At WWDC 2024, Apple introduced its groundbreaking Apple Intelligence—a suite of AI language models designed to enhance user experience across Apple devices and servers. This new AI framework integrates both on-device and server-based models, aiming to deliver faster and more accurate results.
Foundation Language Models
Apple’s new foundation language models are large-scale generative AI systems, utilizing up to 3 billion parameters. These models are designed for general use, enabling more natural and context-aware interactions with Apple's AI systems. Apple has named these models AFM-on-device and AFM-on-server, reflecting their operation on local devices and Apple’s AI servers, respectively.
Key technical features of these models include:
Transformer architecture
IO Embedding Matrix
Pre-normalization
Query-key normalization
Grouped-Query attention
SwiGLU activation
RoPE positional embeddings
Fine tuning
Human adjustments and input
In addition, AppleBot—an automated web crawler—helps Apple Intelligence learn from the web, while open-source software from GitHub aids in refining code-based AI capabilities.
Private Cloud Compute
Apple's Private Cloud Compute (PCC) service leverages these models, providing enhanced speed, accuracy, and privacy for AI operations. PCC also employs Secure Enclave and Secure Boot technologies to ensure data protection and system integrity.
Apple's commitment to Responsible AI is evident in their detailed white paper, which outlines the ethical development and deployment of their models. The company emphasizes that these models are designed to support everyday activities across Apple products while aligning with Apple’s core values.
With the upcoming releases of iOS 18 and the next macOS iteration, Apple's new AI models promise to bring optimized performance to users, both on-device and in the cloud.