SetupAI
EN
來源較早收集於 8m

Apple 於 WWDC 發表 Apple Intelligence 與 Siri AI

Apple 於 WWDC 發表 Apple Intelligence 與 Siri AI
PostLinkedIn
🖥️閱讀原文： Computerworld
#ios-development#on-device-ai#privacy-firstapple-intelligenceapplesiriapple-intelligence

💡Apple 轉向整合式、隱私優先的 AI API，徹底改變了 iOS 與 macOS 開發者的生態環境。

⚡ 30 秒速覽

有什麼變化

推出具備跨應用程式情境感知與任務執行能力的 Siri AI。

為什麼重要

Apple 將 LLM 深度整合至作業系統並強調隱私，為消費者 AI 樹立了新標準。開發者現在可以透過新的原生 API，明確地利用裝置端與雲端 AI 功能。

下一步行動

檢視最新的 Apple Intelligence 開發者 API，找出如何透過原生 AI 整合來增強您應用程式現有的工作流程。

誰應關注：Developers & AI Engineers

關鍵要點

  • •推出具備跨應用程式情境感知與任務執行能力的 Siri AI。
  • •發布全新開發者 API，協助將生成式 AI 功能整合至第三方應用程式。
  • •實作 Private Cloud Compute 以實現安全且逼真的影像生成。
  • •為 Apple Passwords 加入代理式 AI，以自動修復安全性問題。

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 30 個來源。

🔑 增強重點摘要

  • •The enhanced Siri AI leverages a hybrid approach, utilizing a compact 3-billion-parameter on-device model optimized for Apple silicon for many tasks, while offloading more complex reasoning and multi-step planning to Google's Gemini models running on Apple's Private Cloud Compute infrastructure.
  • •Apple's Private Cloud Compute (PCC) is built on custom Apple silicon with hardware security features like Secure Enclave and Secure Boot, running a hardened operating system to ensure stateless computation and ephemeral data processing, meaning user data is never stored and is inaccessible even to Apple.
  • •Siri AI now supports multi-step task chains, capable of executing up to 10 sequential actions from a single natural language request, and maintains conversational memory across up to 50 turns, significantly improving its ability to handle complex user interactions.
  • •Apple employs 'differential privacy' for training some of its AI features, which involves combining synthetic data with anonymized real data from opt-in users to identify patterns without linking to individual personal identifiers.
  • •Apple Intelligence is available for free to users with supported devices, specifically iPhone 15 Pro and newer models, as well as Macs and iPads equipped with Apple Silicon.
📊 競品分析▸ Show
Feature / PlatformApple IntelligenceGoogle GeminiMicrosoft CopilotSamsung Galaxy AI
Privacy ApproachPrivacy-first, device-first with Private Cloud Compute (PCC) for cloud tasks; data is ephemeral, not stored, and inaccessible to Apple.Cloud-first, multimodal, with Gemini Nano for on-device tasks.Enterprise-focused, deeply integrated with Microsoft 365 and Azure; cloud-based.On-device features for translation/summarization, orchestrates multiple agents.
MultimodalityEmphasizes text and images; integrates a vision encoder for visual capabilities.Leads with advanced multimodal support (text, image, audio, video) and long context windows.Supports multimodal inputs via cloud agents.Supports multimodal inputs via cloud agents.
Productivity/EnterpriseSystem-wide integration across iOS, iPadOS, macOS; generative writing tools, smart summaries, agentic password management.Strong in creative and research-oriented tasks; cross-app prompts.Workhorse for professional AI, embedded in Microsoft 365, Copilot Studio for custom agents.Focuses on practical, global features like real-time translation and note-taking.
Device AvailabilityLimited to iPhone 15 Pro/newer, and Apple Silicon Macs/iPads.Widely available on Pixel phones, Android, and Workspace; accessible on more phones and older models.Across Windows, Office, and Azure.On Galaxy devices.
Third-Party Model IntegrationIntegrates Google Gemini for complex reasoning; also partners with OpenAI for ChatGPT integration.N/A (primary model provider)Supports Anthropic's Claude models in addition to OpenAI's GPT-4.Orchestration model that invites multiple agents into everyday devices.

