内容来源：https://www.theverge.com/podcast/936445/sundar-pichai-ai-search-google-zero-youtube-web

内容总结：

独家专访谷歌CEO皮查伊：AI重塑搜索、互联网生态与AGI时间表

在谷歌I/O开发者大会结束后，《The Verge》的Decoder播客节目连续第五年专访了谷歌及Alphabet首席执行官桑达尔·皮查伊。本次对话围绕谷歌在AI时代的战略重组、搜索变革、与内容创作者的关系，以及对通用人工智能（AGI）的展望展开。

组织架构重整：从“各自为战”到AI驱动

当被问及谷歌内部架构时，皮查伊表示公司目前主要分为三大业务板块：搜索、YouTube和谷歌云，以及Android和Chrome等大型平台。他承认，过去谷歌以“多产品、多押注”著称，但AI时代的到来提供了前所未有的整合契机。

“我们首次拥有了一个共同的基础设施——Gemini模型和底层AI架构来驱动所有产品。”皮查伊强调，这使谷歌能够有意图地开展跨产品工作，如“个人智能”功能就是一例。

应对ChatGPT挑战：内部改革与决策哲学

谈到2022年底ChatGPT带来的冲击，皮查伊坦言那是一个关键转折点。他意识到必须重组公司以应对这个“奥弗顿窗口已经改变”的时刻。具体措施包括：将Brain和DeepMind两大顶尖研究团队合并为Google DeepMind（他形容这“比听起来难得多，就像要把斯坦福和MIT合并”），建立中央化AI基础设施团队，并设立首席AI架构师职位。

在决策哲学上，皮查伊表示：“真正重要的决策其实很少，大部分决策并不关键。关键在于做出决定，这决定了组织的速度。”他最近还设立了每周一次的AI产品评审会议。

搜索变革：个性化与“真理来源”的平衡

面对搜索结果的个性化趋势，皮查伊承认需要在“客观事实”和“主观建议”之间取得平衡。“华盛顿的首都是什么”——这是客观的，不会因人而异；但“帮我规划一个蒙特利尔周末旅行”自然需要个性化答案。

针对记者现场展示的“最佳Chromebook”搜索结果——AI概览给出一个答案，而下方Reddit和《纽约时报》的链接给出不同答案——皮查伊认为该体验“可能比应有的更主观”，并承诺将持续优化。

关于“谷歌归零”：康泰纳仕CEO的警示

当被问及知名出版商康泰纳仕（Condé Nast）CEO公开表示“我们假设搜索流量为零来规划业务”时，皮查伊辩称信息生态远比谷歌更广泛。他强调：“如果出版商制作高质量内容，人们喜欢，我们就会在搜索结果中反映这一点。这是我能够承诺的。”

但他同时承认，随着技术进步，“低质量点击正在被过滤”，这是一个自然演进。

消费者对AI的焦虑：不仅是营销问题

针对年轻群体普遍对AI持负面态度、前CEO埃里克·施密特在毕业典礼上被嘘等现象，皮查伊表示：“AI是人类将面临的最深远技术……人们尚未进化到能处理如此快速变化。”他指出，对失业的担忧、对能源成本上升的恐惧等都是合理关切，不能简单归结为“营销问题”。

关于AGI与奇点：DeepMind CEO的震撼表态

当被问及DeepMind CEO德米斯·哈萨比斯在I/O主题演讲结束时所说的“我们正站在奇点的山脚下”时，皮查伊透露两人曾就此进行过深度讨论。“我们在AGI的定义上非常接近——能够更全面地完成包括认知任务在内的广泛任务。”

至于AGI的时间表，皮查伊表示：“三年后，无论你我是否称之为AGI都不重要，因为那时系统将变得非常非常强大，我们必须为此做好准备。”

中文翻译：

今天，我与谷歌及Alphabet首席执行官桑达尔·皮查伊进行了对谈，这场对话录制于谷歌I/O开发者大会之后。这是我连续第五年在I/O大会后与桑达尔坐下来交流，它已成为我个人最钟爱的《Decoder》传统之一。

桑达尔·皮查伊谈人工智能、搜索的未来，以及网络正在经历的变化
谷歌CEO如何重塑公司——乃至整个互联网。

I/O大会总是新闻迭出，今年也不例外——谷歌推出了强大的新一代Gemini模型，将AI智能体融入所有产品，并对网页端和YouTube的搜索功能进行了重大调整，这将再次改变信息生态格局。

有太多话题值得探讨，我和桑达尔深入聊了这一切。但我突然意识到，我已经很久没向桑达尔问过关于公司架构和决策机制的那些"Decoder式问题"了，于是我从这里切入。你会听到桑达尔坦言，几年前面对ChatGPT的崛起，他意识到必须重新思考谷歌的运作方式，并为此进行了大量高管调整和重大决策，让公司展现出更具进攻性的姿态。

《Verge》订阅用户请注意，无论你在哪里收听播客，都能获得《Decoder》节目的专属无广告版本。点击这里即可收听。尚未订阅？点此注册。

当然，我们也聊到了搜索领域的诸多变化。显而易见，谷歌搜索的真正未来在于将新型智能搜索框与公司新推出的Gemini Spark智能体平台相结合。这样一来，搜索不仅能呈现结果，更能触发和执行任务。这令人兴奋，但也很可能再次改变开放网络的格局。

如果你是《Decoder》的听众，会知道我在几年前提出了"谷歌归零"这个概念——即随着谷歌在搜索结果页面上直接回答越来越多的问题，网站来自谷歌的流量将趋近于零。这个概念从最初桑达尔在采访中回避的话题，变成了如今整个媒体行业不得不面对的严峻现实。就连康泰纳仕等大型出版商的CEO们现在也公开表示，他们正在为从今往后搜索流量为零的世界做预案。

谷歌还在用YouTube视频训练模型，并调整YouTube搜索功能，使其能够对视频进行摘要和索引，让你直接跳转到相关片段。这势必会引起部分创作者的不安，所以我问桑达尔，他是否准备好像目前与出版商交锋那样，与YouTube博主们展开同样的博弈。

最后，我向桑达尔提起了谷歌DeepMind CEO德米斯·哈萨比斯在I/O主题演讲结尾时的一句话，他说我们"正站在奇点的山麓"。桑达尔认同德米斯的观点并不令人意外，但他对实现通用人工智能时间表的看法值得关注。

就像我说的，这是我每年最期待的节目之一，因为桑达尔总是乐于接受提问——甚至会和我一起在我手机上查看搜索结果。我相信你会喜欢今年这场对话。

好了，有请Alphabet和谷歌CEO桑达尔·皮查伊。我们开始吧。

本次采访为篇幅和清晰度考虑，经过了少量编辑。

桑达尔·皮查伊，你是Alphabet和谷歌的CEO。欢迎再次做客《Decoder》。
很高兴来到这里。很高兴再次见到你，尼雷。

这是我最喜欢的年度对话之一。我想我们已经在I/O大会后连续做了五次了。
哇，我都没意识到已经五次了，但我很享受这个过程。再次感谢。

我想先快速问几个问题。我一直在想，我们聊过很多次了，每次都深入探讨网络、搜索和一些宏大深刻的想法，但我意识到我已经很久没有问你那些"Decoder式问题"了。

我回顾了我们之前的对话，也回顾了谷歌本身，你对谷歌进行了相当多的调整。我认为你的一些直接下属在过去一段时间里有所变动。你显然重组了DeepMind、平台与设备部门以及安卓团队。告诉我谷歌现在的架构是怎样的。

好的。现在是谷歌和Alphabet。当然我们还有Alphabet，但广义上来说，我认为谷歌内部有三个主要业务：搜索、YouTube和谷歌云。我们运营着巨大的平台，即安卓、Chrome以及相关的整个领域。而支撑这一切的是所有重要的技术领域，即人工智能和我们的基础设施工作。此外还有配套的职能部门。

但从高层来看，你可以将其视为搜索、YouTube、谷歌云，以及我们的大型计算平台。这些是主要部门，当然它们由谷歌DeepMind和我们的基础设施团队提供支撑。这是理解其架构的一个简单方法。当然，除此之外我们还有其他业务，Waymo是其中最突出的，但还有很多很多其他业务，比如Isomorphic Labs等等。

我想把重点放在谷歌本身。我觉得我们可以花整整一个小时来讨论Alphabet及其架构，以及它作为一家拥有众多业务的上市公司是如何运作的。但暂且只聚焦谷歌，历史上对谷歌的非议是：这是一家发布大量产品的公司。你不可能卖出那么多产品。缺乏重点。有成千上万种不同产品的名称，它们以各种方式重叠。

