Giải thích: Apple Foundation Model là gì? Google và Gemini làm gì ở đây? Sao có cả NVIDIA?

Key Concepts

• Apple Foundation Model (AFM): Apple's proprietary large language models (LLMs), now in their third generation.
• On-Device AI: Models running locally on hardware (iPhone, iPad, Mac) for privacy and speed.
• Private Cloud Compute (PCC): A secure cloud infrastructure that processes data without storing it, ensuring user privacy.
• Sparse Model: A model architecture where only a subset of parameters is activated for a specific task, optimizing performance.
• Language Model Protocol: A standardized framework allowing developers to integrate various AI models (AFM, Gemini, Claude) into their apps.
• Apple Silicon Optimization: Hardware-level acceleration for AI tasks.

---

1. Evolution and Collaboration with Google

Apple has transitioned from developing its foundation models entirely in-house to a collaborative approach. The third-generation AFM integrates Google’s Gemini architecture and expertise with Apple’s proprietary data. This synergy aims to combine the strengths of cloud-based processing with on-device efficiency, creating an AI ecosystem that the speaker argues is currently unmatched by Android or Windows.

2. Model Architecture and Deployment

Apple categorizes its models based on where they run and their complexity:

• On-Device Models:
  • AFM Core: A 3-billion parameter "dense" model designed for basic tasks directly on the device.
  • AFM Core Advance: A 20-billion parameter "sparse" model. While trained on 20B parameters, it only activates 3–4 billion parameters during inference to maintain performance without overheating or draining battery.
• Cloud Models (AFM Cloud):
  • AFM Cloud: Standard cloud-based model.
  • AFM Cloud Pro: High-precision model, comparable in capability to Gemini Pro, designed for complex tasks.
  • Apple Diffusion Model (ADM): A specialized cloud model for image generation.

3. Infrastructure and Privacy

• AI Infrastructure: Apple, Google, and Nvidia have collaborated to optimize cloud performance. Cloud models are optimized for Google’s TPUs (Tensor Processing Units) and Nvidia GPUs.
• Private Cloud Compute (PCC): A critical privacy feature. When a request is sent to the cloud, the data is processed and returned to the user, then immediately purged. No logs or user data are stored. Apple provides tools for third-party researchers to verify these privacy claims.

4. Developer Ecosystem and Integration

Apple is simplifying AI integration for developers through new SDKs:

• Unified Protocol: Developers can call various models (AFM, Gemini, Claude) using a standardized "Language Model Protocol."
• Cost Efficiency: Apple is currently offering free access to its cloud-based AFM for developers, lowering the barrier to entry compared to paid third-party APIs.
• Deep OS Integration: AI is being woven into the core of iOS 27, iPadOS 27, and macOS 27. Examples include natural language processing in the Calendar app and a significantly upgraded Siri that can access data across both Apple and third-party apps via Spotlight.

5. Key Arguments and Perspectives

• Superiority of Integration: The speaker argues that Apple’s approach is more cohesive than Android or Windows. Because Apple controls the hardware (Apple Silicon) and the OS, the AI implementation is more uniform and efficient.
• Privacy-First AI: Unlike many competitors, Apple’s model emphasizes that user data is not used for training and is not stored on the cloud, positioning privacy as a core product feature.
• Developer Advantage: The speaker notes that the ease of implementation and the lack of initial costs for AFM will lead to a surge in AI-integrated third-party applications.

6. Synthesis and Conclusion

The new Apple Foundation Model represents a strategic shift toward a hybrid AI model—leveraging both local hardware and secure cloud infrastructure. By partnering with Google for architecture and infrastructure while maintaining strict privacy standards through Private Cloud Compute, Apple is positioning its ecosystem to be more deeply integrated and developer-friendly than its competitors. The ultimate goal is a seamless user experience where AI acts as an invisible, privacy-conscious assistant embedded directly into the operating system.