Apple Intelligence: Users should evaluate the output, Google Tensor chips in use

Apple has sought feedback from users of its developer betas of iOS 18.1, iPadOS 18.1 and macOS 15.1, which are part of Apple Intelligence testing that has been going on since this week. For this purpose, the so-called Feedback Assistant Equipped with appropriate functions. These are about communicating your experience with AI operations and AI outputs to the manufacturer.

Advertisement

“obscene and offensive”

“Help make Apple products and services better! Content such as your emails from the Mail app and messages from the Messages app will be displayed to you with content generated from them through Apple intelligence features. This data always stays on your device unless you choose to submit it to Apple after evaluating your experience,” the company writes Macromers Report,

Users can rate their recent experiences with generated content via thumbs up/thumbs down. If there is a problem, you can also say, among other things, whether the summary of an email (or any other generated text) is “too long”, “too hard to understand”, “contains stereotypes or prejudices” or even “obscene or offensive”. Apple wants Apple Intelligence to assess as many areas as possible. There are also categories for smart replies, other writing tools, the (not yet available) image generator Image Playground or summaries of emails and notifications. Apple also promises that users can collect points that can be gained through more feedback. Confusingly, however, the company does not (yet) specify what these points are used for and/or whether they are purely symbolic.

What hardware did Apple use for AI?

Meanwhile, it became known that Apple uses not only its own Apple Silicon chips for its AI training, but also Google’s hardware. At least that’s what someone thinks Research Paper on Foundation Model of Group From US broadcaster CNBC found out became.

Redmi Watch 2 Lite in the test: Fitness tracker with GPS and 10 days of running time

Accordingly, Apple apparently used no or fewer GPUs from Nvidia, but rather several Tensor processing units from Google. “Cloud TPU clusters” were used for training, Apple writes. The TPUs, in turn, come from Google. So the Apple Foundation Model (AFM) used 2048 TPU-v5p chips for its on-device model. The server version again used 8192 TPU v4 chips (over eight so-called slices). What Apple paid for this is unclear.

(B.Sc.)