多模态AI如何改变搜索 | 谷歌产品副总裁罗比·斯坦

自远古以来，搜索就以谷歌及其辉煌的十条蓝色链接形式存在。

Since seemingly the beginning of time, search has come in the form of Google and their glorious 10 blue links.

无论你提出什么问题，搜索引擎都会迅速整合全球知识并几乎即时呈现给你。

Ask any question in the world, and the search engine goes off aggregating the world's knowledge and presenting it to you nearly instantly.

这简直只能用魔法来形容。

It can only really be described as magic.

但现在交互方式正在转变。

But now the interface is shifting.

它正从这种黑箱模式转向多模态，比如谷歌镜头——现在有了能与现实世界互动的传感器，能在语音、音频、视觉等多维度提供搜索功能。

It's shifting from this black box to multi modality, and things like Google Lens where you now have sensors that engage with the real world and can now give you search in many dimensions voice, audio, visual, all of the things.

埃杰斯，我一直对此很好奇。

And I've been curious about this, Ejes.

我知道你也对此非常好奇，幸运的是还有个人对此深感兴趣，他就是今天来到节目的罗比·斯坦。

I know you've been really curious about this, and thankfully, another person has been really interested in this, and this is Robbie Stein, who is now on the show.

欢迎你，罗比。

Welcome, Robbie.

Robbie是谷歌搜索的产品副总裁，看来你是与我们讨论这个话题的绝佳人选。

Robbie is the VP of product at Google Search, so you seem like the absolute perfect person to discuss this topic with us.

我想从这个问题开始：新版TenBlueLinx会是什么样子？

I think the place I wanna start is, what does the new version of TenBlueLinx look like?

搜索是否会保持不变，在AI时代，谷歌搜索未来的计划是什么？

Is is search going to continue to be the same, and kind of what is the plan for Google Search moving forward in a world of AI?

首先感谢邀请我参加节目。

First of thanks for having me on the show.

很高兴能来到这里进行这次对话。

Really fun to be here and to have this conversation.

我认为，在这个AI时代，它本质上是在扩展的。

I think, ultimately, in this AI moment, it truly is expansionary.

我们讨论的是它比以往任何时候都能为人们做更多事情。

We talk about the fact that it's doing more for people than ever before.

当你思考搜索的未来时，实际上要从它现在的样子开始。

When you think about what the future of search looks like, it actually starts from what it is today.

因为人们日常需要快速获取电话号码、在线支付账单或找到某个直接链接的需求并不会消失，但谷歌搜索能做的远不止这些。

Because the everyday need of someone grabbing a quick phone number, being able to pay a bill online, find a direct link somewhere, it doesn't really go away, but it's that Google search can do so much more.

所以我认为你会发现搜索体验已经提升，变得更为强大。

So, I think what you're finding is that the search experience has enhanced, and it's become, a lot more powerful.

而现在，AI体验也开始显现。

And now there's AI experiences that start to show up.

在页面顶部会有一个小预览，展示AI可能在哪些方面提供帮助。

It's a little preview at the top of the page where AI could be helpful.

你可以想象这样一个场景：当有人提出一个非常困难的问题时，这类问题会越来越多，通过点击那里，你现在就能体验到这种更具对话性的搜索版本。

You can imagine a world that someone asks a really hard question, you get more and more of those, and by tapping into there, you're now having this more conversational version of search.

因此，第一个重要主题是你可以获得AI驱动的体验，而且是双向互动的。

So, the first big theme is that you can go and have an AI driven experience, and one that's a back and forth.

你可以进行优化、提出后续问题、满足更多好奇心，这是一种不同的模式。

You have refinement, you have follow-up questions, you have more curiosity, and that's a different paradigm.

但这确实只适用于特定类型的问题。

But it really makes sense for a specific kind of question.

你只是在随意浏览，并不清楚自己具体想问什么。

You're just kind of browsing, you don't really know what you're trying to ask.

实际上，核心搜索体验很大程度上就是为这类场景优化的，这可能就是最佳方案，因此并非每个查询都需要AI介入。

Actually, lot of the core search experience is really optimized for that kind of experience, and that might be the best, which is why you don't see AI for every single query.

但我认为第一个显著变化将是这种对话式交互。

But I think the first big change you're gonna observe is that that conversationality.

我认为第二个变化是多模态能力。

I think the second is multimodal.

过去AI模式中，人工智能和搜索总体上更局限于类似网页的体验形式。

I think in the past AI mode, AI and search in general has been more constrained to a web experience that feels more like a web page.

你输入文字，返回给你一个页面。

You type text in, you get you get a page back.

但逐渐地，AI将演变为一种无处不在的智能存在——想象一个被编码来理解尽可能多的世界信息、网络内容，并能针对每个问题匹配最有价值网络资源的知识大脑。

But increasingly, it's just this kind of ambient knowledgeable AI that imagine something that has been encoded to understand as much of the world's information as possible, as much of the web and what parts of the web are most helpful for every single given question, and it's in this brain.

你可以直接与它对话。

You can talk to it.

一种方式是通过文本，通过我刚才提到的网页界面，你可以进行完整的对话。

One way to do it is through text, through the web interface I just mentioned, and you can have a whole conversation.

另一种方式可以是实时的。

The other way could be live.

所以现在在我们的应用中，你可以在开车时进行实时对话，使用完全相同的模型，它开始了解这个世界并与你交谈。

So in our apps right now, you can have a live conversation while you're driving with the exact same model, just start learning about the world and just talk to you.

这有点难以置信。

It's kind of unbelievable.

另一个是图像。

The other one is is is images.

我们看到谷歌上视觉搜索量每年增长70%。

And and we see a 70% increase in the amount of visual searching year over year on Google.

这实际上是人们搜索速度最快的方式之一。

It's actually one of the fastest ways people are searching.

人们只想拿出相机，询问他们看到的东西或设备上的内容。

People just wanna take their camera out and ask questions about what they're seeing or ask questions about what's on their device.

所以你基本上是从网页世界转向一个可以直接询问所见事物的世界，它能适应你生活中所需的各种形式。

So you kinda move from this world of a page to a world where you can just ask about what you're observing and it can kind of fit the form factor that you need given your life.

