This seems like a step in the right direction. Much like language translation, doing it on-device is the only way to preserve people’s data agency / privacy.
Of course, otherwise would mean investing in huge data centers for running LLM models, or worse, buying hardware from NVIDIA.
Optimization is the key. Privacy is just an added bonus.
No. Seriously, you underestimate how much these companies can make by transferring data to their server.
You probably underestimate the amount of effort Apple puts into not doing this, to maintain user privacy, and for a good while their services have suffered for it.
As an example I’d highlight the year in review feature between Apple Music and Spotify. “Replay” is significantly worse than “Wrapped” and I believe the difference is data handling is the key differentiator. However, there are some advances in balancing privacy 2ith utility, as highlighted in this post from Apple ML research: https://machinelearning.apple.com/research/scenes-differential-privacy
I’m going to disagree on the quality difference between Wrapped and Replay. I found Replay to actually be more useful when it came to actual data, Spotify doesn’t tell you as much as Apple about your listening patterns over the year and they also don’t even tell you what the whole year was, just January to October. Apple will also provide you with a large CSV file of every song you’ve listened to since 2018 if you go to their privacy page and request it. It takes a week or two, but they clearly have the data on file.
If you want a better Wrapped or Replay experience, try Last.fm. I get a ton of metrics from my listening and it has helped me find so much new music in the process.
Imagine all the stuff that Grammarly gets to see
I think this is the way to go. Rather than paying every random app a subscription to jerry-rig AI into their programming somehow, I’d like to have my own personal, and private, AI that plugs into a framework that each app offers. I’d also like to be able to purchase curated extensions to privately enhance my own AI.
I just want Swype to return. It got pulled out of Play and iOS stores in 2018. Swipe/glide keyboard input via gboard is crap in comparison.
You can already run a llm natively on Android devices.
You can on iOS as well!
The hard part isn’t running ai on a device… it’s doing so while retaining battery life, performance, and privacy.
Privacy is also easy with a local LLM. Performance and battery not so much.
Which one do you use? I tried MLCChat, but all 3 times it either showed a java error or generated giberrish, what’s worked for you?
“Wants” to? They’ve been doing that for years
You’re talking about two very different technologies though, but both are confusingly called “AI” by overzealous marketing departments. The basic language recognition and regressive model algorithms they ship today are “Machine Learning”, and fairly simple machine learning at that. This is generally the kind of thing we’re running on simple CPUs in realtime, so long as the model is optimized and pre-trained. What we’re talking about here is a Large Language Model, a form of neural network, the kind of thing that generally brings datacenter GPUs to their knees and generally has hundreds of parameters being processed by tens of thousands of worker neurons in hundreds of sequential layers.
It sounds like they’ve managed to simplify the network’s complexity and have done some tricks with caching while still keeping fair performance and accuracy. Not earth shaking, but a good trick.
They are the same thing, just bigger models. And many big models already ship with a smaller variant that you can run on an average gaming gpu.
Hard disagree on them being the same thing. LLMs are an entirely different beast from traditional machine learning models. The architecture and logic are worlds apart.
Machine Learning models are "just"statistics. Powerful, yes. And with tons of useful applications, but really just statistics, generally using just 1 to 10 variables in useful models to predict a handful of other variables.
LLMs are an entirely different thing, built using word vector matrices with hundreds or even thousands of variables, which are then fed into dozens or hundreds of layers of algorithms that each modify the matrix slightly, adding context and nudging the word vectors towards new outcomes.
Think of it like this: a word is given a massive chain of numbers to represent both the word and the “thoughts” associated with it, like the subject, tense, location, etc. This let’s the model do math like: Budapest + Rome = Constantinople.
The only thing they share in common is that the computer gives you new insights.
Wouldn’t this absolutely hammer the battery though, or at least give the CPU a hard time? My understanding is that offloading the work to a cloud platform means that the processor-intensive inputting, parsing, generating, and outputting operations are done in purpose-built datacentres, and end user devices just receive the prepared answer.
Wouldn’t this rinse the battery and increase the overall device temperature for “normal” end users?
