（预览）关于WWDC 2026、Fable 5及其护栏，以及Anthropic与Apple共同点的5个问题
Apple因2024年积累的信誉债务，几乎不可能在WWDC 2026上重新点燃真正的兴奋感，尽管Ben Thompson认为发布的软件整体表现扎实。 预录演示浪费了Apple通过高风险现场演示重建信任的机会——主题演讲中约23%的内容似乎都是加载指示符。既然有报道称这些功能运行良好，Apple缺失的核心信息就是“真正做到开箱即用”（It just works）。
Apple持久的AI优势不在于模型领先，而在于它对iPhone数据、操作系统、开发者和分发渠道拥有特权式访问。 App Intents和Apple的索引系统可以调取散落在用户各个App中的信息，即便技术上“并不惊艳”，也能形成差异化能力。这让Apple无需匹敌竞争对手巨额的资本开支，也能提供有意义的AI体验。
对投资者而言，核心判断是：Apple可能不需要赢下AI，维持现有位置就足够了。 极度环境化、代理化的界面最终可能威胁iPhone，但消费者付费购买的是娱乐，而手机仍是一台“极其出色的娱乐设备”。因此，Apple即使错过最佳模型的桂冠，仍“可能拥有使用人数最多的AI”。
Mike Rockwell试图修复Siri，不仅会影响Apple Intelligence，也会带来Vision Pro层面的组织后果。 Thompson认为，成功可能提升Rockwell的权力和影响力，并帮助Vision Pro；他还表示，Rockwell离开可能让该产品陷入困境。Andrew Sharp则将成功描述为让Rockwell的“孩子”继续活下去。Sharp对Apple据称将Vision Pro视为“技术上成功，尽管商业上并不成功”感到意外。
在Thompson看来，Craig Federighi关于Apple依赖Google的说法，在狭义技术层面准确，却有意制造了错误印象。 Thompson估计，Apple大约有5个模型，包括端侧模型；其中一个是200亿参数的混合专家模型，他认为该模型是在Google协助下由Gemini蒸馏而来。云端模型“据我理解”以Gemini为基础，再针对Apple进行了大量后训练。“这一切都与Gemini深度整合。”
Apple的新Private Cloud Compute保住了隐私承诺，却放弃了原先的基础设施叙事。 云端模型运行在Google数据中心的Intel主节点和NVIDIA GPU上，而不是Google面向客户部署的TPU上；安全飞地技术则提供了一条从设备到计算过程、可通过密码学验证的链路。Thompson称这是一次“彻底切换”，但Apple仍然可以可信地维持隐私。
1. Rockwell救援Siri，或可保住Vision Pro
Thompson特别指出，Mike Rockwell主动请缨修复Siri；Rockwell正是他认为对Vision Pro负有最大责任、也最投入的人。若成功，Rockwell的权力和影响力可能上升，也有助于Vision Pro；Thompson称，Rockwell离开可能让该产品“陷入困境”。Sharp则将成功定义为让Rockwell的“孩子”继续活下去。
Sharp认为其中的内部逻辑尤其耐人寻味：据报道，Apple之所以提拔Rockwell，是因为公司认为Vision Pro“技术上成功，尽管商业上并不成功”。Thompson仍表示结果尚不确定，同时提到有报道称Rockwell对向Craig Federighi汇报感到不满。
2. Apple需要一场现场演示来偿还信誉债
Gabriel的听众提问点出了Apple的困局：展示雄心勃勃的AI，没人相信；展示兑现中的旧承诺，又会让所有人失望。Thompson的结论很直接：“这就是2024年的代价”（This is the cost of 2024），因为Apple亲手堵死了办出一场令人兴奋的主题演讲的可能性。
主题演讲异常简短，但Thompson估计“约23%”的内容都是加载指示符。现场演示本可以把无聊转化为悬念：Apple可以放弃预录制形式，冒着公开失败的风险，换来观众一句：“这帮人真有种。”（Balls on those guys）
Thompson将Apple不愿承担这一风险视为Tim Cook时代的弱点。不过，由于报道称已发布的功能运行得“非常好”，他认为产品实质上是成功的：几乎没有新东西，但Apple眼下的信誉考验在于是否“真正做到开箱即用”（It just works）。
3. iPhone为Apple提供差异化AI切入口
Thompson在2024年最初的兴奋，源于Apple所处的结构性位置：用户通常知道某条信息存在于个人数据的某处，却无法将其调取出来。Apple掌握操作系统层，也能推动开发者采用App Intents和索引——“他们拥有真正重要的数据”——无需在前沿模型竞赛中胜出。
Sharp认为，这些功能对90%的Apple客户来说只是标配；Thompson则进一步指出，个人数据检索对高阶AI用户同样重要。它之所以高度差异化，恰恰是因为竞争对手无法复制Apple对iPhone的控制，即便底层工作在技术上并不那么惊艳。
Apple仍将自己锚定在同一套范式上，这套范式已经沿用了10或15年。Thompson承认外界对这一姿态的所有批评，却拒绝顺势得出的看空结论：“有时这就是正确的思路。”如果AI仍然只是软件层，避开数千亿美元开支可能显得聪明，也可能只是走运：“宁可走运，也别只求高明。”