无论你是在散步、开车，还是躺在沙发上玩手机，Google都与你同在。

And whether you're going for a walk, you're driving, or you're just on your couch on your phone, Google's there with you.

我喜欢这种描述，它就像一个试图突破被束缚媒介限制的环绕式大脑。

I love this description of it's kind of like this ambient brain that's trying to break through the confines of whatever medium it's been chained to.

最初，你知道的，就是那10个链接，Google主页的搜索排名，这对我来说充满了怀旧感。

And originally, you know, it's this 10 links, you know, the the homepage of Google, the search ranking, this is something that's pretty nostalgic to me.

对吧？

Right?

因此，如果它是互联网的门户，那么随之演变的任何事物都至关重要。

And so if it's the doorstep to the Internet, anything that evolves with that is is super important.

我一直在思考的是，随着AI通过大语言模型、ChatGPT和Google Gemini流行起来，看到搜索页面演变至今，我很想了解推动我们或你们达到最终媒介的时间线和决策点。

And one thing I've been asking myself is, as AI has become popular through LLMs, ChatGPT, through Google Gemini, seeing the search page evolve and get to where it is today, I would love to understand the timeline and decision points that made us or made you get to that final medium.

我的理解是，我们从搜索引擎起步，获得一个按从最热门到最不热门排序的10个蓝色链接的搜索页面。

So the way I think about it is we've started with the search engine and we get a ranked search page of 10 blue links, right, ranked from most popular to least popular.

接着我们迎来了Google Gemini的出现，这个大型语言模型聊天界面。

And then we had Google Gemini appear, this LLM chat interface.

然后这个大型语言模型聊天界面接入了网络搜索功能。

And then this LLM chat interface got access to web search.

再后来我们有了由Gemini驱动的Google搜索，开始看到更多这类AI概览功能。

And then we had Google Search powered by Gemini where we started seeing more of these AI overviews.

现在我们甚至有了能在后台自主工作的Google智能体。

And now we even have, like, Google agents that are doing autonomous work behind the scenes.

你能简要介绍一下这个发展历程和关键决策点吗？

Can you briefly walk us through this timeline and decision points of how we got here?

好的。

Yeah.

我认为中间其实有几个关键环节需要说明清楚。

Mean, I think that there's a couple of pieces in between actually that are helpful to clarify.

其实Google很早以前就以这个经典搜索页面闻名，大家都熟悉并喜爱那些蓝色链接，但此后它已经经历了多次演变。

I mean, I think Google a long time ago had, you know, blue links as this prototypical search page that everyone knows and loves, But actually it's evolved a lot over the years since then.

我们考虑过通用搜索，比如当你查询本地信息时，地图信息就会显示出来。

We think about universal search and how if you ask for local information, maps information shows up.

或者当你提出视觉相关问题时，会得到通用图像搜索结果，看到相关图片。

Or you ask about visual questions, you get a universal for, image search and you'll see like related images.

又或者查询热门信息时，顶部会出现一个新闻头条单元，展示相关文章。

Or if you ask about trending information, you'll get a top stories unit with, articles at the top.

对吧？

Right?

此外还有些非常具体的问题，我们会在页面顶部特别展示一个带有超大摘要的网页结果，比如问‘有多少...’这类问题。

So and then you also have very specific questions someone might ask where we would feature one web result with an extra large snippet at the top of the page that might be like, hey, how many?

我不知道。

I don't know.

你得提出一个很多人都会问的具体问题，我们就会特别突出显示它。

You gotta give a very specific question that lots of people ask and we just kinda highlight that.

所以我认为上述每个变化都是搜索体验的重大演进。

So each of those I think have been fairly large evolutions to the experience.

我认为AI概览功能正是基于这一背景应运而生的。

And I think off the back of that is where AI overviews came.

因为，你知道，我们已经具备了这些丰富的体验。

Because, you know, we had these rich experiences.

当你查询天气时，页面顶部就会显示天气信息，对吧。

You asked about weather, you get weather information, right, at the top of the page.

在页面最顶端，你可以针对具体问题获取更精确的信息。

Where right at the top, you could get more specific information when you had a specific question.

因此很自然地，AI能进一步解锁用自然语言提出更复杂、更冗长问题的能力。

And so a natural extension of that is how AI could unlock the ability to really ask these harder, longer questions with natural language.

这样即便没有单独网页包含该信息，你也能通过提出具体问题获得AI生成的概览。

So you could ask a very specific question, and even if there's not one specific web page that has that information, you could provide this AI overview.

这就是我们所见证的第一次重大演进。

And so that was the first big move in evolution of what we saw.

接下来，一旦实现这点，我们实际观察到用户正试图更频繁地触发和使用该功能。

And the next one was, okay, well once you had that, we actually observed that people were trying to get it to trigger more and show up.

所以人们实际上会在搜索框里输入‘AI’这个词。

So people would actually put the word AI in the search box.

然后，另一件事是他们并不满足于此。

And so, and then the other thing was they weren't done there.

他们想继续追问，但没有简单的方法实现这一点。

They wanted to ask follow-up questions and there was no easy way to do that.

这最终催生了AI模式的概念，即通过搜索实现对话式交互体验。

And so, really, that led to this idea around AI mode, which is a way to have this conversational experience with search.

而且从AI概览进入这个模式变得越来越容易。

And it's increasingly been, you know, easy to get to from AI overview.

你搜索后会在顶部看到AI预览，展开后就能进入AI模式深入探讨。

So you search, you get the little AI preview at the top, you expand it, and you can go deeper in AI mode and have a poll.

现在你可以和谷歌进行十到二十轮的对话交流。

You can have ten, twenty back and forth conversation with Google now.

以上就是整个发展历程的主要脉络。

And so, all of that was kind of the main arc of how it's evolved.

然后我认为它只是沿着这种复杂性和用户需求的曲线逐步上升。

And then I think it just kind of ascends this kind of curve of complexity and and and and kinda user need.

我们就在想，好吧，用户接下来想做什么？

We're like, okay, what's the next thing the user's trying to do?

你不断提出这个问题，思考我们还能如何提供帮助？

You just keep asking that question, and how else we can be helpful?

