed off, in my opinion, and they're calling out bullshit. I'm really happy about that, and I support that completely!
1) Mark Gurman falsely claimed Johny Srouji would quit.
Johny Srouji said that's blatantly false and untrue, and that he loves Apple and isn't going to quit working
2) Mark Gurman falsely claimed new Siri would run on Google's "TPU" servers.
Tim Cook said on an earnings call that the new Siri would run on Apple's custom hardware, PCC servers.
3) Mark Gurman now falsely claims only DAYS after literally saying it would come in iOS 26.4 that Siri is "delayed" after hitting "snags."
Untrue, and Apple refuted him a third time publicly saying Siri is on track and not delayed.
All I'm going to say to on this is that from the moment Mark Gurman started claiming it was coming in iOS 26.4, and I wish I had written this down, I said "that's bullshit."
But more than that: I said to myself, "he will claim for months, repeatedly, that Siri will 'definitely release in 26.4,' only for 2 weeks before he writes it will launch he will claim massive problems are happening, and it's now delayed."
Guess what happened? Mark Gurman is a very, very predictable and awful person. Anyone who believes his nonsense will 100% be misled. Any site that publishes his shit should never be visited, which unfortunately is most of them. Good riddance to AppleInsider, MacRumors, and 9to5Mac who don't care about anything except 24/7 clickbait
That pattern is identical. That cannot be a coincidence.
He even claimed in that article about 18.4 "the true modernized conversational Siri wouldn't come until iOS 20" (iOS 27) when he was saying it was coming in 19.4 (26.4)
Side note: Oh, and he was publicly refuted by Craig Federighi saying that what they showed off at wwdc, although limited in demo videos, was real and working, not vaporware, despite him characterizing it as such ("compute generated videos").
He also incoherently across multiple articles screws up code names, mixing up efforts and renaming stuff, mixing names like "Glenwood," "Linwood," "Campos," "Campo," "LLM Siri," etc together and literally contradicts himself even in the same articles. What the hell he's ever referring to is unclear, even to him apparently.
Also:
[Earlier in article] For iOS 19, Apple’s plan is to merge both systems together and roll out a new Siri architecture. I expect this to be introduced as early as Apple’s Worldwide Developers Conference in June of this year — with a launch by spring 2026 as part of iOS 19.4. The new system, dubbed “LLM Siri” internally, was supposed to also introduce a more conversational approach in the same release. But that is now running behind as well and won’t be unveiled in June.
.[End of his article] Anyway, the follow-up release, iOS 18.5, is where the good stuff is supposed to wind up. It has the AI-infused Siri that the company showed off last June, as well as support for Apple Intelligence in China. But with all of the AI-related delays at Apple, let’s see what happens.
He doesn't even know he's saying. The "unified architecture" IS "LLM Siri," yet he says the unified architecture he expects will be unveiled at WWDC (which was wrong, by the way lol), but the "LLM Siri" won't? Also the Siri with personal context will come in 18.5, but then actually it won't, but it will be unveiled at WWDC, but actually it won't be unveiled at WWDC still?
This is so extremely confusing.
"AI Infused Siri," "Conversational Siri," "LLM Siri," "personalized Siri," "unified architecture Siri," "chatbot siri," "new Siri," "new new Siri," are ALL THE SAME THING, yet he literally splits all of those into somehow separate features and products lol.
Also I am not unaware of his bad attempt at sleight of hand. In the very same article about 26.4 Siri being "delayed" now, he attempts to weasel himself out of the claim that Siri will use TPU from Google for inference.
[Earlier in his article] Beyond those upgrades, Apple is also developing a major new AI initiative for iOS 27, iPadOS 27 and macOS 27: a fully overhauled Siri that operates more like a chatbot. It will be powered by Google servers and a more advanced custom Gemini model.
[Later in] Chief Executive Officer Tim Cook hinted at even more changes down the road during an all-hands meeting with employees last week, saying that the company was working on new data-center chips to bolster its AI capabilities.
“Apple silicon is enabling us to build data center solutions that are tailor-made for our devices,” Cook said. “I will say that, going forward, the work we’re doing is going to enable an entirely new class of products and services.”
Cook was likely referring to Baltra, a long-running project to develop high-performance chips for cloud-based AI processing.
This can't be simultaneously true, especially considering other outlets (who operate similarly to him, unfortunately) literally say Broadcom is working on developing a custom server with Apple for this specific task of Siri and PCC soon, which was named by THOSE outlets as Baltra too. He's backtracking and not trying to say it.