4. 守住手机或许就够了
Thompson可以想象另一种未来：一种“随需出现、用完即隐的UI”，配合后台代理、预判式行为和环境传感器。Microsoft的Project Solera是他的具体例子，尽管他认为该项目“可能永远不会发布”并最终沦为废案；如果这种计算模式胜出，Apple确实可能暴露在风险之下。
Netflix说明人们愿意付费购买什么——娱乐；而iPhone是一台“极其出色的娱乐设备”，可用于社交媒体、竖屏视频和通信。Thompson本人多数阅读和想法形成都在手机上完成，但他强调，大多数人并非大部分时间都在工作。
Sharp认为，一款可信的手机挑战者可能仍要10年才能出现。因此，Thompson降低了战略门槛：Apple不必成为OpenAI，也不必拥有最佳模型；有时只要“维持住自己的位置”就够了，同时仍有可能提供全球使用人数最多的AI。两人的篮球类比是：领先25分时，投上篮和罚球，不要连续投9个三分。
5. Gemini藏在Apple精细拆分的表述之下
Federighi表示，Apple既不使用Gemini App，也不使用其客户端代码，更没有Google Assistant。Thompson拆解称：这些是调用模型的应用，而不是模型本身。Federighi那句“所以我希望这很清楚”才是笑点——每句话都可以是正确的，同时又刻意遮蔽了双方的关系。
Thompson估计，Apple有5个模型，其中端侧模型被作为Apple自有模型提供。一个模型是200亿参数的混合专家模型，每次查询只选择并加载一个专家，而不是每个token都切换，从而减少必须留在内存中的内容。
Thompson认为，这些小模型是在Google合作下由Gemini蒸馏而来。直接访问模型，使蒸馏的范围更广、效率更高，而不是只能从最终输出中学习；他还指出，商业版ChatGPT和Anthropic用户如今也在使用蒸馏版本。因此，尽管都以Gemini为基础，Apple的版本并不是Google向客户部署的同一批模型。
云端模型“据我理解”以Gemini为基础，但针对Apple的使用场景进行了大量后训练。它运行在Google数据中心的Intel主节点和NVIDIA GPU上，后者的安全飞地能力支持一条可验证的私有计算链；而Google通常通过TPU为客户提供服务。Apple似乎还维护着自己的搜索索引——这让Federighi在Google面向客户的部署栈和Search问题上严格说来是准确的，但整体表述仍然具有误导性。
Ben, how are you doing?
Well, at some point, maybe in the first half or early in the second half—
Hm.
I was thinking about our Sharp Tech Knicks emailer and coming in hot—
He did it to himself.
Too early.
Yep.
I was telling him it was too early. I was thinking about how funny it was going to be when it all fell apart. I'm still not entirely sure what happened.
I mean, I still can't believe what happened. I was thinking of a GOAT emailer, Cameron in Australia, who emailed me and Gulliver about a month ago asking whether he should buy tickets to the NBA Finals and fly to New York City from Sydney, Australia. I recommended at the time, “Do it. Life is short. It's going to be crazy expensive, but if this is a bucket-list item, this is your opportunity. Who knows when the Knicks will be back here? Just make the move.”
And then he did it, and he spent God knows how much on tickets to Game 3 and Game 4. I was sitting there after they lost Game 3, watching Game 4 with them down 30 at halftime, feeling really guilty that I was encouraging Cameron to go to the game. Then it turned into probably the most unforgettable game of his life—and any of our lives. So here we are.