举个例子，假设你想在旧金山这周五晚上7点找个约会餐厅。

So, for example, let's say you want to find a restaurant for date night in San Francisco this Friday at 07:00, specifically.

你可以把这个输入谷歌，我们会给出AI概览，然后你可以在AI模式下反复交流，直到得到一份不错的餐厅列表。

Like, it's possible that you could put that into Google and we get an AI overview and you would go back and forth with AI mode until you, you know, like, got to some list of great restaurants.

但你上谷歌不只是为了找餐厅。

But then you're not going to Google just to figure out a restaurant.

你真正想要的是把预订作为这段旅程的最终目标。

You actually want the reservation as the the ultimate end to that journey.

行动。

Actions.

是的。

Yep.

所以你的下一个问题是，好吧，我们能做些什么来代理性地为你预订那张桌子或提供帮助。实际上，我们现在已经实现了一个功能，让代理可以浏览并与OpenTable及网络上的预订系统交互，然后带回一个列表，不仅包括通过分析和推理为你呈现的优秀餐厅，还在列表中显示可用的餐桌时间，这确实非常神奇。

And so your next question is, okay, well, what could we do agentically to actually book that table for you or help, you know, we actually implemented now in his live, a way for an agent to browse talk and OpenTable and the web for reservations, and then bring back in the list, not only great restaurants that use, you know, analysis and reasoning to present that list for you, but then in that list, times at the table's available, which is really really magical.

希望这能稍微说明我们是如何将体验从感觉像一页纸转变为感觉像一场互动式AI对话，并且能够越来越多地为你和与你一起完成事情的。

And so hopefully that tells a little bit of the story of how we went how how kind of kind of evolved the experience from one that felt like a page to one that felt like an interactive AI conversation and one that could increasingly do things for you and with you.

这种互动对话正是我认为这个新世界的样子，但我有点难以理解这种思维跳跃具体是怎样的。

That interactive conversation is exactly what I think this new world looks like, but I'm kind of struggling to get my head around what that mental leap looks like.

对吧？

Right?

我习惯了搜索。

So I'm used to search.

我习惯了在键盘上敲击按键，输入文字，获取我想要的信息，然后滚动、点击，最终完成一个操作。

I'm used to tapping keys on my keyboard, typing words, getting the kind of information that I want, and then kind of scrolling, clicking, and getting to an action.

这种多模态版本会是什么样子呢？

What does the multimodal version of this look like?

或许我们甚至可以用一种心智模型的形式来实现。

And maybe we can even do it in the form of, like, a mental model.

那么，如果传统搜索像是‘输入文字获取链接’，当我的相机、语音、截图和电视成为交互方式时，这会变成什么样？

So, like, if classic search was kind of, like, type to links, then what is it when my camera, voice, screenshots, and TV are kind of, like, the the way you interface?

那么，这个类比看起来是怎样的？

So, like, what does that analogy look like?

是的。

Yeah.

我想首先要说的是，我不认为人们想要彻底改变他们对谷歌的心智模型。

So I guess the first thing to say is I don't think people wanna dramatically change their mental model of Google.

他们基本上希望——这是我个人的观点，尤其作为用户时的感受——通常你只想说‘我有个问题’。

They wanna basically think and and this is my personal view, and particularly as a user, how I feel, is in general, you kinda wanna just say, I have a question.

我想把这个问题抛给谷歌。

I wanna put I wanna ask Google this question.

你把它输入谷歌。

You put it into Google.

目前，主要的方式还是通过文本，但它正迅速向多模态、视觉、语音等方式扩展。

And right now, the the way main way that happens is text, but it's very quickly growing to be multimodal, visual, voice, etcetera.

但它们都源于同一个根本需求——我有个问题要问。

But they're all just different modes of the same root need, which is I have a question.

可能是关于你正在看的东西的问题，比如在社交媒体上截图问'我想知道这套衣服是什么，怎么买'。

Could be a question about something you're looking at, could take a screenshot on social media and say, I wonder what this outfit is and how I could buy it.

可能是你路过一棵树时拍照想知道那是什么树。

Could be a tree you're walking past, take a picture of, wanna know what that tree is.

但无论如何，核心都是提出一个问题。

But it's a question, never you know, nevertheless.

所以你带着问题，把它交给谷歌，然后选择最便捷的提问方式。

And so you take your question, you put it into Google, and then you you kinda decide what's the easiest way to do it.

如果你在电脑前，可能是复制粘贴作业问题，或直接打字提问。

Well, if you're on your computer on your desktop, maybe you have a homework question, you'd copy and paste it, or you'd ask a question, or you just type it.

如果你在外用手机，就会选择拍照提问。

If you're out and about on your phone, and you'd take a picture.

如果你正使用安卓设备查看手机，可以使用圈选搜索功能。

If you happen to be on an Android device and you're looking on your phone, you use circle to search.

人们已经在以巨大规模进行这种操作。

People are already doing this at enormous scale.

每月约有十亿人使用镜头功能，使用这些可以拍照的产品。

It's like a billion people a month using lens, using these products where you can take a photo.

所以我认为这已经是一种广为人知的模式了。

So I think this is a pretty commonly understood pattern at this point.

然后它就直接跳转到谷歌。

And then it just goes to Google.

从用户角度来说，我认为他们并不希望被打扰太多。

And to from the user's perspective, they don't really want, I think, to be bothered too much more than that.

他们只想把问题输入谷歌，然后谷歌应该负责判断并为你提供最佳信息。

They just wanna get their question into Google, and then Google should do the judge do the work to give you the best possible information.

所以如果你问一个非常基础的问题并想浏览相关信息，比如有个你从未听说过的新首发四分卫，你输入这个人的名字。

And so if you ask a question that's really basic and you wanna browse about it, like, let's say there's a new starting quarterback you've never heard of, you put the person's name in.

当然，你可能想要一个快速描述，但实际上你通常想浏览的是那种针对人物的搜索。

Like, sure, maybe you want a quick description, but you actually wanna browse typically that a per a people search like that.

你想看到这个人的照片，快速的知识面板。

You wanna see photos of the person, quick knowledge panel.

你想看看他们在社交媒体上的近期帖子。

You wanna see their, like, recent posts on social media.

