> “As those things get scaled up, they start to make a meaningful change in the user experience and enable new innovation which has the potential to create a new upgrade cycle in smartphones.”
Interesting read. "Create a new upgrade cycle" is a phrase that's almost too on the nose, but will be interesting to see if the AI-powered feature set genuinely justifies an upgrade. I have a hard time imagining someone wanting to shell out 4 digits to be able to run ex. stable diffusion locally if they weren't already in the market for a new phone, but maybe I'm underestimating the AI hype wave right now. I could see predictive typing or improved voice assistants being the two biggest opportunities for a killer feature, but even then not sure if the value proposition would change that much, especially if the battery life takes any hit whatsoever.
Siri that actually works sounds very appealing to be honest. All the integration and UX is there. It just doesn’t work today. So an LLM based Siri that can understand my request is worth an upgrade (for me)
Exactly, just figure out how to run Whisper on my phone and it would be the best phone on the market. Whisper plus Mixtral on my phone and it could do anything.
On the contrary, Siri already works perfectly (or rather, used to work) for a lot of things... It's just Apple keeps on randomly changing stuff, and every time they change something, shit breaks.
"Hey Siri, good evening" - "I turned on the `good morning` scene."
"Hey Siri, timer five minutes" - "I set an alarm for seventeen hundred hours"
"Hey Siri, weather" - turns on the lights in the bathroom
I wish I was joking or exaggerating... The last thing I need is an LLM, just please fix the obvious problems and then maybe iterate on that
Yeah, I'll get random weeks in which "text <wife>" works fine, and other weeks where it says "which one?" and gives me her email and phone as different contacts. No changes on my end that'd cause it.
My wife is "Jen". I used to work with "Joe Chan". To this day, if I say "text Jen ...", there's about a 20% chance it'll reply with "OK, saying ... to Joe Chan."
Dude. How about optimizing for the option that I've actually communicated with this decade, and 4 times already today?
Additionally, in a multi-person household adding things to a shopping list should take any authentication. Not everybody has an iDevice or needs to have an iDevice - my petsitter, babysitter, maid and random family members should be able to add stuff to the shopping list just because they are in the room with the homePod. On Alexa, it works - but we have to deal with all the advertisements for crappy AliExpress junk. As soon as Apple figures out shopping lists for whoever is in the room, I'll be dropping all my Alexa devices.
Not sure exactly which way round you are asking -- should requests be tied to the person or be generic?
In the HomePod settings there is a section for "Personal Requests". When enabled, adding reminders will then go to the device / account of the person speaking. I'm not sure what the behavior is when set to off. You can change the setting for each HomePod.
In an ideal world, anyone who can access the HomePod physically _should_ have the ability to contribute items to the shopping list, regardless of whether that list is being shared among multiple users or not. At present, in order to add items to a shared "shopping list" in Reminders, each individual must authenticate themselves through voice recognition on their personal iOS device. It's not possible to add items to the "shopping list" without this initial authentication. Additionally, enabling personal requests is a prerequisite for adding items to a reminders list on HomePod devices. Even when it's set up correctly, it takes nearly a 20 second lag to add things to a list. With Alexa, it's instant without the need for authentication.
I spent the past 30+ years typing exact incantations into a computer. I give Siri an exact incantation today and it works, I give Siri the same exact incantation tomorrow and it breaks.
> “Hey Siri, timer five minutes" - "I set an alarm for seventeen hundred hours”
I think I figured out the issue with this one, finally: Siri understands “Set a timer for 6:30” as a request for a timer for however far in the future 6:30 is. And it also understands “set a timer for 6” similarly.
And Siri doesn’t seem at all intelligent about when to stop transcribing.
So Siri has a rather large chance of interpreting “set a timer for 5 minutes” as “set a timer for 5” and ending up with utter nonsense.
A LLM that jointly handles language and audio might do much better.
Something's not right. I say "hey siri 10 minute timer" or somesuch like five times a day. I'd play with your timing and word order. For instance my kid has trouble making music work with siri in the car, but if I say "hey siri play the album technique by the band new order" just like that it works 100% of the time.
Not sure what you're needing that you're note getting -- there is certainly a lot that it doesn't do -- but Siri works like a champ for me these days for maybe 70% of the things that I would do with my phone or ever computer without it. I initiate all calls with it, do tons of texting with it, cue up all my music in the car with it, constantly add things to reminders lists ('hey siri... add soy sauce to the grocery list' results in soy sauce on my grocery list... in the condiments sublist)
I hadn't noticed the improvements until I started using my apple watch a lot... I had given up on it a few years back, I guess.
It didn't seem like training will occur locally, just serving prompt responses from a GPT model (maybe I misunderstood the article). My personal privacy calculus places a lot of less value on keeping prompts local, especially if some kind of E2EE can be provided between me and the serving model. I understand that not everyone feels this way, but from the perspective of market adoption I don't see the average user caring too much whether a prompt is kept local or sent to the cloud (if encrypted or similar). To me it seems like the most critical determinants would be response latency and battery life, plus the basic performance of the thing.
If it's just serving prompt responses I'm not sure if the IP protection is a huge dealbreaker for me as a downstream user. The most critical use case I can currently imagine is better predictive text for my keyboard, and that one seems pretty light on the IP violations (for me personally). The biggest thing for me here is E2EE if the text is leaving my phone.
There aren't many more single new features that are going to justify upgrading that will exist on the mobile platform. Apple is playing a game where they try to add enough features to prevent the upgrade cycle from becoming too long. Encourage people with 2 or 3 year old phones to upgrade this cycle instead of delaying a generation.
Honestly, Apple's business model is harder if everyone wants to upgrade to a certain new version. The M1 Macs were so good, everyone upgraded their Intel Macs and then mac sales dropped. It's better logistically if the sales are more predictable and consistent year-to-year.
That's an interesting point, so more of a game of amortizing upgrades over a few years rather than instigating a massive one-off purchasing craze. I could see how a drip of sufficiently useful AI features and improvements over like '24-'27 would bolster the market. Even then though, this seems like an opportunity to hold rather than really reverse the trend; at some point I think Apple is going to have to reckon with the fact that the market is more or less saturated, though I could definitely be wrong here.