Also you have to love this "juicy scoop" (his opinion) from the 18.4 article:
I attended the debut of Alexa+ in New York City, and it felt like seeing the first ChatGPT demonstrations three years ago: This is going to change everything. Of course, we’ll have to see how it plays out in practice. Amazon won’t actually begin rolling out Alexa+ for a few weeks and the company failed — under prior leadership — to get the software out the door last year. We’ll need to see it operating at full scale, but there’s reason for optimism.
I don't know if it's extremely stupidity, extremely laziness, or extreme malice, but Gurman has directly misquoted Tim Cook, and then used that as a frame for his articles
Now, Chief Executive Officer Tim Cook says that Apple has been working on generative AI technology for years. But I can tell you in no uncertain terms that Apple executives were caught off guard by the industry’s sudden AI fever and have been scrambling since late last year to make up for lost time.
He is referring to a paraphrased quote in one of his articles with a hyperlink under "Tim Cook says." However, and I am relying on the accuracy of a transcript so might be incorrect, Tim Cook did not say this.
Every three months after releasing their corporate earnings, Apple’s CEO and CFO get on the phone and chat with financial analysts. By which we mean, they read from prepared statements and th…
sixcolors.com
Shannon Cross, Credit Suisse: Tim, can you talk a bit about AI? You know, obviously that’s more than the topic of the day, it seems like the topic of the year. Just how do you think about it through your products and services? I know you use it in different ways. But also if you can just give us any thoughts you have on generative AI and, and I don’t know where you see it going. Not sure what you want to say on it, but I’m, I’m really curious as to your take. Thank you.
Tim Cook: Yeah. Thanks for the question, Shannon. You know, as you know, we don’t comment on product roadmaps. I do think it’s very important to be deliberate and thoughtful in how you approach these things. And there’s a number of issues that need to be sorted, as is being talked about in a number of different places. But the potential is certainly very interesting. And we’ve obviously made enormous progress integrating AI and machine learning throughout our ecosystem, and we’ve weaved it into products and features for many years, as you probably know. You can see that in things like fall detection and crash detection and ECG. These things are not only great features, they’re saving people’s lives out there. And so it’s absolutely remarkable. And so we view AI as huge and we’ll continue weaving it in our products on a very thoughtful basis.
He constantly paraphrases in his articles. That a recorded statement on an earnings call is not at all what Gurman claimed in his paraphrase in meaning or intent should concern you. How many times has he misquoted somebody from his "sources?" No one can know except him and his "sources," but suffice to say his work has a lot of problems in general.
It's also completely nonsensical. The same company that used neural networks to perform Face ID authentication and created the industry's first consumer neural network chip somehow doesn't know about transformer models? Transformer models are literally just neural networks. A specific subtype, but not so far removed from the NN's they used that they wouldn't be aware of them. It was a public paper published in academia by Google.
Let me just say that there is some partial truth to what Gurman is saying. Apples ML division is not very healthy at the moment. Buying model weights from Google was a good business decision, at the same time it does not fix the fundamental structural issues Apple is facing.
I too was a bit puzzled why Gurman was pushing the line so hard about Siri running on Google infra.
Let me just say that there is some partial truth to what Gurman is saying. Apples ML division is not very healthy at the moment. Buying model weights from Google was a good business decision, at the same time it does not fix the fundamental structural issues Apple is facing.
Well, I think Face ID, Spatial Personas, Neural Engine are some major examples among many that demonstrate not just overall ML competency, but industry leading competency.
Are transformer models currently Apple's strong suit? No. But TMs are just one piece of a broad, important field of ML; and people like Gurman aren't exactly trying to inform people on who's leaving chatbot companies among others to come to Apple, so it looks a little more chaotic than what's probably happening.
On top of that, Apple has produced genuinely novel and innovative features of TMs (such as ASTC encoding of weights, parallel-track mixture of experts models, Guided Generation, etc), not to mention Gurman himself says that Apple developed internal trillion parameter models "on par" with leading TMs at the time he wrote it.
I also must strongly reiterate Private Cloud Compute is truly revolutionary and completely slept on by almost everyone, despite the fact that's it's working and deployed. They can, in theory, just swap in an Opus 4.6 class model and instantly be at the front. Chatbot companies cannot replicate PCC as easily (or at all). Even Google's "Private AI Compute" is absolutely nothing compared to PCC. That's a different discussion, but absolutely true.
While it's debated on this website (and that's fine), MLX is the strongest open source TM platform out there right now. The community (I just watch them) is really strong, and nothing even comes close. Awni Hannun and others are doing really amazing stuff.
TMs at this point are dime a dozen. They're a commodity. Anyone can rent cloud GPUs and produce TMs at this point (this method is used for some major open source models), and hobbyists indeed do for training data they want to make a TM for.
So with that said, I do welcome you to explain more of what you're thinking as I've explained what I'm thinking (sort of).
