谷歌CEO桑达尔·皮查伊：AI的无限可能、算力稀缺困境与未来AI产品

本期节目由Genesis赞助播出。

This episode is brought to you by Genesis.

要让智能体AI真正在整个企业中发挥作用，到底需要做些什么？

What does it actually take to put agentic AI to work across an entire enterprise?

我最近和Genesis的董事长兼CEO托尼·贝茨聊了聊，为什么编排正在成为客户体验领域的竞争优势，以及企业在AI时代需要做好哪些事情。

I recently spoke with Genesis Chairman and CEO Tony Bates about why orchestration is emerging as a competitive advantage in customer experience, and what companies need to get right in the AI era.

之后，我和Voya金融企业业务解决方案负责人亚当·米切尔坐下来聊了聊，深入探讨了一次大规模的智能体客户体验转型，包括其中的业务决策、取舍和一路走来的经验教训。

Then, I sat down with Adam Mitchell, head of Enterprise Business Solutions at Voya Financial, to go inside a large-scale agentic CX transformation, including the business decisions, trade-offs, and lessons along the way.

你现在就可以在我的YouTube频道上观看这两场对话。

You can watch both conversations now on my YouTube channel today.

获取前沿计算能力从未像现在这样困难。

It's never been more challenging to get frontier compute.

你怎么知道该去哪里找、该信任谁，或者该付多少钱？

How do you know where to look, who to trust, or how much to pay?

Akash是一个开放云，提供便捷访问、灵活性和安心保障。

Akash is an open cloud that provides easy access, flexibility, and peace of mind.

这就是像雷蛇这样的公司能把推理成本降低15倍的原因。

That's how companies like Razer were able to cut their inference cost by 15 times.

访问akash.network/alex获取免费咨询。

Visit akash.network/alex for a free consultation.

对，就是akash.network/alex，可以免费咨询。

That's akash.network/alex for a free consultation.

我们正处于有史以来最具影响力技术的早期阶段。

We are at early stages of the most consequential technology ever.

这不像光纤，它不是黑暗的。

This is not like fiber, it's not dark.

我们正在使用它。

We are using it.

对于AI研究人员来说，计算机越大，AI就越好。

AI researchers, for them, it's like the bigger the computer, the better the AI.

你给研究人员最大的计算机，即使这会损害今天的利润，比如云业务。

You give the researchers the biggest possible computer, even if it is to the detriment of today's profits, like the cloud business.

所以我们今天拥有的算力，是我在2024年做的决定的结果。

So the compute we have today is a function of decisions I made in 2024.

今天，显然，当你坐下来看任何事情时，每个人都很短缺，不只是我们，其他公司也一样，每个人都希望你有计划的两倍。

Today, obviously, when you sit and look at anything, everyone is short, not just us, other companies, and everyone wishes you had 2X of what you had planned for.

谷歌的Muse或Dots产品在哪里？

Where's Google's Muse or its Dots product?

用户，无论是消费者还是企业，都将能够使用某种多智能体框架，让它们全天候工作，处理各种事务。

Users, both in consumer and enterprise, will have access to some multi-agent framework for them working all the time, doing things.

所以，这感觉像是一个充满无限可能的时刻。

And so, it feels like a moment of infinite possibilities.

Alphabet首席执行官桑达尔·皮查伊加入我们，讨论谷歌在AI前沿竞争的计划，尖端AI研究与公司云业务之间的平衡，以及未来界面会是什么样子。

Alphabet CEO Sundar Pichai joins us to talk about Google's plan to compete on the frontier of AI, the balance between cutting-edge AI research and the company's cloud business, and what the future of the interface looks like.

今天，桑达尔就在山景城的工作室里和我们在一起。

And Sundar is here with us in studio in Mountain View today.

桑达尔，很高兴见到你。

Sundar, great to see you.

欢迎来到我们的节目。

Welcome to the show.

谢谢，亚历克斯。

Thanks, Alex.

很高兴来到这里。

Great to be here.

所以谷歌凭借Gemini 4 Argon重新回到了前沿。

So Google is back on the frontier with Gemini 4 Argon.

简单来说，这意味着它是一个世界领先的模型。

In plain speak, that means it is a world-leading model.

它打破了基准测试的记录。

It's benchmark breaking.

它在各种基准测试中击败了OpenAI和Anthropic的公开模型，包括编程、数学、科学和知识工作。

It beats OpenAI and Anthropic's public models on a variety of benchmarks, including on coding, math, science, and knowledge work.

之前大家普遍认为谷歌基本上已经放弃了前沿，只打算做那些够用的快速模型，用在自家产品里，不再追求前沿AI了。

Now, the conventional wisdom had been that Google had basically given up on the frontier, that it was going to make these fast models that were good enough and use them in its products, and it wasn't pursuing frontier AI anymore.

