#头条首发大赛#
Quick guide
Apple officially released the smart beta this week, detailing the development process of its AI capabilities. In particular, Apple's two base models are trained using Google's Tensor chips instead of its own Apple silicon chips. This choice marks an important step forward for Apple in the field of artificial intelligence and machine learning. Apple is generally cautious about the product development process, but the recently released article, "Apple's Intelligent Basic Language Model," provides valuable information about model training. The AFM-on-device model uses 2,048 TPUv5p chips, while the larger AFM-server model requires 8,192 TPUv4 chips. Although the paper doesn't elaborate on why Google silicon was chosen, the decision raises questions about Apple's reliance on Google technology to drive smart advancements in the iPhone.
Apple Smart beta released
Apple Smart officially entered beta this week, with the company releasing a thorough overview of how some of its AI features were created. One of the notable details is that Apple's two base models were developed using chips made by Google. This marks a major step forward for Apple in artificial intelligence and machine learning.
Develop insights into your practice
In general, Apple is cautious when disclosing the specifics of its product development process. However, the company has always shared research results related to artificial intelligence and machine learning for easy public scrutiny. The recently published article is titled "Apple Intelligence Basic Language Model", and its release follows the introduction of Apple Intelligence at WWDC. Although this literature is primarily aimed at researchers, it contains valuable information about Apple's intelligent language model training.
Use Google Chip to train the model
The two models, AFM-on-device and AFM-server, were not trained using Apple's own Apple silicon chips. Instead, Apple has taken advantage of Google's Tensor chips, which are similar to those used in Pixel phones. According to the literature, the training process requires a large number of Tensor chips: the AFM-on-device uses 2,048 TPUv5p chips, while the larger AFM-server model requires 8,192 TPUv4 chips. It is worth mentioning that Apple's choice of Google's Tensor chips over the Nvidia chips commonly used by other companies is thought-provoking. Although this choice is not stated in the paper, questions are raised that may be answered in future publications. Ultimately, while Apple has historically favored in-house solutions, it's clear that the iPhone's smart advancements thanks to Apple's intelligence rely in part on Google's technology.