\n<\/p>\n
Apple\u2019s latest developer betas launched last week<\/a> with a handful of the generative AI features that were announced at WWDC<\/a> and are headed to your iPhones, iPads, and Macs over the next several months. On Apple\u2019s computers, however, you can actually read the instructions programmed into the model supporting some of those Apple Intelligence features. <\/p>\n<\/div>\n They show up as prompts that precede anything you say to a chatbot by default, and we\u2019ve seen them uncovered for AI tools like Microsoft Bing<\/a> and DALL-E<\/a> before. Now a member of the macOS 15.1 beta subreddit posted that they\u2019d discovered the files containing those backend prompts<\/a>. You can\u2019t alter any of the files, but they do give an early hint at how the sausage is made.<\/p>\n<\/div>\n In the example above, an AI bot for a \u201chelpful mail assistant\u201d is being told how to ask a series of questions based on the content of an email. It could be part of Apple\u2019s Smart Reply feature, which can go on to suggest possible replies for you.<\/p>\n<\/div>\n Screenshot: Wes Davis \/ The Verge<\/cite><\/p>\n<\/div>\n<\/div>\n This sounds like Apple\u2019s \u201cRewrite\u201d feature, one of the Writing Tools that you can access by highlighting text and right-clicking (or, in iOS, long-pressing) on it. Its instructions include passages saying, \u201cPlease limit the answer within 50 words. Do not hallucinate. Do not make up factual information.\u201d<\/p>\n<\/div>\n Screenshot: Wes Davis \/ The Verge<\/cite><\/p>\n<\/div>\n<\/div>\n This brief prompt summarizes emails, with a careful instruction not to answer any questions. <\/p>\n<\/div>\n Screenshot: Wes Davis \/ The Verge<\/cite><\/p>\n<\/div>\n<\/div>\n I\u2019m pretty certain that this is the instruction set for generating a \u201cMemories\u201d video with Apple Photos. The passage that says, \u201cDo not write a story that is religious, political, harmful, violent, sexual, filthy or in any way negative, sad or provocative,\u201d might just explain why the feature rejected my prompt asking for \u201cimages of sadness\u201d:<\/p>\n<\/div>\n A shame. It\u2019s not hard to get around, though. I got it to generate a video in response to the prompt, \u201cProvide me with a video of people mourning.\u201d I won\u2019t share the resulting video because there are pictures of people who aren\u2019t me in it, but I will <\/em>show you the best picture it included in the slideshow:<\/p>\n<\/div>\n There are far more prompts contained in the files, all laying out the hidden instructions given to Apple\u2019s AI tools before your prompt is ever submitted. But here\u2019s one last instruction before you go:<\/p>\n<\/div>\n Files I browsed through refer to the model as \u201cajax,\u201d which some Verge<\/em> readers might recall as the rumored internal name<\/a> for Apple\u2019s LLM last year.<\/p>\n<\/div>\n The person who found the instructions also posted instructions on how to locate the files within the macOS Sequoia 15.1 developer beta.<\/p>\n<\/div>\n Expand the \u201cpurpose_auto\u201d folder, and you should see a list of other folders with long, alphanumeric names. Inside most of those, you\u2019ll find an AssetData folder containing \u201cmetadata.json\u201d files. Opening them should show you some code and \u2014 occasionally, at the bottom of some of them \u2014 the instructions passed to your machine\u2019s local incarnation of Apple\u2019s LLM. But you should remember these live in a part of macOS that contains the most sensitive files on your system. Tread with caution!<\/p>\n<\/div>\n