但事实并非如此。

But that was wrong.

完全不是。

Very wrong.

你看，我觉得我们非常执着于站在前沿。

Look, I think we are so committed to being at the frontier.

我觉得这已经融入了我们的基因。

I think it's in our DNA.

我们不会以其他方式行事。

We won't do it any other way.

我们已经思考AI前沿超过十年了。

We've been thinking about the AI frontier for well over a decade now.

前沿领域变化非常快。

The frontier is very dynamic.

竞争也非常激烈。

It's very competitive.

这非常令人兴奋。

It's very exciting.

我对Argon的整合成果再满意不过了。

I couldn't be more pleased with how Argon has come together.

我们花了时间慢慢打磨。

We took our time with it.

我们确保进行了安全测试，并谨慎部署。

We made sure we tested it safely and deployed it carefully.

考虑到它的能力，我们比其他模型花了更长的时间。

We took a bit longer than our other models, given its capabilities.

但我觉得无论是智能体工作、知识工作、编程还是创意任务，Argon都是顶尖的前沿模型之一。

But I think across the board, be it for agentic work, knowledge work, coding, creative tasks, I think Argon is one of the really good frontier models.

而且自发布以来的头72小时内，它已经使用了近25万亿个token，这大约是Gemini 3.0头72小时使用量的3倍。

And in the first 72 hours since we launched, it's almost up to 25 trillion tokens used, which is about 3X what Gemini 3.0 had in the first 72 hours.

我们在提供服务时也遇到了容量限制。

And we are capacity constrained as we are serving it.

所以，虽然很兴奋，但总有一些领域，前沿技术本身也就那样。

So, excited, but there are always areas, both the frontier is just less.

这个领域变化非常快，你必须不断更新模型。

It's very dynamic, and you have to keep evolving the models.

而且我们也看到一些需要改进的地方，所以我们会继续在这些方面努力。

And there are areas which we can see we need to improve as well, so we'll be at work on all of that as well.

哪些方面？

Which areas?

你看，就拿任何一款模型来说，如果你只看两个月前的版本，Argon在各方面都会更好。

Look, on any given model, Argon will be across the board better at everything if you just look even two months back.

但这个领域发展太快了。

But the space is so evolving.

不过即使在编程方面，也有一些领域和薄弱环节，我们知道需要做得更好，而且有明确的反馈，所以我们会改进它。

But even on coding, for example, there are areas, pockets, where we know we need to do better, and there's clear feedback, and so we'll make it better.

让我至少为谷歌辩护一下，为什么标准模型可能就足够了。

Let me at least make the argument for why standard models might be good enough for Google.

我来转述一下大家是怎么说的。

I'm going to channel what people say about this.

不。

No.

就是说，现在AI的前沿技术已经非常强了，基础模型也很强，所以前沿模型其实只对特定任务特别有价值。

And that is that the frontier of AI today is getting so good, and the baseline of AI is getting so good, that the frontier is really only valuable for a specific set of tasks.

如果你有一个标准模型，你其实可以用它做出很好的产品。

And if you have a standard model, you could actually build great products with a standard model.

比如Meta的Muse。

Take Meta Muse, for example.

Meta是用一个不错的模型做的Muse，但不是最好的模型。

Meta built Muse on a good model, but not the best models.

如果你要做生物研究，可能前沿模型确实很必要。

If you need to do biological research, maybe the frontier is really necessary.

但那些会说“Google其实不需要前沿模型”的人，他们会说：“你只需要那些标准模型，然后把它们用到地图、邮箱和搜索这样的产品里就行了。”

But the people who would say, "Google doesn't really need the frontier," they would say, "You just need those standard models, and you can put them in products like Maps and Gmail and Search."

再说回Meta那个例子，那就足够了。

And again, take that Meta example, that would be sufficient.

你怎么看？

What do you think about that?

说实话，我完全不是这么想的。

Look, it's definitely not how I have thought about it at all.

我觉得我们非常致力于推动最前沿的技术，真正去探索这项技术能走多远、能去哪里，从Transformer开始一路走来。

I think we are deeply committed to driving the state-of-the-art and actually push what this technology can go, where it can go, all the way from transformers.

这是我们一直在走的路。

That's a journey we've been on.

所以你会看到我们继续这么做。我觉得当你处在最前沿时，你会比别人更早看到那些新兴能力，这有助于你更好地为未来做准备。

So you will see us continue to do that I think when you're at the frontier, you get an earlier look into the kind of emerging capabilities earlier than others, which helps you better prepare for the future.

我们很幸运，在消费端和企业端都有很多机会。

We are fortunate to have a wide set of opportunities across consumer and enterprise.

