正文：
苹果公司在2024年全球开发者大会上宣布了其全新的个性化智能系统——Apple Intelligence，该系统旨在为iPhone、iPad和Mac用户提供更加个性化和实用的智能服务。经过一个多月的等待，这项技术终于在iOS 18.1开发测试版中落地，让拥有iPhone 15 Pro或iPhone 15 Pro Max的用户得以体验其功能。近日，一份长达47页的技术报告详细介绍了Apple Intelligence背后的秘密武器。
报告详细介绍了Apple Intelligence所使用的两款基础模型——AFM-on-device和AFM-server。AFM-on-device是一个约30亿参数的语言模型，而AFM-server则是一个更大的基于服务器的语言模型，两者均基于Transformer架构构建，并在设计上采用了共享输入/输出嵌入矩阵、RMSNorm预归一化、分组查询注意力等技术，以提高效率和训练稳定性。
在训练过程中，研究团队注重效率和数据质量，以确保模型能够支持Apple Intelligence的多种功能，并满足苹果对保护用户隐私和遵循Responsible AI原则的要求。后训练工作包括监督微调（SFT）和来自人类反馈的强化学习（RLHF），以及两种新的后训练算法，以进一步提升模型性能。
此外，Apple Intelligence还采用了基于运行时可交换适配器的架构，能够使单一基础模型适应数十个特定任务。通过这种方式，苹果公司不仅提升了小模型的性能，还确保了用户体验的一致性和安全性。
苹果CEO库克曾表示，Apple Intelligence是苹果创新的新篇章，将改变用户使用产品的方式。随着技术的不断发展和用户体验的提升，Apple Intelligence有望引领智能系统的新潮流。
英语如下：
Title: “Exposing Apple AI: A 47-page Document Deep Dives into the Next-Generation AI System”
Keywords: Apple AI, Personalized Services, Technological Breakthrough
News Content:
Apple Inc. unveiled its new personalized intelligent system, Apple Intelligence, at the 2024 Worldwide Developers Conference. The system is designed to offer more personalized and practical intelligent services to iPhone, iPad, and Mac users. After months of anticipation, this technology finally landed in the iOS 18.1 development testing version, allowing users of the iPhone 15 Pro or iPhone 15 Pro Max to experience its features. Recently, a 47-page technical report detailed the secret weapons behind Apple Intelligence.
The report delved into the two foundational models used by Apple Intelligence: AFM-on-device and AFM-server. AFM-on-device is a language model with approximately 30 billion parameters, while AFM-server is a larger server-based language model, both built on the Transformer architecture and employing techniques such as shared input/output embedding matrices, RMSNorm pre-normalization, and grouped query attention to enhance efficiency and training stability.
During the training process, the research team focused on efficiency and data quality to ensure the model could support the various functions of Apple Intelligence and meet Apple’s requirements for protecting user privacy and adhering to Responsible AI principles. Post-training work included supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), as well as two new post-training algorithms to further improve model performance.
Additionally, Apple Intelligence employed an architecture based on runtime swappable adapters, capable of adapting a single foundational model to tens of specific tasks. In this way, Apple not only enhanced the performance of smaller models but also ensured consistency and security in user experience.
Apple CEO Tim Cook has stated that Apple Intelligence represents a new chapter in Apple’s innovation and will change the way users interact with their products. As technology continues to develop and user experience improves, Apple Intelligence is poised to lead the new trend of intelligent systems.
【来源】https://www.jiqizhixin.com/articles/2024-07-31-7
Views: 2