I don't know. All I know is I was writing a far longer comment before you replied, and I've put it to the side for one reason: his articles are literally incoherent. He claims all versions of an idea so he can never be wrong. This is why he will unequivocally state something, then paragraphs later contradict himself completely. He gets away with it because everyone regurgitates his gossip.
He constantly mixes up code names, he has numerous typos, he mischaracterizes stuff, and he will say two versions of something and claim them both as his. It's very difficult to even understand what he's saying. The only reason he's in business from Bloomberg is they don't give a shit as long as it produces effects in the industry and clicks. Blogs without thought regurgitate everything he says even if it makes zero sense. This is why I dislike him
I will say my personal theory as to why he keeps claiming that Apple will use Google TPUs is because Apple already does use TPUs. They use TPUs to train TMs (so does Google for Gemini). Apple publicly stated this. Him saying that Apple's moat with PCC will vanish, and Apple will give up privacy because they're so behind feeds into his narratives.
Well, I think Face ID, Spatial Personas, Neural Engine are some major examples among many that demonstrate not just overall ML competency, but industry leading competency.
Oh, there is no doubt that Apple has some of the best engineering teems in the world. I happen to know several of them and spoke to even more, and I can attest that they are truly impressive professionals. As far as customized solutions, tooling, and infrastructure goes, Apple is the absolute top.
Nevertheless, that excellency does not seem to translate well to building large general-purpose AI models, and Apple has been lagging behind here for some time. My suspicion is that this is a leadership and culture issue, and it is possible that the same reason why Apple is so good at many things (small, focused teams working on their own set of problems) is the very reason why they struggle with revamping Siri. Communication between teams is alas not Apple's strongest suit.
I don't know. All I know is I was writing a far longer comment before you replied, and I've put it to the side for one reason: his articles are literally incoherent. He claims all versions of an idea so he can never be wrong. This is why he will unequivocally state something, then paragraphs later contradict himself completely. He gets away with it because everyone regurgitates his gossip.
He constantly mixes up code names, he has numerous typos, he mischaracterizes stuff, and he will say two versions of something and claim them both as his. It's very difficult to even understand what he's saying. The only reason he's in business from Bloomberg is they don't give a shit as long as it produces effects in the industry and clicks. Blogs without thought regurgitate everything he says even if it makes zero sense. This is why I dislike him
I will say my personal theory as to why he keeps claiming that Apple will use Google TPUs is because Apple already does use TPUs. They use TPUs to train TMs (so does Google for Gemini). Apple publicly stated this. Him saying that Apple's moat with PCC will vanish, and Apple will give up privacy because they're so behind feeds into his narratives.
I have the feeling that he got stuck on the that part of the deal where was Apple committing to Google as their cloud provider (they mostly used Amazon before that AFAIK). Now, Siri did use to run on third-party cloud. But now with PCC the context has changed. Apple is still using third-party cloud for many services. Not sure why he was insisting on this particular interpretation that much. I agree with you that Apple will likely continue to use Google servers for training and other infra stuff, but that is entirely compatible with them running their own PCC.
Nevertheless, that excellency does not seem to translate well to building large general-purpose AI models, and Apple has been lagging behind here for some time. My suspicion is that this is a leadership and culture issue, and it is possible that the same reason why Apple is so good at many things (small, focused teams working on their own set of problems) is the very reason why they struggle with revamping Siri. Communication between teams is alas not Apple's strongest suit.
They've been working on improved Siri for years, and failed to deliver on pretty much every account. They also failed to train a functioning advanced large LLM, despite working on it for a very long time and spending tons of money. All this led them to pretty much abandon their in-house efforts to develop these models and instead make a deal with Google buying Gemini weights. There was also a lot of restructuring in Apple's ML division, and quite a bit of confusion.
They've been working on improved Siri for years, and failed to deliver on pretty much every account. They also failed to train a functioning advanced large LLM, despite working on it for a very long time and spending tons of money. All this led them to pretty much abandon their in-house efforts to develop these models and instead make a deal with Google buying Gemini weights. There was also a lot of restructuring in Apple's ML division, and quite a bit of confusion.
They also failed to train a functioning advanced large LLM, despite working on it for a very long time and spending tons of money. All this led them to pretty much abandon their in-house efforts to develop these models
Apple has literally deployed multiple large LLMs at this point. I understand most people here tune out anything related to this stuff, but even according to Gurman himself (who knowingly sat on the info and mischaracterized it) Apple already had developed an LLM literally before ChatGPT became public. This info is now likely given that Awni Hannun confirmed that MLX was being worked on before ChatGPT. MLX is a general framework, yes, but it's a framework designed for TM inference.