你知道的，你想看看关于他们的报道文章。

You know, you wanna see articles that have written about them.

在很多情况下，对于这种需要浏览类的问题，AI的回答其实并不理想。

In many cases, an AI response is actually not great for that kind of a question that's like a browser kind of need.

但如果你问一个非常具体的问题，砰，你会在顶部直接获得AI提供的概览，并有深入探讨和来回交流的方式。

But if you ask a really specific question, boom, you're gonna get AI right at the top with an overview and a way to dive in and have them back and forth.

所以你某种程度上依赖谷歌总能根据问题给你最佳的呈现形式。

And so you kind of rely on Google to always give you the best format given the question.

如果这是可预测的，那么你大概知道，好吧，如果你要求一些鼓舞人心的图像类内容，你就会得到图片。

And if it's predictable, then you kinda know, okay, if you you ask for something inspirational and image y, you get images.

对吧？

Right?

你会得到视觉内容。

You get visual stuff.

如果你问一个具体的知识性问题，就会得到AI的回复。

If you ask a specific question about knowledge, you get AI.

如果你查询某个人物，就能看到他的照片和人物简介。

If you ask a question about a person, you get see photos of that person, you get his description of people.

这就是我的理解，也是我们从用户那里听到的反馈。

So that's how I think about it, and I think what we hear from users.

你一直在讨论形式因素，这让我很感兴趣，因为我觉得包括我们在内，当人们谈论AI时，往往只想到聊天机器人，在搜索场景下就是一个聊天框加文本框。

So you're talking about the the form factor a lot, which is something I'm interested in because I think a lot of the times when people talk about it, ourselves included of AI, we kind of imagine, well, there's there's a chatbot, or in terms of search, there's a chatbot, and then there's a a text box.

中间形态很少，我好奇的是你如何看待这类产品最终形态的演变方向。

And there's not really much in between, and I guess what I'm curious about is the way you see the the final form of this kind of evolving over time.

因为你描述的是谷歌的文本框，而我们还有带摄像头的多模态交互。

Because you you describe a text box with Google, and then we have the multimodality with cameras.

当我听到这个，它让我想起了早期的计算机，那时我们刚开始只有基于文本的东西，这比打孔卡要好，然后我们发展到了图形用户界面，接着又进化到了网页浏览器。

When And I hear this, it it kind of reminds me of the early computer where we kind of started with just text based stuff, which was better than punch cards, and then we evolved to a graphical user interface, and then we evolved to web browsers.

这种自然的演进过程在当时并不明显，但事后看来却如此清晰。

And there was this kind of natural evolution that wasn't obvious, but in hindsight, it was so clear.

我很好奇你对这种演进接下来会走向何方有什么看法，因为我知道你是Google Lens的先驱，这非常了不起。

And And I'm curious if you have any ideas on where that evolution goes from here, because I know you pioneered Google Lens, which is amazing.

我一直在使用这个产品。

I use the product all the time.

这是我最喜欢的与世界互动的方式，因为它就像一根魔法棒，我可以看着任何东西并从中获得答案。

It's my favorite way of engaging with the world because it just feels like this magical wand that I could look at anything and get answers from.

我很好奇你认为这种技术未来会如何发展。

I'm And curious where you see that kind of progressing to over time.

是的。

Yeah.

嗯，我是说，你之前问过我关于模式的问题，比如，心理模型是什么。

Well, I mean, you were talking I mean, earlier you asked me about what's the pattern, like, what's the mental model.

听你重新阐述后，我想最简洁的说法是：这是一种对话的心智模型。

And I guess in hearing you rearticulate it, the the the most succinct way of saying it is it's the mental model of of conversation.

所以如果你想想，当你与某物或某人对话时，会有文本输入框、照片上传功能，还能实时连线看到对方。

And so if you think about, you know, when you have a conversation with something or a person, then you have a text field, you have a way to upload photos, you can go live with that person and see them right now.

对吧？

Right?

你可以和那个人进行实时视频聊天。

You can have, like, a live video chat with that person.

这些都只是接触那个人的不同方式而已。

They're all just different modes of accessing that person.

我认为技术正朝着这个方向发展：与人交流会变得像和真人对话一样简单。

And I do think that technology is moving in a direction where it's as simple to communicate with as talking to a person.

它能接受所有交流模式和讨论方式。

Mean, it accepts all of the modes and all the ways of discussing with the person too.

不过现在的情况是，对大多数人来说，大多数需求并不需要长时间对话。

Now it just turns out that most people, for most needs, don't want need to have a long conversation.

就像，它们更像是时间线上的单次瞬间。

Like, they're kinda single shot moments in time.

就像你坐在电脑前，突然想起来：哦，我完全忘了。

It's like you're sitting at your computer, you're like, oh, I just totally forgot.

我得给这个地方打个电话。

I need to call this place.

你快速谷歌一下电话号码，随便哪个。

You just Google really quick phone number, whatever.

打给他们。

Call them.

对吧？

Right?

所以你不需要——我不需要和它进行长达二十分钟的对话。

And so you don't need I don't need to have like a twenty minute conversation with the thing.

但如果你想想那个小瞬间，你其实是在与谷歌交流，表达需求并获取反馈信息。

But if you think about in that little moment, you were kinda communicating with Google, you're expressing a need, you're getting information back.

因此，我认为从心智模型的角度来看，这越来越是正确的方式。

And so I think increasingly that is the the right way to think about it from a mental model perspective.

所以如果你思考事物的发展方向，可以大致表达出，至少对谷歌而言，我们非常关注信息需求。

And so if you think about where things are going, you can kind of just articulate, you know, what else, at least for Google, we think a lot about informational needs.

那么，比如，我们还能提供哪些新型信息任务来帮助你？可能是新模式，如果有用的新模式出现，也可能是在特定问题上为你做更多。

So, like, what other kind of informational task that we could do to be helpful for you could be new modes, like if new modes come up that could be useful, but it could also be doing more for you given a given a certain question.

就像你可能会发短信找人帮忙说：嘿，能帮我个忙扫描这些东西并告诉我需要了解的内容吗？

The same way you might text someone to help you with something and say, hey, can you do me a favor and scan all this stuff and tell me what I should know about it?