🛠️ 技術深入

  • On-device LLM Architecture: Apple utilizes a compact, approximately 3-billion-parameter Transformer-based model specifically engineered for the resource constraints of mobile hardware.
  • Optimizations for Apple Silicon: Key optimizations include Grouped-Query Attention, shared embedding tables, a smaller vocabulary (49,000 tokens), low-bit quantization (Palettization), KV-cache sharing, and 2-bit quantization-aware training to balance model size and computational efficiency.
  • Private Cloud Compute (PCC) Infrastructure: PCC nodes are custom-built server hardware running on Apple silicon, incorporating hardware security technologies like the Secure Enclave and Secure Boot.
  • PCC Operating System: A hardened subset of iOS and macOS foundations is used, tailored for Large Language Model (LLM) inference workloads with an extremely narrow attack surface, leveraging iOS security technologies such as Code Signing and sandboxing.
  • Data Handling in PCC: PCC is designed for stateless computation, meaning user data is processed ephemerally (in memory only) and deleted after fulfilling the request, with no persistent storage, profiling, or logging.
  • Verifiable Transparency: Apple allows independent security researchers to inspect the cryptographically signed software that runs on PCC servers to verify its privacy promises.
  • Anonymous Requests: PCC uses Oblivious HTTP, a protocol developed with Cloudflare, to route requests through a third party, ensuring Apple receives requests without knowing the individual user's IP address.
  • Neural Engine Evolution: First introduced in the A11 Bionic chip in 2017, Apple's Neural Engine is a dedicated AI accelerator that performs matrix multiplications efficiently. It has evolved from 0.6 TOPS (A11) to 38 TOPS (M4 chip).
  • Server LLM Architecture: The scalable server model for PCC is a Mixture-of-Experts (MoE) system built on a novel Parallel-Track Mixture-of-Experts (PT-MoE) transformer, combining track parallelism, sparse computation, and interleaved global–local attention.
  • Siri's Reasoning Backend: For complex reasoning, multi-step planning, and natural language understanding, the new Siri routes queries to Google's Gemini models, while Apple maintains control over the user interface, data routing, and privacy enforcement.

🔮 前景展望基於引用來源的 AI 分析

Apple's deep integration of AI across its ecosystem will strengthen user lock-in and differentiate it from competitors.
By embedding AI directly into the operating system and core apps, Apple creates a seamless, privacy-centric experience that third-party AI solutions cannot replicate due to their lack of full-stack control.
The partnership with Google Gemini for Siri's reasoning layer could lead to increased scrutiny regarding data flow and privacy, despite Apple's assurances.
While Apple emphasizes on-device processing and Private Cloud Compute, the reliance on a third-party model for complex tasks may raise questions about the extent of data sharing and control, particularly for privacy-conscious users and enterprises.
Apple's AI strategy, combined with leadership changes, will define its market position in the post-Tim Cook era.
The rollout of Apple Intelligence and the transition of CEO leadership to John Ternus mark a pivotal moment, with the success of these AI initiatives being crucial for Apple's future hardware, services, and developer ecosystem growth.

⏳ 時間線

2005
Siri development began at SRI International Artificial Intelligence Center.
2010-04
Apple acquired Siri, Inc.
2011-10
Siri launched as a built-in feature on iPhone 4S.
2017-09
Apple introduced the A11 Bionic chip with the first dedicated Neural Engine.
2024
Apple Intelligence framework was launched, bundling AI features.
2026-03
Apple confirmed Siri's reasoning layer would be powered by Google Gemini for complex tasks.
🖥️閱讀原文： Computerworld
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #ios-development

同產品

更多關於 apple-intelligence

同來源

最新來自 Computerworld

Apple 的 Siri AI Recaps 引發隱私疑慮

Apple 的 Siri AI Recaps 引發隱私疑慮

Computerworld•Sep 9
AI 會議筆記工具面臨日益升高的隱私訴訟風險

AI 會議筆記工具面臨日益升高的隱私訴訟風險

Computerworld•Sep 9

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： Computerworld ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。

每週電子報

每週一封，可隨時退訂。