Apple knows the cell phone market is saturated. That's why they have iPads, AirPods, Apple Watches, MacBook Airs, etc. Their plan is to create new accessories (and services) that pair with the iPhone. This generates a moat around existing iPhone users.
I guess if they start layering LLMs or similar into their service or accessory interop stack I could see how having the phone running the model could be a big win for latency and availability. Like if Apple Calendar uses a GPT to create a .ics from the invitation text my friend sent me then it could be very handy to do that locally, moreso if I'm using my Apple Watch to actually signal that I want the .ics created.
I think people (and the market) make this assumption but the correction isn't Apple doing something differently it is the market and public assumption seeing that when you sell something that lasts longer than a year and works really well you need to accomodate the ebbs and flows of sales across a 5 year period instead of quarterly and YoY.
I agree, I think this is what they meant. A feature for the people who have an iPhone 8 or 9 who say “why would I upgrade, my current phone is fine”.
If you can use this tech to make a feature that they really want you can trigger upgrades that otherwise may not happen for many more years. We’ve seen that before with things like when Apple first released a bigger phone. That triggered a LOT of “early” upgrades that wouldn’t have otherwise happened.
Of course they “want” it to just like they want Siri and everything else to be computed locally since best for privacy and offline performance. Whether they compete with a cloud AI time will tell. Will presumably be far more limited than ChatGPT 4 but hopefully light years better than Siri which is what is most important.
Also light years ahead when the cloud access isn’t available or spotty in an elevator, the middle of the lake, a parkade, travelling with cellular turned off and wifi is on.
> Whether they compete with a cloud AI time will tell.
But they're not competing with cloud AI. Why would a person need to go to the cloud to give you a reminder or download an app?
They're competing against the current local assistant, Siri.
Large models are great but they can't fit on 8 or 16gb of ram. And that's a very big deal.
They don't need to put all of the world's information locally, just the relevant bits. It doesn't need to know every celebrity's full history for example
You can have the basic stuff on-device with the "smarts" of an LLM that can have conversations with the user and have context to previous questions.
The other stuff can be fetched from the cloud (with the user's permission OFC) and optionally saved locally.
This brings up the question how efficient the human brain is.
As AI keeps catching up more and more with human intelligence, will it also need more and more hardware? Can it achieve human and superhuman intelligence and still run on a device like the iPhone, which weights about 170g? The human brain is 8 times heavier.
> still run on a device like the iPhone, which weights about 170g [1]? The human brain is 8 times heavier.
Why are you comparing the weight? I'm sorry but this is a bizarre comparison. This isn't even apples to oranges, this is apples to a telephone pole.
I'll also throw in that a single Nvidia H100 is 1200 grams. Unless you have a "Bracket with screws" which will add 20 grams (who wouldn't want an extra 20grams of intelligence?).
Like 70%+ of the human brain is water. The human brain needs a massive network of systems to transport nutrients/oxygen which is irrelevant to logical processing.
Similarly the majority of the iphone weight is the battery and frame. The weight of the processing chip is _grams_.
Besides the conflicting variables with the weight, the way that ML works on a physical level is completely different from the human brain.
They are going to want to start putting more RAM in their machines then or they are going to find it a bit tricky to compete given how big even moderate AI is.
They will ship hardware (a new iphone for example) that has a co-processor and flash storage dedicated to the Apple model (which many are calling Apple GPT). They aren't going to make it easy to run your own model. That's completely against their style
There are apps in the app store that are allowed to use the new co-processor that I'm predicting they'll develop? That seems physically impossible since it doesn't exist yet.
I think that running AI on device is the clear next step in the world of super powerful portable phones/devices. I have a 2021 model phone with very powerful specs (granted it is the high end model that I got used). With nearly the same RAM and storage as some of my laptops from a few years ago, I see that Apple might want to offload some of their expensive AI apps to these powerful user devices.
I see this also as a privacy win as I don't want any of my data being used for training. I also like to have control over what models I use and the ethics the models follow.
I am cautiously optimistic, but the reality is these models are huge. LLama 7b is 13g. Even if there is enough cpu cycles, the phone's storage can't support multiple models (and there will need to be multiple models).
Compare that to cloud, you can run the latest and greatest model and effectively unlimited storage.
But the cloud would cost apple money, while if ran in your own device they can sell you the hardware for a healthy margin, and market privacy. I think a single useful model on device or many tiny ones are feasible. Even today, the photos app on my iphone knows what pictures are dog pictures, can remove backgrounds, etc.
A reasonable quantization (Q4_K_M) is 4GB on disk and uses 6.5 GB of RAM. It will run at reasonable speed on an iPhone 15 Pro today. Using a smaller model or a more efficient quantization on a phone two years from now (aka whatever's in their lab today) will be easier than running the latest First Person Shooter.
Welcome to the future.
Edit: I see a few comments about hundreds of GB. That works for storage, but DRAM consumes battery just by existing, and that's been keeping RAM sizes from increasing the way storage has.
13 GB a lot? What year is it? 1995? If Apple or Google wanted to add 1 TB to their phones, it wouldn’t cost them anywhere near what they charge us to add 256 GB.
Inference/RAM is 100% the bottleneck, not storage, and it isn't close.
These models all have to be fully loaded in memory to run. I think the max RAM on an iPhone is 8GB, max storage is 1TB, so we're talking 2 orders of magnitude more ability to store models vs actually running them.
And of course you'd really rather not use the full RAM for inference.
You're 100% correct on the technical challenges here, but the fully loaded requirement is what Apple is trying to address as per the paper they just published. I'm somewhat skeptical that they've solved it since it's a hard problem. We'll see!
I'm currently running dolphin 2.2.1 mistral on an M1 Mac Mini with 16GB of RAM. Zero optimisations, just running base LM Studio, I haven't even bothered with GPU acceleration.
It thinks about 5-10 seconds before answering and writes the answer about the speed I can read it.
The current top of the line iPads are more powerful than this and an optimised modern iPhone implementation using the ML hardware should be faster than my current setup.
Just as an example, my phone has 512gb of onboard storage and 16gb of RAM. I think it is safe to say that in the not-to-distant future, we will see the option to run multiple models on device with only a slight storage hit.
I wager there is literally zero chance that Apple will let you have any freedom in the model used, and there's zero chance that you'll be using anything except the Siri_Model unless you hack your iDevice.
