29 comments

[ 3.5 ms ] story [ 81.1 ms ] thread
(comment deleted)
Did Apple PR write this?
No. But a shareholder, for them hope springs eternal...
In a year or two, everyone is going to realize that Apple is a decade behind everyone else in AI and the stock will crash. nVidia is going to fight it out with Microsoft for the title of the highest market cap. It’s going to get ugly. The question is, when will Warren Buffett pull out of the Apple stock?
> In a year or two, everyone is going to realize that Apple is a decade behind everyone else in AI and the stock will crash.

https://machinelearning.apple.com/research/vision-transforme...

> The question is, when will Warren Buffett pull out of the Apple stock?

So much to unpack here.

Vision Transformers are from 2020. A decade behind is an exaggeration, but OP's general idea is roughly right.
If I had a penny for each of the Apple doomsayers since the day AAPL went IPO, I would be a Billionaire by now.
I don't expect Apple to product an answerbot clone of ChatGPT. That would not play to Apples strengths and would not provide much unique benefit to Apple, it's platform, or its customers.

What I do expect is for them to use LLM-style AI to enhance the use of their existing features. Make photos easier to cleanup. Make music playlists better. Make it easier to create calendar events. Suggest code snippets in Xcode. Make it easier to find what you want in Spotlight or in the App Store. Make Siri smart again. Give Siri enough ability to understand intent and to clarify requests and then be able to match that to available services. Maybe even be able to do more than one thing in a sequence. Make it easier to automate all kinds of tasks.

Using AI for that gives more control to the users and surfaces functions that are otherwise hidden or hard to use.

Agreed. Also think this is the most probable path for them to take.

How do you imagine these features getting exposed to end-users? (When LLMs are discussed people default to thinking in terms of chat interfaces but to your point that doesn't have to be a hard and fast rule.) Do you think there will some sort of primitives that'll get exposed as APIs to developers, like with CoreML[1] for example?

[1]: (https://developer.apple.com/machine-learning/core-ml/)

I hope that there are standard interfaces like intents that Apple and third parties can publish that expose their functionality. That would make it easier to use them in a more modular fashion, through automation, through Siri voice, or through Siri text prompts. Some of that intents interface exists already but they should greatly expand it with AI in mind.

Another route would be for Apple to provide an LLM accessible to app so that they could offer that to interpret use request within those apps. If you're in Calendar, make it easy to add an event with voice or text.

I would love to be able to chat with Siri, just like one can with ChatGPT, Gemini, Ollama, etc. It would be nice if Siri had a very streamlined LLM built-in to be used for that, and/or more with the ability to "offload" to a remote server if no local answer is found (that last behind a user's setting).
I just never use Siri, every time I use it I feel like I still have to use very stilted and clearly enunciated speech, even though the speech transcription for voicemail can apparently deal with a drunk pilot flying at altitude with the windows open.

That and it still seems to be very command driven and have no contextual link between sequential statements means that if it gives you a useless response you can’t give it a further instruction to refine the answer. Maybe that particular issue would be improved by LLMs or something, but I literally never want an LLM to be the source of an answer.

After a day of heavy gpt interaction it’s jarring to tell Siri to turn on “private mode” and be met with a head-scratchy “I can’t do that.”

“Turn on do not disturb” worked fine.

Apple tried to give Siri a humane sound but the system is rigid and dumb. Given what we know is possible these days, and the vibe Apple likes to project, this situation truly grates.

Siri is 10 years old, and even then, the tech was probably not cutting edge. Apple bought much of Siri from another company and we have heard that they ended up having to do a lot of hand coded cases and that it was so complicated that they stuggled to make any changes. It was clearly a dead end, but until recently there was not a lot of demand for more. Alex and Google were a little bit ahead but neither one of those is an honor student either.

I have to confess when you said to turn on "private mode" I wasn't sure what you meant. The browser's private window came to mind until you said "do not disturb".

It could be useful in the car but it's too dumb without a LLM behind it.
My point was that “private mode” was indeed a mistake on my part, I would just expect a modern llm-driven system to react in some way that’s more productive than “I can’t do that”. Almost anything would be better.
That list of expected items is so versatile that is might as well be ChatGPT. The main problem for Apple is how to differentiate themselves in marketing terms to make it sound unique and prestigious so it can drive sales of their hardware and hopefully even subscription to their (AI?) services. Now they cannot just create a clone of ChatGPT or one of the other services because that will not sound unique even if their most loyal fans might invent reasons to paint it that way...and they certainly cannot publish something much worse than ChatGPT since even the fans would not be able to ignore that. So they just wait until the excitement of other services has passed and then release something bland but covered in shiny coat and let the marketing do the rest. Amazing - AppleGPT can schedule appointments all by itself without your input or requests based just on content of your iMessage chats and emails. No other AI can do that! ...and just ignore the glaring privacy nightmare you live in where ONE company controls all your digital life.
They should however get a good enough Siri LLM out the door.
For real. I can't believe how little improvement there has been to Siri over the years.
I feel like it gets worse and worse. I used to be able to get it to turn the tv off and on reliably, but now I have to say it twice or more.
It's schizophrenic. I have a shortcut called "What's the spa temperature" that ran fine for years. Of a sudden, now it can't match to it.

"Hey Siri, what's the spa temperature?" => "The temperature outside is 75°"

"Hey Siri, run the shortcut 'What's the spa temperature'" => "I'm sorry, I didn't quite get that!" (x3 repeat, of course, because I can't believe what's happening)

=> rename to: "spa status"

"Hey Siri, spa status" => "...yada yada... it's currently 81 and set to 93, please check back in 15 minutes" (that's the original shortcut that I'd put together).

Suuuuper basic stuff throughout HomeKit, Siri, are horrible. Renaming things doesn't always take, "I'm sorry, I'm having trouble connecting to the internet." Can't properly set an Airplay target by voice.

"Hey Siri, turn off the light!" => (iPhone hears it instead of HomePod) => "Did you mean master bedroom? Kids room? Guest room? Kitchen? Living room? Family room? Office? Dining room?" => cancel, cancel, cancel! :-(

(comment deleted)
A funny thing I remembered was that when ChatGPT 4 was announced, Google panicked-released Bard in a half-complete state as a response

Apple's response during that time was to release the yellow iPhone.

Still wanting a universal translator on the iPhone a la Meta Seamless. Language would no longer be a barrier given global mobile phone saturation.