People have speculated that with FPGAs or even an ASIC you could take some smaller models we have today, turn the chip into essentially that model-as-hardware, and get some really compelling performance/power-usage characteristics.
At that point running a GOOD quantized Q8 model on a phone is not beyond the realm.
That's what I was hoping for with Taalas, ultimately ending up with a microsd-ish card that's hot swappable intelligence.
The code for all of those things is there, but not enabled if for most iOS users. There's a non-zero chance this code is there for Apple to try and convince the EU to let them release a Siri AI, and then just like sideloading have some impossible-to-actually-use alternative path to make the DMA happy.
There's profit to be made by offering customers choices here, I'm sure Apple will be ready to scoop up all of those delicious slop dollars.
I've been using the new siri, and it seems more or less stuck in instant-answer mode, which gives really terrible answers with any provider.
I wish more people were talking about this, but if you can't crank up the models to extended thinking, the technology is more or less stuck at 2004 levels of usefulness.
So many other actors in the industry would have just doubled down and follow the race to the bottom
Apple Intelligence was an outlier in how they historically operate. I mean they literally just put out a folding phone after waiting what, a decade, for others to perfect it first.
This looks more like a participation medal, if at all.
This would be insulting if it even mattered. Apple has fumbled AI while having record breaking quarterly profits. The market does not care.
It turns out that voice assistants aren’t all that important. Customers aren’t willing to pay for them and they’re more of a nice to have convenience feature.
Adding quotes around “genius” as it seems you misunderstood what I meant
Almost everyone developing their own AI is doomed to fail. I would not be surprised that when the dust settles, Microsoft CoPilot and Google Gemini will be left standing while OpenAI and Anthropic fade away. They will get what they need and then throw them away.
"Embrace, extend, extinguish" is the same playbook for decades, and you can only do that when you make money. Yes, making money still matters and post-capitalism is a big fat lie.
Handle the query on device. If the device cannot handle the query, push the query to Private Cloud, If Private Cloud cannot handle the query, push the query to a third party of the users choosing.
Which is pretty much the same as what they announced a few years ago.
Apple underestimated the amount of time and work needed to get to work and really the only bit that I think changed is that Private Cloud (and possibly the on device model) is a white label Google model rather than being developed in-house by Apple.
It also requires a relatively complex distributed architecture.
I worry about how successful they can be with this approach, in terms of best user experience.
Maybe apple can even have something similar to MCPs for tool invocations of apps and publish it so all models can actually be trained to use that framework
https://www.macrumors.com/guide/apple-command-center/
Mark Gurman seems to get extraordinarily accurate and detailed leaks. I wonder how Apple got this leaky. Pretty much everything about the Duo was known months in advance.
I built an app that lets me run terminals on my MacBook. It’s a Next app so I can go to that website on my phone using Tailscale. It has a built-in MCP server, so I can connect it to ChatGPT and have it create terminal sessions, orchestrate terminal sessions, all that good stuff.
Down the line, I’d like to be able to connect it to Siri and just talk to my phone and it can control terminal sessions and projects that I’m working on on my laptop.
With mobile devices especially where ideally models could be shared across several apps.
But this specific code I'd assume is targeted to comply with EU regulations more than anything.
OT:
Since I've moved to Spain my principal Apple account is still in my home country but I've made another one for local app-store (since some local apps aren't available if you're not in the Spanish store).
iPhone is acting really strange, I'm logged to my principal account for iCloud+ and anything and local App Store.
If I log back to my principal account, I loose HCE mode and my building NFC app stops working until I re-log in the Spanish Apple Store.
If I'm logged at the Spanish account, I loose Siri AI which is blocked in the EU...
I wonder what the margin on this stuff is for them:
Some Apple Intelligence features in the Home app, such as
video summaries, are available for a limited number of
cameras depending on your iCloud+ plan — up to 1 camera
in the 2 TB plan, up to 2 cameras in the 6 TB plan, and
up to 5 cameras in the 12 TB plan. Supported camera
counts and other iCloud+ plan features may change over time.
