How Apple Fumbled Siri’s AI Makeover

Last June, at Apple’s annual developers conference, the company offered a dazzling demonstration of how artificial intelligence could supercharge Siri, the iPhone’s virtual assistant. But behind the scenes and in the months that followed, the Siri team couldn’t make up its mind about the basic technology needed to make it all work.

One option they considered was to build small and large language models, which they dubbed Mini Mouse and Mighty Mouse, according to a former Apple employee familiar with the effort. The small model would run on a user’s iPhone and handle simple tasks, such as setting an alarm using Siri. The large model would run in the cloud and perform more-complex chores, like automatically booking an Uber to get an iPhone user to their next meeting.

Then Siri’s leaders decided to go in a different direction and build one large model that could handle everything. That would have required running Siri’s software in the cloud, a reversal of an earlier approach when most of Siri’s software was moved on device for privacy. Other significant technical pivots followed, frustrating Siri staffers and prompting some to leave Apple. That put its goal of releasing the new Siri features this year in jeopardy.