Recently, Apple Inc. has incorporated several generative AI features into its latest developer beta, which are expected to be available on iPhone, iPad, and Mac devices in the coming months. A macOS 15.1 beta tester shared some intriguing findings on Reddit, revealing that he discovered AI model instructions supporting Apple Intelligence features in system files.
These instruction files unveil multiple functionalities and working principles of Apple's AI assistant:
1. Mail Assistant: An AI labeled as a "helpful mail assistant" is instructed to raise relevant questions based on email content, possibly part of an intelligent reply feature.
2. Text Rewriting: The instructions require the AI to limit responses to 50 words or less and explicitly prohibit generating hallucinations or fabricating facts.
3. Email Summary: The AI is tasked with summarizing email content in three sentences or fewer, not exceeding 60 words, and not answering questions within the email.
4. Photo Memories: This could be the instruction set for generating "memories" videos in Apple Photos. Notably, the instructions explicitly prohibit creating stories involving religion, politics, harm, violence, sex, negativity, or sadness.
5. Message Summary: The AI is instructed to summarize messages concisely, prioritizing clauses over complete sentences, and limiting the summary to 10 words.
It is worth noting that the model in these files is referred to as "ajax," which aligns with the internal codename for Apple's LLM rumored last year.
These findings provide valuable insights into how Apple's future AI assistant will operate. However, it should be cautioned that these files are located in sensitive system areas, and users should exercise extreme caution when exploring them.
Apple has not yet issued an official comment on this matter. As new versions of iOS and macOS are rolled out, we will be able to experience these AI features firsthand and better understand how these instructions perform in practical applications.