Just like there is 1 app store and 1 assistant, Apple will fight forcefully to make sure there is 1 AI model.
Well, then if you can convince the EU to force Apple to allow uncensored models that will happily teach you how to make bombs and all manner of things like that, then you can have those models on your iDevice. Not sure the EU is going to jump on board that kind of free speech though.
You could make the same (unconvincing) argument for apps on third-party stores, teaching the same dangerous things or doing even worse things like (gasp) allowing pornography :P
Generally speaking, Apple has maintained rigid control over apps and have banned everything you mentioned.
If EU app stores fill up with pornography and terrorism apps, you can imagine their move will be decried as a debacle and used by Apple to demonstrate how the EU failed by forcing them to open up
Apple already has the CoreML framework anyone can use to do on device processing. They even have tools to help convert things into the format the iPhone uses. As others have pointed out you can already find apps with public models on the Apple App Store.
Is Apple going to ship these LLMs as part of the OS? I expect app developers to bundle them in their apps. Also Llama has a specific license clause that it’s available to anyone except Apple and Google
I find it very difficult to believe that Apple is going to incorporate AI hardware into their device and then open it up to everyone freely. That would be very off brand. Even in the browser space Apple famously forced all competing browsers to use Apple Webkit as their engine. They certainly have history around lockdown and lock-in.
I think there are two things that you can’t do on iOS:
- have “Hey Siri” send audio to your code
- have your code reliably listen in to the microphone so that your code can listen for its own trigger phrase (AFAIK you can start an audio recording session, and have that continue while your app is in the background, but that requires users to open your app first. I also am not sure that recording will reliably survive device sleep and automatic app shutdown because of memory pressure)
I think #1 is fine. It doesn’t seem fair to me that Apple should be forced to allow other developers link “Hey Siri” to their voice assistant because that assistant may be detrimental to the value of the “Hey Siri” mark or Apple’s values, for example because it is racist, sexist, homophobic, etc.
The second IMO isn’t fine, but I can see Apple argue they are in their rights to have some control over how that rigger phrase gets detected. “Hey Siri” uses a small on-device network specifically trained for that, in order to conserve battery life (https://machinelearning.apple.com/research/hey-siri), and they may argue having good battery life is essential to their brand.
My next step at least is various models running at home that i can talk to remotely with my phone. Join a discord channel and talk to my bots. My bots respond or do things back in the channel or with whatever else they have access to.
I'm totally aligned with this line of thought and I think running local is the best possible thing to do with all of the privacy risks AI presents. In the future I want my cognition 10x amplified by AI and the only way I can imagine that working is if info is being fluidly exchanged with my brain at the speed of thought. No way am I going to plug into a cloud data sink for advertisers and spy agencies to snoop my unfiltered thoughts. I want my cognition air-gapped on hardware I own.
I expect that locally installed models will quickly resemble how we think about drivers today. In the short term you'll install a notes app that requires Mixtral-8x7B so you pull that in through ollama. Down the line all operating systems will ship with various models preinstalled.
I've noticed that despite Apple's stated on-device aspirations, they've put very little incremental die space toward the neural engine in the most recent generations. I think there's a number of possible explanations for the disconnect. Curious if anyone with a better perspective knows which are true.
1) They think the existing hardware is powerful enough.
2) The utilization of the ANE doesn't justify increased resources.
3) They plan to re-generalize AI compute through things like vector operations.
4) (the most pessimistic) They are saving large increases for future releases when they need to force upgrades.
I could see the math for on-device just not working for things like massive LLMs. The amount of silicon you'd need to make it possible would be large, and the frequency of use low. The same silicon for one person's phone could likely support dozens if it were in a datacenter instead.
I'm not clear on what precisely the Neural Engine can do, but Apple has put plenty of resources into improving their GPUs, which can definitely do compute well (see eg, MacWhisper).
Yeah, it could be that they plan to do #3 through use of their chunky GPUs. There is definitely a lot of focus on the whole APU concept. From look at some of the dies and what projects like ggml have been able to do through mixed computation, I'm guessing the neural engine is ~3-5x the throughput/area vs GPU for the kinds of operations its doing.
Trying to map neural nets onto graphics seems reasonable, but personally I would bet on neuralizing graphics to be the better strategy long term.
There wasn't a "killer app" for on-die AI accelerators until roughly this March. Until about nine months ago, the only widely-discussed ideas were re-aiming pupils to make your video call feel more natural, and applying background filters.
So today's on-die AI accelerators were, as far as I can tell, a cautious bet. It turns out caution was warranted, because LLM's need large amounts of memory bandwidth and capacity far more than they need a specialized neural compute engine.
I wonder if there will be a separate product line, e.g. iPhoneAI or aIphone, that will have the necessary hardware upgrades with corresponding jump in price, to not have such a cost hit across their entire lineup when the cloud model will do fine for most people.
On my iPhone it works, and provides very decent performance. The only catch is that it needs pretty much the entire phone's memory to work - so if you switch out of it to another app and back again it resets the state and has to load the model from scratch.
There is no paging in iOS. At least paging in the Linux sense. Your apps memory will get evicted when you switch to another app if the system needs it.
It is up to each app to store their state and resume. One of the most annoying apps for not doing this well is Google Maps. Holy crap I hate when I search something and get a route set and tab out to check something and com eback to a fresh Google Maps like 5 seconds later. They will fix it for a while then break it again several times a year. Once you start navigating it seems pretty reliable but until you hit start it is a shitshow.
I really want to like it, but it's pretty terrible at many things imho.
I've tried really hard to switch to Apple Maps but even after three months it simply can't seem to remember I basically drive to 5-6 addresses on a weekly basis. It's always to the parking lot near the office on Monday and Wednesday morning, daycare to pick up the youngest on the way back, swimming pool on Saturday, etc. etc. but navigating to those addresses is always a pain (even selecting it from recent addresses in CarPlay) and it did sent me to the same street as the daycare in a completely different town several times now.
While I value not giving more data to google, after three months I decided it simply wasn't worth it. I found the experience (especially the first time it sent me to a completely different town I never go to and I didn't catch it) pretty infuriating.
I'd recommend setting up some Shortcuts with the addresses that can be triggered from Siri. For example, "Siri, we're going to the office", "Siri, let's go swimming",
One nit with this is that I haven't been able to make it automatically transition into navigation.
