Chance Miller <https://9to5mac.com/author/chmiller44/>
Apple is widely expected
<https://9to5mac.com/2023/10/22/apple-ai-features-ios-18-iphone-siri/> to
unveil major new artificial intelligence features with iOS 18 in June. Code
found by 9to5Mac in the first beta of iOS 17.4 shows that Apple is
continuing to work on a new version of Siri powered by large language model
technology, with a little help from other sources.
In fact, Apple appears to be using OpenAI's ChatGPT API for internal testing
to help the development of its own AI models.
According to this code, iOS 17.4 includes a new SiriSummarization private
framework that makes calls to the OpenAI's ChatGPT API. This appears to be
something Apple is using for internal testing of its new AI features.
There are multiple examples of system prompts for the SiriSummarization
framework in iOS 17.4 as well. This includes things like "please summarize,"
"please answer this questions," and "please summarize the given text."
The system prompts also make mention of what to do when it is given input in
the form of an iMessage or SMS. This aligns with previous reporting from
<https://9to5mac.com/2023/10/22/apple-ai-features-ios-18-iphone-siri/>
Bloomberg, which said Apple is working on AI integration in the Messages app
that can "field questions and auto-complete sentences"
Given user received SMS(s) containing sender, content, send_time fields,
suggest an appropriate action for a voice assistant to take, including the
action type, action value, action value type, and confidence score in JSON
format. The possible action types are MessageReply, GetDirection, Call,
SaveContact, Remind, MessageContact, and None. The possible ActionValueType
are message, address, phoneNumber, contact, reminder. The possible score
value ranges from 0 to 1 which represents the confidence score of the
suggested action.
Apple is unlikely to use OpenAI models to power any of its artificial
intelligence features in iOS 18. Instead, what it's doing here is testing
its own AI models against ChatGPT.
For example, the SiriSummarization framework can do summarization using
on-device models. Apple appears to be using its own AI models to power this
framework, then internally comparing its results against the results of
ChatGPT.
In total, iOS 17.4 code suggests Apple is testing four different AI models.
This includes Apple's internal model called "Ajax," which Bloomberg has
previously reported
<https://9to5mac.com/2023/07/19/apple-gpt-chatbot-generative-ai-plans/> .
iOS 17.4 shows that there are two versions of AjaxGPT, including one that is
processed on-device and one that is not.
Other models referenced by iOS 17.4 include the aforementioned ChatGPT as
well as FLAN-T5, which you can learn more about here
<https://huggingface.co/docs/transformers/model_doc/flan-t5> .
The biggest takeaway from these findings is that Apple is ramping up its
efforts to integrate large language models into iOS. It's also notable to
see Apple simultaneously developing its own system and comparing the results
of that system to things like ChatGPT and FLAN-T5.
In October, Bloomberg's Mark Gurman gave an in-depth rundown
<https://www.bloomberg.com/news/newsletters/2023-10-22/what-is-apple-doing-i
n-ai-revamping-siri-search-apple-music-and-other-apps-lo1ffr7p?srnd=undefine
d> on some of Apple's goals for AI in iOS 18. The report outlined that
"there's an edict" within Apple and Craig Federighi's software team to fill
iOS 18 "with features running on the company's large language model."
9to5Mac.