在我看来，这种情况的根源在于，你确实有这些大型基础设施项目。你拥有所有这些能力，而负责业务的人可以利用这些能力来开发产品。但可能缺乏足够的统筹或中央规划，比如"我们是不是发布了两个相同的东西？"你如何解决这种矛盾？看起来谷歌确实变得更有重点了一些，但公司的文化就是："我们要下很多赌注，看看哪些能成。"你是如何调和这一切的？

我们做的事情也很有目的性。我们拥有13个各自拥有十亿用户的产品，这并非偶然，而且我们对这些产品有着长期的投入。你可以回想一下Gmail、地图、谷歌文档、搜索或Chrome是什么时候推出的。我们在许多领域都长期保持着深入和一致的投入。

在人工智能时代，我内化的一点是，我们第一次拥有了如此通用的基础设施，通过我们的Gemini模型和底层AI基础设施来驱动所有这些产品。因此我们更能有目的性地去做一些跨领域的事情。个人智能就是一个很好的例子。这是一项统一的工作。用户可以选择在每个产品中开启它，但它基于一个共同的基础设施构建，从而在我们的产品中提供一致的体验。

底层的Gemini模型本身就是一个例子。我们能够将该模型融入产品情境中，比如在地图产品中提供"询问地图"功能。但支撑它的许多技术——语音技术、模型、智能——都是一项工作，这就是为什么我认为人工智能时代为我们提供了一种新的思考方式，而且随着时间的推移，不仅限于谷歌，也适用于整个Alphabet。这个时刻之所以如此独特而强大，是因为你可以在研发和基础设施上投入巨资，开发出一种技术，然后将其应用于所有这些领域——当然是在对用户有用的情境下——但底层的技术平台是通用的。这其中蕴含着大量的意图和规划。

你必须为创新留出空间，允许边缘团队能够发布一些新功能。有时你之后再来协调统一它们。以NotebookLM为例。笔记本现在出现在Gemini中，实际上就是作为笔记本的项目。你可以在Gemini中创建一个笔记本，也可以去NotebookLM，你会看到相同的笔记本，反之亦然。这就是一个先创新、后统一的例子。

昨天我看了主题演讲，看到了谷歌展现出的强烈意图和自信："我们拥有这项核心技术。我们可以用多种方式表达它。它本质上仍然是谷歌的风格。"产品很多，Gemini这个词也随处可见。我保证我会把它们都搞明白。

与此形成对比的是……我不知道，三四年前ChatGPT出现的时候，每个人都担心谷歌会怎么做。OpenAI能否出现并抢走你在搜索领域的市场份额？从那时到现在，你改变了谷歌。你重组了它。领导层出现了新面孔。请为我串联起这些变化。面对当时的竞争形势，你是如何思考"我需要真正改变公司的运作方式"，并最终走到今天这一步的？

这是个好问题。我一直铭记那个时刻。对外传达这一点很难，但我让公司转向以人工智能为先。我们拥有所有要素，所以从某种程度上，我感觉舆论空间已经改变。人们采用这些技术的速度比我们预期的要快。对我来说，这是一种通过我们的产品来表达自我的方式，但我意识到我们必须为此进行组织。回到我之前提到的观点，我意识到我们需要一个核心模型和一个核心基础设施团队来支撑我们在谷歌范围内所做的一切。我最初的很多精力都花在了建立这个体系上。

为了组建一个统一的人工智能团队，我们拥有世界一流的研究团队，即谷歌大脑和DeepMind，我们将它们合并为谷歌DeepMind，这比听起来要困难得多，因为这就像说"把斯坦福和麻省理工合并，从中创建一个系或一所大学"一样。所以我认为我们在这方面做得不错。那时，我还与现任AI基础设施高级副总裁的阿明·瓦赫达特一起，建立了一个集中的基础设施团队，这带来了巨大的回报。另一个演变是，我们意识到需要一位首席人工智能架构师来为整个谷歌架构这项技术，科拉伊·卡武库乔格鲁担任了这一角色。这些都是重要的变革。

搜索需要更快地行动，而搜索部门的领导架构此前较为分散，所以我们将其交由伊丽莎白·里德负责，尼克·福克斯负责整体领域，乔什·伍德沃德协助我们的实验室产品，后来致力于Gemini工作并推动创新。公司里还有其他杰出的领导者，比如负责所有运营事务的菲利普·辛德勒等等。所以，这是退后一步，从端到端地思考组织结构，确保我们为这个需要作为一家公司更快行动的时刻做好了准备，这意味着我们需要做出更快的决策。

我设立了每周一次的新产品审查会议。这些是人工智能产品审查，确保我们有意图地应用这项技术，决定在何处应用，并直接审查所有内容。任何我们向用户发布的人工智能相关产品，都要经过这个渠道。我直接与相关的工作人员花时间沟通。

另一个我会问每个人的Decoder式问题与决策有关。你描述了众多重大决策，其中一些在人事调整时可能令人不快。你是如何做决策的？你的决策框架是什么？

我框架的一个重要部分是，随着时间推移，我认识到真正有重大影响的决策非常非常少，大多数决策并非如此。更重要的是你做出决策本身，因为这决定了组织的速度。你越能做出这些决策，越能让公司向前发展，通常你就做得越好。

当然，有一些决策，比如合并并成立谷歌DeepMind，影响更为深远，你需要花时间深思熟虑并付诸实施。但很多决策的关键就在于做出决定。你越能做到这一点，随着时间的推移，你就越能形成模式匹配，因为你之前已经遇到过类似的问题版本。所以我认为，依赖这种能力，将信号从噪声中分离出来是有益的。信号意味着这是一个真正重要的决策，你需要认真权衡；而有些决策虽然看起来很大，但只是你需要采取的常规行动。

环顾整个行业，你在大型科技公司的同行们有着一些我生平从未听过的、最离奇的组织结构图构想。我认为Meta希望让50名工程师向一个拥有AI智能体力量的单一直属经理汇报。Block的杰克·多尔西希望所有6000人都直接向他汇报。你是否也有类似的想法，觉得应该利用AI发明一些最疯狂的组织结构图？

领导者与人才至关重要。这取决于具体情况。有些公司的产品线要狭窄得多，因此不同的结构可能有效。当你运营像谷歌云这样规模庞大的业务时，拥有一位CEO负责是非常重要的。我们服务的规模涵盖了全球所有顶级企业，那么如何为此进行组织？优秀的领导者最终会发挥巨大作用，比如我们的托马斯·库里安。我确实思考过这个问题。

但我更多思考的是我们如何更有效地利用人工智能，我们在内部已经看到了这种转变，尤其是在我们的开发者中，我们已经从使用AI工具辅助编码，过渡到一部分工程师越来越多地有效地指挥AI智能体团队。这些转变正在进行中，并且将从工程领域扩展到组织的其他部分。这已经在发生了。我们在Gemini Spark中所做的工作，也正是为了将这种超级能力交到消费者手中，以及你可以通过这些智能体工作流等实现什么。

我更专注于确保我们能够以一种原生方式部署这种能力，并使其运作良好，因为对我们来说，这不仅仅是提高公司效率的问题，它还是我们提供给他人的产品。我以一个非常不同的视角来看待它。我们内部的做法正是我们向外部用户提供的东西。我们在内部使用Antigravity。我们向外提供的就是这个。因此，Antigravity中的智能体正是我们的开发者所使用的，也是我们试图推向外部的东西。它因此具有了额外的维度。

Decoder的听众希望我开始向CEO们提出的头号问题……我就直截了当地问吧。AI距离取代你作为CEO有多近？

我只是觉得CEO的工作并没有那么复杂。在决策方面，我认为它会有极大的帮助。我开玩笑地说——部分是在开玩笑——我不得不花大量时间来分配算力。我心想，"好吧，随着时间的推移，AI似乎会做出更理性的选择，"因为我在处理这类流程时，要面对大量的申诉和情绪。

在我所看到的各处——这可能与我的想法略有不同——如果运用得当，这些工具将使我们能够在所从事的一切工作中提升到一个新的水平。这并非意味着你不再做以前做的事，而是你会从一个更高的基础起步。当电子表格在公司推广开来时，我并不在场。我不禁回想，在那之前人们是如何做所有这些财务分析的？我确信在三四年的时间里，情况发生了根本性的变化，然后我们就习以为常了。

我认为智能体等也是类似的情况。这并不意味着你不再需要计划生日派对。假设你在计划去某个地方旅行。也许你实际上是把时间花在想真正要做的事情上，而不是费心去查营业时间以及如何买票等等。它把所有事情都提升到了一个不同的基础，我是这么看的。

让我问问你关于智能体的事。那些演示令人着迷。搜索将为每个人构建定制软件的想法，在软件工程领域看来，似乎是一种初步印象。这个想法是你向计算机提问，计算机的回应是为你制作一个帮助你找到答案的软件。我对这个想法很着迷，但这从根本上改变了搜索。