There's no dollar figure that could have made it feel too expensive after last night.
Oh my God. I'm so happy for him.
You messaged me about how worried you were about him, and I think you messaged it exactly at the nadir.
Right.
I think they were down 29, and the game turned thanks to your care. So—
The tide was turned. Oh my God.
That's right.
The rest is history.
I was going to counsel you, “Look, you need to not take responsibility for other people's decisions. Cameron is his own person.”
And he made that clear.
“He can make his own decisions.”
To be clear, he didn't put it on me, but I still felt guilty during Game 4.
No, but maybe I'm wrong. If you were able to turn the tide, maybe you do need to care more about your advice. So—
Indeed. And every Knicks fan—
Um.
—in the audience, thank me over the next several days.
Well, let's let them win first. Let's let them finish the series.
It is the Knicks.
Because now a Finals loss would be even funnier, so brace yourself.
Absolutely.
For now, Ben, we'll begin with WWDC. We're going to talk through Anthropic and its new Fable 5 model later in the show. But first, I look forward to WWDC conversations every year.
As a tech guy?
As a tech guy, as a Vision Pro fanatic. The Vision Pro got a couple of mentions in this year's keynote, so I love to see that.
Not dead yet.
It's not dead yet. We got Siri on the Vision Pro. But in general, it's sort of an Apple State of the Union. It's also how I mark the time on Sharp Tech. I still think back to the initial WWDC with Vision Pro—
Well—
—Apple Intelligence.
I have an extra take here.
Okay.
This was actually an incredibly important WWDC for you.
Why is that?
Because the person at Apple most responsible for and most passionate about the Vision Pro is Mike Rockwell, who basically put up his hand and said, “Let me fix Siri, please.”
Mm-hmm.
“Let someone else—like, we need to get rid of the clowns who are in charge. Put me in charge.” Despite the fact that he was, I think, from Dolby Logic originally, with no obvious connection to building AI. If WWDC went well, and if he pulls it off, he's going to rise in power.
He'll keep his baby alive.
He's going to rise in power and prominence at the company. There have actually been questions about that, some political intrigue around his position there. If he were to leave, I think your Vision Pro might be in trouble.
Mm.
And if he were to succeed, it might be good for your Vision Pro, so there's a lot at stake.
It's funny because there was the line in Mark Gurman's report about that, and about the elevation of Rockwell, that prompted a double take as I was reading it. Apple looked to Mike Rockwell because Vision Pro was considered a technical success, if not a commercial success. I'm surprised that's the way it's interpreted internally. I think it's a technological marvel, but I'm surprised that Rockwell is getting any credibility out of what Apple has committed to over the last several years.
Well, maybe. We'll see. There were reports a few months ago that he's not super happy about reporting to Craig Federighi. That came out again the week before. If this works out and he gets put in charge of more stuff, that might be good for everyone. We'll see.
There you go. Well, the palace intrigue will continue indefinitely here. For now—
Yeah, it's not really my beat, to be clear. But—
Yeah. Well—
Because I know how important it is to you, I thought I'd bring it up.
I look to Gurman. I enjoy reading his tea-leaf reading and insider reporting. But I have a six-pack of questions that we can run through quickly at the top of the show here. We got a lot of emails on the heels of WWDC, but I'll begin with Gabrielle, who wrote on the way into WWDC and sent this email last week:
“Ben and Andrew, is there any way for Apple to come out of WWDC leaving a good impression? If they play it safe and only deliver on features that were promised back in 2024, people will think it's just warmed-over news. On the other hand, if they show off new capabilities for Apple Intelligence, people will be wary of trusting them. Which of these approaches should Apple focus on? Is there any third option where they can leave the public excited about what they presented without the bitter aftertaste of the whole ‘something is rotten in the state of Cupertino’ debacle?”
There was no third option. A lot of people were pretty underwhelmed. What did you think? You were in Cupertino.
Yeah, I think Apple foreclosed the opportunity. What Gabriel is putting his finger on is that they foreclosed the opportunity for people to leave this keynote excited, for the exact reasons he said. They could show exciting things and no one would believe them, or they could show that they were actually doing basic stuff, but it worked this time.