我们的一些云客户当然想要最前沿的模型。

Some of our cloud customers for sure want the model at the frontier.

我觉得前沿模型在递归自我改进的时代，会更擅长开发下一代模型。

I think the frontier models will be better at developing the next set of models in an era of recursive self-improvement.

所以我认为这一点很重要。

So I would argue it matters there.

在内部，我们通过向开发者部署Argon，确实看到了一个明显的拐点。

Internally, we have definitely seen a steep point of inflection by deploying Argon to our developers.

它确实解决了我们之前的模型无法解决的问题。

It definitely tackles problems which our previous models could not do.

所以站在前沿，就意味着站在创新的前沿，也能推动创新。

And so being at that frontier puts you at the innovation frontier and helps you drive that as well.

各种规模和能力的模型都有其存在的空间。

There is room for models of all sizes and capabilities.

我们一直是这样做的，所以推出了Flash、Flash Light、Flash等系列。

We have always, that's why we did Flash, Flash Light, Flash, et cetera.

我同意你的看法，不一定非要使用最前沿的模型才能打造出非常好的产品体验，因为那些模型往往更慢、也更贵。

And I agree with you that you can develop very good product experiences with not necessarily the most frontier model because they tend to be slower and more expensive too.

所以我觉得，但我们很在意前沿模型的进展，我们致力于把这一切都做好。

So I think, but we care about the parade of frontier and so we are committed to doing all of that well.

你说过，开发前沿模型能让你看到未来的趋势。

You said that developing the frontier gives you a look into what's coming.

那你觉得未来会有什么？

What do you see on the horizon?

最让人兴奋的是，你辛辛苦苦做了几个月，终于推出一个产品，但你已经看到前面还有很大的提升空间，对吧？

The exciting, what makes this moment so exciting is you build something which you think is a culmination of months of work, and you put it out, you know you already see clear headroom ahead, right?

我觉得，前沿模型会朝着多个智能体协作的方向发展，我们一直在研究这些长周期的智能体能力，但显然，未来的前沿模型会让多个智能体一起合作完成任务，对吧？

I think, the frontier is going to, we've all been working on these agentic long-horizon capabilities, but clearly the frontier is going to have multiple of these agents working together to do things, right?

这就是一个令人兴奋的新范式的例子。

And so that's an example of an exciting paradigm underway.

而且我也很清楚，无论是消费者还是企业用户，都会拥有一个多智能体框架，让它们全天候工作，处理各种事情。

And it's also clear to me that users, both in consumer and enterprise, will have access to some multi-agent framework for them working all the time doing things.

所以，这感觉就像是一个充满无限可能的时刻。

And so, it feels like a moment of infinite possibilities.

你提到了递归自我改进。

You mentioned recursive self-improvement.

这就是模型有效地让自己变得更好——

This is the models effectively making themselves better-

……也许没有人类参与其中。

... maybe without humans in the loop.

人们称之为智能爆炸，或者期待听到像“奇点”这样的说法，然后突然之间，AI就拥有了这种超人般的能力。

People call it like an intelligence explosion or expect to hear this sound like foom, and then all of a sudden, AI is just like this superhuman capability.

事实上，我记得谢尔盖在DeepMind和研究人员进行了一次圆桌讨论。

And in fact, I think Sergey was doing this roundtable with researchers at DeepMind.

他谈到了Argon，说Argon的核心创新之一就是更多地依赖模型来帮助开发模型，对吧？

And he talked about with Argon, that one of the core innovations for Argon was relying more on the model to help develop the model, right?

我们最近看到了AI模型的发展情况。

We've seen what's gone on with AI models recently.

它们对齐得很好，但并非完美对齐，而这些不完美之处是个大问题。

They are well-aligned, but they're not perfectly aligned, and the imperfections are a big problem.

显然，科技界，现在更广泛的世界，都知道Hugging Face事件，OpenAI的机器人突破了限制，入侵了Hugging Face，做了一些疯狂的事情。

Obviously, the tech world, and now the world more broadly, knows about the Hugging Face incident where OpenAI's bots broke containment and hacked into Hugging Face and did some crazy stuff.

谷歌也遇到过类似情况，它的机器人获得了访问权限。

Google has also had a scenario where its bots got access.

我想它们是在猜密码，然后进入了一个本不该进入的网站。

I think they were guessing passwords, and they got access to a site they shouldn't have.

它们停了下来。

They stopped.

但这些模型，有足够的证据表明它们并没有完全按照我们的意愿行事，所以追求递归自我改进，或者更快地改进模型，这种想法有点吓人。

But these models, there's enough evidence that they're not doing exactly what we want them to be doing, that this idea of pursuing this recursive self-improvement or even making the models better much faster, it's sort of freaky.

