Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Apple Intelligence Foundation Language Models Tech Report 2025

machinelearning.apple.com

Post reply on HN
Print view View original

191–200 of 210 posts

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#191

Post by alwillis »

martinald wrote: post #162
alwillis wrote: post #94

Earlier quoted context omitted.

> Private cloud is used AFIAK for virtually 0 use cases so far. Applications using Apple's foundation models can seamlessly switch from on-device models to Private Compute Cloud. Research is already showing the use of LLMs for people's most intimate relationship and medical issues. The usual suspects will try to monetize that, which why Private Cloud Compute is a thing from the jump. > Then they've got OpenAI/Gemini/…

8GB RAM is not enough for a semi-decent model IMO. 12/16GB is better (4GB for model and 8GB for OS) and really if you were going hard on device you'd probably want more like 32GB (24GB for model + 8GB for everything else - you'd be able to run a 13b param model with larger context size with that). Even still though people are used to the quality of huge frontier models, so it will feel like a massive downgrade on man…

> 8GB RAM is not enough for a semi-decent model IMO.

Apple's ~3 billion parameter on-device model is about as good as it gets on a smartphone, especially for the functions it was designed for: writing and refining text, prioritizing and summarizing notifications, creating images for conversations, and taking in-app actions.

Every Mac comes with at least 16 GB of RAM; while every iPhone comes with 8 GB of RAM, some models of the iPhone 17 will have 12 GB.

Remember, an app using the on-device model can seamlessly shift to a much bigger model via Private Cloud Compute without the user having to do anything.

If the user enables it, Apple's Foundation Model can use ChatGPT in a privacy preserving way. By the fall, Gemini and Sonnet/Opus could be options as well.

Again, ChatGPT is used in a privacy preserving way; you don't need an account: "Use ChatGPT with Apple Intelligence on iPhone" [1].

[1]: https://support.apple.com/guide/iphone/use-chatgpt-with-appl...

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#192

Post by simonw »

hn_go_brrrrr wrote: post #189
simonw wrote: post #188

Earlier quoted context omitted.

I stand by what I said. Apple are saying you can opt out of their training data collection using robots.txt. But... they collected their training data before they told people how to opt out. I don't understand why me pointing that out as "eyebrow raising" is controversial here.

It's not controversial, it's just not how the ecosystem works. There has never been an expectation that someone make a notification about impending crawling. It might be nice if there were categories that well-behaved bots could follow, as noted above, but even then the problem exists for bots doing new things that don't fall into existing categories.

My complaint here isn't what they did. It's that they explain it as "here's how to opt out" when the information was too late to allow people to opt out.

I think that's disingenuous of them.

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#193

Post by al_borland »

stogot wrote: post #175
al_borland wrote: post #166

Earlier quoted context omitted.

Sure, what’s not reasonable is expecting Siri to be a modern LLM, when they know it’s not. They asked a question they knew Siri couldn’t handle just to slam it. I’m not critical of a 5-function calculator for not one-shotting complex equations like a computer. While Siri only does one thing at a time, I trust the answer more, because it’s doing the actual math and not just guessing what the most likely answer is, lik…

It’s not unreasonable Amazon already reworked Alexa to be backed by a LLM months ago, and they were delayed doing it. You’re telling me that Apple isn’t capable of the same to Siri?

The unreasonable part is acting like Siri got its big LLM update, when they know it didn’t. Just like it would be unreasonable to expect any famously delayed, or unannounced, feature to magically start happening.

Amazon just needs a generic LLM. Apple, from the sound of it, is trying to create deep integration with the OS and on-device data. That’s a different problem to solve. They also seem to be trying to do it while respecting user privacy, which is someone most other companies ignore.

I don’t see what the big deal is. I’d rather wait for something good than have them rush out a half-ass “me too” chatbot, that is indistinguishable from the dozens of other chatbots I can simply download as an app for that.

If we believe what Craig Federighi said, they had something, it just wasn’t up to their standards when talking about rolling it out to a billion devices. Which is fair, I run into bad data from ChatGPT and other LLMs all the time. Letting it mature a little more is not a bad thing.

ChatGPT spent a couple months getting my dad pumped up for an elective open heart surgery; he was almost arrogant going into it about how the recovery would go, thinking ChatGPT gave him all the info he could possibly need and a balanced view of reality. Reality hit him pretty hard in the ICU. He sent me some of the chats he had, it was a lot of mutual ego stroking. He was using ChatGPT to downplay the negatives from the doctors and boosting the positives. While it’s good to feel confident, I think it went too far. I spent the whole week in the hospital trying to pull him out of his depression and recalibrating the unrealistic expectations that ChatGPT reinforced. I hope Apple finds a way to be more responsible. If that takes time, great.

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#194

Post by bigyabai »

OldfieldFund wrote: post #186
bigyabai wrote: post #108

Earlier quoted context omitted.

Apple is behind. People forget that Google was shipping mobile-scale transformer-based LLMs in 2019 : https://github.com/google-research/bert By the time Apple has an AI-native product ready, people will already associate it with dehumanization and fascism.

I think it's a great move that Apple is cautious with including a hardcore LLM in everything. They are not that useful to the regular user.

Nobody is forcing "hardcore LLM" features on anyone, besides maybe Microsoft. This is that same cope as "I'm glad Apple Car can't crash and threaten people's lives" despite the fact that... yunno, Apple really wanted to bring it to market.

Siri, sideloading and AI features all all the same way; give people options and nobody will complain.

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#195

Post by sxg »

al_borland wrote: post #178
stogot wrote: post #173

Earlier quoted context omitted.