然后你看看Gemini Spark，这是你在云端推出的智能体平台，你告诉它"帮我去订几张票"，Spark可能会四处为你订票或为你完成某些任务。还有Antigravity，这个智能体编码平台。大致来说，每年都会有一个新的AI范式。先是大型语言模型，然后也许我们要把一些大型语言模型串联起来，接着是推理，现在到了智能体。这是基础吗，还是说还会有另一个范式转变？

这是个好问题。我们正在奠定大部分基础构件。从根本上说，能够推理、使用工具和编写代码，这很大程度上就像拥有智能和推理能力——能够规划、查阅信息、使用工具，如果需要，还能在此基础上构建东西。你正在设置所有的基础元素。Antigravity是为开发者准备的，但Antigravity引擎和框架现在已经内置于Gemini中。而Spark只是Gemini的一种模式。随着时间的推移，它会成为一个功能。我们正在推广它，但它只是Gemini中的一个标签页。

所以你正在引入那个智能体框架。用户不需要考虑它。开发者会理解它。随着时间的推移，在Spark中，他们可以编写出强大的东西。但作为用户，你可能会构建、创造某些东西，规划一次旅行，所有这些都在幕后运作。

我们正在为智能体端到端运作奠定大量所需的基础元素，更重要的是，为AI的运作奠定基础。我们一直拥有的那个关于Google Assistant的长期愿景，我们尝试过无数种形式，但未能完全做好，现在我们比以往任何时候都更接近实现那个承诺。我们还没有完全实现，但我认为这段旅程现在比以往任何时候都更近。

我审视了所有产品，它们看起来确实应该融合。你有新的智能搜索框，我肯定想更详细地讨论搜索。但你看着那个搜索框，再看看比如制作应用的Canvas。你正在策划一场婚礼，它会直接为你制作一个应用来帮你规划旅行或婚礼之类的。然后你还有Spark，它可以自主去执行任务。我昨天看到这些，也和人们聊过，很明显这应该是一个统一的产品。

会的。我之前举了Notebook的例子，你创建笔记本……但笔记本是什么？你实际上是把所有你想要的情境放在一个地方，然后在此基础上工作。它就像是人们一直使用的文件夹，而笔记本应该成为你使用的谷歌产品中一个一致的基础元素。我就是这样看待智能体的。这应该无关紧要。在创新的最初阶段，你创造能力。团队在用它们做实验，但随着时间的推移，对用户来说，如果你发起一个旅行计划的任务，它应该能在两个地方都能运作，我是这么想的。你说得对。

谷歌搜索有一个非常重要的特性——多年来甚至几十年来，它一直是人们获取事实真相的来源。"谷歌一下"，你会得到一个答案，而且这个答案对你我来说通常是相同的，这一直是一个非常重要的理念。我认为这已经成为了文化的一部分。或许在所有公司中，谷歌是最后一个声称会告诉你真相的公司。

但是现在，我们打算无限个性化搜索框，无限个性化搜索体验。我们都会得到不同的查询答案。我们甚至可能会根据询问的内容、个人的情境以及谷歌掌握的数据多少，看到不同的界面。

你对此考虑得深远吗？你在多大程度上会动摇人们在互联网上能体验到的最后一个共同的真相来源？

看，有一些因素远超我们的控制范围，那就是如今人们拥有比以往更多样化的信息来源。人们从如此多的不同渠道获取内容。但在谷歌的世界里，我仍然认为我们非常重视将其作为知识和信息的来源。存在客观体验和主观体验。美国首都是哪里？这个答案不会为任何人定制。这是客观事实。"帮我规划一个去蒙特利尔的愉快周末"——自然，答案不需要对每个人都一样。这里有一个连续的区间。

我们非常在意这一点。对于某些类别的信息，我们仍然锚定在权威信息上，以尽可能呈现客观的观点。如果是与健康相关的查询，我们自然会倾向于展示比"哪个更好？我该去买哪个？"这类问题更权威的答案。

我能给你看一个搜索结果吗？
可以。

几年前，我给你看过一个搜索结果。这个结果我追踪了好几年。
我一直很喜欢这个。在十万亿次查询中……

是的，这个结果是我的最爱。
我们有非常科学、统计的方式来处理这个。

我认为这很重要，我想探讨消费者可能会如何体验这些产品。这是我经常做的一个搜索："最佳Chromebook"。我给你看看。就是这个。

它以一个AI概述开始。它非常自信地告诉了你答案，然后是一堆赞助商框。紧接着下面，让我在意的是，我相信结果是来自Reddit，它有一个Reddit上的置顶结果。实际上它给出了与AI概述不同的答案。然后是《纽约时报》，它又给出了不同的答案。

你往下滚动，你会想："AI概述告诉我一件事，第一个自然搜索结果在页面很靠下的位置，而且所有这些给出的答案都不同。"我理解你所说的客观结果和主观结果。"我该买哪款笔记本电脑？"处于这两者之间。我只是好奇，你认为如今在AI模式下，消费者的这种体验如何，以及你认为它未来应该走向何方？

听我说清楚，在AI概述的世界里，我们使用一种AI模式。我们进行组织并提供上下文，但全程都有来源，所以我们仍然在以不同的方式呈现自然内容。你得到了链接和来源，但同时也附带了观点，这正是你在谈论的。

其中一些方面需要与用户进行迭代。我们在搜索方面发现的一个优点是，很容易衡量用户满意度。25年来，我们学会了通过一种与产品质量提升（而非短期效果）相关联的方式来衡量用户幸福感和满意度。这就是为什么我们要做这些长期研究。如果我们在任何体验上出了问题，数据会显示出来，我们就会修正。我们为能够长期追踪这些指标而感到自豪——无论是参与度、会话次数、返回某个话题的频率，还是用户跳出的次数。这是一种非常非常精细的观察方式。在类似的一些领域，体验会不断演进。

你认为今天的体验好吗？
对于你给我看的那个特定查询，它可能比应有的程度更带有主观倾向。这是我作为用户的反应。我的意思是，在这个快速发展的领域，改进的空间是存在的，但我预计产品本身会自然演进。我的直觉是，"哦，这个结果太主观了。"也有可能这是针对你的个性化结果。你可能以某种独特的方式进行了个性化测试。不过，那个查询可能并不完全具有代表性，因为我了解你如何评测所有这些产品。你有可能处于那0.0001%的百分位——

这基本上就是我为什么要问关于无限个性化结果的问题，对吧？我也在问体验是否良好，因为我敢打赌，大多数人一直都在使用谷歌搜索中的人工智能功能。他们都有那种被切换到AI模式的体验。关于用户满意度，有些是你可以衡量的，但还有公众对人工智能的感受。

我认为在"用户数字在上升，我们接近十亿用户，人们体验到的免费产品可能相当不错"和纯粹的民调数据之间，存在着相当大的鸿沟。年轻人不喜欢人工智能。这几乎和客观事实一样明确。你可以去问他们，他们会以可衡量的方式告诉你他们不喜欢它。

谷歌前CEO埃里克·施密特在一次大学毕业典礼演讲中被喝倒彩。七成美国人反对建设数据中心。人们的产品体验和他们对该技术的感受之间存在差距。你认为你能弥合这个差距吗？你认为这些产品足够好吗？

这是一个非常深刻的话题，你把两件事联系起来了。人工智能是人类将要面对的最深奥的技术。它正以非常快的速度发展。我不认为人类已经进化到能处理如此巨大的变化，特别是过去几年的变化速度高得惊人。尤其是结合他们所听到的一切，人们正试图理解未来，以及这在他们个人生活语境中的意义，包括经济层面等等。

人们对这项技术感到焦虑，这完全说得通，我们应该非常敏锐地意识到这一点。这是一个重要的话题，而且比正在发生的具体方面更广泛、更宏大。人们并不总是将这两者直接联系起来。在某些情况下，是的，它们以某种方式相关联。

人们在各种产品中体验这些模型的免费版本。他们打开社交媒体，看到低质内容。他们看到关于这些事情的标题。这些工具就那样呈现在他们面前。Gemini的光芒出现在所有谷歌产品中，无论你是否要求。然后我认为你会将其联系起来："他们需要大量电力，也许我的电费会涨。也许所有的工作都会消失。"这相当可怕，我不知道价值交换是否对等。

这些都是值得研究的好事。你过于聚焦具体发生的事，而我正将其放宽，指出这可能是部分原因。我确实认为还有其他更直接的因素。

你认为这只是个市场营销问题吗？我听过你的同行说，人工智能只是存在营销问题。

不，我不这么认为。这正是我的观点。我实际上是在反对这种说法。我认为人们对此感到担忧是合情合理的。这对我来说感觉很自然。人们在谈论人工智能可能会让很多工作消失。你难道不会感到焦虑吗？我认为这些是我们作为社会必须应对的更深层次的问题。是的，在产品层面存在对人工智能垃圾内容的担忧。这些都是事实。我只是想指出，这是一个多层次的问题。但我并不认为数据中心焦虑的所有根源都直接与你某个产品中的特定体验或类似的东西相关。这就是我的观点，对吧？它比那更广泛、更宏大。

