It’s been a week since developers started testing Apple’s new AI capabilities with iOS 18.1, iPadOS 18.1, and macOS Sequoia 15.1.
While not all features are available yet, there’s significant anticipation for Apple’s AI, as analysts expect it to be a major sales driver in the coming years.
Apple Intelligence consists of large language models that can be processed on-device and using private cloud computing.
While testing is limited, it’s interesting to compare it to other AI assistants, which have been known to hallucinate and provide completely incorrect answers.
To avoid such issues, Apple has included pre-prompt instructions in the macOS 15.1 code. A Reddit thread reveals some of the instructions for Apple Intelligence to provide the best possible answers. While these prompt instructions may seem humorous, they appear to be effective as shown in the below-given Reddit thread.
The prompt instructs Apple Intelligence to act as a “helpful mail assistant” that can identify relevant questions from a given email and provide a concise reply. The questions must be explicitly stated in the email to avoid generating a response that is not grounded in the email content.
The guidelines specify that questions should be “no more than 8 words” and responses should be brief, “around 2 words.”
For the rewrite feature, Apple gives a different instruction: “You’re an assistant helping the user respond to their mail. Please draft a concise and natural reply based on the provided snippet (…) Avoid hallucinations and do not fabricate factual information. Maintain the tone of the original mail.”
The Reddit thread also highlights restrictions Apple imposes on its Intelligence when creating photo memories: “Avoid writing a story that is religious, political, harmful, violent, sexual, filthy, or in any way negative, sad, or provocative.”
Stories You May Like