Apple insists its AI models avoid Gemini components
AI
Compute
Cloud
Gemini
Apple
Apple Intelligence
Google
TLDR: Craig Federighi said Apple Foundation Models use none of Google Gemini apps, client code, or Google Search infrastructure, while training used Gemini outputs for refinement. Apple Intelligence users get new on device and private cloud models, plus AFM Cloud Pro running on Nvidia GPUs in Google hosting.
Key Takeaways:
- Apple’s AFM family powers Apple Intelligence with a model lineup split between two on device models and three server models.
- Federighi said Apple uses none of Gemini models, Gemini client code, or Google Search for the knowledge backbone; training still used Gemini frontier outputs for refinement.
- AFM Cloud Pro shifts from standard Private Cloud Compute by extending it to Nvidia GPUs in Google cloud using ambiguous confidential compute, to keep Apple server contents private.
Apple is drawing a bright privacy line: no Gemini code, no Search plumbing, yet Gemini outputs helped refine the brains. It is a distinction that matters to lawyers and also to what users assume is happening behind the screen.
Apple is drawing a bright privacy line: no Gemini code, no Search plumbing, yet Gemini outputs helped refine the brains. It is a distinction that matters to lawyers and also to what users assume is happening behind the screen.
Q&A
If Apple used Gemini outputs during refinement, what stops the models from inheriting Gemini style biases or weaknesses?
Apple did not describe a full bias audit in this material, but refinement via distillation or reinforcement learning can still transfer behavioral patterns. Expect scrutiny on evaluation suites, red teaming, and ongoing monitoring for regression.
How might the System Orchestrator change user experience when requests need both on device and private cloud reasoning?
By routing queries based on complexity and personal context, Apple can reduce latency for quick tasks while reserving heavy reasoning for private cloud calls. The user sees steadier results, not a menu of model choices.
Why does Apple treat ambiguous confidential compute as a privacy breakthrough instead of a normal cloud setup detail?
Because it targets the trust boundary: Nvidia GPUs can be hosted elsewhere while compute privacy reduces the risk of reading Apple server contents. That shifts trust from hardware operators to cryptographic and access controls.
What does it mean for competition if AFM Cloud Pro quality is similar to Gemini frontier models but runs under different deployment constraints?
Comparable quality with a distinct privacy and deployment architecture raises the bar for rivals who rely on more public or shared infrastructure. It also pressures other model providers to justify both performance and data handling.
Could Apple’s insistence that knowledge comes from its own World Knowledge Service reshape expectations about citations and sources?
It suggests Apple can tightly control freshness, provenance, and retrieval logic for current events. Users may get a more consistent source policy, but they may also notice fewer cross ecosystem links than rivals.
No comments yet. Be the first to share your thoughts!