Yeah.
And then everyone's underwhelmed. So that's Apple's own fault. This is the cost of 2024: no one is going to leave 2026 genuinely excited and feeling like Apple is changing the world.
Mm-hmm.
Given that constraint, I thought they did a great job. The most important aspect of doing a great job was the extent to which this was a fairly painful keynote.
Yeah.
On one hand, it was shockingly short, and it was shockingly short despite the fact that it felt like approximately 23% of the keynote was watching—
Watching the AI work.
—loading indicators. This was the keynote of all keynotes that could have benefited from being live.
Mm-hmm.
Because what they needed to do was rebuild their credibility, and you rebuild your credibility by having stuff actually work. Now, I guess they're never going to do that again.
They're allergic to that at this point, but I was texting Dummen[?] during the presentation that all the AI demos were pretty tedious. However, if you were doing them in front of a live audience and then watching them work—
Right.
—in real time—
The anticipation. The “What if it fails?” Right?
Yeah. Then it becomes—
No, exactly.
—kind of thrilling, and they could have done that—
Honestly—
—but didn't.
It's actually a testament to a weakness of the Tim Cook era. To go out there and do it live and say, “We're going to roll the dice. We think we've actually got it, and we're going to go out there and risk total embarrassment,” is itself an acknowledgment of the embarrassment they should already feel, given what happened in 2024.
And number 2, that is the third way. That's the way where you come out of it and you're like—
We're hanging it all out there.
“Balls on those guys.” Yeah, they went out there. I use that term carefully, but that would've been the response. They went out there, they threw away their prerecorded keynotes, and they said, “We are doubling down. We're putting our chips on the table, and we're going to do live demos. Screw it.”
Yeah.
And it actually works this time. In that case, given what they shipped, it's totally what they said in 2024.
There was almost nothing new. And you know what? That’s fine.
Mm-hmm.
What they showed in 2024 was pretty compelling.
Yeah.
That’s why we were excited in 2024. It’s compelling because—
Well, crucially, it’s compelling for 90% of Apple customers, for whom high-end, expensive AI is not really what they want. They’re giving people table stakes.
It’s compelling for high-end AI customers. You knew something was in your data somewhere, and you wanted to surface it.
Mm-hmm.
And they can do it because they’re the operating system. It’s because of their position with the iPhone that they are uniquely positioned to do something with AI that no one else can do. Even though it’s not technically impressive, it’s still differentiated.
Yeah.
Highly differentiated. And that’s why I was so enthusiastic about 2024, because, aha, they have an angle here. They don’t need to do a bunch of capital expenditure spending. They can actually deliver something meaningful given their preexisting position. They’re leveraging their position as—
They own the data that matters, yeah.
They own the data that matters, and they have a position to corral developers to use the App Intent system and their index system to give them their data, even though the developers don’t like Apple and all these bits and pieces. It was a very smart strategic play, and I was very enthusiastic about it in 2024. And if I step out of the “it’s 2026” framing and look at what Mythos—
They’re so far behind.
—or Fable or—
They should be so ashamed.
—Gemini’s doing—
Yeah.
—it remains impressive and compelling.
Mm-hmm.
And now I think where they fell short is that they didn’t go all the way. I think a live demo would’ve transformed this. It would’ve been a smash hit, especially because almost universally, the reports are that it works, and it works super well.
Hm.
I have not installed it. I usually never install betas. I actually do plan on installing this one. This is part and parcel of the fact that I was still unpacking last year. I have to find my iPhone from last year.
Okay.
My spare iPhone that I carry with me is actually an iPhone from 2 years ago. It’s been a busy week, so I apologize. I was not properly prepared for this show. But by all accounts, everyone who’s talked about it says it just works.
Hm.
And it just works. That’s what Apple needs to get back to: It just works.
Yeah.
And it seems like it does.
Well, we shall see. Boyan, responding to your WWDC article on Tuesday, wrote in and said, “Apple is not thinking different. It is living in the past and becoming a victim of the ‘when all you have is a hammer, everything looks like a nail’ mentality. That is exactly the mistake that Microsoft made with mobile. They tried to shoehorn everything into a Windows viewpoint. Apple is replaying the same playbook. If AI changes the world like its biggest proponents think it will, Apple will be in deep trouble. Fortunately for them, it is very likely that AI becomes just another very useful software layer, and Apple dodges a bullet.” What do you agree with there?