Yes, I’m aware. I shouldn’t have used that specific verbiage but that was what I meant. Which is worse than paging from a UX perspective since in practice no app does it perfectly (esp if there’s unsaved user input in a textbox).
I'd love to know why it's that big. I'm pretty sure it doesn't include the actual LLM model files - the first time you use a model you have to click to download it, and those are likely even larger than 3.3GB.
I've been playing with it on the recent iPhone Pro, and it is pretty fast. The mistral 7B is very wordy though and is prone to making up stuff. It's fun to talk to, but it's not very useful currently.
Based on this demonstration of a local LLM, with optimizations, it definitely the way of the future on apple devices at least.
Don't have a link right now, but I read earlier they have a model developed internally already called colloquially "Apple GPT" that (supposedly) rivals OpenAI in parameter size, etc.
They'll release a hardware co-processor similar to Google's when they're ready to ship this.
I think it's a great idea but... local will always not be as good as the an AI connected to the cloud. I find myself going from conversations I left off on google/chatgpt/notes between my desktop, phone. Would that it could work offline great but then what of the divergent thoughts brought up in our mutual thinking?
Apple's biggest problem here is that they've been consistently stingy with RAM, with only the latest iPhone 15 Pro featuring 8GB. If they want to make a really gutsy move for on-device AI, they could up that to 16GB at the very least.
That eats into Tim's precious margins, but it seems like a price worth paying if they want to do exciting, useful things with this tech.
I do wonder how useful significantly more RAM would be to the majority of iPhone owners until now. I’m sure Apple has tons of metrics collect to decide with.
Adding more would certainly be a price/margin issue. Plus DRAM uses battery even if sitting unused, and battery is precious on a phone. Unless they came up with some sort of hot-plug style option to be able to turn off unused RAM chips completely, which would be cool.
DRAM that is unused according to the OS doesn't need to use battery, as it doesn't need to be refreshed. If this selective no-refresh feature isn't already implemented, I would expect Apple to be working on implementing it in their SoCs. But it's such an obvious and simple power saving feature that I would be surprised and disappointed if it wasn't implemented many years ago.
It's crazy to say Apple is late to AI or missing the boat.
Yes, OpenAI gets a lot of attention to the extent LLM's could displace search - except search is based on ads from links and selling data, which are even less promising for AI.
"Big" AI now is stuck with organization-defining cloud bills for training and huge players are scrambling to push software into hardware. They are barely getting started on the integration.
OpenAI has been talking about AGI as it lines up commercial partners across the globe and up and down the stack. But AGI is no more realistic than crypto taking over from central banks. It's possible in theory, but not in fact.
Meanwhile, Apple has had neural processors on devices for 4+ years; AI features have figured in every subsequent marketing campaign. And VisionOS augmented reality provides a whole new domain for AI utility, explicitly targeted not just for play, but also for work, as remote work becomes the rule instead of the exception. Plus, Apple is the only safety- and privacy- enforcing ecosystem in existence.
> OpenAI has been talking about AGI as it lines up commercial partners across the globe and up and down the stack. But AGI is no more realistic than crypto taking over from central banks. It's possible in theory, but not in fact.
At some level it feels like the pearl clutching about GPT4 as it relates to AGI is astroturfed.
I'm inclined to agree with you on that to an extent. When I hear Sam Altman talking about it like it's inevitable and he carries the weight of the world on his shoulders, just seems a little cringey. Don't get me wrong. I like the guy, but it is cringey.
The idea that they could create such a thing both enhances their reputation and is something they can use to try and get legislation to prevent anyone from catching up with them.
“How could you trust random people to do this safely? It needs to be illegal and something only we can be trusted to make money o… I mean use responsibly.”
I still think LLMs displacing search will have hugely negative consequences for the heath of the internet. Yes, a lot of Google results are seo garbage now, but there's value in interacting directly with the source of a piece of information. LLMs alienate people from those sources, and that threatens the structure of the information economy on the internet. Why share anything if OpenAI is just going to train on your data and cut you out of the transaction?
> but there's value in interacting directly with the source of a piece of information.
Well, who said they're just gonna abandon it altogether. I'm not going to. Those of us that make use of all the google operators sure as hell ain't abandoning it. If anything, it will provide augmented results where appropriate and explicit results where appropriate. Otherwise, we'll switch to a competitor that does.
Yes, my Apple devices are the first where I felt AI was integrated for something actually useful. I can just select and copy text in pictures as if it was the most normal thing. I can search pictures based on their content, again including visible text. And it runs locally. Too bad it only seems to work on the latin alphabet. With greek letters, chinese symbols etc I get nothing at best.
> Yes, my Apple devices are the first where I felt AI was integrated for something actually useful. I can just select and copy text in pictures as if it was the most normal thing. I can search pictures based on their content, again including visible text
Apple’s so good at pushing out these kind of “little” features that immediately feel like a kind of bare minimum a system should do. Having to use a system where these two things weren’t automatic, quick, and practically flawless would feel like taking a big step back.
See also: Live Photos. I wouldn’t even consider any kind of phone or camera without that feature, now that I’ve experienced it.
[edit] context of how good it is: some days back there was a link here that went to an image with a bunch of text on it. I followed the link, read it, and did my usual compulsive-text-selection thing I do when reading on the web.
I didn’t realize I’d been looking at an image, rather than html, until I came back to the thread and read complaints about it. Double checked, yep, it was an image.
I was really surprised and happy to be able to take a picture of my son’s hand written Christmas list and the phone made the text selectable with very few errors.
No, I believe it only works with two or three alphabets. It’s a limitation that was documented somewhere. At least I think that was the case when first released.
Just like Siri/voice recognition I’m sure Apple will add additional scripts over time.
Worth trying again. I just cut and paste from a screenshot of my daughter's Chinese homework. (edit: and iPhone photo translate worked for her Chinese homework too!)
A recent iOS update greatly improved identification. It would often get the breed of our dogs wrong. Recently it was not only was more accurate, it started being able to identify both dogs in a single image, which I'd never seen before two days ago.
A Flickr contact posted a picture of an animal I'd never seen or heard of earlier today. I took a picture of my monitor, and it was able to identify it as a coati.
> It's crazy to say Apple is late to AI or missing the boat.