Apple has deployed two large scale models to the public, with the second being competitive with 4o. The second gen of which deployed unique innovations (that I just wrote in the previous comment), which allow users to use their cloud model without needing $500 billion data centers, nuclear reactors, or privacy invasion.
Gurman has also stated, and mischaracterized on multiple occasions, that Apple has developed an even larger trillion parameter LLM that executives saw as competitive with leading models at that time. He continues to mention this, saying that this work -- and forgive me I'm trying to paraphrase his work to make it coherent -- continues and is even more advanced now.
What Apple has yet to do is develop personalized Siri with a single TM architecture. That is what Craig said they're working on now, after (rightfully and overdue) ditching the V1 hybrid model.
Apple did not buy Gemini models or weights. As I tried to say in my previous post titled "Apple confirms Google technology, not Gemini, will help power Siri," they only partnered with them to develop models. What they will ultimately do with them is unclear. Gurman says that these are custom models developed specifically for Apple. That makes sense to me. Apple has never used off the shelf parts, even if they're going to use off the shelf parts. They always go to OEMs and customize the off the shelf parts.
AppleInsider insists that Apple is distilling the models Google gave into Apple's internal models. This is also plausible and likely.
This is true, but what the structure is remains to be seen because Giannandrea is still working on Siri and wants to see it through, plus Gurman is the source of that confusion. They said he will retire "in spring," and the other guy isn't even on the webpage.
Assuming you mean personalized Siri, then yes, it's been delayed once. They've maintained repeatedly it's on track now, which makes sense. Everything I've seen corroborates this. Only they know why they were pursuing V1 then V2. Simplicity dictates V2 should have been done from the beginning, but again, there are valid reasons for going V1 to V2. What they have said is they learned, and are returning back to status quo: announce when finished.
And I'm not saying there hasn't been a bumpy road regarding all of this. I'm not saying everything is amazing. Siri has been one feature (fairly or unfairly) that has always been expected more of than technology ever allowed, even today. So when Apple announces and shows off significant enhancement to Siri built on nascent technology, I can understand the disappointment when a delay happens. I've never denied this
my comment about Gurman's incorrect paraphrase of Tim Cook wasn't meant to be "Apple is the leader in generative AI." It was meant to point out something I just noticed, so I wrote about it. This entire post is centered around Gurman constantly being incorrect, and he is. Much of his work is built on paraphrasing sources. So when he incorrectly paraphrases a public, verifiable statement by the CEO, then proceeds to build framing around that in future articles, it seriously calls into question if we should believe what he says. It wasn't a mistake at all
But in addition, I also feel that a lot of catastrophe around this is invented by clickbait outlets and journalists and people reposting it without thought. A lot of it comes from him. Most of it, actually.
He was the one that misled the public to believe the timeline, narrative, and features before it was even announced.
They have trained a bunch of models, yes, and have used them in highly specialized contexts. I wouldn’t say that any of their models was state of the art. From my perspective, the main achievement was distilling/optimizing the model do that it can run on an ultracompact device. I feel that you are overestimating the performance of these models.
Apple did not buy Gemini models or weights. As I tried to say in my previous post titled "Apple confirms Google technology, not Gemini, will help power Siri," they only partnered with them to develop models. What they will ultimately do with them is unclear. Gurman says that these are custom models developed specifically for Apple. That makes sense to me. Apple has never used off the shelf parts, even if they're going to use off the shelf parts. They always go to OEMs and customize the off the shelf parts.
What do you think this „Google technology“ is? It’s the models collectively known as Gemini. What’s important here is that Apple does not use the Gemini service. They have the model weights themselves , and they can further finetune/distill/build their own services around them as they want. The main point is that they don’t have to train the foundation model anymore.
They have trained a bunch of models, yes, and have used them in highly specialized contexts. I wouldn’t say that any of their models was state of the art. From my perspective, the main achievement was distilling/optimizing the model do that it can run on an ultracompact device. I feel that you are overestimating the performance of these models.
What do you think this „Google technology“ is? It’s the models collectively known as Gemini. What’s important here is that Apple does not use the Gemini service. They have the model weights themselves , and they can further finetune/distill/build their own services around them as they want. The main point is that they don’t have to train the foundation model anymore.
Just to make it clear, nothing what I said is based on Gurman.
I've spent a long time writing and rewriting a response multiple times, quite thorough with multiple sources cited.
Respectfully, I don't know how to respond to this, because analysis of Apple's "AI" really isn't the point of my post. I think there may be a misread of the intention of my post, and you're claiming that I claimed things that I didn't claim, and you're saying stuff is my opinion. I tried not to just cite my opinion, but public articles, research, announcements, etc. If you disagree, I do welcome you and want you to tell me what you think is my opinion (about Apple's "AI") in the previous 2 comments.