或者，比如，你能帮我预订这个活动吗？

Or, you know, can you, you know, book this event?

我正在想办法买这场比赛的票，但过程很复杂。

I'm trying to get tickets to this game and it's complicated.

在某些情况下，你甚至可能雇人来帮你，比如为超级碗比赛或其他难买的票找到最佳位置。

You know, you might even hire, like, someone to help you, in some cases, get the perfect ticket for a Super Bowl game or some, you know, hard to get ticket.

或许可以请个代理帮你完成这项工作。

You know, maybe an agent could go do that work for you.

在后台，这就像发短信给一个人，对方回复说，给我几分钟时间。

And on the back end, it's almost like getting texting a person, the person getting back to you saying, give me, like, a few minutes.

我会查一下，然后他们会给你一些结果。

I'm gonna look into this, and then they give you, you know, some result.

实际上，这两张票是有的，它们相邻，价格是这个，但AI能做到大多数人无法付费或对人来说非常困难的事情，因为这需要你搜索数百个东西才能回复你。

Like, actually, these two tickets are available, they're next to each other, and they're at this price point, except that the AI can do things that, you know, most people, you know, couldn't pay for or would be really hard for a person to go do because it would require you to search, like, hundreds of things to get back to you.

所以我认为这是一种有帮助的思考方式。

So I think that's that's kind of a a helpful way of thinking about it.

关于最后一点，搜索数百个东西，我们询问了观众他们想问你的问题，其中一个问题是，当人们认为AI正在蚕食搜索时，谷歌将如何生存，一个有广告模式，一个没有，我想问你这个问题，因为我觉得你会说什么，我想让你帮我验证这个零拇指理论，或者说零和理论，关于聊天机器人，我听过你说，也听过谷歌整体说，AI概览和AI整体让人们搜索更多，某种程度上我们看到杰文斯悖论，随着数据增多，人们更有动力去搜索更多，创建更多查询，我想让你帮我们理解这个现象，在有AI的情况下，实际上有更多的搜索发生。

To that last point, the the searching hundreds of things, we we asked our audience what they wanted to ask you, and one of the questions is, how how is Google going to survive when the perceived thing is that, well, AI is kinda cannibalizing search, and one has an ad model, one doesn't, and I kind of wanna ask you, because I have a feeling what you're gonna say, I want you to help me falsify this zero thumb theory, or zero sum theory, as it relates to chatbots, and kind of like, I know I've heard you say, and I've heard just Google as a whole say, that AI overviews and AI as a whole, it kind of makes people search more, and in a way we're seeing Jevan's paradox where as there becomes more data, there becomes more of an inkling to search more, to create more queries, and I kinda want you to help help us understand that phenomenon, I guess, where in the case that there is AI, well, there's actually a lot more search happening.

搜索数量并没有下降，实际上是在上升。

The the number of searches don't go down, they actually go up.

是的。

Yeah.

所以我认为AI是一种扩张性的东西。

So I think AI is is an expansionary thing.

就像人们有各种各样的问题，他们本可以问更多，但由于种种限制而没有问。

It's like people had all these questions and they could ask so much more, but they they didn't because there were limitations.

因此我认为最好的例子就是类似Lens结合AI这样的应用，这是最佳方式。

And so I think the best example of that is something like Lens and with AI, which is the best way.

对吧？

Right?

现在你可以拍一张书架的照片然后说，我已经读过这些书了。

So you can now take a picture of your bookshelf and say, I've given these books.

接下来我该读什么？

What should I read next?

或者你可以截屏某个名人的穿搭然后问，哪里能买到这件夹克之类的东西？

Or you can take a screenshot of, you know, a celebrity outfit or something and say, where could I buy this jacket or something?

十年前可能也有人尝试在谷歌上输入这类问题，但当时要做到这点相当困难。

It's possible someone would try to put that kind of a question into Google ten years ago plus, but it'd be pretty hard to do that.

同理，要让人输入一个长达20个句子的问题也是相当困难的。

In the same way, it would be pretty hard for someone to type in like a 20 sentence question.

就像说，我要去旅行了。

That's like, I'm going on this trip.

我有个孩子。

I have a kid.

这孩子有这种过敏症。

The kid has this allergy.

他们有这样的需求。

They're like, this need.

我得去住酒店。

I have to go to a hotel.

酒店需要满足这些条件——以前根本没法这样查询。

Hotel needs to be far you just like couldn't do that.

所以人们往往就会放弃或者不去尝试。

And So people would just kind of give up or not do it.

当你为用户解锁新需求时就会带来增长，这正是我们当前看到的情况。

So you get growth when you're unlocking new needs for people, and that's what we're seeing.

我认为最好的总结方式是：人们通过搜索获取快速高效信息的基本需求并未改变，但现在可以向技术提出更多问题，这正是增长来源所在。

So I think the best way I can summarize it is the everyday need I just mentioned for people getting fast, efficient information from search isn't really changing, but now you can ask technology so many more questions, and that's where the growth's coming from.

要知道，我们讨论过这个，比如AI概览功能，我们确实看到了增长。

Know, and we talked about this, like, you know, AI overviews, you know, we're seeing growth.

我们讨论过AI概览功能出现的场景。

You know, we we talk about where AI overview shows up.

它通常出现在用户提出更具体问题时，尤其是那些更详细、更具体的查询。

It's around when people have a more specific question, but they put like a longer, more specific question.

当用户把这些输入谷歌时，就会获得AI生成的回答。

They put that into Google, they get this AI response.

这类查询在美国等主要市场增长了约10%，以谷歌的规模而言，这样的年增长率相当可观。

Those kinds of questions are up about 10% in in large markets like The US, which at Google scale is a pretty enormous kind of growth number year over year.

而我提到的这类视觉搜索功能，年增长率更是达到了70%。

And then these kind of visual searches I mentioned are up 70% year over year.

因此在这些市场扩张最快的领域出现了巨大增长，这正是我们观察到的现象。

So huge growth in these areas where the market's expanding the fastest, and that's exactly what we're seeing.

但显然，市场蛋糕正在扩大，人们有大量机会从多方获取信息，这令人振奋。

But obviously, the pie is growing, there's lots and lots of opportunities for people to get information from lots of people, and that's exciting.