Apple is always "late to X". Their approach to all things ML so far has always been "can we run this on device". And very few, if any, of the latest LLMs can be run on the device reliaby, and without sacrificing device resources.
> OpenAI has been talking about AGI
Of course they have. Precisely because they "line up commercial partners"
---
Edit: previous version said, confusingly, "AI is always late to X". Also fixed typos and cleaned up sentences
> Plus, Apple is the only safety- and privacy- enforcing ecosystem in existence.
You mean the company that wanted to ship CSAM detection for iCloud uploads with iOS, which was only abandoned because of the huge backlash? Or the company that's implementing privacy features which neatly don't affect its own apps, effectively boosting their ad platform? [1]
Apple wasn't required to implement any CSAM detection, so why did they do it? Saying it was pushed onto big tech by governments when there is no law demanding it feels apologetic to me.
> Apple wasn't required to implement any CSAM detection, so why did they do it?
As things were, Apple was able to serve up peoples images when presented with a warrant.
If they went to end to end encryption for photo libraries they wouldn’t be able to do that. And I can basically guarantee something would happen and there would be claims about how “Apple protects child molesters”. It’s just too juicy, and we saw the ones about Apple protecting terrorists by not unlocking phones.
So I think they tried to come up with a solution. Something that would let them do end to end encryption but be able to show they’re not protecting child pornographers.
Yes, for photos stored in the cloud, the company that implemented a privacy-preserving CSAM detection algorithm on-device rather than just running it on unencrypted photos in the cloud like everyone else.
>You mean the company that wanted to ship CSAM detection for iCloud uploads with iOS
If someone wants to store stuff on YOUR servers YOU get to inspect it for whatever you want for whatever reason you want, arbitrarily and/or capriciously within the terms of whatever agreement you enact with your users. And the law, of course. If you don't want to inspect it, that's fine and that's your business. If you do want to inspect it, that's fine and that's your business.
If users want to keep stuff on their device you don't get to inspect it.
Can they just give an option "I'm OK with my data going to Apple cloud for better LLM inference and better Photo library search, etc".
I'm all for local inference, but until it gets good enough on limited devices, I would like to share you my data, please.
Keep in mind that Apple is charging you money for the hardware that you are using for this inference. You are paying for the electricity that it uses, and you have paid for the storage space and memory.
Who pays (and how) if Apple starts doing this for you, on their equipment?
I'm sure it's part of their ML with the newer iOS but this does make me happy to think of the AI running on the hardware. Yesterday when driving I got a message and usually it would just say "Jane says 'blah blah' and sent a photo." But last night Siri in CarPlay said "Jane says 'blah blah' and sent a photo of a dog in a blanket on a couch". And when I checked the photo they sent it was exactly that.
If it's E2E encrypted, I don't think it'd be able to identify that during transit, so it had to have been locally done.
Thought it was interesting, and a nice QoL update. Excited to see how this can grow if they introduce more things to be processed with AI on the hardware.
I love how Apple obsessively does things on the edge, on the device itself rather than on the cloud.
Sometimes my lowly iPhone X gets super hot during charging. Next morning I either get the facial recognition updated, a new memory generated or something else has happened. It's nice, confidence evoking even. The features you like are run on your device, with your data not leaving your device, on its free time.
This is the AI I like. Personal, confined and private.
They're in the best possible position to innovate on the literal edge, because they've got a fleet of billions of devices that mostly auto-update. As hardware advances, Apple will have the clearest path to deploying truly personal and locally hosted AI.
More so these days you buy some expensive hardware but keeping subscribing to apps and buy TV shows and cloud storage from them due to convenience. And then you buy the next model after a few years because why would you change a working formula for some Samsung/Google android BS that won’t be as clean….
> If it's E2E encrypted, I don't think it'd be able to identify that during transit, so it had to have been locally done.
Yes. Image description has been done on-device for quite a while now, and it is great. You can look for pictures using a description, and you can look for any text in any picture as well. It works very, very well.
Yeah, but I agree with the parent; I think something has changed recently with Siri. The other day I asked it some questions as a joke that I knew would just say "here's what I found on the web", but it answered them. It made me wonder what they're changing under the hood.
I had always assumed that search box in Photos.app would be useless, but once in desperation I typed something in and it found exactly what I was looking for (a photo of one of my kids' passports I knew I'd taken but could not for the life of me remember where).
Now I use it more and more, it makes my tens of thousands of photos dramatically more useful to me. That time we at at the restaurant with the big shrimp platter? The airbnb we stayed at where I took a picture of the cool coffee table? All right there.
Maybe it’s not especially amazing in 2023 but I totally agree. Finding random documents you snapped a photo of because you’re looking for some text that happened to appear on the edge of the page in a tiny font, it’s really useful. L
I’ve used the feature to find my cars VIN, passport numbers, random places etc.
Apple photos is wonderful for this. I can search for “cows” and I get photos of cows. All on device in the middle of a mountain with no internet 50 miles in any direction.
Plus I believe it indexes the text in images too. So if you remember something on a sign that you took a picture of it can find the picture without you having to know where/when it was.
It is amazing for accessibility users too, if you turn on VoiceOver, the Camera app and the Photos app tells you the description of what’s currently on display. And yes, it has been a feature for a while.
This is the possibility I'm most excited for in the AI space. Cloud models are great, but a lot will be unlocked once we get decent models running on consumer devices. If I never have to look at a web page again and use an LLM as my browser, it will be a game changer for the industry.
I think that no matter what you're still going to need to connect over WAN to interact with remote services. GPTs are super capable but I don't see them generating arbitrary webpages without at least fetching some basic resources from another machine somewhere. Like if a restaurant updates its menu, that data needs to make its way to the model in one way or another before it can generate their webpage. At the very least I think the "website" might change from an HTML page to an engineered prompt that gets served to you and which generates the page locally. But even in this case it's not clear to me why I'd do that generation on my phone rather than have the server gen the page for me and send the finished product.
I kinda wonder if it’s a derisking maneuver. If the AI runs on the client hardware you gain certain characteristics: the compute happens on the client hardware (saving apple $$) and if a bad actor succeeds at jail breaking from the AI controls then only the client hardware is owned (instead of a multi-tenant apple server).
And seems to be aligned well with Apple too: their focus on privacy, their hardware performance gains, their ai assistant products.