And if you have specific questions about Apple's "AI" performance, I'm willing to answer them based on what I know, but I only have access to what you have access to: public research, articles, announcements.
This is an earnest reply by the way. I'm not angry.
I've spent a long time writing and rewriting a response multiple times, quite thorough with multiple sources cited.
Respectfully, I don't know how to respond to this, because analysis of Apple's "AI" really isn't the point of my post. I think there may be a misread of the intention of my post, and you're claiming that I claimed things that I didn't claim, and you're saying stuff is my opinion. I tried not to just cite my opinion, but public articles, research, announcements, etc. If you disagree, I do welcome you and want you to tell me what you think is my opinion (about Apple's "AI") in the previous 2 comments.
And if you have specific questions about Apple's "AI" performance, I'm willing to answer them based on what I know, but I only have access to what you have access to: public research, articles, announcements.
This is an earnest reply by the way. I'm not angry.
Sorry to hear you feel that way. I looked through the last couple of posts again and I'm not quite sure what might have prompted such a response. It's perfectly fine with me if we do not continue this conversation at this point. .
Sorry to hear you feel that way. I looked through the last couple of posts again and I'm not quite sure what might have prompted such a response. It's perfectly fine with me if we do not continue this conversation at this point.
I'm so confused at this point lol. Can you elaborate a little more? Feel what way?
I offered you to explain why you thought I was just purely speaking opinion in my previous comments about Apple's models.
I tried to make it clear I wasn't angry and was genuinely inviting you to explain.
I also offered to answer any questions based on what I know of Apple's efforts, because I don't think most people are aware that Apple produced a 4o quality model with 50 billion less parameters? This is by definition an advanced, large LLM that is competitive with a leading model when released.
I allowed the conversation to deviate from my original topic (of pointing out flaws in Mark Gurman's articles), because I thought it was an earnest conversation, and I wanted to contribute insight. It was clear after your last response that there's mismatch between what you think I'm saying and what I'm trying to say. I want us discussing the facts of the matter, so I offered to elaborate on my end where I'm getting all these pieces of info from and wanted to earnestly explain.
If you want to discuss, let's discuss! I just ask we stick to sourced information, as I've tried to. I did not mean insult.
Please consider that some of us might have information which is not sourced from press releases or rumor mills. If you only want to discuss the "official" information, that's fine, it's just that I wouldn't have much to add to that.
Please consider that some of us might have information which is not sourced from press releases or rumor mills. If you only want to discuss the "official" information, that's fine, it's just that I wouldn't have much to add to that.
I did acknowledge that. I "liked" your comment. And that's why I've stuck to public information.
But now you're doing it again. I'm not referring to "press releases." I've cited multiple things, with or without direct sources, but I'm happy to provide them upon request: 1) public statements on earnings calls, 2) public statements by the company to news, 3) Gurman's articles, 4) published research, and 5) press releases of work.
That said, you're basically asking me to dismiss the fact that they produced a 4o quality model with 50 billion less parameters (150 vs 200B), without resorting to building $500 billion data centers, without resorting to funding nuclear reactor companies, without invading people's privacy; in the context that Apple engineers on the AI/ML have found it extremely difficult to train models because of privacy and not being able to train with customer data. So even with all these things in mind, they still produced that.
And as much as I appreciate your claimed connection to Apple employees, this is meaningless to my discussion. I cannot verify that, and even if I could, basing an entire judgement of a team's work from one or two people is insufficient. It suffers the same thing Gurman's work does.
Not to mention the inherent flaw in "trust me bro." I could claim that too. I don't know anyone at Apple, but I could claim it anyways and somehow you're supposed to trust me? I think that isn't useful to a discussion, respectfully. Finally, you never cited your connections, so I'm unclear why you're now bringing it up.
Most importantly, my post is about Gurman and his inaccuracy. It isn't about Apple's "AI," their team, or trying to say everything is fine. I feel like I've been refreshingly objective, and you're sort of taking my thought out comments citing multiple things, and you're kind of just saying, "so what?" I don't understand what the motive is here.
A couple of replies ago you attempted to paint me as overly reactive, when I feel I've been anything but. That's why I'm genuinely confused by the statements like, "sorry you feel that way." Well, I didn't express anything I felt. My comment was dry, and I tried to accommodate you and include you instead of being dismissive. I asked what you meant multiple times.
I said earlier you misinterpreted the point of the post. This isn't about saying "all is well" for their TM team. I'm saying that there are serious discrepancies in Gurman's reporting, and that's validated in his written work.
I once again invite you to go through my comments and specifically tell me what you believe my opinion is. I also invite you to ask me questions based on what I have read, not just of Apple's work but other companies work.
