During the keynote on June 10, it was announced the features of the HE is the apple of 18, iPadOS 18 and macOS Sequoia, referred to collectively as well as Apple's Intelligence. The company's Cupertino has published details of two of the models will be used for the creation on the device, and the server (via a Private Cloud Compute).
The pattern of IT on the device and in the cloud
The model, who works in the device, there are approximately 3 billion parameters. It is therefore an SLM (the Language of the Smaller), but it requires at least an iPhone 15 Pro, or an iPad, or a Mac with an M1. The model, which works in servers, it is much larger (the size is unknown). An algorithm automatically determines which is to be used for the repair of specific.
Both of them were trained by the AXLearn, is a framework of open-source, which allows for efficiency and to scale to the different harduerësh, and cloud computing. The details of the licensing and public data are collected on the web by zvarritësi the web AppleBot, were used for the training. The owners of the site can get him in the robots.txt. Apple's use of the user data and remove any sensitive information that is available on the internet.
After the improvements, and the optimizime. The latter are particularly important for the model of the device. The objective is to minimize the use of RAM in it, and increased performance. On the 15th of the Pro's in the reach of the 30 soil in the second and a delay of about the 0.6 milliseconds, so that the answers are practically instant.
The training of the model is achieved through the unit to a small network of nerves, called the adaptorë, which are loaded on the fly, based on the functionality of IT. The assessment of the human, the model is the one that stands with the best user experience. Apple then tested the generation of a summary for emails and announcements, taking the results to be higher than those of the Phi-3-mini-model of the Microsoft.
The estimate is similar to the human being conducted on the ability of the model, including the answer to the questions are open-ended, write the code, and is the solution of mathematics. Model of the device was estimated at more than the Phi-and 3-mini, Gemma-2B, Gemma-7B, and the Mistral-7B, as the model of the cloud exceeds the DBRX-Instruct, Mixtral-8x22B, and the GPT-3.5-Turbo (4 GPT Turbo provides high performance the most).
The two models of the Apple is also the “the mighty” when you use an application to generate the content to be rendered, or to be dangerous. Of course, it will be improved further for the answers to the most suitable. CEO Tim Cook said that the models are not immune to the halucinacioneve.
Discussion about this post