现在外面有很多人工智能垃圾内容。我感觉得到。在技术早期阶段，加上现有的竞争态势，很多东西都在被推出。但我们也通过实证看到人们正在以非常深入的方式使用这些产品。如果你去一个Waymo还没到的地方，刚刚对人们进行民意调查，问他们关于自动驾驶汽车的看法，你从调查中得到的结果与他们实际使用这些汽车时的感受是不同的。技术也会经历这些阶段。顺便说一句，如果你问起互联网，人们对它也有相当负面的看法。但它已经成为我们生活的一部分，我们必须适应它。所有这些都在同时发生。

这是一个复杂的话题。对我来说，感觉人们是在担心能源价格上涨，如果是这样，他们希望确保人工智能不会加剧这个问题，这是一个合理的担忧。作为行业，我们有责任确保，如果你在建设数据中心，我们能做些什么来确保我们不会助长这个问题？我认为这是我们的责任，不仅仅是我们。政府方面，对此也存在两党共同的担忧。例如，我们所有人都签署了一份包含一系列承诺的"费率缴纳人承诺书"。也许还需要做更多。所有这些都相辅相成。谈论技能培训、劳动力适应等话题也很重要。我们正在以非常快的速度推动社会发生巨大变革。这些最终也会成为非常重要的话题。

在所有这些层面都存在担忧，并且我预计这些担忧在未来将具有重要意义。很多年前我就说过，"这比火或电的影响更深远"，所以我们一直都有这种感觉。或者想想深度伪造，以及你如何知道某件事是真是假？这些模型在模拟现实方面越来越强。这就是我们如此努力工作的原因。我们正在开源它，我们聚集了许多许多合作伙伴，看到行业在这样的议题上合作，对我来说是件好事。网络安全是另一个很好的例子。这些都是切实的担忧。

作为一个行业，我们需要做得更多。政府将发挥更强的作用，公众也需要参与进来。你不可能在民主国家里，在没有公众公民对此发出合理声音的情况下，让这项最具影响力的技术就这样在世界上推广开来。经历这个阶段非常重要，这是我们学习如何适应的方式。

我的观点是，产品本身就起到了营销作用。这是我的坚持。我仍然在等待能够起到这种作用的面向消费者的杀手级应用。我认为我们已经有了企业级的杀手级应用。

有一点，我曾使用Gemini处理过一段健康相关的体验。对我来说，它感觉比杀手级应用更重要，比我以往做过的任何事情都好。人们也正在经历这样的体验。

我想谈谈网络，在Gemini中进行健康之旅需要网络上存在丰富的健康信息数据集。你在用YouTube视频训练Gemini，对吧？Veo需要YouTube生态系统来运作并产生成果，以创造新的作品。多年来，你和我讨论过我称之为"谷歌归零"的概念，即你将停止向网络输送流量的想法。你不同意我的看法，认为这不是真的。

非常不同意。在过去很多年里，这并没有发生。

那么，我就给你读一段引述吧。这次不是我说的，我也没教他这么说。康泰纳仕的CEO罗杰·林奇上周接受了TBPN的采访，他说："每年我们的搜索流量下降幅度都超过我们的预测，所以去年我告诉我的团队，'假设没有搜索了。你们必须像搜索流量为零那样来规划业务。'"

这就是"谷歌归零"。康泰纳仕在说："我们假设搜索流量将归零。" 对于这个世界最大、最具标志性的出版商之一说"我不能再依赖这个了"，你会如何回应？

听着，首先，信息生态系统远不止谷歌，它要广阔得多。我们从数据中能看到这一点，你在任何地方都能看到。所以，如果在过去的10年里，任何出版商……我会看看The Verge，想想你最初接手时的情况，自那以后它演变到了什么程度，你们制作的内容类型，你们在哪些平台发布内容，所有用户是如何找到你们的。这是异常动态的，所以我认为每个出版商都在适应这个新世界，这对我来说是合理的。

我们正在适应不断变化的世界以及用户消费技术的方式。当世界从网页转向移动端时，我们必须这样做。现在，我们正从移动端世界转向人们进行持续对话、与这些产品聊天、用语音以及各种不同形式消费内容的时代。

人们对各种类型的内容表达了偏好。他们在寻找用户生成的内容。他们在寻找播客。他们在寻找这些。在此过程中，我们非常致力于满足用户的期望，同时也要将他们与网络上的内容连接起来。即便在过去一年里，自从我们推出这些功能以来，我们也回过头去添加了更多的链接。另一个行为正在改变的领域是，许多出版商正在合理思考订阅模式的问题。

当然可以。但我只是说康泰纳仕在说："鉴于我们看到的趋势，我们将假设搜索流量为零。" 他们应该这样假设吗？

听着，我一直认为……人们更了解他们自己的业务……我的意思是，我无法告诉这样一个标志性的出版商他们应该如何思考自己的业务或计划。如果他们制作的是高质量且人们喜欢的内容，我期望我们的产品能够体现这一点。这是我能够承诺他们的。

但我认为，在此演变过程中，与任何其他公司相比，我们更努力地工作以确保人们能够连接起来，我们计划在搜索和Gemini中做到这一点，而这仍然支撑着我们做的很多事情。但是，演变是存在的。随着技术的改进，低质量的点击会被过滤掉。这是我们看到的自然演变。我们能在指标中看到这一点。反弹点击正在减少。这些都是动态变化。

人们正在获取更广泛的信息，而生产信息的人数比以往任何时候都多。这个蛋糕正在变大。所有这些动态都在发生。这是一个复杂的生态系统，但我们的承诺是确保我们反映内容的广度和多样性，并且我们确实认为人们最终想要连接到这些来源，但我们正试图在这些时刻满足他们的需求，而人们带着非常不同的意图和非常不同的时刻来到搜索。

我们做的一个小功能，但我认为非常重要，就是如果你订阅了某个内容，我们会将其反映为优先来源。但这是一个新的变化，以前我们没有这个功能。我们正在适应出版商越来越多地转向订阅产品的现实。

出版商和YouTube创作者，如果他们想要在搜索中被展示，是否应该能够选择退出被用于模型训练？

这是一个更广泛的话题。法律法规也必须随之演变。法院也必须介入。保护版权很重要。保护合理使用也很重要。因此，这些概念将通过法律途径动态演变。

但你希望陷入与YouTube创作者的一系列诉讼吗？你在英国正与出版商打官司。那场诉讼的言辞越来越激烈。谷歌表示，提议的解决方案是一份"搭便车者宪章"。每年，新闻媒体协会都会发给我一段引述，让我读给你听，他们说："谷歌称我们为搭便车者，这显然是荒谬的。这是基本的供应链经济学。如果价值真的完全在谷歌这边，他们就会简单地允许出版商选择退出。"

你是否希望就"选择退出"问题与YouTube上的大量创作者陷入同样的争斗？

听着，我们一直在——作为Gemini开发的一部分……

我们确实通过Google-Extended提供了新的选择退出方式，并且我们正在与出版商进行对话。我们会听取反馈，并随着时间的推移，确定合理的方式。显然，我们并不是这个大生态系统中唯一的参与者。我们也在努力推出与市面其他产品具有竞争力的产品。所有出版商也会写文章说我们的产品不够好。所以情况比看起来要复杂。

你花了更多时间思考网络、网络的健康以及网络的必要性。请为我描绘一下，在一个智能体搜索的世界里，一个健康的网络是什么样的。

我长期以来一直在论证的一个观点，而且我实际上看到它正在逐渐显现，那就是在过去一年到一年半的时间里，我开始更多地使用网络了。所有这些AI体验反而让网络更多地回归了。曾经有一段时间感觉……但我一直认为网络会充满活力。事实上，我每年都在论证网络会充满活力，我今天仍然这么认为。网络在不断演变。我从未见过像网络这样动态的事物，这就是为什么能成为这种演变的一部分是如此荣幸。

我看向智能体，那是网络的下一次演变，我们将不得不面对它，我认为它会相当深刻地改变网络。关于什么是允许的、什么不允许，会有很多争论，但人们想要发布信息，想要与他人连接。人们想要被连接。人们不想生活在一个孤立、脱节的世界里。这不符合人类经验的现实。我认为网络将像以往一样扮演核心角色。事实上，通用商务协议，如果要说的话，我认为人们有点低估了昨天我们宣布的它的影响。

