Apple Intelligence is already here, but its premiere has not been as brilliant as many expected. The company’s new proposal still has improvement marginboth in English and Spanish. Although it has been presented as a clear sales argument, its impact among people has been rather discreet. The first impressions are mixed and, for now, it has not managed to generate groundbreaking enthusiasm.
From Cupertino they already move to reinforce one of the greatest software bets in their recent history. Among the movements in progress are two fronts: the rumored restructuring of the team responsible for Siri, whose improved version has been delayed until 2026, and the creation of new techniques designed to improve their language models, with the mission of not neglecting its focus on privacy.
One step beyond synthetic data
Apple usually training its models with synthetic data and data labeled by humans, a solution that has been effective to some extent. It does not always represent the real world. Consequently, it limits the functioning of AI products. This has led the technological led by Tim Cook to develop a new solution that combines synthetic data with anonymous signals of participating devices.
As explained in an article published this week, everything starts with a synthetic message, that is, an email invented by Apple itself with a format that simulates the real emails. For example: “Would you like to play tennis tomorrow at 11:30?” From there, several variants are generated that change some elements, such as sport, schedule or tone, to try different possible structures.
These phrases are sent to a part of the devices whose users have agreed to share analytics with Apple. There is something key: each iPhone, iPad o Mac Take a handful of real emails transformed into embeddings local, that is, mathematical representations that convert each message into a set of numbers that reflect their theme, style and length. The important thing is that these emails never leave the device.
Thus, the system compares embeddings synthetic, which Apple has previously generated, with the embeddings Real emails, to see which ones are more similar. This resemblance is reduced to an anonymous signal, a simple “this version coincides better”, which is sent to Apple without revealing the original mail or the embedding of the user. With this, Apple intends to learn which synthetic variants better reflect the real use of language, but without seeing a single fragment of private content.
The idea is that this helps improve Apple Intelligence functions such as email summaries or writing tools.
This approach is based on the same differential privacy techniques that Apple already uses in other functions such as Genmoji. In that case, the company collects anonymous signals about what prompts They are more popular, such as “a dinosaur with a hat”, to improve the results without registering which user made what request.
The idea is simple but very interesting. Improving without using user data data allows you to maintain the privacy approach that the company has been defending so many years.
This new technique will begin to be implemented in the next betas of iOS 18.5, Ipados 18.5 and Macos 15.5. It should be noted that only those who have activated the option to share analytics from the privacy settings are activated. So, if you don’t want to be part of this system, You can deactivate it whenever. You just have to go to Settings > Privacy and safety > Analysis and improvements and deactivate the option “Share iPhone analysis”.
Images | Apple | APPSHUNTER
In WorldOfSoftware | Public administrations turn the tortilla: Midni promises to be another success with Midgt and my citizen folder
