Understanding Apple’s on-device and client base concepts

Siri icon in a datacenter
Apple announced new AI languages models at WWDC. These models are available on Apple devices as well as Apple’s own Apple Silicon AI servers. Artificial Intelligence (AI), which is based on language models, provides knowledge input for AI to produce answers to prompts (questions). Language models can be used to train computers in specific topics so that they act as domain experts. AI alignment is the process of designing AI systems to conform to human values, goals, and desired outcomes. Alignment is designed to keep AI focused and prevent it from becoming dangerous by deviating from its original purpose. Apple Intelligence, Apple’s AI, will be available on both devices and servers. Apple’s AI can become more accurate, focused, and faster by using new models. Foundation language models
Apple calls its general generative AI models foundation language models. These models are Large Language Models, which have up to 3 billion parameter and are designed to be used for basic generative artificial intelligence. Apple Foundation Models.Apple refers to these two models as AFM-on device and AFM on server. Apple Intelligence also includes other general-purpose models. These models can be run on Apple’s devices and servers. Apple provides a detailed 47-page white paper on its foundation language models. Apple’s Foundation models are based on a set of AI techniques that include: Transformer architecture
IO Embedding Matrix
Pre-normalization
Query-key normalization
Attention grouped queries
SwiGLU activation
RoPE positional embeddings
Fine tuning
Apple Intelligence uses an automated web crawler named AppleBot. Sites can tell AppleBot to not use their content in their robots.txt file. Apple Intelligence, for code AI, also learns by analyzing open-source software hosted at GitHub. It then condenses the information and removes duplicate cases. Apple’s white paper explains in detail how the models are created and the training methods, with some advanced math included at the end. Private Cloud Compute Apple Private Cloud Compute is a remote AI services that uses all of the models above, plus has access additional models for increased intelligence. According to a blog post that describes PCC by Apple, the company has several goals for PCC. These include speed, accuracy and privacy. PCC uses the same Secure Enclave, Secure Boot and Apple consumer devices in order to ensure that the operating system and data cannot be tampered. PCC, like many other AI offerings by tech companies, allows remote execution of AI commands, but at a faster performance. Apple’s Machine Learning Research Page on its models includes a discussion about how Apple approaches Responsible AI. Apple summarizes the foundation models by saying: “Our models were created to help users perform everyday tasks across Apple products. They were developed responsibly at each stage and guided Apple’s core value.” We look forward sharing more information on our broader family generative models, which includes language, diffusion, coding models. Apple admits using Google Tensor hardware for Apple Intelligence training. Apple Intelligence promises iOS and Mac users faster, optimized AI both on devices and the cloud. We’ll just have to wait and watch how it all plays out when iOS 18 and macOS 10.1 are released.