I think Apple missed the boat on LLMs and was surprised by the success of ChatGPT.

They are no doubt working on an LLM-based Siri now, but they don't rush things, and they don't preannounce things, so we'll only hear about it once it is ready. And they are probably going to use a different term for marketing it rather than using a generic term like LLM.

Yeah, it seems plausible that they'll do it eventually, but it'll be "the new improved Siri, powered by the Neural Engine and transformer models".

While trying to come up with an Apple buzzword I realized we are not far away from Apple announcing Thought Control.

They better quadruple iPhone RAM for that.

Siri, now powered by Speech Synthesis Engine™.

Not sexy enough. My money’s on „Apple Voice“

I don't think they missed the boat, I think they recognize that LLMs aren't Apple-ready yet.

They don't roll out half-baked products, and they won't be comfortable placing the big disclaimers that ChatGPT relies on. They'll incorporate LLMs if and when they can guarantee that what Siri says will be consistently accurate and extremely difficult to abuse.

I agree, a quote I heard from a former Apple employee from the onboarding materials: “Apple doesn’t do things first, we do things right.”

Generally agree, but Siri is out and is awful

Or in the case of the Touch Bar: they did it first and they did it wrong.

Siri is the very definition of a half-baked product.

Perhaps even quarter baked. and has been sitting there in the turned-off-oven for YEARS.

LLMs as they are now are not a drop-in replacement for what people use voice assistants for.

If you ask it a question it doesn't know the answer to, "I don't know" is better than making up a bullshit answer. That could be potentially hazardous at iPhone levels of scale, giving everyone easy access to what appears to be a very intelligent AI chatbot without them understanding that it shouldn't be relied on for a lot of information people will ask it.

They would get an immediate benefit from using an LLM for intent recognition - i.e. knowing that a user wants to search or set a timer, etc. Current Siri is really bad at this. I'm sure even a fine-tuned Llama 7B would be a massive improvement over Siri.

Agree that Siri isn't very good, and have been quite stagnant. It was released in 2011, that's 12 years ago!? Sure it's probably better now, but not 12 years of development better?

But, at the same time. ChatGPT and the likes have a very different failure mode, that feels even less Appley. Like saying actual fake stuff, or saying it understand that you want to turn the AC on when your son leaves school coming home - but then just not doing it, or doing something completely different. With current Siri the failings are at least very obvious…

Anecdotally, I feel like Siri was usable to send text messages and emails when it first came out. Now it struggles to do even that.

Siri was actually once a better product than it is now.

Just like the apple keyboard on iPhones.

You’re not imagining it; it is actually worse.

This is the first ever integration with iPhone after acquisition, it never got better than this https://youtu.be/SpGJNPShzRc?si=a5JgHgnK6Nf2p7BY

It is phone voice assistant and if you learn how to use it, it works great as intended. The problem is when you want too much interaction. ChatGPT can fake it, Siri - cant.

No half-baked products? Have you ever tried using the translation feature in the book app? That is the most infuriating, dumb piece of software I have seen in a long time.

They definitely missed the boat in my eyes. The speed at which others have been able to produce meaningful assistants with LLM is astonishing. Considering Apple also have an approach where they announce and release in November / early the following year, to have nothing revealed seems like a missed opportunity to me.

> They don't roll out half-baked products

You must be really young if you don't remember when they rolled out Apple Maps.

there was nothing wrong with apple maps, you were just holding it wrong

It’ll be something like Siri+… now only $9.99/month

Apple has consistently missed the boat on every major technical advancement, when compared to android. But when they do catch up, the quality is unmatched! Hoping the same with LLMs. I am sure they’re already working on it.

It takes a while to roll out fundamental changes to things at that scale.

There were zero rumors about LLM from Apple being announced today. It wouldn't be the time and the place for such an announcement.

Such a feature would be included with iOS 18 and that's not going to be talked about until next June's WWDC 2024.

When they talked about Siri on-device on Apple Watch 9 they briefly said Siri had a ‘new transformer model’ and talked about improved dictation accuracy.

Honestly was half expecting more along those lines as part of the new chip announcement for the iPhone 15 Pro, but apparently not.

Reliability, probably? Apple is not going to allow for hallucinations.

Siri in its current form is constantly "hallucinating" all sorts of stuff as it is. "Hey Siri, turn off the lights in the kitchen." -- "Sure, now playing a continuous stream of U2 on all of your speakers!"

LOL as an iPhone user whose only music on the phone is That One Free U2 Album That We Never Wanted, I laughed too hard at that one!

Because they can't control it not talking extremist stuff, and that would be used by media and normies for months

I think the last thing people want is a voice assistant that babbles on and on forever when asked a simple question.

I really hope that when they release something, the responses will be 5% of the endless stream of words ChatGPT & others produce.

Agree - fwiw, I added "ChatGPT should keep answers brief and direct." as part of my Custom instructions and it's a massive improvement.

I usually say things like “please be extremely succinct. Response should be one sentence at most”.

“Sure! I will try to be succinct.

[4 paragraphs]”

I will remind you that most every parent and child both want this, albeit for totally different reasons.

From rumours [1] it is being worked on.

Apple isn't the sort of company to rush something out just to be competitive.

[1] https://www.macrumors.com/2023/09/06/apple-conversational-ai...

Not going to happen until Apple can fully control the bounds of the output generated by the LLM (which no one is capable of doing right now). A handful of instances of it spouting...unfriendly stuff is enough to dent Apple's reputation.

There is a rumor that Apple is working on an "embedded" "SiriGPT". Imagine "ChatGPT" but on device, no API calls, no extra latency. And all apps on your device can use this "SiriGPT" instantly to enrich the user experience.

LLM? Everyone has an LLM. They'll announce a PLM (personal language model) in a year and act like they were first.

it will be more profitable if they sell the Siri entry point to the one who bothers to do the research , the same way they sell the search box to google.

May be Apple care about privacy?

Siri has been less performant than OK Google for something like 5 years, in my estimation.

You can set up a Siri Shortcut linked to chatgpt and fire away your questions.

Apple's own LLM and major Siri update is supposed to be coming in iOS 18 next year. So says the rumour, https://www.macrumors.com/2023/09/06/ios-18-siri-improvement...

LLMs can't guarantee factuality. They're great for the conversational component, but they're a seriously bad and responding with reliably factual information. More important that its ability to deal with complexity would be an assistant's ability to respond to a request correctly.

Nothing and no one can guarantee factuality. That's a red herring.

That's abusing a turn-of-phrase to provide a counter-argument. Of course no one can guarantee universal factuality. But relative to information we're seeking from an assistant (schedules, email data, messages, music preferences) it is very easy to obtain highly accurate data. But LLMs don't work like that.

The conceptual model behind LLMs utilizes frequency to generate a reasonable response to a query. But that's a bad fit when you're looking for data accuracy that isn't represented by the model. If your model says most dental appointments occur on the 15th and responds so but my appointment is on the 17th, it's not only useless to me it's detrimental.

You underestimate how long it takes to create the marketing materials.

If I ask Siri what the weather is going to be I want a deterministic response, not some hallucinated answer based on some weather blog from 2020.