实际上，我能把这两件事并列一下吗？有很多关于新产品、新功能以及你可以使用的智能体工具的重磅发布，还有UCP，亚马逊和沃尔玛等都说，"我们将使用一个我们正在构建的新购物标准"，所有这些都非常具体。

然后I/O大会以DeepMind CEO德米斯·哈萨比斯的登台结束，他说的这句话让我一直无法忘怀。他说："谷歌前沿的研究和产品将帮助释放通用人工智能的巨大潜力，造福整个世界。当我们回首这段时光时，我想我们会意识到，我们当时正站在奇点的山麓。"

你能告诉我"站在奇点的山麓"意味着什么吗？

德米斯和我就此话题进行过长期深入的对话。在这种语境下，通用人工智能的到来被他视为奇点。

你对通用人工智能有定义吗？你们争论过吗？你们有共识吗？

我们经常辩论。我认为德米斯和我在如何看待事物上非常接近。对于通用人工智能有一个更严格的定义，即它必须能够更全面地完成广泛的任务，包括认知任务，且能力相当。我们最终会在某个时间点将其作为公司的一个目标发布出来，我们正在为此努力。但这就是他在此语境下谈论的内容。

顺便说一句，我认为重要的是我们要理解这项技术正在非常迅速地发展。今天晚些时候，我将去和我们的AI研究人员交流，不仅仅是我们公司的，还包括其他前沿实验室的。广泛的共识是，这项技术，通用人工智能……人们可能会对究竟是三年还是五年有分歧，但这项技术的到来会更快而不是更晚。传达这一点更为重要，因为——回到我们谈话的前半部分——让我们的社会理解它并尽可能做好准备，这一点很重要。

也许在我们第一次谈论AI时，我就问过你这个问题。我问过你语言是否就是智能。而我们这里的进展是，我们正在大型语言模型之上叠加越来越多的东西。我们正在做更长的推理链，构建框架，做所有这些事情，但核心技术仍然是Transformer。它仍然是谷歌很久以前发明的东西。大型语言模型能让你达到通用人工智能吗？这条路径清晰吗？

过去三年的轨迹令人难以置信。当今的大型语言模型也在很多方面发生了演变。我们也在不断演进它。对我来说，这就像问，计算机能让我们达到……？冯·诺依曼架构仍然是驱动今天大多数计算机的基础，但他不会认出我们现代的TPU Pod。又或许他能。它们仍然有很多共同点。底层技术在不断地、深刻地演变。我看我们每年都有重大突破。我的意思是，你刚刚看到我们在Antigravity中演示了通过提示创建操作系统的能力。

谷歌能够制造新的操作系统是非常危险的。

我们必须确保不要为了创造……而过度消耗token。我承认这一点。这很公平，但这就是这些东西正在做的事情的力量，对吧？世界上最顶尖的数学家、最顶尖的物理学家正在与这些工具互动，并以重要的方式使用它们，但这些工具能否从根本上自主做出新颖的科学发现？还不能。

它的进展已经非常显著。我确实认为它还需要发生重要的演变，而且世界上存在强烈的观点，认为要迈出下一步飞跃，需要对世界有真正深刻的理解。我对我们将继续取得巨大进步相当乐观。

你的时间表是什么？是三年，还是五年，达到通用人工智能？你站在哪一边？

我一直是这样回答的：我认为这个时间表并不重要，因为进步的速度意味着你正在以深刻的方式处理越来越智能的系统。所以，我用这种方式回答这个问题：从现在开始三年后，无论你和我是否将其称为通用人工智能，这并不重要，因为它将非常非常强大，我们必须为此做好准备。