That said, you're basically asking me to dismiss the fact that they produced a 4o quality model with 50 billion less parameters (150 vs 200B), without resorting to building $500 billion data centers, without resorting to funding nuclear reactor companies, without invading people's privacy; in the context that Apple engineers on the AI/ML have found it extremely difficult to train models because of privacy and not being able to train with customer data. So even with all these things in mind, they still produced that.
I am not asking you to dismiss that, I am just skeptical of the claim. Apple's large foundational model is not publicly accessible, and the we only ever saw it in (fairly disappointing, IMO) Xcode workflows. And even if their public claims about the model performance are true, it's not like they have much to show for it, as the model is clearly not widely used.
That's fair. At the same time, we are not in a court of law. It's not like some information is admissible and some not. We are just talking. I (in a friendly manner) maintain that some of your understanding (e.g. about Apple's foundational model and how they use Goggle technology) might be incorrect or incomplete. And no, I can't deliver any proof. Do with that what you wish.
Most importantly, my post is about Gurman and his inaccuracy. It isn't about Apple's "AI," their team, or trying to say everything is fine. I feel like I've been refreshingly objective, and you're sort of taking my thought out comments citing multiple things, and you're kind of just saying, "so what?" I don't understand what the motive is here.
That is also fair. At the same time, my posts are about Apple AI and the team. Sometimes conversations diverge. Then again, it's your thread, if you prefer to merely talk about Gurman, that's fine with me. Only I don't have anything to say about Gurman.
I am not asking you to dismiss that, I am just skeptical of the claim. Apple's large foundational model is not publicly accessible, and the we only ever saw it in (fairly disappointing, IMO) Xcode workflows. And even if their public claims about the model performance are true, it's not like they have much to show for it, as the model is clearly not widely used.
First, thank you for specifically referencing something I said. I can now proceed with where I got that information from.
This is structured into three parts: 1) a published study by Apple's TM team, 2) Gurman's articles, and 3) context of ChatGPT releases.
I need to clarify something first. The AFM PCC 2nd gen model is not at all related to Xcode or what was previously called Swift Assist. Now, if you have knowledge otherwise, I'm not asking nor do I want it. But according to all public info (granted I may make a mistake) these are 2 separate models.
Yes, Swift Assist was not up to par, but that's different. And Apple didn't ship it. Also, coding models are usually tuned for coding. That doesn't mean the model is useless, so if it is actually one and the same, then that doesn't mean it's inherently bad altogether.
With that in mind,
The first part:
Apple said in their testing
(which I have zero reason to doubt is false for a multitude of reasons, including but not limited to: Apple is a more up front company than most, they've published results even if it shows them very far behind (AFM 1st gen was this), and they've characterized their own progress as behind even in the paper I will cite)
that their AFM PCC model 2nd generation ties or matches with ChatGPT 4o 68% of the time.
Otherwise, they equally won.
What did I originally claim? I said it was competitive with leading models at time of release. That performance matches my claims, at least in terms of EN/US responses.
The second part:
Gurman has made hyper specific claims about what models Apple had at WWDC. He said that the largest model was 150 billion parameters. Parameters do not inherently translate to sophistication anymore, but it is a metric. He repeated this claim again. Given his inaccuracy, this could be completely made up, but something he said does seem similar to what Apple AFM team claims: the 2nd generation PCC model is scalable.
Here's what he said, and what they said:
Gurman:
That’s not to say Apple hasn’t made recent progress in AI. Internally, it has a broad range of models of various complexity. Versions with 3 billion, 7 billion, 33 billion and 150 billion parameters are now in active use. The 150 billion parameter model, which relies on the cloud like OpenAI and Google, is far more powerful than Apple’s on-device technology and capable of more nuanced reasoning. Internal benchmarks have shown that this model approaches the quality of recent ChatGPT rollouts. But the company has held off on using the technology to offer its own chatbot due to concerns over hallucinations and philosophical differences among company executives.
We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: (i) a ∼3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and (ii) a scalable server model built on a novel Parallel-Track Mixture-of-Experts (PT-MoE) transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global–local attention to deliver high quality with competitive cost on Apple’s Private Cloud Compute platform.
The custom Gemini system represents a major advance from the 150 billion parameter model used today for the cloud-based version of Apple Intelligence. The move would vastly expand the system’s power and its ability to process complex data and understand context.
...Apple executives believe it can reach a similar quality level as the custom Gemini offering.
ChatGPT 4 was a trillion parameter model, but less sophisticated than 4o. 4o brought improvements to benchmarks, as well as multi-modality. Microsoft inadvertently all but confirmed that 4o was a 200 billion parameter model, likely mixture of experts like Apple's. Given their involvement, i am inclined to believe that.
ChatGPT 4.5 was not as well received as 4o. It was often cited as underwhelming.
