Apple wants to train AI on your private personal data

Apple aims to train its AI on your private personal data. The company's next-generation Apple Intelligence is deeply integrated into its operating systems and driven by a new architecture with privacy as its foundation. A key component of this architecture is the third generation of Apple Foundation Models (AFM), a family of five models custom-built in partnership with Google.

These models include on-device and server-based options, all optimized for Apple silicon. The on-device models, such as AFM 3 Core Advanced, use a novel sparsely activated architecture that allows for powerful AI capabilities without storing or sharing user data, even with Apple. This architecture enables efficient inference and minimizes data movement, making it suitable for consumer hardware.

On the server-side, AFM 3 Cloud leverages Private Cloud Compute to enhance multimodal reasoning capabilities. Training these models is based on a mix of publicly available data, licensed or purchased data, open-sourced data, data obtained through dedicated studies, and synthetic data. Apple does not use users' private personal data or interactions for AI training.

They also acknowledge the rights of web publishers to opt out of foundation model training. The models are trained using a mix of supervised fine-tuning, multi-stage reinforcement learning, and hardware optimization for Apple silicon and NVIDIA GPUs. This results in responsive, high-quality AI experiences for Apple users.