The OP is making a point that Apple is behind. They might be publishing research, but it’s completely useless to the end user buying their products.

A plethora of LLMs are available on Apple platforms. If someone wants a chatbot, they can get a chatbot on Apple products. It’s not hard. Are all Android users using Gemini exclusively? Are all Windows users only using Copilot? Where is the native Linux desktop LLM? I really don’t understand this criticism. Would it be nice if Siri could do more, sure. Do I have tolerance for Siri to start hallucinating on simple pro…

Siri is the default and only voice assistant that has access to all the data on your phone. It doesn't matter if I have ChatGPT, Claude, Gemini, or another SOTA model on my iPhone—I can't easily activate them in the car or another handsfree situation or use them with any other app or data on my iPhone.

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#196

Post by imoverclocked »

lxgr wrote: post #170
imoverclocked wrote: post #145

Earlier quoted context omitted.

Excellent question. I don’t speak for Apple but certainly you can appreciate that there is a fine balance between providing basic functionality and providing apps. Apple works to provide tools that developers can leverage and also tries to not step on those same developers. Defining the line between what should be built-in and what should be add-on needs to be done carefully and often is done organically.

Are we talking about the same Apple whose behavior resulted in the expression “to Sherlock”?

Yup, exactly the same.

Q: Should search be core behavior or third-party functionality?

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#197

Post by wat10000 »

simonw wrote: post #192
hn_go_brrrrr wrote: post #189

Earlier quoted context omitted.

It's not controversial, it's just not how the ecosystem works. There has never been an expectation that someone make a notification about impending crawling. It might be nice if there were categories that well-behaved bots could follow, as noted above, but even then the problem exists for bots doing new things that don't fall into existing categories.

My complaint here isn't what they did. It's that they explain it as "here's how to opt out" when the information was too late to allow people to opt out. I think that's disingenuous of them.

It's been common knowledge for anyone running a web server since 1994.

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#198

Post by simonw »

wat10000 wrote: post #197
simonw wrote: post #192

Earlier quoted context omitted.

My complaint here isn't what they did. It's that they explain it as "here's how to opt out" when the information was too late to allow people to opt out. I think that's disingenuous of them.

It's been common knowledge for anyone running a web server since 1994.

I don't think you are reading my posts in full.

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#199

Post by ninkendo »

lxgr wrote: post #129
ninkendo wrote: post #97

Earlier quoted context omitted.

People have been giving Siri a few years for a decade now. Siri used to run in a data center (and still does for older hardware and things like HomePods) and it has never supported compound queries. Siri needs to be taken out back and shot. The problem with “upgrading” it is the pull to maintain backwards compatibility for every little thing Siri did, which leads them to try and incorporate existing Siri functionalit…

As a user, I'd gladly opt into a slightly less deeply integrated Siri that understands what I want from it . Build a crude router in front of it, if you must, or give it access to "the old Siri" as a tool it can call, and let the LLM decide whether to return its own or a Siri-generated response! I bet even smaller LLMs would be able to figure out, given a user input and Siri response pair, whether the request was res…

> Build a crude router in front of it, if you must, or give it access to "the old Siri" as a tool it can call, and let the LLM decide whether to return its own or a Siri-generated response!

Both of these approaches were tried internally, including even the ability for the LLM to rewrite siri-as-a-tool's response, and none of them shipped, because they all suck. Putting a router in front of it makes multi-turn conversation (when Siri asks for confirmation or disambiguation) a nightmare to implement, and siri-as-a-tool suffers from the same problem. What happens when legacy siri disambiguates? Does the LLM try to guess at an option? Does it proxy the prompt back to the user? What about all the "smart UI" like having a countdown timer with Siri saying "I'll send this" when sending a text message? Does that just pass through? When does the LLM know how/when to intervene in the responses the Siri tool is giving?

This was all an integration nightmare and it's the main reason why none of it shipped. (Well, that and the LLM being underwhelming and the on-device models not being smart enough in the first place. It was just a slower, buggier siri without any new features.)

The answer is that they need to renege on the entire promise of a "private" siri and admit that the only way they can get the experience they want is a _huge_ LLM running with a _ton_ of user context, in the cloud, and don't hinder it all with backwards compatibility with Siri. Give it a toolbox of things it can do with MCP to your device, bake in the stock tools with LoRA or whatever, and let it figure out the best user experience. If it's a frontier-quality LLM it'll be better than Siri on day one, without Apple having to really do anything other than figure out a good system prompt.

The problem is, Apple doesn't want to admit the whole privacy story is a dead-end, so they're going to keep trying to pursue on-device models, and it's going to continue to be underwhelming and "not meeting our quality bar", for the foreseeable future.

Re: Apple Intelligence Foundation Language Models Tech Report 2025

#200

Post by astrange »

albumen wrote: post #164
jonplackett wrote: post #160

Earlier quoted context omitted.

Yes, and these are all pointless. ‘Writing tools’ is there in the text pop up now instead of ‘look up’ even if you’re selecting text on a webpage. It’s just in the way and useless.

I agree. You can revert this UI frustration. Go to your settings. Then, go into screen time. Next, go into content and privacy restrictions and enable that. Under Apple intelligence in the restrictions you can then disable individual features.

My anonymous friend who wrote that settings pane would like to say that is not what they meant it for, but have fun.

(Since it's meant to restrict your children, using it to restrict yourself will disable some features that'd let you escape it. I forget what exactly, but you might not be able to change the time or something like that.)

Post reply on HN