Version 5 was even less well received. A study, supposedly double blind, claimed 48% of users preferred 4o over 43% of users 5 in private testing. Public backlash was so strong they brought back a model (4o) for the first time ever.
5 was the one to beat, supposedly. It was so hyped up that it was going to be THE thing for "AGI," but it wasn't the case. Actually it was so terrible both in responses and objectively even during demos (and also shortly after demos) that people finally said stuff like this:
[5] isn’t artificial general intelligence (AGI). You might have had valid reasons to assume that GPT-5 was going to be AGI since we have been teased over several years that GPT-5 was going to move mountains and finally attain true, across-the-board human-level intelligence...
We mainly have a somewhat better generative AI [5] that gets accolades for various notable advances and enhancements. That’s worth a hearty cheer. By and large, you will find it a useful and handy tool. Again, it isn’t pinnacle AI by any stretch of imagination.
These are the qualifications and evidence to my claim "leading models."
At the end of the day, benchmarks are a metric but not the most important one. The fundamental premise of TMs is, "what does the user want?" If users hate the model, no amount of benchmarks (which are largely gamed at this point, intentionally or not) matter. This is why versions 5.1, 5.2, and 5.3 were made using that feedback. Eventually they somewhat rectified it, and users moved over to 5.1, etc. There are still petitions to bring 4o back.
So 4o was the model to beat. Apple didn't beat them. They said as much. I didn't claim they did. What I did say, though, was they made a model that was competitive with it, and competitive it is indeed. 68% of responses being matched or tied by persons evaluating it, with either model winning otherwise equally, is competitive. It's a dead heat.
With this in mind,
That isn't to say the PCC model is perfect. There are two caveats to my claim:
1) the metric I used is EN/US written testing. In other languages and regions, it wasn't a dead heat. 4o was more often winning, less tied -- but less so than AFM 1st gen vs ChatGPT 4.
I said "competitive with," despite that, because Apple intelligence is predominantly used by that language and region, although they've expanded. Nevertheless it's not a small metric or category. Apple often does English and US first, especially with brand new paradigms.
2) 4o was much superior in image analysis. So while Apple did update their model and become multi modal, it still has road left to run. In less uncertain terms, it's behind. Apple literally says this:
We found that Apple’s on-device model performs favorably against the larger InternVL and Qwen and competitively against Gemma, and our server model outperforms Qwen-2.5-VL, at less than half the inference FLOPS, but is behind Llama-4-Scout and GPT–4o.
The problem I had with the responses I received was largely it seemed/felt like you didn't even look up if I might have cited real stuff. You would see I wasn't citing opinion but sourced info.
Yes, Sourced info doesn't equal accuracy, but sourced info isn't a personal opinion either. I was citing articles and studies and observation. You also see that I had a flaw in my argument too, which I just said now.
To your point that "no one uses them," well, it's partially true. It's true they don't make a chatbot that scores high on GPQA Diamond but then tells someone to commit suicide, so there's that. But I will say, in the "macOS 26.2 adds Infiniband over Thunderbolt support" post, I called this out:
I have read multiple people on Twitter saying that Apple Intelligence is much faster and more accurate on iOS 26.2. Do you think this may be due to Apple rolling out their custom server, and do you think RDMA is used in that custom server? Do you think they're putting multiple high end chips onto a single board, then using RDMA for each of these to talk to each other?
Apple's PCC model already does this with ATSC, and from recent reports on social media, PCC model seems to be upgraded significantly with 26.2, producing far faster, more accurate responses.
Given the reports of 26.2 Apple Intelligence being faster and more accurate, something seems to be improving but no one here has commented on whether their PCC model is faster and smarter on 26.2 yet.
So I don't just take people's word for it. That said, and it's getting extremely long now, I've seen YouTube videos showing the PCC model very capable in Shortcuts app. Plus, Apple does use this model for multiple AI features for users.
GPT-5 was released on Aug 7, 2025. The swift removal of all legacy models from the ChatGPT UI was met with an even swifter backlash: some people online felt that GPT-4o was more personable, human, and engaging, whereas GPT-5 was stiff and robotic. This viral meme encapsulated the faction’s thesis:
I take a look at GPT-5, newly released today, and provide some initial comments. It is a handy improvement. But it surely isn't AGI or ASI. Here's the AI Insider scoop.
That's fair. At the same time, we are not in a court of law. It's not like some information is admissible and some not. We are just talking. I (in a friendly manner) maintain that some of your understanding (e.g. about Apple's foundational model and how they use Goggle technology) might be incorrect or incomplete. And no, I can't deliver any proof. Do with that what you wish.
Well yes I have an incomplete picture! Haha I don't work there, and I don't know anyone there.