I’m sorry, was this one Boyan or 2 Boyans?
One Bojan.
Because Boyan’s last sentence completely refutes the first 5 sentences. If, in fact, Apple AI is just another very useful software layer, then that means Apple is, in fact, thinking differently by not spending hundreds of billions of dollars—
Yep.
—and going their own way and sticking with the iPhone paradigm. So which is it, Bojan? I feel I don’t even need to answer. The guy answered himself.
I mean, I would say they are demonstrably thinking differently than the rest of tech, and have been over the last several years.
And they’re also demonstrably anchoring themselves in the iPhone paradigm.
Mm-hmm.
That doesn’t mean it’s wrong. That was a point I was trying to make. I didn’t love my article this week for a few reasons, one of which was this: I had to rewrite the ending multiple times because I was trying to say, look, all these critiques, like Boyan’s critique, are totally legitimate. Apple’s totally doing what’s best for them. They’re totally doing what makes sense for their model. They’re totally not investing in things that they’re probably not good at. And all those reasons—
And they’re not thinking any differently than they were over the last—
That’s right.
—10, 15 years.
That’s right.
Yeah.
When you articulate that companies are doing what is good for them and in their interest, it’s usually followed by a but.
Mm-hmm.
The world’s different now, and this is totally wrong thinking. Sometimes it’s just right thinking. You’re just fortunate enough that the model actually works. And the fact of the matter is, I have talked about this multiple times. Boyan should give me a little credit here—or one of the Boyans, Boyan 1—
Yeah.
—that there’s—yes, you can imagine—
You’ve been warning about a downside.
—a future of a different—
Sure.
—UI paradigm: just-in-time UI, there when you need it, gone when you don’t; agents working in the background on your behalf, much more anticipatory of your needs; all these ambient sensors and input devices. This is why I talked about it in the context of Microsoft’s Project Solera a bit, which, again, I love that demo, even though it will probably never ship. It’s going to be vaporware. But it was a tangible articulation of a completely different model of computing that is not like the iPhone at all. It’s the exact opposite. It’s what Boyan 1 says is coming, and Apple is screwed.
Yeah.
But then Boyan 2 is here and saying, “Actually, the iPhone’s pretty phenomenal.”
Mm.
And in a context where—what do consumers pay for? Give me 1 service that consumers pay for.
What do you mean? Like a software service?
Yeah.
I don’t know.
Netflix.
Mm. Yep.
What is Netflix?
Entertainment.
I should have put Netflix in the article because that actually helped make my point more. I just said customers don’t pay for productivity. They don’t pay for productivity; they pay for entertainment. The iPhone is a phenomenal entertainment device. That is its number-one functionality. It’s social media, vertical video, talking to your friends.
Mm-hmm.
Now, is it also a productivity device? Absolutely. It’s one of the most useful productivity devices in the world, but it’s used by people to do work. I do most of my work on the iPhone. I do my actual writing—
Yeah.
—and maybe some Excel spreadsheets. Someone was asking me for my software stack, and my software stack is funny. I have this superpowered MacBook Pro, which I rarely use as a text editor, and maybe for some Excel spreadsheets. I do most of my reading, most of my thinking, and a lot of my sketching out ideas or whatever on my phone. I use an iPhone like a Gen Z maniac.
Mm-hmm.
And it is definitely a productivity device. But for most people, they’re not doing work most of the time. Entertainment is just as important—more important, in fact—than productivity. And on that vector, from that perspective, is any of these devices better than a phone?
We’ll see. It might be 10 years before any sort of device emerges that could actually challenge the primacy of the phone.
So I’m with Boyan 2 here. I’m coming around to the idea that maybe Apple’s lucky to be in this situation. Better to be lucky than good.
No, exactly. And I look at what they announced this past week, and Apple doesn’t need to hit home runs in AI to win in AI and to double down on all those advantages.
Well, they also don’t need to win in AI. They need to—sometimes it’s sufficient to maintain your position.
Yeah.
Right? They’re not going to win in AI. They’re not. They’re not going to be the perfect AI.
They’re not going to have the best AI—
They’re not going to be OpenAI.
—the best model. No, exactly.
They might have the most-used AI, though.
