We ask Gemini and Alexa to track cats and give advice | The Vergecast

《The VergeCast》播客，由詹妮弗·帕特森-图伊主持，嘉宾尼莱，深入探讨了智能家居语音助手的现状，特别是亚马逊的Alexa Plus和谷歌的Gemini for Home。在生成式AI承诺革新家庭语音控制一年之后，人们曾期望这些助手能更智能、功能更强大，能理解自然语言并无缝管理智能家居，但现实却是进展与显著挫折并存。 在深入主要话题之前，播客简要提到了其他科技新闻：《GTA 6》泄露事件、针对FCC大疆禁令的创意规避方法，以及Meta眼镜与公共服务人员被拍摄所引发的隐私担忧。 回到智能家居AI的讨论，尼莱分享了一个“阴谋论”：他的家人主要使用Google Home是为了其Google相册集成，而非助理本身的优点。他认为，Gemini for Home似乎只是向旧版Google Assistant发出指令。这种架构分层导致了明显的性能下降，并且在执行基本智能家居任务时，旧版助理偶尔会重新出现。旧版Google Assistant曾快速且可预测，而Gemini却常常“更糟、更慢”。尼莱指出的一个根本性矛盾在于，生成式AI的非确定性本质导致其提供多变的响应，这与智能家居对一致、精确操作的需求相冲突——“当我三次要求智能家居开灯时，我希望它三次都能做到，”他强调。他举例说，Google Home未能控制Chromecast电视，或者难以执行“打开我办公室的灯”这样的简单命令，尽管房间标记清晰。 詹妮弗，主要使用Alexa Plus，指出Alexa在AI集成到单一体验方面做得更好，而Gemini独特的“Gemini Live”聊天模式需要单独的命令才能激活。过去一年里，她看到Alexa Plus有了显著改进。然而，Alexa Plus的一个显著缺点是它不遗余力地推销亚马逊产品，将诸如寻求育儿建议等互动变成了“半遮半掩的推销”。詹妮弗表示，这“几乎让我不想再使用它了”，凸显了亚马逊的盈利模式是降低用户体验的一个因素。 两位主持人都认同*迟钝*是主要问题——与旧版非AI助理相比，这是一个显著的退步，尤其是在执行链式命令时。此外，这些AI模型会“自信地犯错”，坚持提供不正确的信息（例如世界杯比赛日期），这在无需动手操作的家庭环境中尤为令人沮丧。谷歌的“Gemini Live”聊天功能被形容为“滑稽地崩溃”，有时会自我反馈，或自问自答。尽管两位助理都能管理日历，但Alexa以更具对话性、不那么生硬的表达方式而著称，不过两者目前都无法集成工作日历，这限制了它们对许多用户的实用性。 话题转向了苹果或OpenAI等潜在竞争对手是否能提供更好的解决方案。尼莱认为挑战巨大，不仅包括AI技术实力，还要整合差异巨大的控制系统、解决专利纠纷，并实现100%的可靠性。尽管HomeKit被广泛采用，但苹果的Siri在用户识别和日历集成（如果不是iCloud）方面仍存在问题。其最新的AI测试版甚至在智能家居场景控制方面有所退步，例如发出“晚安”命令时，Siri可能会回复用户“你也要晚安”，而不是激活相应场景。OpenAI虽然在通用AI方面强大，但缺乏智能家居基础设施，也没有从零开始构建（例如Matter支持）的动力。 主持人质疑其价值主张，尤其是考虑到这两种服务现在都需要付费订阅（Alexa Plus与Prime会员绑定，Gemini for Home分级收费），而用户体验却常常有所下降。“环境智能家居”的愿景严重依赖摄像头（外部如Ring/Nest，或可能部署在室内），尽管这些摄像头在物体/人物检测方面有所改进，但引发了严重的隐私担忧。尼莱讲述了一个事件：手机上关于快速去买牛奶的提醒让他感到“不舒服的监视程度”，凸显了高度互联的智能家居系统可能带来的“监视噩梦”。 最终，尽管提及了同时设置多个计时器等微小改进，但无缝、直观的AI驱动智能家居的核心承诺仍未兑现。用户为那些更慢、更不可靠且可能具有侵入性的系统支付更多费用，这使得“价值主张”变得模糊，“监视噩梦”则日益令人担忧。主持人总结说，他们对任何一家公司能通过一次性整合“所有功能”而“做得正确”并不抱太大希望。

The VergeCast, hosted by Jennifer Patterson-Tui with guest Neelai, delves into the current state of smart home voice assistants, specifically Amazon's Alexa Plus and Google's Gemini for Home, a year after generative AI promised a revolution in in-home voice control. The expectation was smarter, more capable assistants understanding natural language and running the smart home seamlessly, but the reality has been a mixed bag of progress and significant frustrations. Before diving into the main topic, the podcast briefly touches on other tech news: GTA 6 leaks, creative workarounds for the FCC's DJI import ban, and privacy concerns regarding Meta glasses and public-facing workers being filmed. Returning to the smart home AI discussion, Neelai, whose family uses Google Home primarily for its Google Photos integration rather than the assistant's merits, shares a "conspiracy theory": Gemini for Home seems to merely issue commands to the legacy Google Assistant. This architectural layering, he argues, results in noticeably slower performance and occasional re-emergence of the old assistant for basic smart home tasks. While the old Google Assistant was fast and predictable, Gemini is often "worse and slower." A fundamental friction point identified is that generative AI, by its non-deterministic nature, provides varied responses, which clashes with the smart home's need for consistent, precise actions – "when I ask my smart home to turn on the lights three times, I want it to do that three times," Neelai emphasizes. He provides examples of Google Home