桑达尔，这次对话很棒。再次非常感谢你抽出时间。
谢谢，尼雷。很高兴。

有疑问或评论？请发送邮件至 decoder@theverge.com。我们真的会阅读每一封邮件！

《Decoder》与尼雷·帕特尔
来自The Verge的一档关于宏大想法和其他问题的播客。

英文来源：

Today, I’m talking with Google and Alphabet CEO Sundar Pichai, in a conversation we recorded just after the Google I/O developer conference. This is the fifth year Sundar and I have sat down after I/O, and it’s become one of my favorite Decoder traditions.
Sundar Pichai on AI, the future of search, and what’s happening to the web
How Google’s CEO is reshaping the company — and the internet.
There’s always a lot of news at I/O, and this year was no exception — Google has powerful new Gemini models, it’s putting AI agents in everything, and it’s making huge changes to Search on both the web and YouTube that will once again reshape the information ecosystem.
That’s a lot to talk about, and Sundar and I got into all of it. But I also realized it’s been a long time since I’d asked Sundar the Decoder questions about structure and decision making, so I started there. You’ll hear Sundar say he realized he needed to rethink how Google worked a few years ago in response to ChatGPT, and he made a lot of executive changes and big decisions to get the company in a more aggressive posture.
Verge subscribers, don’t forget you get exclusive access to ad-free Decoder wherever you get your podcasts. Head here. Not a subscriber? You can sign up here.
Of course, we also talked about all those search changes, and how it seems obvious that the real future of Google Search is bringing things like the new intelligent search box together with the company’s new Gemini Spark agent platform. That way, searches can set off tasks, not just deliver results. That’s exciting, but it seems likely to yet again change the dynamics of the open web.
If you’re a Decoder listener, you’ll know that I coined the term Google Zero a few years ago — that’s the idea that Google traffic to websites would fall to zero as the company answered more and more queries directly on the search results page. That’s gone from an idea Sundar batted away in previous interviews to something the entire media industry is grappling with. Even the CEOs of major publishers like Condé Nast are now publicly saying they’re planning for a world of zero search traffic from now on.
Google is also training its models on YouTube videos, and changing YouTube search to summarize and index videos so you get dropped right into the relevant parts. That’s sure to cause some creator angst, so I asked Sundar if he’s ready to fight the same battles with YouTubers as he currently is with publishers.
Finally, I asked Sundar about Google DeepMind CEO Demis Hassabis ending the I/O keynote by saying we’re “in the foothills of the singularity.” It’s no surprise that Sundar agrees with Demis, but his thoughts on the timeline to AGI are worth paying attention to.
Like I said, it’s one of my favorite episodes to do every year, because Sundar is always game to actually take the questions — and even look at search results on my phone with me. I think you’re really going to like this year’s conversation.
Okay: Sundar Pichai, CEO of Alphabet and Google. Here we go.
This interview has been lightly edited for length and clarity.
Sundar Pichai, you’re the CEO of Alphabet and of Google. Welcome back to Decoder.
It’s great to be here. Nice to see you again, Nilay.
This is one of my favorite yearly conversations. I think we’ve done it at I/O now five times.
Wow. I didn’t quite realize it’s been five times, but I enjoy it. Thanks again.
I want to start with a little bit of a lightning round. I was thinking about this. We’ve talked a lot. We always get deep into the weeds of the web and search and big, heady ideas, and I realize I have not asked you the Decoder questions in quite some time.
I was just looking back at our previous conversations, and at Google itself, and you’ve made quite a lot of changes to Google. I think a number of your direct reports have changed over time. You’ve obviously restructured DeepMind, platforms and devices, and Android. Tell me how Google is structured right now.
Okay. It is Google and Alphabet. Obviously we have Alphabet as well, but broadly I think about it as there are three main businesses in Google: Search, YouTube, and Google Cloud. There are enormous platforms we run, which is Android, Chrome, and the whole area to do with it. And powering it all is all these important technology areas, which is AI and our infrastructure work. And then you have the functions to go with it.
But at a high level, you can think of it as Search, YouTube, Google Cloud, and then our big computing platforms. Those are the main groups, and obviously powered by Google DeepMind and our infrastructure teams. That’s one simple way to get a mental model around it. And of course, we have other bets beyond that, Waymo being the most prominent of them all, but there are many, many other bets, like Isomorphic Labs and so on.
I want to stay focused on the Google of it. I feel like we could do an entire hour on Alphabet and how that’s structured and how that works as a public company with many bets. But just to stay focused on Google for one second, the knock on Google historically is this is a company that ships lots and lots of products. You can’t sell lots of products. There’s not tons of focus. There are thousands of names of different products that are overlapping in different ways.
Where that comes from, at least in my view, is that you do have these big infrastructure bets. You have all these capabilities, and the people running the businesses can use those capabilities to spin up products. And there’s maybe not a lot of overlap or central planning like, “Did we launch two of the same thing?” How do you resolve that tension? It does seem like Google has gotten a little more focused, but that is the company’s culture: “We’re going to make a lot of bets and see which ones work.” How does that resolve for you?
There’s a lot of intent in what we do too. I think it’s not an accident we have 13 products with a billion users each, and we’ve been committed to those products longer term. You can go back and think about when Gmail launched or Maps launched or Google Docs launched or Search launched or Chrome launched. We’ve been deep and consistent in many, many areas over a long period of time as well.
One way I’ve internalized it in the AI moment is for the first time, we have such a common infrastructure powering all of them with our Gemini models and the underlying AI infrastructure. So we are more able to, with intent, do things which cut across things. Personal intelligence is a great example of it. It’s one effort. Users get a choice to turn it on in each of the products, but it’s built with one common infrastructure so that it works consistently across our products.
The underlying Gemini model itself is an example of it. We are able to bring that model in the context of the products, like Ask Maps in the context of the Maps product. But a lot of the technology powering it — the voice tech, the model, the intelligence — is all one work, which is why I think the AI moment offers us a new way to think about it, and not just across Google, but across Alphabet too over time. What makes this moment so uniquely powerful is that you can invest so much in R&D and infrastructure and develop a technology, which then you can apply across all these areas, obviously in a context in which they are useful for users, but the underlying technology platform is common. There’s a lot of intent that way and so on.
You have to give room for innovation, so allowing room for innovation where teams on the margin are able to ship some new features. Sometimes you later work to harmonize them. Take NotebookLM. Notebooks are now showing up in Gemini, and it’s effectively projects as Notebooks. And you can create a Notebook in Gemini, you can go to NotebookLM, you will see the same Notebooks, vice-versa. So that’s an example of where you innovate it first, and then you’re harmonizing later.
I was watching the keynote yesterday and I saw a lot of intent and confidence from Google: “We have this core technology. We can express it in lots of ways. It’s still essentially Google-y.” There are lots of products, lots of Gemini words. I’m going to figure them all out, I promise.
I would contrast that with… I don’t know, three, four years ago when there was the ChatGPT moment, everyone worried about what Google would do. Could OpenAI show up and take your market share in search away? Between that and now, you have changed Google. You have restructured it. There are new people in leadership roles. Connect those dots for me. How did you think about, “I need to actually change how the company works,” with the competitive moment you were in that got you here?
That’s a great question. I always internalize that moment. It was tough to convey it outside, but I pivoted the company to be AI-first. We had all the ingredients, so in some ways I felt like the Overton window had changed. People were adopting these technologies faster than we had expected. To me it was a way to go and actually express ourselves through our products, but I realized we had to organize ourselves for it. And going back to my earlier point, I realized we need a core model and a core infrastructure team to power everything we are doing across Google. A lot of my initial energy was to go set that up.
To get one AI team, we had world-class research teams in Brain and DeepMind and brought those together as Google DeepMind, which was harder than it sounds because it’s like saying, “Go put Stanford and MIT together and create a department out of it or a university out of it.” So I think we’re doing that well. At that time I also set up with Amin Vahdat, who’s now our SVP of AI infrastructure, a centralized infrastructure team, which has paid great dividends. Another evolution was realizing we need a chief AI architect to architect this technology across Google, and Koray Kavukcuoglu took on that role as well. Those were important changes.
Search needed to move faster, and Search was split across many leaders, so we put it under Elizabeth Reid, with Nick Fox being responsible for the overall area, Josh Woodward coming to help with our Labs product and working on Gemini later and driving innovation. I have other extraordinary leaders in the company as well, leaders like Philipp Schindler who runs all our operations and so on. So it is stepping back, and thinking end to end about the structure and making sure we are set up well for this moment where we need to move faster as a company, which means we need to make faster decisions.
I set up these new product reviews once a week. They were AI product reviews, making sure we are intentional about how we apply this technology, where we apply it, and to review everything firsthand, that anything to do with AI, which we were shipping to users, went through that channel. I spent time directly with whoever was working on it.
The other Decoder question I ask everybody is about decisions. You’re describing a lot of big decisions, some of them uncomfortable as you change people around. How do you make decisions? What’s your framework?
A big part of my framework is over time understanding that there are very, very few decisions which are really consequential, and most decisions aren’t. What matters much more is that you make the decision, because that’s what determines the velocity of an organization. The more you’re able to make those decisions and keep the company moving forward, you’re generally better off.
Of course, there are a few decisions like combining and setting up Google DeepMind that are more consequential, and you want to take your time deliberating and doing it. But a lot of decision-making is about just making them. The more you’re able to do that, the more you do develop over time some pattern matching and you’ve seen a version of the problem before. So I think it’s good to rely on that and separate the signal from the noise so that the signal is that this is a really important decision and you want to really deliberate around it versus it may look big, but it is more a normal course of action you need to take.
Looking around the industry, your peers in Big Tech have some of the wildest org chart ideas I’ve ever heard in my entire life. I think Meta wants to have 50 engineers report to a single manager with the power of agents. Jack Dorsey at Block wants all 6,000 people to report to him. Are you having similar thoughts that you should invent some of the craziest org charts with AI ever?
Leaders and people are incredibly important. And it depends. Some companies have a much narrower suite of products, and so different structures may work. When you’re running something at the scale of Google Cloud, it’s important that there is a CEO in charge. We are serving all the top enterprises in the world at a scale, and so how do you set up for that? Great leaders end up mattering a lot, like we have Thomas Kurian there. I do think about it.
But what I do think about it is how we are using AI more effectively, and we’ve seen the transition internally, particularly amongst our developers where we have transitioned from using AI tools to assist coding to them, a portion of the engineers directing teams of agents effectively more and more. Those are transitions underway, and that will flow beyond just engineering into the rest of the organization. It’s already happening. Even the work we are doing in Gemini Spark is to put that superpower in the hands of consumers, and what you can do with these agentic workflows, et cetera.
