Re:Apple iPhone 15 Pro and iPhone 15 Pro Max
by quitit on Wednesday September 13, 2023 @07:40AM
(#37495396)
Also interesting was the large number of Neutral engine cores.
These two together seem to be very much about enhancing AI and Siri to be mostly on-device (something which they already started on).
A frequent complaint I read on HN is "why aren't there LLM based assistants as yet from any vendor" - but the answer to this is straightforward, it's not realistic to deploy cloud-based LLM at Siri-size user bases. Most queries also don't need LLM-style responses. (Also the rapid development of the field means that processing requirements are rapidly decreasing.)
In a prior keynote Apple noted that Siri was the most used assistant in the world. I don't have 3rd party data to support that, but it's a big enough claim that we could take their word on it just for conjecture purposes about the processing power needed to deploy LLM-based Siri.
So if we're going to see LLM-based usability enhancements to Siri, both in the types of data it's allowed to work with and enabling natural, context-driven interactions. That will have to be fully on device, and ultimately that will be a better experience for users as internet connectivity is an unreasonable barrier for such interactions (especially as it can stutter natural-language conversations).
These two together seem to be very much about enhancing AI and Siri to be mostly on-device (something which they already started on).
A frequent complaint I read on HN is "why aren't there LLM based assistants as yet from any vendor" - but the answer to this is straightforward, it's not realistic to deploy cloud-based LLM at Siri-size user bases. Most queries also don't need LLM-style responses. (Also the rapid development of the field means that processing requirements are rapidly decreasing.)
In a prior keynote Apple noted that Siri was the most used assistant in the world. I don't have 3rd party data to support that, but it's a big enough claim that we could take their word on it just for conjecture purposes about the processing power needed to deploy LLM-based Siri.
So if we're going to see LLM-based usability enhancements to Siri, both in the types of data it's allowed to work with and enabling natural, context-driven interactions. That will have to be fully on device, and ultimately that will be a better experience for users as internet connectivity is an unreasonable barrier for such interactions (especially as it can stutter natural-language conversations).