Apple is good at launching new product categories, ahead of competition. The apple watch had no competition for the first couple of years after launch, ipads are a class of its own, I wonder if they’re planning a similar move with ai too.
Apple has been using their Neural Engine for a couple of years now, and it’s obviously been noticed. Other companies are now starting to include similar blocks in their CPUs.
They could be a match or even better, or they could end up looking like an old integrated graphics card against a discrete GPU.
Time will tell us how much of a difference all of this makes.
> While Microsoft and Google have largely focused on delivering chatbots and other generative AI services over the Internet from their vast cloud computing platforms, Apple’s research suggests that it will instead focus on AI that can run directly on an iPhone.
Isn’t there supposed to be a version of Gemini specifically for Pixel phones? It seems like Google see this as a USP for their hardware as well. I wonder how companies like Samsung are thinking about it.
The metal support for pytorch is a real sh*t show at this point - all APIs are either legacy or unfinished. Also, app memory is about 3GB even for the largest phones.
It really still feels that it‘s all in its infancies.
Apple already has their own environment in CoreML and the neural engine. They have ways to train/run models on it in highly optimized formats and unified memory.
PyTorch came from PCs running on GPUs with their own pool of dedicated memory and a slower link to main memory.
I know people have been speeding it up a lot on Apple processors, but I’m not sure it’s a good indicator of how ready the iPhone is for running models for end user programs.
210 comments
[ 4.5 ms ] story [ 239 ms ] threadInteresting read. "Create a new upgrade cycle" is a phrase that's almost too on the nose, but will be interesting to see if the AI-powered feature set genuinely justifies an upgrade. I have a hard time imagining someone wanting to shell out 4 digits to be able to run ex. stable diffusion locally if they weren't already in the market for a new phone, but maybe I'm underestimating the AI hype wave right now. I could see predictive typing or improved voice assistants being the two biggest opportunities for a killer feature, but even then not sure if the value proposition would change that much, especially if the battery life takes any hit whatsoever.
"Hey Siri, good evening" - "I turned on the `good morning` scene."
"Hey Siri, timer five minutes" - "I set an alarm for seventeen hundred hours"
"Hey Siri, weather" - turns on the lights in the bathroom
I wish I was joking or exaggerating... The last thing I need is an LLM, just please fix the obvious problems and then maybe iterate on that
Dude. How about optimizing for the option that I've actually communicated with this decade, and 4 times already today?
In the HomePod settings there is a section for "Personal Requests". When enabled, adding reminders will then go to the device / account of the person speaking. I'm not sure what the behavior is when set to off. You can change the setting for each HomePod.
I think I figured out the issue with this one, finally: Siri understands “Set a timer for 6:30” as a request for a timer for however far in the future 6:30 is. And it also understands “set a timer for 6” similarly.
And Siri doesn’t seem at all intelligent about when to stop transcribing.
So Siri has a rather large chance of interpreting “set a timer for 5 minutes” as “set a timer for 5” and ending up with utter nonsense.
A LLM that jointly handles language and audio might do much better.
I hadn't noticed the improvements until I started using my apple watch a lot... I had given up on it a few years back, I guess.
https://www.notebookcheck.net/MrWhosetheboss-video-reveals-G...
> Google's Pixel 8 Pro Tensor G3 off-loads all generative AI tasks to the cloud
I’d like my data to stay local.
I probably don’t need cloud processing for any regular usage.
The privacy to life benefit equation is broken in society atm.
This will change.
Honestly, Apple's business model is harder if everyone wants to upgrade to a certain new version. The M1 Macs were so good, everyone upgraded their Intel Macs and then mac sales dropped. It's better logistically if the sales are more predictable and consistent year-to-year.
If you can use this tech to make a feature that they really want you can trigger upgrades that otherwise may not happen for many more years. We’ve seen that before with things like when Apple first released a bigger phone. That triggered a LOT of “early” upgrades that wouldn’t have otherwise happened.
Other kinds of transformers could too
Fix Apple autocorrect and detect replacements for misspelled words better
But they're not competing with cloud AI. Why would a person need to go to the cloud to give you a reminder or download an app? They're competing against the current local assistant, Siri.
Large models are great but they can't fit on 8 or 16gb of ram. And that's a very big deal.
You can have the basic stuff on-device with the "smarts" of an LLM that can have conversations with the user and have context to previous questions.
The other stuff can be fetched from the cloud (with the user's permission OFC) and optionally saved locally.
You can't make a cloud model work 100% locally for privacy or during connection issues.
As AI keeps catching up more and more with human intelligence, will it also need more and more hardware? Can it achieve human and superhuman intelligence and still run on a device like the iPhone, which weights about 170g? The human brain is 8 times heavier.
Maybe in 24 years the iPhone 39 will have the computing power that OpenAI has today? Hopefully our compression gets much better in the meantime.
But they also have to carry around their resource gathering and self-repair systems, something phones don't have to worry about.
Why are you comparing the weight? I'm sorry but this is a bizarre comparison. This isn't even apples to oranges, this is apples to a telephone pole.
I'll also throw in that a single Nvidia H100 is 1200 grams. Unless you have a "Bracket with screws" which will add 20 grams (who wouldn't want an extra 20grams of intelligence?).
Like 70%+ of the human brain is water. The human brain needs a massive network of systems to transport nutrients/oxygen which is irrelevant to logical processing.
Similarly the majority of the iphone weight is the battery and frame. The weight of the processing chip is _grams_.
Besides the conflicting variables with the weight, the way that ML works on a physical level is completely different from the human brain.
I see this also as a privacy win as I don't want any of my data being used for training. I also like to have control over what models I use and the ethics the models follow.
Compare that to cloud, you can run the latest and greatest model and effectively unlimited storage.
Hell, that’s likely what will happen anyway. The models run on device with cloud assistance.
Welcome to the future.
Edit: I see a few comments about hundreds of GB. That works for storage, but DRAM consumes battery just by existing, and that's been keeping RAM sizes from increasing the way storage has.
A copy of GTA V takes like 100 GB.
These models all have to be fully loaded in memory to run. I think the max RAM on an iPhone is 8GB, max storage is 1TB, so we're talking 2 orders of magnitude more ability to store models vs actually running them.
