We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Lobsters·Jun 14·5 min read
The future of Siri, or: why private inference isn’t private enough
TL;DR
Voice assistants like Siri have struggled with performance and privacy concerns. Apple is integrating Google Gemini models with its Private Cloud Compute (PCC) system to enhance Siri's capabilities while ensuring data privacy.
✦ Why It Matters
Engineers can learn to balance AI performance improvements with stringent privacy measures in their applications.
Key Takeaways
- Explore integrating advanced AI models in applications.
- Implement stateless data handling to enhance privacy.
- Utilize trusted hardware for secure data processing.
- Stay informed about evolving privacy standards in AI.
Full Summary
Apple announced integration of Google Gemini models into Siri to improve voice assistant capabilities while addressing privacy concerns—a critical gap since deploying frontier AI models typically requires sending user data to remote servers. Apple developed Private Cloud Compute (PCC), a stateless inference system running on Apple Silicon hardware in Apple datacenters, combined with Google's Confidential Inference technology to process queries and private device data without persistent storage.
The technical approach encrypts data end-to-end from phones to dedicated servers, uses trusted hardware security modules to prevent even Apple from accessing inputs, and deletes data immediately after generating responses. The system theoretically achieves privacy-preserving inference by eliminating data persistence and leveraging hardware-based isolation, though the article raises questions about whether these architectural guarantees are sufficient in practice.
Engineers and researchers should evaluate whether stateless design and hardware security modules adequately protect against side-channel attacks, inference-time data leakage, and the fundamental tension between model capability and privacy preservation in production systems.