failing to control a Chromecast TV or struggling with simple commands like "turn on the lights in my office" despite clear room labeling. Jennifer, whose primary assistant is Alexa Plus, notes Alexa's better integration of AI into a single experience compared to Gemini's distinct "Gemini Live" chat mode, which requires a separate command to activate. She has seen significant improvements in Alexa Plus over the past year. However, a significant drawback of Alexa Plus is its relentless attempt to sell Amazon products, turning interactions like seeking parental advice into thinly veiled sales pitches. This, Jennifer states, "has found almost enough to make me not want to use it anymore," highlighting Amazon's revenue model as a degrading factor in the user experience. Both hosts agree on the major problem of *slowness* – a significant degradation from older, non-AI assistants, especially for chained commands. Additionally, these AI models can be "confidently wrong," insisting on incorrect information (like a World Cup game date), which is particularly frustrating in a hands-free home environment. Google's "Gemini Live" chat function is described as "comically broken," sometimes feeding back to itself or asking and answering its own questions. While both assistants can manage calendars, Alexa is noted for a more conversational and less stilted delivery, though neither currently integrates work calendars, limiting their utility for many users. The conversation turns to whether potential competitors like Apple or OpenAI could offer better solutions. Neelai argues the challenge is immense, encompassing not just AI prowess but also integrating wildly disparate control systems, navigating patent disputes, and achieving 100% reliability. Apple's Siri, despite HomeKit's widespread adoption, struggles with user identification and calendar integration (if not iCloud). Its latest AI beta has even regressed in smart home scene control, where a "Good Night" command might result in Siri wishing the user "a good night too" instead of activating the scene. OpenAI, while powerful in general AI, lacks the smart home infrastructure and motivation to build it from the ground up (e.g., Matter support). The hosts question the value proposition, especially since both services now require paid subscriptions (Alexa Plus with Prime membership, Gemini for Home with tiers) for an often degraded experience. The vision of an "ambient home" relies heavily on cameras (outside, like Ring/Nest, or potentially inside), which, while improving in object/person detection, raise significant privacy concerns. Neelai recounts an incident where phone alerts about a quick milk run felt like an "uncomfortable level of surveillance," highlighting the "surveillance nightmare" that could arise from highly interconnected smart home systems. Ultimately, while minor improvements like setting multiple timers simultaneously are noted, the core promise of a seamless, intuitive AI-driven smart home remains largely unfulfilled. Users are paying more for systems that are slower, less reliable, and potentially intrusive, leaving the "value proposition" unclear and the "surveillance nightmare" a growing concern. The hosts conclude that they are not deeply optimistic about any single company getting it "right" by integrating "everything" at once.