What about Apple as the Spurs in the second half of Game 4 last night? They’re up 25 points, and the key for them here is not to miss 9 threes in a row.
Right. Why are you taking 9 threes in a row like you’re behind? Just getting layups and free throws—
And you’re set.
—would get the job done.
Yeah. And I think that’s what WWDC was this year, and it’s perfectly sensible as far as I’m concerned. Jonathan says, “Goats.” I would love to hear some clarification on the Apple and Google AI deal. I thought Apple was using a white-label version of Gemini and/or using Google Cloud for off-device compute.
This quote from Craig Federighi, though, sounds like Apple is using nothing from Google? Here’s what Federighi said:
“Of course, we don’t have the Gemini app as our app. In fact, none of that client code is part of how we run on iOS. For these models, we use none of the models that Google deploys to their customers, nor do we use the infrastructure and means by which they deploy models to their customers. And then when it comes to the knowledge base, we of course don’t use Google Search or anything like that as the foundation of our system. So I hope that’s clear. The amount of the Google Assistant that we use is none.”
And Jonathan says, “I suspect the truth lies somewhere in the middle, but please help a normie make sense of this bizarre corporate speak.” Can you help Jonathan, Ben?
The funniest line that Craig Federighi said there is, “So I hope this is clear.”
Mm-hmm.
Because he does not hope it’s clear. He did not lie anywhere in this statement, and he also created an impression that is totally wrong.
Okay, great.
I don’t know. Should I break it down sentence by sentence?
Yeah, go sentence by sentence. Fire Joe Morgan here.
I’m not sure how long this is going to take, but okay. “Of course, we don’t have the Gemini app as our app.” Duh. Of course. Gemini is an app that you actually download on your phone or go to a website to use. They don’t have that.
Yeah.
Okay, fine. That’s clear. “In fact, none of that client code is part of how we run on iOS.” This is saying the same thing again. The code of the Gemini app supports the Gemini app. So, no, they did not take the code of the Gemini app, in addition to not taking the Gemini app, and put it in their app.
Great.
Glad we’re clear.
Yes.
Yes. Okay. Again, the code of the app has nothing to do with the model.
Yeah.
The app is a means to access the model, okay? It’s just like my website is a way to access my writing, right? It’s like a container for it, okay?
Uh-huh.
What’s next? “For these models, we use none of the models that Google deploys to their customers.” They don’t use the exact same model that Google gives their customers. That is correct. “Nor do we use the infrastructure…” But what are they actually doing?
Apple has, I think, 5 models, several of them on-device and then 1 that’s in the cloud.
Mm-hmm.
All the ones that are on-device are Apple’s own models. Those models are, of course, because they’re on-device, very small. One of them is pretty interesting. It’s actually quite large, a 20-billion-parameter model in which only part of it is loaded in memory at the time.
It’s called a mixture of experts. Usually, for mixture of experts, you load or use a different expert every token, which means you have to have the whole thing in memory. In this case, you do an expert per query. So when you make a query, it’s like, “We need this part of the model.” It loads only that part of the model into memory, and then that’s in memory for all the tokens.
Regardless, all those on-device models are Apple models.
Mm-hmm.
But how do you get small models?
Distill it down.
Distilling, exactly. They won’t say “distilling.” I think they said they were Google-derived or made in conjunction with Google. Basically, what you do is ask the big model questions—
Yeah.
—and then you get answers, and that’s how you build your model. These are Google-distilled models.
This is legal, aboveboard distilling that’s happening between Apple and Google.
Right. And by the way, the reason it’s not just legal is that having actual access to the model—the whole idea of generating tokens—there are actually multiple layers underneath. If you’re only getting access to the final output, you can distill from that, but it’s way more inefficient, and you’re not getting as broad a coverage and understanding of the parent model.
If you have actual access to the model, your distilling is far more effective. And by the way, when you’re accessing most of these models online, including ChatGPT and Anthropic, you’re accessing a distilled model as it is.
Sure.
No one’s getting access to the real model because it would be astronomically expensive and impossible to scale. So even the big models that are online are distilled to an extent, and they can do a good job of it because it’s their model. They can distill it better than anyone else can.
Mm-hmm.
But all those models on Apple’s devices are not the models that Google serves. They are the models Apple serves because both Google and Apple distill from Gemini for the models they serve.