I like tinkering with this stuff, everything around me is a perpetual lab, so I have more than 5 cameras that I'd like to hook up to this. But I REALLY don't NEED or WANT 12TB or storage.
I want all of these integrations to be possible locally, without having to send my data to the cloud.
But I cannot. A lot of these nice integrations aren't offered for fully-local setups.
iPhone is different, one button hold on phone without touch the phone is arguably Apple's last and only moat.
Self promotion blog article:
https://braw.dev/blog/2026-04-06-mobile-os-as-agentic-broker...
Besides, the modern concept of "agent" is what, some year old? Bit fast to ossify stuff in the OSes already.
Ostensibly they'll activate this in the EU when it's ready. Given the current functionality, it looks ready (or close to being ready). My guess is it is ready and Apple is doing their usual negotiation tactics with the EU to see who blinks first.
This is not how politics works here.
The EU said “build it then we’ll tell you”, which is crazy.
The problem is not having Siri use different models beneath. That’s not what the EU wants. They want Siri to be replaced by Gemini/etc completely. Which fits with the DMA.
Apple’s proposal was for something akin to an agent that could gather the necessary data for the 3rd party AI to use. Because what they refused to do is let any other AI have unfettered access to everything whenever it wants.
So I suspect this is really for Apple’s benefit. I don’t think the EU would be happy if you could still use Siri, but it was grok underneath. Because that would still be Siri.
This is not that. This is expanding the Extensions that you already have. You can look at it in your Settings already.
All of which is why EU users are being left behind on the new Apple Intelligence (for now).
But it will never prompt you anymore, because Siri AI can handle that stuff now when it couldn’t before.
Whether you think it’s better or not is up to you. I do. I also found the version of ChatGPT it used to be worse than the one in the app.
[PhraseSpotter] Detected wake word “hungry” with a score 0.942.
[com.apple.voiceactivation] Proceeding with full wake.
They could instead make a good assumption that you'll search the web, open the Yelp or OpenTable app. Take your phone to your favorite restaurant and/or pay with a credit card at the grocery store.
If you walk to a taco truck without your phone and pay with cash, you probably don’t do that every time and you generate enough data through other means to be accurately advertised to.
All the big companies know how suicidal that would be. They already get accused of it by people who don’t understand how targeted web advertising works.
Some of the new AI start-ups don’t seem to understand that, but Apple absolutely will not make a mistake that stupid.
High rep accounts like yours always come to defend, and you will probably dead this comment.
You claim that the AOP (always on processor) - which must always run in order to listen for “Hey Siri” - doesn’t actually listen for “Hey Siri”.
Or you claim that it does but they would never ever use the PhraseSpotter for that purpose because they’re too benevolent.
I’m not claiming benevolence. Apple is less benevolent now than ever, however much you think that was.
They care about money. And I think they care about selling $1000 phones + tie ins far more than small amounts from advertising on all their phones by surveilling what people are talking about and risking those sales.
Trust me. They’re already shoving ads everywhere without that, much to my annoyance.
This would’ve been useful during their bake off where they were trying to choose if they wanted to go with Google, Anthropic, or OpenAI. But even without that, they would absolutely not want to be in a situation where they had to redo everything to get off a bad partner.
I do not want my personal data or the queries against it to leave my network boundaries if at all possible. This should be a default option.
Make a Shortcut that will open your local voice assistant app of choice when you press the action button.
Never underestimate the power of a bored programmer
Now that's an agent I'd use. Can we start with Siri for the terminal?$ Siri 'List the files in this folder'
“AI as a chat bot” has always been cool and fun in sci-fi movies, but it’s the least imaginative real world use for it. We need to move past it.
> $ Siri 'List the files in this folder'
Why in the world would I ever use this instead of just ‘ls’?