The problem I have with Gurman is he is constantly imprecise. I try to be precise.
That said, my claim that Apple is using custom models and not Gemini weights directly rests on two pieces of evidence: 1) Gurman's articles, and 2) established history of Apple products and development.
The first part:
I'm not going to go into a comprehensive list of every time he's mentioned it, but I'll give some examples from articles
First, he has repeatedly stated that Apple has been seeking out three providers to develop models for evaluation. Why would they need to develop models if they can just give the models to Apple to evaluate? It's cheaper (zero cost), faster, easier to use off the shelf stuff
Apple Inc. is considering using artificial intelligence technology from Anthropic or OpenAI to power a new version of Siri...
The iPhone maker recently approached Alphabet Inc.’s Google to explore building a custom AI model that would serve as the foundation of the new Siri next year
The chatbot will run a higher-end version of the custom Google model, comparable to Gemini 3, that’s known internally as Apple Foundation Models version 11.
Down the road — as part of a full Siri overhaul included in the iOS 27 and macOS 27 operating systems — the custom models would be enhanced and run directly on Google’s infrastructure.
So you see, the idea of a custom model is constantly said by him. He tries to say both things: Apple is using Gemini, and Apple is using custom models.
Gemini models are a specific thing. The weights of Gemini and architecture are a specific technology. If Apple used Gemini weights, they would be using Gemini itself. If they were using custom models built from Google's entire infrastructure and dataset that they use, then that's a different thing but related to Gemini.
The second part:
Honestly this needs little explaining. Appleshistory is extremely rich and documented by books, articles, interviews, public statements that they specifically go to suppliers and work on highly custom stuff. I won't cite sources, because I feel that's as inherent to Apple as simplicity is to them. For me to question something fundamental like that would basically make this conversation meaningless. So I won't cite anything.
Is it possible theyre using off the shelf Gemini? I mean, yeah? Is it likely, for something as important as Siri? Absolutely not. Not even remotely likely. Does it mean they aren't? No. But still, extremely, extremely, extremely unlikely.
So with those two pieces, I said Apple is using Google tech, not Gemini. This imprecision is due to MacRumors and the rest wanting clicks. "Apple is using Gemini" is more provocative than "Apple is getting Google to train custom models."
It cuts both ways: if Siri ever has a hiccup, they'll cite the numerous issues with Gemini and pin that on Apple's choice. If it doesn't, then they'll cite recent public sentiment stating Google is "so advanced" now. Seems like Apple loses no matter what happens in the view of MacRumors
Am I accurate? Who knows. But what I said wasn't just my opinion, that was what I was trying to explain
That is also fair. At the same time, my posts are about Apple AI and the team. Sometimes conversations diverge. Then again, it's your thread, if you prefer to merely talk about Gurman, that's fine with me. Only I don't have anything to say about Gurman.
That's absolutely fair, and I think that would be fun to discuss it. It deserves a post. I tried to keep in line with that spirit ("sometimes conversations diverge") by asking you what you thought and asked for clarification at times. I didn't write immediately that Apple "AI" is the best and you're stupid and wrong.
And if I ever seemed to give off that impression, I want to say categorically think you are very smart and I enjoy talking with you, so don't ever think that I think about you like that! (I don't think it about anyone else here either.)
My post was trying to point how unreliable he is. I think that the cycle goes like this: Gurman had a reputation, which he used to ingratiate himself at Bloomberg. This inherently brings credibility deserved or not. It could have been a hiring mistake.
Every so often, he's correct about something, which isn't hard when you shit out 1000 version of the same story and claim all of them yours. Half of the times that he's right, which isn't often, it's because of educated guessing. I would know because I've predicted the same stuff launched/etc that he's claiming he has people telling him confidential secrets about.
The way he writes (once you analyze it) is extremely weird. It's written almost like it's made to skim and cherry pick quotes. This leads to dumbass blogs like MacRumors, who for the moment have sunken to complete clickbait chaos, to cherry pick stuff out.
Because he claims 1000 versions of something, and MacRumors is more than willing to publish snippets without actually realizing how incoherent his work is, he will eventually be right about something.
This cycle is self reinforcing: Gurman appears right, blogs cite him, Bloomberg keeps him because they cite him, and so on.
No one stops to question what the hell he's even saying. So I tried, a little anyways
I know it's long, so feel free to read it all, parts of it, whatever on your own time. It wasn't anything against you, so don't feel that the length has anything to do with you. This is partly what I had researched beforehand and was going to reply, but decided I would ask instead specifically what you wanted me to explain.
But it was a bitch to write, so if you could let me know you've seen it I'd much appreciate it lmfao!
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.
www.cnbc.com
I don't work there, and I don't know anyone there.