I’m more focused on making sure we are actually deploying that capability in a native way and that it’s working well, because for us it’s more than just making the company efficient because it’s the products we provide to others. I look at it with a very different lens. How we do it internally is what we are giving to users outside. We use Antigravity internally. That’s what we are providing outside. So the agents in Antigravity are what our developers are using, and so that’s what we are trying to put outside. It has that extra dimension to it.
The number one question Decoder listeners want me to start asking CEOs... I’ll just ask it straightforwardly. How close is AI to replacing you as the CEO?
I just think the CEO job is not that complicated. There are aspects of it where I think it’s going to be very, very helpful in terms of decision-making. I joke around that — partially joke around — that I have to spend a lot of time allocating compute. And I’m like, “Well, that seems like the AI is going to make more rational choices over time,” because I deal with a lot of appeals and emotions as part of working through a process like that.
Everywhere, what I see — which is maybe a bit different than how I think — is that done correctly, these tools are going to allow us to operate at the next level in everything we are doing. It’s not like you won’t do what you were doing before. You will start from a higher foundation. I wasn’t there when, I don’t know, spreadsheets rolled out to companies. I have to think back to how did people do all this financial analysis before? And I’m sure it changed over a period of three to four years fundamentally, and we got used to it.
I think agents and so on are a version of it. It’s not like you’re not going to plan birthday parties. Let’s say you’re planning a trip somewhere. Maybe you’re actually spending your time thinking about the actual things you want to do with your time versus chasing opening times and how to get tickets and so on. It elevates everything to a different foundation is how I think about it.
Let me ask you about that and agents. Some of those demos are fascinating. The idea that Search is going to build custom software for everybody seems like an idea in software engineering, a first impression. The idea is that you’re going to ask the computer a question, and the response will be for it to make you software that helps you get to an answer. I’m fascinated by this idea, but that is fundamentally changing Search.
And then you look at Gemini Spark, which is your agent platform in the cloud where you will say, “Go book me some tickets,” and Spark might run around and book you some tickets or do some task for you. And then there’s Antigravity, the agentic coding platform. Broadly, every year there’s a new paradigm for AI. There were LLMs first, and then maybe we’re going to chain some LLMs together, then there’s reasoning, and then now we’re at agents. Is this the foundation, or is there another paradigm shift to come?
It’s a great question. We are laying most of the building blocks in place. Fundamentally being able to reason, use tools, and code is a lot like having intelligence and reasoning — being able to plan, being able to look up things, use tools, and, if you need as part of that, to build something. You are laying all the primitives. Antigravity is for developers, but the Antigravity engine, the harness, is built into Gemini now. And Spark is just a mode of Gemini. Over time, it’s a feature. We are positioning it, but it’s just a tab within Gemini.
So you’re bringing that agentic harness. Users don’t need to think about it. Developers will understand it. Over time, in Spark, they can code powerful things. But as users, you may be building something, creating something, planning a trip, and all that is working behind the scenes.
We are laying a lot of the primitives of what we need for agents to work end to end, and more importantly, for AI to work. This long-running vision of Google Assistant we’ve all had and worked through myriad forms of it and failed to fully do it well, we are closer than ever before to delivering on that promise. We haven’t delivered it yet, but that’s the journey which I think is now closer than ever before.
I look at all the products, and they do seem like they should converge. You have the new Intelligent Search box, and I definitely want to talk about Search in more detail. But you look at that search box and then you look at, say, Canvas which makes you the apps. You’re planning a wedding, and it’ll just make you an app to help you plan a trip or a wedding or something. And then you have Spark which can go off and do things. I looked at that and I was talking to people yesterday, and it just seems obvious that that should be one product.
It will. I gave the earlier Notebook example of like, you’re creating Notebooks... but what are Notebooks? You’re effectively putting all the context you want in one place and then working off it. It’s folders as they’ve always existed for people, and Notebook should be a consistent primitive across the Google products you use. I just view agents that way. It shouldn’t matter. When you’re at the earliest stage of innovation, you create the capability. Teams are experimenting with it, but for a user over time, if you fire off planning a trip, it should work across both places is how I would think about it. You’re right in that.
There’s something very important about Google Search — it is a source of truth for people for however many years or even decades now. Go Google it, and you’ll get an answer, and that that answer is the same for you and me generally has been a very important idea. It is, I think, a fixture in the culture. Maybe Google is the last company saying it will just tell you the truth, out of all the companies out there.
Okay, but now we’re going to infinitely personalize the search box, and we’re going to infinitely personalize the Search experience. We’re all going to get different answers to queries. We’re all going to maybe even look at different interfaces depending on what we’re asking, what our personal context is, how much data Google has.
Do you think about that profoundly? How much can you destabilize the last common source of truth most people experience on the internet?
Look, there are factors well beyond our control, which is that people today have a wider variety of sources than ever before. People are getting content from so many different sources. But within the world of Google, I still think we deeply care about this being a source of knowledge and information. There are objective experiences and subjective experiences. What’s the capital of the USA? It’s not going to be custom-created for anyone. These are objective things. “Help me plan a nice trip to Montreal for a weekend” — naturally, the answers don’t need to be the same for everyone. There is a continuum there.
We deeply care about it. For certain categories of information, we do still anchor around authoritative information to present as much of an objective view as possible. And if it is health-related queries, we naturally tend to show more authoritative answers than if you’re saying, “What’s better? Should I go buy?”
Can I show you a search result?
Yeah.
A few years ago, I showed you a search result. I’ve been tracking this one for years.
I always love it. Amongst the 10 trillion queries...
Yes. Well, this one’s a favorite.
We have a very scientific, statistical way of doing this.
I think this is important, and I want to get into how consumers might be experiencing these products. So this is a search I just do all the time: “best Chromebook.” I’ll just show it to you. There it is.
So it starts with an AI overview. It just very confidently tells you the answer, and then there’s a bunch of sponsored boxes. And then the one that gets me is right below that, I believe the result is Reddit, and it has a top result in Reddit. It’s actually a different answer than the AI overview. And then there’s The New York Times, which has a different answer.
You scroll this and you’re like, “The AI overview is telling me one thing, the first organic result is fairly down the page, and all of these are different answers.” I hear what you’re saying about objective results and subjective results. “What laptop should I buy,” is somewhere in the middle of those things. I’m just curious how you think that experience for consumers is today in AI Mode and where you think it should go.
Look, to be very clear, in the world of AIO, we use an AI mode. We are organizing and giving context, but there are sources throughout, so you’re still presenting organic content in a different way. There are links and sources you’re given, but there is an opinion to go with it too, which is what you’re talking about.
Some of this will be iterative with users. One of the great things we find with search is it’s easy to measure user satisfaction. Over 25 years we’ve learned to measure user happiness, user satisfaction in a correlated way with improving the quality of the product, not for short term. That’s why we do these long-term studies. If we get any experience wrong, it shows in the metrics and we course-correct. We pride ourselves on the ability to track this over the long term — be it engagement, sessions, returning to a topic, the number of bounce-backs they do. It’s a very, very sophisticated way of looking at it. In some areas like that, the experience will continue to evolve.
Do you think that experience is good today?
It’s probably more opinionated than it should be for the particular query you showed me. That was my reaction as a user. That’s the scope for improvement is how I would say it, in a fast-evolving space, but I would expect that to happen in the product. My intuition there is, “Oh, that’s way more opinionated.” There is some chance that’s personalized to you. You may be testing it in a way that you’re uniquely personalizing. The reason that query might not be exactly representative, though, is that I know how you review all these things. There is some chance you’re in the .0001 percentile–
This is kind of why I’m asking about infinitely personalizable results, right? And I’m also asking if the experience is good, because I would bet that most people experience AI in Google Search all the time. They have that experience where they’re kicked to AI mode. There’s the stuff you can measure about user satisfaction, and then there’s how the public feels about AI.
I think there’s a pretty yawning gap between, “There’s these user numbers going up, and we’re close to a billion users, and the free products people are experiencing, how good they might be,” and then just the polling data. Young people dislike AI. It’s as objective as that gets. You can go ask them, and they will tell you in measurable ways they dislike it.
Eric Schmidt, the former CEO of Google, was booed at a college graduation speech he was giving. Seven in 10 Americans oppose data center construction. There’s some gap between the product experiences people are having and how they feel about the technology. Do you think you can close that gap? Do you think these products are good enough?
It is a very profound topic, and you’re linking the two things. AI is the most profound technology humanity’s going to deal with. It’s happening at a very fast pace. I don’t think humans have evolved to process this much change, and the rate of change particularly over the last few years is incredibly high. And particularly with all that they’re hearing, people are trying to understand the future and in the personal context of their lives, including what it means at an economic level and so on.
It really makes sense why there is anxiety around this technology, and we should be very attuned to that. That’s an important topic, and that’s much broader and bigger than the facets of what’s happening. People don’t directly associate these two all the time. In some cases, yes, they are linked in certain ways.
People experience the free versions of these models in various products. They open their social media feeds and they see slop. They see headlines about all that stuff. They have the tools just presented to them. The Gemini sparkle shows up in all the Google products, whether you ask for it or not. And then I do think you link it to, “They’re asking for a lot of electricity, and maybe my rates will go up. And maybe all the jobs will go away,” and that’s pretty scary, and I don’t know if the value exchange is there.
These are good things to study. You’re being too specific on what’s happening versus I’m just broadening it out and saying that might be part of the explanation. I do think there are other cheaper factors too.
Do you think it’s just a marketing problem? I’ve heard your peers say that AI just has a marketing problem.
No, I don’t think so. That’s the point I’m making. I’m in fact arguing against it. I think it makes sense to me why people would feel concerns about it. It feels natural to me. People are talking about how AI could make a lot of jobs go away. Why wouldn’t you feel a sense of anxiety about it? I think those are deeper issues which we have to tackle as a society. Yes, there’s concern about AI slop at a product level. All that is true. All I’m pointing out is it’s a multilayered problem. But I don’t think all the source of the data center angst is directly related to one specific experience you’re having in a product or something alone like that. That’s all the point I’m making, right? It is broader and bigger than that.
There’s a lot of AI slop out there. I feel it. In an early phase of technology with the competitive dynamic that exists, a lot of things are getting rolled out. But we also see empirically how people are using these products in very deep ways. If you go to a place where Waymo hasn’t come and you’ve just polled people, talking about self-driving cars, what you get in the polls is different from how they feel when they use these cars. Technology also goes through these things. People have pretty negative views of the internet too, by the way, if you ask about the internet. But it’s a fabric of our lives, and we have to adapt to it. All of that is simultaneously happening.
