Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple Rumored to Follow ChatGPT With Generative AI Features on iPhone as Soon as iOS 18
- Thread starter MacRumors
- Start date
- Sort by reaction score
Or it might launch maps and give directions to Cuba... 'fun times ahead' no doubt..
Siri sucks. It's laughably bad compared to Alexa and Google. Why? The privacy thing. The more information these "assistants" have, the better they work, which makes perfect sense. Apple has painted themselves into a corner when it comes to privacy, the result being that their assistant will always be brain-dead compared to others.
Whatever Apple does with AI, it will likely be underwhelming and limited, just like Siri.
Absolutely and good example. There is no inherent 'confirm validity of what is being stated' in LLMs - false history, poisonous recipes, non-existent court cases, etc.
Of course there's work being done on this but LLMs themselves are quite 'large' with billions of hyper parameters, and not insignificant resources. Yes, there are several smaller implementations, which are effectively pre-trained reduced hyper-param models more focused in a specific area/less general. 'Training' is not really a real-time thing on LLMs, although 'prompt refinement' can be. There are ways to 'fine-tune' versus the enormous effort of training from scratch (e.g. ingest the internet).
Without some type of significant breakthrough, the data collection loop might well turn into for 'updates to 'SiriGPT' with OS releases, or quarterly backend updates using the aggregate user data for additional fine-tuning.
I don't work for them, but you can play around with 'real LLMs' using H2) Studio if anyone's interested..to get a better feel for what LLMs really are and how they are typically 'fine-tuned' etc.
H2O LLM Studio
I'm sure Apple's priorities is running models directly on devices. They may be well ahead of competition in that metric. Often Apple is just not in the same race as everyone else. There is a lot of room to optimize models to make smaller models work as well as large ones. The race to create ever larger models like everyone else might not be in Apple's interest.
Apple's lag with releasing the new tech (OLED iPad/Mac, USB-C on the phone, periscope lens and now ChatGPT) on average is climbing fast and might be approaching 5-7 years.
I don't disagree with a second that they can do well. What I'm saying si that they can't possibly have the breadth of input of one that use a wider knowledge base as a starting point?
Come on, think outside the box. The number of people that use, let's say Google and other usual suspects for research or just general info has a significant effect on your life whether you like it or not.
It massively increases the amount of 'input' you receive whether that input is good/bad, or welcome/unwelcome.
I can have privacy and ALSO a full life becasue other are willing to give up theirs.
It's like a fight for your rights. People that sit at home because they can't be bothered or are in fact too restricted in time, money or other resource benefit from those that do the hard work.
Those people you might see on YouTube 'flexing their rights' are often the ones that ensure yours don't get trampled on even though you may not be interested in what they're doing.
"as the company considers how to use and process personal data in a way that aligns with its commitment to customer privacy" -
This means that Apple needs to figure out how only it can use the customer data and monetize while denying the same to competitors (all others).
This means that Apple needs to figure out how only it can use the customer data and monetize while denying the same to competitors (all others).
Oh, how wrong you are.
Today might be VR/AR, but tomorrow will be true Spatial Computing, with no headset required. Technology evolves, and needs to start somewhere. You can't jump to the end of the race.
When you search for the hardware requirements for running ChatGPT models, you come across different estimates including the ones like this: "The largest model, GPT-3 “175B,” requires 600 GB of RAM and a specialized GPU with 250 GB of VRAM.".
For iPhone to be able to run a ChatGPT model, it would need to be foldable (10 times?) and cost (based on Apple upgrade strategies) a few million dollars.
You must be kidding.
Confirmed: Apple Caught In Siri Privacy Scandal, Let Contractors Listen To Private Voice Recordings
Atherton Research's Principal Analyst and Futurist Jeb Su unpacks the ramifications of Apple letting outside contractors access private conversations of Siri users.
www.forbes.com
It's early, and the research is still evolving, but yes: it seems that you can get small models that can understand a broad input and produce high quality narrow output. You might also have a "coordinator" that is trained on picking the right small model given an input. All of the leading edge papers that I've seen are showing this approach to be very, very effective.
I still trust Apple way more than Google...You must be kidding.
Confirmed: Apple Caught In Siri Privacy Scandal, Let Contractors Listen To Private Voice Recordings
Atherton Research's Principal Analyst and Futurist Jeb Su unpacks the ramifications of Apple letting outside contractors access private conversations of Siri users.
Untrue. The only advantage Google Maps has, at least here in the US, is that Google's point of interest database is a lot better. Navigation, Apple Maps is better designed and has much better voice prompts and graphics. Routing and traffic info is constantly updated thanks to millions of iPhone out there on the road.
LLM models improve very fast. Your particular example is incorrect. Out of curiosity I just asked ChatGPT the following question: "Tell me about the biography of George Washington's youngest grandchild Sara". And the answer was:
I’m sorry for the confusion, but there seems to be a misunderstanding. George Washington, the first president of the United States, did not have any biological children of his own1. However, he was a stepfather to Martha Washington’s two younger children from her previous marriage: John Parke Custis and Martha Parke Custis1.
John Parke Custis had four children: Elizabeth Parke Custis, Martha Parke Custis, Eleanor Parke Custis, and George Washington Parke Custis2. There is no record of a grandchild named Sara in the historical data available. If you have any other questions or need information on a different topic, feel free to ask!
But I am sure SiriGPT will get confused.
POI is at least half of the map app functionality (location, work hours, how busy, reviews etc). Opinions about the voice prompts and graphics are very subjective. Out of those millions of iPhone users on the road, how many use Google Maps? Google have access to their data in addition to the data of Android users (about 50% in US?) and Apple has access only to iPhone data.
How could it be based on Siri even if Apple wanted to?
Siri has nothing to do with AI, as all of its content is human curated.