我很好奇。

I'm curious.

在构建这一愿景时，你们对消费级硬件设备有多重视？

How much importance do you place on consumer hardware devices when you're kind of thinking about building out this vision?

你知道，Meta一直在尝试通过眼镜重塑设备形态。

You know, Meta's been attempting to, you know, revamp the form factor with glasses.

有传言称OpenAI正在打造自己的消费级设备。

OpenAI is rumored to be building their own consumer device.

天知道那会是什么样子。

Goodness knows what that looks like.

当然，谷歌很重视与苹果的合作关系，通过Safari提供主要搜索功能。

And, you know, obviously, Google values their partnership with Apple being the primary search function through Safari.

你们有考虑过这方面吗？

Is this something you consider at all?

这是谷歌想要自己掌控并塑造的东西，还是说更像是‘让我们静观其变，让搜索功能自然融入各处’？

Is this something that Google kinda wants to own and kind of, like, form themselves, or is it just kind of, like, let's see what happens and search will be integrated everywhere?

我认为搜索已经无处不在，我们视其为一种服务，应当能在人们需要帮助的任何设备或场景中触手可及。

I mean, think search is so ubiquitous that we think of it as a service that should just be accessible to any device or context that someone needs help in.

无论你是用手机时对所见内容产生疑问，还是在拍摄某物时——谷歌希望人们能在任何情境下获得谷歌的价值。

You know, whether you're on your phone and you have a question on what you're looking at, when you're taking a picture of something, like Google wants to we want people to get the value of Google in whatever that context looks like.

我个人非常认同‘最好的相机就是你口袋里的那部’这种说法。

I personally am a big fan of the the kind of saying the best camera is the one in your pocket.

嗯。

Mhmm.

我觉得这是个非常贴切的比喻。

I think it's just like a really apt point.

因此，未来可能会涌现许多酷炫的新事物。

And so, like, there's lots of cool new things that will probably be created.

但这可能需要时间。

They'll probably take time.

但真正重要的是人们日常使用的便利性。

But really, there's just the convenience of what people use every day is important.

懂吗？

Know?

人们依赖某些技术，依赖特定的提问方式。

People are people rely on certain technologies, certain ways of asking questions.

就像他们的相机一样。

Similar to their camera.

比如，他们有相机，随身携带，并且经常拍照。

Like, they have a camera, they have it in their pocket, and they're constantly taking photos.

所以如果你能贴近这种体验，那就太好了。

And so if you can be close to that experience, great.

这意味着你只需轻点一下就能把照片发送给谷歌，比如说，然后询问你看到的是什么。

That means that you can be one tap away from sending your photo to Google, let's say, and asking a question what you're looking at.

我认为这就是很多人想要与技术互动的方式。

And I think that's how a lot of people wanna interact with with technology.

显然，我认为搜索领域一直渴望融入每一个令人兴奋的新技术交互方式，而现在确实出现了一个激动人心的突破性硬件类别。

And I've some obviously, there's an exciting breakout hardware category, you know, that would be something that I think search historically has always wanted to be a part of every new exciting new way that people are interacting with technology.

罗比，我们的观众和听众最常问的一个问题是：这听起来太神奇了。

Robbie, the the number one question we get from viewers of our and listeners of our show is this sounds magical.

不过，这是怎么运作的呢？

How does this work, though?

帮我理解一下，背后的原理是什么？

Help me understand, like, what's happening under the hood.

我想把这个问题指向Gemini搜索和谷歌搜索，包括你刚才解释的所有AI概览功能。

I would like to point that question towards Gemini search and Google search in general with the AI overviews and everything you've explained so far.

能否给我们一个高层次的解析，说明背后的运作机制？

Can you give us maybe a high level breakdown as to what's happening under the hood?

在谷歌，我们在AI领域的努力方向是打造世界上知识最渊博的人工智能。

So at at Google, what we've done with AI is we've really tried to create the world's most knowledgeable AI.

这个AI不仅要深度连接谷歌的海量信息，还要整合整个网络上的知识。

So one that really is connected to the vast information that obviously, like, sits at Google, but also around the web.

对吧？

Right?

就像，关键在于连接和理解这个世界，这是独一无二的。

Like, it's about connecting and understanding the world, and that's unique.

因此我们有机会打造一个AI，要知道，购物目录中有数十亿商品，每天有数亿商家在更新本地信息。

And so we have an opportunity to have an AI that, you know, obviously, there's there's, say, billions of products in the shopping catalog, hundreds of millions of places that businesses are updating with local information every day.

知识库中实时更新着数万亿条事实数据。

There's trillion facts in the knowledge base updated all the time.

还有实时金融数据、体育赛事、旅行信息、航班动态价格等各类资讯。

There's information about live finance data, sports, travel information, live prices for flights, all that stuff.

我们希望让每个人都能轻松快捷地获取这些信息。

Like, we want to be able to make that easily and quickly accessible to anyone.

当然还有我们想为你连接的浩瀚网络资源。

And then you obviously have the vastness of the web that we wanna connect you to.

所以我们构建的模型——基于Gemini基础架构的AI模型，本质上是个能理解自然语言和多模态问题、生成回答，并能消化所有知识的大型语言模型。

And so the models that we've built, there's an AI model that is kind of based on Gemini's foundational model, which is the one that kind of under is a large language model that understands, you know, natural language and multimodal questions and is able to generate responses, is able to also understand all that knowledge.

因此，当你提出一个问题时，后台的AI模型会开始生成谷歌搜索来进行相关研究。

And so you can ask a question, and what'll happen on the back end is the AI model will start to actually generate Google searches to start researching that.

根据问题的复杂程度，它可能会花时间进行思考、推理和研究。

And given the complexity of the question, may actually spend time thinking and reasoning and doing research.

比如你询问应该选择哪种太阳镜，想了解偏光与非偏光镜片的区别及其优点，可能会衍生出数十个相关问题。

And so if you ask a question about what kinds of sunglasses you should get and learning more about polarized versus not and its benefits, there might be dozens of questions connected to that question.

在底层机制中，模型实际上正在进行搜索。

And what would happen under the hood is that the model's actually searching.