Okay.
So they’re doing their own distilling with Google’s help, but it’s not Google’s. It’s not the same ones that Google has.
Right.
The one in the cloud, as I understand it, is some sort of Gemini model at its base. But the way models are today, there’s so much post-training and so many things you layer on after the initial model is done that it’s effectively—not a different model, but there are multiple variations of models that can be accessed.
When you see all these companies saying, “Oh, new model, new model,” they’re releasing models all the time. Now, there are genuinely new models, like, say, a Mythos, that are a new training run. But lots of the new model releases are not new training runs. They’re new post-training and new reinforcement learning. You do 1 pre-training a year or a couple a year, and then you just iterate on the one that’s there.
Yeah.
Apple has done that to Gemini. That’s what’s being served in the cloud. It is different from Google’s, and appropriately so. They’re tuning the model for their use case.
Mm-hmm.
So all of this is intimately integrated with Gemini, and Craig Federighi went out of his way to make you think that wasn’t the case.
Let’s obfuscate this. Sure.
But he didn’t lie, right?
Mm-hmm.
“So I hope that’s clear.”
No, it’s not clear, and you did that on purpose. Now, do we use the infrastructure and means by which they deploy their models to their customers? No. They have their own setup with Intel head nodes and NVIDIA GPUs.
Okay.
This is different from a standard NVIDIA rack, where NVIDIA has its own head node with its Grace GPU and then a Blackwell. That’s why it’s called Grace Blackwell. Blackwell is actually the GPU. Grace is the CPU. And then you have Vera Rubin. Vera is the CPU. Rubin is the GPU.
They’re using the NVIDIA GPU part—
I never knew that. Okay.
—with Intel as the front part. The reason they have those is that both Intel and NVIDIA have secure-enclave technology that lets you ensure you get the Private Cloud Compute aspects, because the whole chain is fully verifiable: what got passed there, whether it got flushed, where it went.
It’s a whole cryptographic handshake between your device, the Intel head node, and the NVIDIA GPU. That whole chain is fully secured.
Mm-hmm.
So they can continue to make their promise around Private Cloud Compute, even though it’s totally different from the Private Cloud Compute they tried to sell 2 years ago, which was their servers on their chips.
Right.
But it is Google infrastructure. It’s running in Google data centers on Intel and NVIDIA chips. This is a total switch from what Private Cloud Compute was.
It’s still credibly Private Cloud Compute because of this technology embedded in these Intel and NVIDIA chips that lets you cryptographically guarantee that it’s totally private.
Yeah.
But it’s a total change in infrastructure, and that’s not what Google uses. Google uses fleets of TPUs. They don’t have the cryptographic guarantees around it being private and all those sorts of things.
As for the knowledge base, we don’t use Google Search. Yes, apparently they have their own search index, which is sort of interesting. Then we get to, “So I hope that’s clear.”
The amount of the Google Assistant we use is none. Google Assistant is an app. We get it: you don’t use the Google Assistant; it’s the old app. So it’s like saying, “We don’t use the new app, and we don’t use the old app.” Thanks, Craig.
Yeah.
We figured that out.
So nonsense, nonsense. The first clause of the third sentence is also pretty misleading. They don't use the infrastructure and means by which they deploy models to their customers. That seems to be the most accurate claim there, and then—
No, everything in here is correct. It's just—
The least misleading claim—
—meant to obfuscate.
Yeah.
Yes.
So, in any event, Jonathan, I hope that clears things up for you. Paul says, “If Siri AI is some sort of consumer harness for white-label Gemini, although the underlying model itself is an Apple, given Ben’s note today about the integration value of harnesses, is that circumstance enough to give Apple AI breathing room for a while to come, given their distribution advantages, OS device integration mode, installed base, et cetera?
“And assuming Siri AI was developed all in-house, if Apple was able to create a good harness lite for Gemini, isn’t that skill valuable enough to let them keep some competitive tension among the leading models? And finally, if Apple was able to make a smaller Gemini-derived model run on device, does that further help maintain competitive tension among model makers, all to Apple’s advantage?” Any thoughts there?
Yeah. I’m sorry, Paul. I’m going to have to pick on you a little bit, but I think this question stands in for a lot of people who don’t have an awareness of how advanced AI is right now.