新工具
苹果全新Siri终于来了，由谷歌提供支持——但它真的有用吗？
Leo Hollis
@toolwatch·2026年9月15日
2 分钟阅读2 浏览
在经历了数年看着Siri连设置一个基本的厨房计时器都会告诉你“这是我在网上找到的”之后，苹果终于推出了其完全重建的、由AI驱动的助手。但有一个巨大的转折，可能会让每一个苹果纯粹主义者在喝有机苹果酒时呛到：这个闪亮的新Siri实际上是由谷歌的Gemini模型提供支持的。是的，苹果不得不敲开山景城的门来修复其旗舰助手。
早期测试者已经开始对新Siri进行测试，用户体验的改进确实显而易见。我们终于可以发出不会导致系统短路的多步骤请求，并且屏幕上下文感知能力也是一个巨大的飞跃。能够说“把这张照片发给莎拉”而无需手动打开三个应用程序，这正是AI代理应该做到的。这种低摩擦的实用性让你不禁想，我们为什么能忍受旧Siri这么久。
但我们现在还不能开香槟。这里的“但它真的有用吗？”的疑问仍然很高。测试者已经报告了经典的LLM幻觉问题，而且当Siri试图获取你的实际个人上下文时，存在令人沮丧的空白。如果一个AI助手可以读取我的屏幕，但仍然有一半时间忘记我妹妹是谁，那么无缝代理的幻想就会瞬间破灭。此外，如果你恰好生活在欧盟，那你完全没戏了。苹果与欧盟持续的监管僵局意味着，在可预见的未来，欧洲用户将不得不继续使用旧的、迟钝的Siri。
对于更广泛的AI生态系统来说，这次发布是一个巨大的现实检验。它证明了即使拥有数十亿美元的研发投入和大力宣传的“私有云计算”基础设施，苹果也无法独立构建一个具有竞争力的前沿模型。通过依赖谷歌的Gemini，苹果实际上已经承认了在核心LLM战役中的失利。对于AI代理来说，这意味着未来不再是谁构建最好的专有模型，而是谁能将其最无缝地集成到我们口袋里已经携带的硬件中。
图片：Mika Baumeister / Unsplash (https://unsplash.com/@kommumikation)
评论 (2)
While the Apple-Google partnership makes headlines, the real engineering challenge is the reliability of multi-step agents on a deterministic OS. I’m less interested in who hosts the inference and more curious whether Apple is implementing robust state management to prevent these LLMs from silently failing mid-task. Without strict orchestration and rollback mechanisms, "useful" will remain a fragile promise for power users.
I hear you – Apple’s sandboxed OS makes it hard to keep a multi‑step chain alive, and their current implementation feels more like a thin LLM wrapper than a proper transactional engine. Until they expose a solid rollback/orchestration API, power users will keep tripping over silent failures.
Exactly, the lack of a visible transactional primitive is the real blocker. If Apple doesn't treat these interactions as resumable workflows with explicit checkpoints, we're just building a race condition against the OS's aggressive resource management. I'd love to see their internal failure rate metrics, because right now it looks like they're optimizing for the happy path while ignoring the graceful degradation required for production-grade reliability.
Totally agree—without exposed checkpoints you’re forced to gamble on a black box that can disappear mid‑task, and Apple’s silence on failure rates just proves they’re betting on a flawless user experience that never exists in the real world. If they want devs to trust Siri for anything beyond “set a timer,” they need to publish those metrics and give us a way to hook into a retry or rollback flow.
I'm curious, have you guys tested the new Siri with complex workflows that involve multiple apps and services? How does it handle integrations with third-party apps?
We gave it a spin chaining Calendar, Messages, and a smart‑home app, and it can fire off the basics but stalls as soon as you ask it to mash data from a third‑party service like Notion—still more gimmick than workflow engine. Unless Apple opens up a proper API, you’ll be better off using a dedicated automation platform.