I've never understood this obsession with trying to invent the talking ship computer from Star Trek TNG. The purpose of the talking computer in the series was not to be a "cool 24th century technology advancement". It was a plot device to keep the story going and explain things to the audience. Having a voice conversation with a computer is terrible, imprecise, plodding UX. It's not a good interface, it's not a good invention, it's not good technology. Yet $billions are being spent to try to invent it.
The best example I can think of in favor of what you are saying, where the computer does kind of serve the story and helps the audience, is the episode where Doctor Crusher gets stuck in the warp bubble and she has a lot of conversations with the computer about what is happening. But even then, she is out of her element, needs a lot of help from the computer, and it seems more convenient to do so via voice than anything else. She still travels to different areas of the ship to carry out actions, even though she probably doesn’t have to in case. The computer is clearly at its best for simple actions and answering questions, even if it can do more. The characters understand that and use the technology accordingly.
Instead of quickly iterating to find how to take voice to the next level, both projects quickly locked in on whatever that was and became too costly to pivot.
The thing that nobody really realized is that setup is a big factor that isn't worth to a lot of people. For example, to turn off lights with Alexa, you have to do a whole bunch of setup that involves installing special lights, installing app, connecting the app to Alexa. Even though relatively this is simple, its just not worth the effort to a lot of people.
Amazon or Apple could easily overcome this by offering Alexa or Siri connected devices as incentive to builders building houses or apartments. That way, when you plug in Alexa in your new apartment, it can auto detect everything it can control including lights, thermostat appliances, e.t.c.
cobolcomesback: "ls, granny"
Granny: "it said No such file or directory"
cobolcomesback: "no, that's... just write ls, granny.
Granny: "I wrote ls granny it said no such file or directory."
cobolcomesback: "just.. from now on... just write Siri 'list all files in this directory', granny"
Granny: "I wrote siri list all files in this directory granny and it told me all the files and then it also asked me who's granny and I told it and it said nice to meet you"
cobolcomesback: "SIGH"
and more often than not isn't reliable unless you have a midwestern US accent.
Siri wasn't good. It still isn't good. Apple hasn't delivered on its promises with it and that's why it's being replaced with Gemini.
For major platforms: “iOS 10.0+ / iPadOS 10.0+ / macOS 12.0+ / watchOS 3.2+”
Maybe I’m misunderstanding what you’re missing?
It already exists in macOS 27: fm(1)
It is an extremely powerful tool, but here are the basics:
https://developer.apple.com/videos/play/wwdc2026/334/
Google assistant on the nvidia shield left Siri in the dust years ago. Siri can't even chain basic commands in your smart home, it can't retain any kind of context (e.g. dim the lights...a bit more).
This was frustrating for me as it always prompted “who cares? Just turn off the lights.” As well as would drive my parents and other guests batty.
I guess you should be able to create a bucket of low-risk tasks that require zero authentication, like letting your guests turn off the lights. I do get why they are cautious, though, because in some cases this could be a security risk, like a burglar turning off lights from outside the home.
So, progress?
Any agent with the Home Assistant MCP hooked up can do it fine.
Example: "I can use AI to combine multiple PDFs into a single PDF."
Okay let me just install a tool to do that because the OS doesn't usually have it built in in 2026.
Oh, it's riddled with ads. Oh it's paid for. Etc.
It's the same thing with the recent opening of knowledge thanks to AI.
Before: search for recipe with search riddled with ads, find website riddled with ads cookies notification requests forced sign up and a million other dark patterns, then an entire fucking life story before the recipe.
Now: "find me a good recipe for pasta that uses the stuff I have left over from the party last weekend, I don't mind buying a few more bits".
I am waiting for this function to sort my folders
https://youtube.com/watch?v=W63oJtX0dP8
https://apfel.franzai.com/
"Dammit, Bobby! Turn off the dang lights!"