Fair warning: I haven’t read the two papers outlined in the article.
CPUs can have special hardware accelerators for stuff like this, and you’d be surprised how powerful our little phone CPUs are and how optimized stuff like this can become.
Awesome, thanks for the insight.
I’m showing my age here, but much like we had math coprocessors running beside the 286 and 386 gen CPUs to take on floating point operations; then graphics cards offloaded geometry-based math operations to GPU’s - are we looking at AI-style die or chips to specifically work on AI functions?
Excuse my oversimplification, this isn’t my field of expertise!
Well, your not too off. Like ASICs are made for mining cryptocurrency. Specialized processing designed for specific computations. This indeed make it’s efficiency greater than a general purpose CPU.
They already have dedicated hardware they call the neural engine, and use for coreML, ARKit, some of the magic they do to turn terrible sensors and lenses into passable images, etc. There’s a lot of processing that already happens on your device. Being able to search your images by subject might be something Google does too, but Apple does it locally.
So my guess is they’ll just adjust the architecture of the neural engine to accommodate any new requirements, rather than adding a “new core”. But it’s kind of all semantics. There will be new hardware components and intercommunication at a low level.
Apple added (a while back) what they call a “Neural Engine,” which is hardware dedicated to efficient execution of ML workloads.
https://en.m.wikipedia.org/wiki/Apple_A11
They have been refining it ever since. I would not be surprised if they made advancements in both the hardware and software used for local GAI.
And Google did the same with the Tensor Processor Unit in the Pixels.
Yes!
not a dedicated chip per se, the trend is to build it directly into the SoC (mobile devices) or the dedicated GPU
Yup, technology and especially phones have come a disgustingly long way in such a short amount of time. Running AI efficiently on them is the next step, one that we probably won’t struggle with too much.
Apple already does a lot of this stuff. For example, it’ll do offline face recognition for your photos while your phone is charging overnight.
Plus, Apple is ahead of the curve when it comes to performance on this stuff. You don’t want to be running Stable Diffusion on your iPhone, but smaller AI is perfectly fine. Plus, unlike on Android, there are huge amounts of devices with ML accelerator chips that can run these models efficiently, allowing for power consumption optimisations by not having to provide a CPU fallback.
We’ll have to see how effective this will be in practice, but Apple generally doesn’t bring these types of features to their newer devices until they’re ready for daily use.
It’s a technical challenge but I wouldn’t rule it out. Apple has been using a “neural engine” in their SoC for faced id, etc. for a while. So it’s something they’ve been working on. It will need to get better, but AI models are also getting more efficient.
Running AI is pretty low power and efficient, especially if you have purpose built chips.
Training AI is another can of worms
If the scope of “Ai” isn’t wide, I’d imagine the battery and cpu usage would be minimized.
I believe that the decentralisation of computing power is both good for our privacy but also for the environment.
Finally Knight Rider’s KITT comes as reality
While Apple has its shares of problems, it feels like they at least care about their users, unlike say…Google…
🤖 I’m a bot that provides automatic summaries for articles:
Click here to see the summary
Apple’s latest research about running large language models on smartphones offers the clearest signal yet that the iPhone maker plans to catch up with its Silicon Valley rivals in generative artificial intelligence.
The paper was published on December 12 but caught wider attention after Hugging Face, a popular site for AI researchers to showcase their work, highlighted it late on Wednesday.
Device manufacturers and chipmakers are hoping that new AI features will help revive the smartphone market, which has had its worst year in a decade, with shipments falling an estimated 5 percent, according to Counterpoint Research.
Running the kind of large AI model that powers ChatGPT or Google’s Bard on a personal device brings formidable technical challenges, because smartphones lack the huge computing resources and energy available in a data center.
Apple tested its approach on models including Falcon 7B, a smaller version of an open source LLM originally developed by the Technology Innovation Institute in Abu Dhabi.
Academic papers are not a direct indicator of how Apple intends to add new features to its products, but they offer a rare glimpse into its secretive research labs and the company’s latest technical breakthroughs.
Saved 74% of original text.