它就像人类一样发起大量谷歌搜索，并可能同时使用购物图谱等API进行研究。

Mean, it's super it's issuing a bunch of Google searches as if a person would, and it's potentially using APIs like the shopping graph to do research as well.

随后它会检索所有相关信息，借助搜索领域的知识和信号，对特定问题所需的最佳信息形成准确判断。

It would then retrieve all of the relevant information and then because of, you know, all of the search knowledge and signals that that are available in search, kind of a good understanding of what information is great for a given question.

所有这些信息都会被反馈给模型，模型经过推理后生成包含深入链接的回应，你可以通过这些链接了解更多信息，甚至购买所需商品，继续探索之旅。

All of that is is brought back into the model and the model reasons about it and and generates a response with links to dive in, learn more, potentially buy the thing you're looking for, and continue your journey.

这一切都是通过这个AI体验实现的。

And and that all happens, you know, through this AI experience.

这基本上就是本次AI概览中展示的主要内容。

And so this is this is largely previewed in this AI overview.

如果你现在输入一个难题，比如‘如何去除浅色白沙发上的番茄酱污渍？’

So if you just put a hard question right now, you go, how do I get ketchup stains out of a light white couch?

在谷歌搜索这个问题，你很可能会在顶部看到一个小型的AI预览——AI概览。

Put that into Google, you're very likely to get a little AI preview at the top, AI overviews.

如果展开并点击AI模式，你就能进行完整的对话交流。

And if you were to expand it and click AI mode, you can have a whole back and forth with that.

而实现这一切的正是这个AI模型。

And and that's the model that's doing that.

立即吸干多余番茄酱，罗比，然后涂抹冷

Block the excess ketchup immediately, Robbie, and apply solution of cold

水和洗洁精的混合溶液。

water the mold dish soap.

我现在真的说得太久了。

I literally am now too long.

罗比，我得坦白一件事，其实我有点像是你的粉丝，直到最近才意识到这一点，因为你曾在Instagram工作过，而且你刚才提到'最好的相机就是你口袋里的那部'。

Robbie, I have to I have to confess something, which is that I I'm a bit of a fanboy and I didn't quite realize until recently because you had a stint at Instagram, and you just mentioned a little earlier the best camera is the one in your pocket.

我热爱摄影，喜欢拍照，尤其钟爱Instagram故事功能，因为这是我最喜欢的分享内容方式

I love photography, I love taking photos, and particularly I love Instagram stories because it's my favorite way of sharing content

在世界上

in the world

在互联网上。

on the internet.

后来我发现，你就是Instagram上实现这一功能的关键人物，这太了不起了。

What I found out is that you are the guy responsible for implementing that at Instagram, which was amazing.

这是个绝妙的产品应用，可能也是我至今仍频繁使用Instagram的主要原因

It's such a great product application, and probably the single reason why I still use Instagram for a lot

多年。

of years.

有很多人共同努力才让这个功能取得成功，但我也很荣幸能参与其中。

Lot of people worked on making that successful, but I was very privileged to have a chance to work on that as well.

是的，它

Yes, it

确实是个非常酷的功能。

was Oh, really a really cool feature.

所以我想请教你的是，在涉及搜索功能时，你是如何考虑实现这类功能的？因为我觉得这并非原创概念。

So what I kind of wanted to ask you about that is how you think about implementing features like that kinda as it relates to search, because I imagine this wasn't an original idea.

众所周知，Snapchat最早推出类似功能，但Instagram的版本我认为做得更好，这也是我现在使用的版本。

Famously, Snapchat kind of had it first, but then Instagram implemented it, what I believe to be much better, and that's the one that I use.

我记得在另一次采访中你解释过思考过程，比如Snapchat不允许上传已有照片，必须使用内置相机拍摄，体验很糟糕。

And I remember hearing you on another interview kind of describing the way you reasoned through, which is like, on Snapchat I couldn't upload my own photos, had to use the inline camera, and that was just a poor experience.

但在Instagram上可以上传我自己觉得更精美的照片，我更喜欢这种方式。

But on Instagram, could upload my own photos that I thought were much more beautiful, and I much prefer that.

所以我很好奇这背后的思考过程，以及你如何将这种思路应用到像谷歌这样的公司，现在你正在为这项新AI技术开发产品。

So I guess I'm kind of wondering the thought process that was behind that, and how you apply that to companies like Google, where now you're developing product for this new AI technology.

是的。

Yeah.

有趣的是，我觉得从那次经历中能学到很多关于产品构建的相似之处。

Interestingly, I feel like there's a lot of similarities in terms of what to learn about product building from that experience.

我认为最主要的教训是，如果一个产品深受大众喜爱并被广泛使用，你就不应该彻底颠覆它，因为用户已经形成了一种自然的、习以为常的使用路径，每天都有数十亿次这样的行为。

And I think, you know, the main one is is that if you have a product that's beloved and used by lots and lots of people, you don't wanna dramatically upend that on people, because there's a there's a natural, well worn path that people are traveling every day, billions of times a day.

你不能某天突然出现，让用户感觉世界都颠倒了。

And you don't wanna just show up one day and just have it feel like upside down world.

那样对任何人都没有好处，只会制造一堆问题。

Like, that's just not a service to anyone and will create a bunch of problems.

话虽如此，如果你正在开发的领域与用户原本的需求直接相关——比如谷歌的信息服务，但用户希望能做得更多——

Now that said, if if you're building in a space where your need for your product is directly connected to what people already come for, in this case, information at Google, but people just wanna do more for it.

那就存在一个自然的机会来扩展服务范围，就像人们来到Instagram是为了分享照片，而后来发现通过'故事'功能与朋友互动可能是更好的方式，因为这种轻松短暂的格式允许你收到私信、与朋友愉快交流并保持联系，整个系统运作良好，但它并没有取代Instagram原有的功能。

There's a natural opportunity to expand what you can do for people in the way that people came to Instagram to share through photos, and there's turns out there was a potentially even better way to do that for friends through stories because it was this low pressure kind of ephemeral format and allowed for this you to get a DM and have a fun conversation with your friends and feel connected and that whole system really worked well, but it didn't replace Instagram.

