Apple Intelligence has not had the best year so far, but if you think Apple is giving up you are wrong. It has big plans and goes on with new model education strategies that could greatly improve its AI performance. However, the changes involve a closer look at your data-if you record.
In a new technical paper from Apple’s Machine Learning Research, “Understanding Total Trends for Apple Intelligence using Differential Privacy”, Apple outlined new plans to combine data analysis with user data and synthetic data rating to better train the models behind many of Apple Intelligence features.
Some real data
Until now, Apple has trained its models on purely synthetic data trying to emulate what real data can be, but there are limitations. In Genmoji’s, for example, Apple’s use of synthetic data does not always point to how real users engage in the system. From the paper:
“For example, it helps to understand how our models work when a user requests genmoji that contains multiple devices (such as” Dinosaur in a cowboy hat “) that we improve the answers to those kinds of requests.”
Essentially, if users record, the system can tune the device to see if it has seen a data segment. However, your phone does not respond to the data; Instead, it sends back a noisy and anonymized signal, which is apparently enough for Apple’s model to learn.
The process is somewhat different for models that work with longer texts such as writing tools and summaries. In this case, Apple uses synthetic models, and then sends a representation of these synthetic models to users who have chosen data analysis.
On the device, the system then performs a comparison that appears to compare these representations with samples of recent e emails.
“These most frequently selected synthetic embedders can then be used to generate training or test data, or we can run additional curation steps to further refine the data set.”
A better result
Those are complicated things. However, the key is that Apple uses differential privacy on all user data, which is the process of adding noise that makes it impossible to connect this data to a real user.
Still, none of this works if you do not choose Apple’s data analysis, which usually happens when you first create your iPhone, iPad or MacBook.
This does not put your data or privacy at risk, but this training should lead to better models and hopefully a better Apple Intelligence experience on your iPhone and other Apple devices.
It can also mean smarter and more sensible rewrites and summary.