And of course you'd really rather not use the full RAM for inference.
It thinks about 5-10 seconds before answering and writes the answer about the speed I can read it.
The current top of the line iPads are more powerful than this and an optimised modern iPhone implementation using the ML hardware should be faster than my current setup.
Just like there is 1 app store and 1 assistant, Apple will fight forcefully to make sure there is 1 AI model.
If EU app stores fill up with pornography and terrorism apps, you can imagine their move will be decried as a debacle and used by Apple to demonstrate how the EU failed by forcing them to open up
Apple has no right to decide what users do with the hardware they sell.
I really want to start playing hardcore porn games on my PS5 as soon as possible. Sony has no right to decide what I do with my hardware, right?
Sure.
I really want to start playing hardcore porn games on my PS5 as soon as possible.
And if somebody figures out how to do that, they should be able to.
Sony has no right to decide what I do with my hardware, right?
Pretty much, yes.
Absolutely.
> Sony has no right to decide what I do with my hardware, right?
That is correct.
Or is there a common API for llm constitution (or however it is called)?
They’re not doing anything to censor any of this.
Nothing would need to change in the existing terms and conditions with developers: we only accept binaries that use our official APIs.
The Core ML API for using the neural processing features of Apple's chips has been in there for a while and plenty of developers are using it.
It doesn’t benefit Apple to restrict access to silicon.
https://developer.apple.com/documentation/coreml
- have “Hey Siri” send audio to your code
- have your code reliably listen in to the microphone so that your code can listen for its own trigger phrase (AFAIK you can start an audio recording session, and have that continue while your app is in the background, but that requires users to open your app first. I also am not sure that recording will reliably survive device sleep and automatic app shutdown because of memory pressure)
I think #1 is fine. It doesn’t seem fair to me that Apple should be forced to allow other developers link “Hey Siri” to their voice assistant because that assistant may be detrimental to the value of the “Hey Siri” mark or Apple’s values, for example because it is racist, sexist, homophobic, etc.
The second IMO isn’t fine, but I can see Apple argue they are in their rights to have some control over how that rigger phrase gets detected. “Hey Siri” uses a small on-device network specifically trained for that, in order to conserve battery life (https://machinelearning.apple.com/research/hey-siri), and they may argue having good battery life is essential to their brand.
[0] https://news.ycombinator.com/item?id=35711680
[1] https://news.ycombinator.com/item?id=35581777
1) They think the existing hardware is powerful enough.
2) The utilization of the ANE doesn't justify increased resources.
3) They plan to re-generalize AI compute through things like vector operations.
4) (the most pessimistic) They are saving large increases for future releases when they need to force upgrades.
I could see the math for on-device just not working for things like massive LLMs. The amount of silicon you'd need to make it possible would be large, and the frequency of use low. The same silicon for one person's phone could likely support dozens if it were in a datacenter instead.
Trying to map neural nets onto graphics seems reasonable, but personally I would bet on neuralizing graphics to be the better strategy long term.
So today's on-die AI accelerators were, as far as I can tell, a cautious bet. It turns out caution was warranted, because LLM's need large amounts of memory bandwidth and capacity far more than they need a specialized neural compute engine.
I've been running it on an iPhone 15 using this app: https://llm.mlc.ai/#ios - App Store link: https://apps.apple.com/us/app/mlc-chat/id6448482937
The same team have an Android version too which I've not yet tried myself: https://llm.mlc.ai/#android
On my iPhone it works, and provides very decent performance. The only catch is that it needs pretty much the entire phone's memory to work - so if you switch out of it to another app and back again it resets the state and has to load the model from scratch.
This has been my experience anyway ever since upgrading to iOS 16 (and now 17) anyway. Everything is always paging, even on the latest hardware.
I've tried really hard to switch to Apple Maps but even after three months it simply can't seem to remember I basically drive to 5-6 addresses on a weekly basis. It's always to the parking lot near the office on Monday and Wednesday morning, daycare to pick up the youngest on the way back, swimming pool on Saturday, etc. etc. but navigating to those addresses is always a pain (even selecting it from recent addresses in CarPlay) and it did sent me to the same street as the daycare in a completely different town several times now.
While I value not giving more data to google, after three months I decided it simply wasn't worth it. I found the experience (especially the first time it sent me to a completely different town I never go to and I didn't catch it) pretty infuriating.
One nit with this is that I haven't been able to make it automatically transition into navigation.
Based on this demonstration of a local LLM, with optimizations, it definitely the way of the future on apple devices at least.
https://arxiv.org/pdf/2312.11514.pdf
I do wonder what foundation model Apple will eventually use.
Both training their own as well as licensing a 3P one seem at odds with their claimed data privacy stance.
My bet though is a licensed 3p model that’s fine tuned locally with on device data.
They'll release a hardware co-processor similar to Google's when they're ready to ship this.
iMessage, apple photos, Siri inputs? These all seem problematic beyond the standard open crawl type datasets. They don’t have a web search corpus.
It’s a very good question though. That could make all the difference, huh.
That eats into Tim's precious margins, but it seems like a price worth paying if they want to do exciting, useful things with this tech.
Adding more would certainly be a price/margin issue. Plus DRAM uses battery even if sitting unused, and battery is precious on a phone. Unless they came up with some sort of hot-plug style option to be able to turn off unused RAM chips completely, which would be cool.
Yes, OpenAI gets a lot of attention to the extent LLM's could displace search - except search is based on ads from links and selling data, which are even less promising for AI.
"Big" AI now is stuck with organization-defining cloud bills for training and huge players are scrambling to push software into hardware. They are barely getting started on the integration.
OpenAI has been talking about AGI as it lines up commercial partners across the globe and up and down the stack. But AGI is no more realistic than crypto taking over from central banks. It's possible in theory, but not in fact.
Meanwhile, Apple has had neural processors on devices for 4+ years; AI features have figured in every subsequent marketing campaign. And VisionOS augmented reality provides a whole new domain for AI utility, explicitly targeted not just for play, but also for work, as remote work becomes the rule instead of the exception. Plus, Apple is the only safety- and privacy- enforcing ecosystem in existence.
At some level it feels like the pearl clutching about GPT4 as it relates to AGI is astroturfed.
“How could you trust random people to do this safely? It needs to be illegal and something only we can be trusted to make money o… I mean use responsibly.”