It’s a complex topic. To me, it feels like people are worried about rising energy prices, and if so, they want to make sure AI is not exacerbating the problem, and that’s a valid concern. And it’s up to us as an industry to make sure that if you’re building data centers, what can we do to make sure we aren’t contributing to that problem? I view it as our responsibility, not just us. And the government, there are bipartisan concerns around some of this stuff. For example, there’s a rate payer pledge we all signed up to with a set of commitments. Maybe there needs to be more done. All of that goes hand in hand. It’s important to talk about topics like skilling, workforce adaptations. We are driving a lot of change very fast through society. Those end up being very important topics as well.
There are concerns at all those levels, and I expect those concerns to be meaningful as we go forward. Many years ago I said, “This is more profound than fire or electricity,“ and so we have always felt that. Or think about deep-fakes and how do you know whether something is real? These models are getting better at simulating reality. This is why we’re working so hard. We are open-sourcing it, we are pulling many, many partners together, and it’s great for me to see the industry collaborate on a topic like this. Cybersecurity is another good example. These are all real concerns.
As an industry we need to do more. Governments will have a stronger role to play, and the public needs to be involved. You cannot have the most consequential technology rolling out the world in a way in democracies without public citizens rightfully having a voice around it. It is really important that we go through this phase, and that’s how we learn how to adapt.
My argument is that the products do the marketing work. That’s my push. I’m still waiting to see the killer app for consumers that does it. I think we have the killer app for enterprise.
One point, there are times I’ve gone through a health journey in Gemini. It feels more than like a killer app to me, better than anything I’ve ever done before. People are going through those experiences too.
I want to talk about the web, the health journey in Gemini that requires a rich data set of health information on the web to exist. You’re training Gemini on YouTube videos, right? Veo requires the YouTube ecosystem to operate and to be fruitful, to make new work in. You and I have discussed the concept I call Google Zero for many years, the idea that you will stop sending traffic to the web. You’ve disagreed with me that this is real.
Very much so. It hasn’t happened in the last many years.
Well, I’m just going to read you a quote. This time it’s not me, and I didn’t feed this to him. Roger Lynch, the CEO of Condé Nast, did an interview with TBPN last week and he said: “Every year our search traffic was down more than we had forecast, so last year I told our teams, ‘Assume there is no search. You have to have your businesses planned as if search is zero.’”
That is Google Zero. Condé Nast is saying, “We’re assuming that search will go to zero.” How would you respond to that, the idea that one of the biggest, most iconic publishers in the world is saying, “I can’t depend on this anymore”?
Look, first of all, the information ecosystem is so much broader beyond Google, by far. We see it in the data, you see it everywhere. So if any publisher over the last 10 years… I would look at The Verge and I would say where you were when you first took over, how much it’s evolved since then, the types of content you make, where all you put that content out, how all users are coming to you. It’s exceptionally dynamic, and so it makes sense to me every publisher is adapting to this new world.
We are adapting to the evolving world and how users are consuming technology. We had to do this when the world shifted from web to mobile. We are shifting it from a world of mobile to people having ongoing conversations, chatting with these products, talking to them, consuming it in voice and many different form factors.
People are expressing preferences for various types of content. They’re looking for user-generated content. They’re looking for podcasts. They’re looking for that. Through it all, we are very committed to both meeting user expectations, and also connecting them to what’s out on the web. Just even in the last year, even since we’ve launched these features, we’ve gone back and added more links. Another area where behavior is changing is that many publishers, rightfully so, are thinking about subscription models.
Sure. But I’m just saying Condé Nast is saying, “We’re going to assume our search traffic is zero, given the trends that we see.” Should they assume that?
Look, I always view... People understand their businesses better... I mean, I’m not in a position to tell such an iconic publisher what they should think about their business or plan. If they are building content that is high-quality and people like it, I expect us to reflect that in our products. That much I can commit to them.
But I think more than any other company through this evolution, we are working very hard to make sure people can get connected, and we are planning to do it in Search and Gemini, and that still underpins a lot of what we do. But there is evolution. As the technology improves, low-quality clicks get filtered out. That’s a natural evolution we see. We see it in our metrics. Bounce clicks are going down. And so those are all dynamics.
People are going to a wider array of information, and there are more people producing information than ever before. That pie is growing. All these dynamics are happening. It’s a complex ecosystem, but our commitment is to make sure we reflect the vastness and diversity of the content, and we do think people want to connect ultimately to these sources, but we are trying to meet them in those moments, and people come with very different intent and very different moments.
One of the small features we have done, but very important I think, is if you’ve subscribed to something, we reflect that as a preferred source for you as a user. But that’s a new change which we didn’t have before. We are adapting to the fact that publishers are increasingly turning to subscription offerings too.
Publishers and YouTube creators, should they be able to opt out of training to get surfaced in Search?
This is a much broader topic. Both laws and regulations will have to evolve. The courts will have to be in. It’s important to protect copyright. It’s important to protect fair use. And so these are constructs which will evolve dynamically through that.
But do you want to be in a bunch of lawsuits with YouTube creators? You’re in a lawsuit with publishers in the UK. That rhetoric in that lawsuit is getting increasingly heated. Google has said that the proposed solution is a “free rider charter.” Every year the News Media Association sends me a quote to read to you, and they say, “Google calling us free riders is obviously ridiculous. It’s basic supply chain economics. If the value were really all on Google’s side, they would simply allow publishers to opt out.”
Do you want to be in that same fight with a bunch of creators on YouTube about opting out?
Look, we are constantly — as part of Gemini developing…
We did offer a new opt-out with Google-Extended, and we are in conversations with publishers. We’ll take feedback and over time work through what makes sense. Obviously we are not the only player in a big ecosystem. We are also trying to put out products which are competitive to other products out there. All the publishers will also write an article saying the product is not very good. So it is more complicated than it looks.
You have spent more time thinking about the web and the health of the web and the necessity of the web. Paint me the picture for what a healthy web looks like in an agentic search world.
One of the arguments I’ve made over time and I actually see it playing around a little bit more, is I’ve started using the web more again over the last year to year and a half. All these AI experiences have brought the web back more. There was a time when it felt like... But I always felt the web would be vibrant. In fact, I’ve argued the web is going to be vibrant every year, and I would still argue it today. The web is constantly evolving. I’ve never seen anything as dynamic as the web, which is why it’s been such a privilege to be part of that evolution.
I look at agents, and that is the next evolution of the web, which we will deal with, and I think it will evolve the web pretty profoundly. There will be a lot of debates about what’s okay, what’s not, but people want to put out information, to connect with other people. People want to be connected. People aren’t trying to be in a siloed world, detached. That doesn’t reflect the reality of the human experience. I think the web is going to play as central a role on it as ever before. In fact, the Universal Commerce Protocol, if anything, what we announced yesterday, I think people are slightly underestimating the impact of it.
Actually, can I juxtapose that? There are a lot of muscular announcements about new products, new features, and agentic tools you can use, and UCP and Amazon and Walmart and everyone saying, “We’re going to use a new standard we’re building for shopping,” and all that is very tangible.
And then I/O ended with Demis Hassabis, the CEO of DeepMind, coming out, and he said this thing that I have not been able to stop thinking about. He said, “Google’s cutting-edge research and products will help unlock AGI’s incredible potential for the benefit of the entire world. When we look back at this time, I think we will realize that we were standing in the foothills of the singularity.”
Can you tell me what it means to be in “the foothills of the singularity”?
Demis and I have had long, deep conversations on this topic. In this context, the advent of AGI is what he thinks of as the singularity.
Do you have a definition of AGI? Have you debated it? Do you have an agreement?
We debate it a lot. I think both Demis and I are very close in how we think about things. There is a harder definition of AGI, which is that it has to be more comprehensively able to do a wide range of tasks, including cognitive tasks, in a way that’s comparable. We’ll at some point actually put it out as a company, and we are working on that. But that’s what he’s talking about in this context.
By the way, I think it’s important for us to understand that this technology is progressing very rapidly. Later today, I’ll be going and spending time with our AI researchers, not just in our company, but also amongst the frontier labs. There’s wide consensus that this technology, AGI, is... people may quibble around whether it will be three years, but the technology’s coming sooner rather than later. It’s more important to communicate that because — to an earlier part of the conversation — it’s important that we as a society understand it and are preparing as much as possible.
I asked you this question maybe the first time we ever talked about AI. I asked you if language was intelligence. And the progression here is we’re layering more and more on LLMs. We’re doing longer chains of reasoning, we’re building harnesses, we’re doing all this stuff, but the core technology is still transformers. It’s still the thing Google invented so long ago. Can LLMs get you to AGI? Is that path clear?
The trajectory over the last three years has been incredible. The LLMs of today have evolved in many ways too. We are constantly evolving it. To me, it’s like asking, can computers get us to the way—? The von Neumann architecture is still what powers most computers today, but he won’t recognize the modern one of our TPU pods. Or maybe he would. There’s still a lot of commonality to it. The underlying technology keeps evolving so profoundly. I look at every year we have had major breakthroughs. I mean, you just saw us demo in Antigravity an ability to prompt and create an operating system.
It’s very dangerous for Google to be able to make new operating systems.
We’ll have to make sure we don’t token max on creating... I’ll give you that. It’s fair, but that is the power of what these things are doing, right? There are the top mathematicians in the world, top physicists in this world who are interacting with these tools and using them in important ways, but can these tools fundamentally make novel scientific discoveries on their own? Not yet.
It’s remarkable how much it’s progressed. I do think it has important evolutions to happen, and then there are strong opinions out there in the world about how much of a real understanding of the world you need to take that next leap. I’m pretty optimistic that we will continue to make a lot of progress.
What is your timeline? Is it three years, or five years, to AGI? Where are you at?
I have always answered it this way: I think that timeline doesn’t matter because the rate of progress means you’re dealing with ever more intelligent systems in a profound way. So the way I would answer that question, three years from now, whether you and I call it AGI or not doesn’t matter because it’ll be very, very powerful, and we have to prepare for it.
Sundar, this was great. Thank you so much for taking the time yet again.
Yeah, thanks, Nilay. Pleasure.
Questions or comments? Hit us up at decoder@theverge.com. We really do read every email!
Decoder with Nilay Patel
A podcast from The Verge about big ideas and other problems.