它成为了Instagram帮助用户的另一种方式。

You know, it it became an additional way that Instagram could help you.

我认为AI与搜索的关系也是如此。

And I think of AI and search in the same way.

人们每天数十亿次地使用谷歌获取信息，实际上在AI出现之前，人们就已经尝试在搜索框中输入各种奇怪的内容了。

People come to Google every day for information, billions and billions of times a day, And actually people have tried typing in crazy stuff into search even before AI existed really.

而以前你确实做不了什么，我甚至可能翻到搜索结果页的底部，告诉你我们找不到任何相关信息。

And before you couldn't really do much, you know, I might even get to the end of the search results page and say we couldn't find anything.

但现在你几乎可以帮忙解决任何问题。

But now you can really help with almost anything.

所以这感觉是件自然而然的事。

And so that feels like a natural thing to do.

但基于同样的经验教训，你必须真正为用户需求而设计。

But from the same learning, you have to really design for the needs of your user.

就像搜索领域一样，还记得你问今天发生了什么时，模型过去会回答'我不知道'吗？

And so in the same way for search, like, you know, remember when you asked the question about what was happening today and models used to be like, oh, I don't know.

它可能会说'我的知识截止到一年前'之类的话。

I was trained up until a year ago or something.

这对我来说总是显得很荒谬——仅仅因为技术发展路径，你竟然无法实时获取世界正在发生的事，哪怕只差100毫秒。

It's like, it just like, like, it always seemed crazy to me that you couldn't get information within a 100 milliseconds of what was happening in the world just because of the way technology evolved.

但现在，你知道，这已经发展得非常不同了，特别是在谷歌，它能在近乎实时的基础上从我们所有知识库中查找信息。

But now, you know, this has evolved to be very different, but particularly in Google, it's finding information in near real time basis across all of our knowledge.

所以我们认为这是我们能做得特别出色的地方。

So that's something that we think we can do uniquely well.

另一个例子是关于视觉和灵感的。

And, you know, another example is around visual and inspiration.

人们总是来搜索。

People come to search all the time.

他们会搜索图片。

They search for images.

他们会使用图片搜索，这本身就是一个巨大的搜索引擎。

They go to Image Search, which is a huge search engine in of itself.

但人们也会寻找设计。

But people look for design.

他们想要壁纸。

They want wallpapers.

他们想要灯光设计的创意灵感。

They want, uplighting ideas.

他们想重新装修孩子的卧室，于是浏览这些图片。

They want to redecorate their kids' bedrooms, and they browse for these images.

如果你问AI这类问题，它会用文字描述如何设计卧室，我一直觉得这很奇怪。

And if you ask AI these kinds of questions, it'll, like, describe in text how to design a bedroom, which I always thought was really weird.

现在有了视觉AI模式，你可以询问'帮我设计我女儿的卧室'，寻找灵感和创意。

And so now with Visual AI mode, you can ask, help me design, you know, at my daughter's bedroom, looking for ideas, looking for inspiration.

可以是关于任何主题的。

Could be about, you know, anything.

你可以在选购时尚连衣裙时，AI模式会找到启发性的图片，然后你可以进行多轮对话。

You could be shopping for fashion dresses, and the AI mode will actually go find inspirational images, and then you can have a multi turn conversation.

你可以说'实际上，我想要极繁主义的暗色调和超级忧郁的主题'。

You could say, actually, I want, you know, maximalist dark tones and super brooding theme.

它能理解这种需求，并运用我们大量的镜头和图像处理技术。

It will like know what that means and using a lot of our lens and and technology for imagery.

将整个网格从加州那种通透轻盈的风格转变为这种阴暗木屋氛围，它能在视觉上理解这种转变的含义。

Go change the whole grid from something airy and light in California to like this dark lodge vibe, and like it knows what that means visually.

我认为这些方式能让谷歌根据用户需求，为世界增添独特价值，而不是仅仅试图打造另一个通用聊天机器人，这并非我们的初衷。

And I think these are ways that Google, based on what Google users need, can add needs, can add unique value to the world, you know, versus just trying to implement another kinda general purpose chatbot, which isn't what our intention is.

我很好奇谷歌的独特优势是什么，以虚拟背景为例，我用它是因为墙上什么都没有。

I'm curious to understand what the advantage that is uniquely Google's because to that example, the reason I'm using a virtual background, I have nothing on my walls.

很希望能在这方面得到些帮助。

Love some help getting some assistance on that.

我知道谷歌擅长这个，但我想了解为什么谷歌在这方面具有独特优势。

And I understand Google's good for that, but I'd love to kind of understand why Google's uniquely good to that.

因为如果我询问其他模型，比如x AI或Grok，它们实际上会去搜索互联网——而我认为大部分内容都是谷歌索引的。

Because if I ask another model, if I ask x AI or Grok, they'll they'll actually go and search the Internet, which I assume is mostly indexed by Google.

谷歌作为谷歌本身，相比需要向谷歌发起查询，是否存在独特优势？

Is there a unique advantage to Google being Google versus having to query against Google?

比如独特的数据集，谷歌特有的用户画像索引方式，这些是否让你们在同领域公司中脱颖而出？

Like, the unique dataset, the unique kind of profile indexing that Google does that separates you from a lot of other companies in the same space?

是的。

Yeah.

我是说，我认为有很多因素让我们在这些情况下能提供真正独特的帮助。

I mean, there's a bunch of things that I think allow us to be really uniquely helpful in these cases.

比如，我认为其中之一就在于技术本身及其输入方式。

I mean, I think one is just in the technology and the inputs itself.

我们在图像识别和视觉理解的多模态能力建设上已经投入了很多很多年。

Like, there's been many, many, many years of building multimodal capabilities for image recognition and visual understanding.

所以我们的模型能够分割你体验中的背景，将注意力集中在物体的正确部分。

So our models are able to segment the background of your experience, put attention on the correct parts of the object.

比如你说'嘿，我身后架子上的那棵小树或者那株小植物怎么样'。

If you were to say, like, hey, like, what's up with the you know, I want, like, a the the little tree behind me on the ledge or the the little, you know, plant.

'架子是什么？'

Like, well, what's a ledge?

明白吗？

You know?