It’s an attempt at regulatory capture.
Well, who said they're just gonna abandon it altogether. I'm not going to. Those of us that make use of all the google operators sure as hell ain't abandoning it. If anything, it will provide augmented results where appropriate and explicit results where appropriate. Otherwise, we'll switch to a competitor that does.
Apple’s so good at pushing out these kind of “little” features that immediately feel like a kind of bare minimum a system should do. Having to use a system where these two things weren’t automatic, quick, and practically flawless would feel like taking a big step back.
See also: Live Photos. I wouldn’t even consider any kind of phone or camera without that feature, now that I’ve experienced it.
[edit] context of how good it is: some days back there was a link here that went to an image with a bunch of text on it. I followed the link, read it, and did my usual compulsive-text-selection thing I do when reading on the web.
I didn’t realize I’d been looking at an image, rather than html, until I came back to the thread and read complaints about it. Double checked, yep, it was an image.
Either way I also get that feature on the laptop, where it's more useful to me overall.
Works with Japanese on my MacBook Pro. I use it quite frequently.
Just like Siri/voice recognition I’m sure Apple will add additional scripts over time.
https://www.apple.com/ios/feature-availability/#live-text-li...
A recent iOS update greatly improved identification. It would often get the breed of our dogs wrong. Recently it was not only was more accurate, it started being able to identify both dogs in a single image, which I'd never seen before two days ago.
A Flickr contact posted a picture of an animal I'd never seen or heard of earlier today. I took a picture of my monitor, and it was able to identify it as a coati.
Have you had non-Apple devices?
Apple is always "late to X". Their approach to all things ML so far has always been "can we run this on device". And very few, if any, of the latest LLMs can be run on the device reliaby, and without sacrificing device resources.
> OpenAI has been talking about AGI
Of course they have. Precisely because they "line up commercial partners"
---
Edit: previous version said, confusingly, "AI is always late to X". Also fixed typos and cleaned up sentences
You mean the company that wanted to ship CSAM detection for iCloud uploads with iOS, which was only abandoned because of the huge backlash? Or the company that's implementing privacy features which neatly don't affect its own apps, effectively boosting their ad platform? [1]
[1]: https://adguard.com/en/blog/apple-tracking-ads-business.html
Apple wasn't required to implement any CSAM detection, so why did they do it? Saying it was pushed onto big tech by governments when there is no law demanding it feels apologetic to me.
As things were, Apple was able to serve up peoples images when presented with a warrant.
If they went to end to end encryption for photo libraries they wouldn’t be able to do that. And I can basically guarantee something would happen and there would be claims about how “Apple protects child molesters”. It’s just too juicy, and we saw the ones about Apple protecting terrorists by not unlocking phones.
So I think they tried to come up with a solution. Something that would let them do end to end encryption but be able to show they’re not protecting child pornographers.
If someone wants to store stuff on YOUR servers YOU get to inspect it for whatever you want for whatever reason you want, arbitrarily and/or capriciously within the terms of whatever agreement you enact with your users. And the law, of course. If you don't want to inspect it, that's fine and that's your business. If you do want to inspect it, that's fine and that's your business.
If users want to keep stuff on their device you don't get to inspect it.
This isn't complicated.
Who pays (and how) if Apple starts doing this for you, on their equipment?
Would you pay $5/mo for it?
If it's E2E encrypted, I don't think it'd be able to identify that during transit, so it had to have been locally done.
Thought it was interesting, and a nice QoL update. Excited to see how this can grow if they introduce more things to be processed with AI on the hardware.
Sometimes my lowly iPhone X gets super hot during charging. Next morning I either get the facial recognition updated, a new memory generated or something else has happened. It's nice, confidence evoking even. The features you like are run on your device, with your data not leaving your device, on its free time.
This is the AI I like. Personal, confined and private.
If we can combine this with a neuro bill of privacy/rights [0] then the future might not be as bleak as it seems.
[0] https://www.cirsd.org/en/horizons/horizons-winter-2021-issue...
> Sean Carroll & Nita Farahany on Ethics, Law, and Neurotechnology
https://www.preposterousuniverse.com/podcast/2023/03/13/229-...
They do stuff on-device because they want to keep selling expensive devices.
It's fantastic for them that they sell the hardware, unlike Google/OpenAI that have to spend on hardware.
Yes. Image description has been done on-device for quite a while now, and it is great. You can look for pictures using a description, and you can look for any text in any picture as well. It works very, very well.
https://findthatmeme.com/blog/2023/01/08/image-stacks-and-ip...
I had always assumed that search box in Photos.app would be useless, but once in desperation I typed something in and it found exactly what I was looking for (a photo of one of my kids' passports I knew I'd taken but could not for the life of me remember where).
Now I use it more and more, it makes my tens of thousands of photos dramatically more useful to me. That time we at at the restaurant with the big shrimp platter? The airbnb we stayed at where I took a picture of the cool coffee table? All right there.
I’ve used the feature to find my cars VIN, passport numbers, random places etc.
It also has other accessibility features like helping you find doors.
You may need to turn something on in settings before all of this shows up though, I don’t remember.
It’s new for iOS 17.
LLM in a Flash: Efficient LLM Inference with Limited Memory - https://news.ycombinator.com/item?id=38704982 - Dec 2023 (51 comments)
Just my naive thoughts
And seems to be aligned well with Apple too: their focus on privacy, their hardware performance gains, their ai assistant products.
Apple is good at launching new product categories, ahead of competition. The apple watch had no competition for the first couple of years after launch, ipads are a class of its own, I wonder if they’re planning a similar move with ai too.
Apple has been using their Neural Engine for a couple of years now, and it’s obviously been noticed. Other companies are now starting to include similar blocks in their CPUs.
They could be a match or even better, or they could end up looking like an old integrated graphics card against a discrete GPU.
Time will tell us how much of a difference all of this makes.
Isn’t there supposed to be a version of Gemini specifically for Pixel phones? It seems like Google see this as a USP for their hardware as well. I wonder how companies like Samsung are thinking about it.
PyTorch came from PCs running on GPUs with their own pool of dedicated memory and a slower link to main memory.
I know people have been speeding it up a lot on Apple processors, but I’m not sure it’s a good indicator of how ready the iPhone is for running models for end user programs.