这让我感到害怕。

It makes me scared.

你不觉得这种担忧是合理的吗？难道不应该采取一些措施让科技行业冷静下来吗？

Don't you think there's reasonable worry here and there should be something done to have the tech industry chill out about this?

嗯，你的问题可能分三部分。

Well, maybe three parts to your question.

首先，咱们先聊聊RSI本身，然后——

First, let's talk about RSI itself and-

你说得对，Sergey提到RSI是开发Argon时的关键推动因素。

... you're right, Sergey mentioned RSI being critical contributor as we are developing Argon.

但要说清楚，今天，在发展的过程中，我们离RSI的最终状态或类似的东西还差得很远，对吧？

But to be very clear, today, in a continuum of, we are pretty far from the end state of RSI or something like that, right?

所以今天，正确的描述方式是，在模型开发的各个方面，RSI循环和AI研究人员是并存的，对吧？

So today you have the right way to describe it is you have RSI loops along with AI researchers in various aspects of model development, right?

所以，我可以说，几乎有智能体以某种自动化的方式帮助AI研究人员一起构建这些模型。

So, I would say you almost have agents helping AI researchers, in some automated way, work together to build these models.

所以我觉得至少了解我们现在的状况是好的，而且这非常令人兴奋，我预计随着我们进入明年，这个过程会不断取得进展。

So I think it's good to at least understand where we are today, and it's super exciting, and I expect that continuum to keep making progress as we go into next year.

第二个问题是关于安全性的。

The second question is around safety.

你看，我们的方式，安全性来自于纵深防御。

Look, the way we are, security comes from a defense in depth.

你说的是对齐问题。

You're talking about alignment.

显然，我们都在做很多工作，确保模型尽可能对齐，但这是一个活跃的研究领域，而且还有明显的差距。

Obviously, we all are doing a lot of work to make sure the models are as aligned as possible, but it's an active area of research, and there are clear gaps.

但这并不是我们唯一依赖的东西，对吧？

But that's not the only thing we rely on, right?

所以，你们在对齐方面做了工作，但除此之外，我们在模型运行时的实时监控上投入了大量精力。

So you have the work you do on alignment, but beyond that, we put a lot of effort into real-time monitoring as these models are working.

也就是说，要确保对正在发生的事情有良好的遥测，这样如果出现意外行为，就能实时干预。

So making sure you have good telemetry into what's happening, and so that you can intervene in real time if you see unexpected behaviors.

而你们如何做到这一点，采用什么样的架构，如何确保一切都处于沙盒和隔离状态，以便在实时监控期间继续正常工作？

And how you do that, the architecture with which you do that, how do you make sure it's all sandboxed and isolated in a way that it can continue doing its job during real-time monitoring?

这些都是我们正在投资的领域。

These are all areas of investment we are doing.

显然，当我们部署这些模型时，给客户和企业提供控制权，包括实时监控，以及他们决定在哪里部署、模型能访问什么类型的数据等等。

And obviously, when we deploy these models, giving customers, enterprises Control, both with real-time monitoring, but their ability to decide where they deploy, what type of data these models ever have access to, et cetera.

所以这些工作都非常重要，我们和Argon在这方面做了很多。

So all that work is very, very important, and we are doing a lot of that with Argon.

所以我觉得作为行业，我们需要不断提升自己，花时间把这些做好，包括投入必要的计算资源，确保这些模型能安全部署。

So I think as an industry, we need to constantly up our game and take the time to do all this well, including dedicating the compute required to make sure you can safely deploy these models.

第三部分是，这些够了吗？

The third part is, is it all enough?

所以我觉得这就是为什么行业里有很多讨论，担心是合理的，但对待担忧的正确方式是尽早思考还能做些什么。

And so I think this is why I think there's a lot of conversations in the industry, and I think it's reasonable to be worried, but the right way to approach something you worry about it is early on, thinking about what more we can do.

我觉得我们所有人，作为行业，确实看到在部署这些模型时，对齐、安全测试等方面的工作量在增加。

And I think all of us are, I think as an industry, I think you're definitely seeing an increased levels of work on alignment, testing for safety, et cetera, as we deploy these models.

我觉得下一步是制定更好的标准，形成跨行业的标准，包括用独立评估机构来评估。

I think the next step would be developing better standards so there are cross-industry standards of how we do these things, including ability to evaluate using independent evaluators.

看看金融控制领域，我们有会计准则，有独立审计师等等。

If you look at areas like financial controls, we have accounting standards, we have independent auditors, et cetera.

这些都是我们行业接下来要努力的方向，而监管也会在这个过程中发挥作用。

So those are all the next natural steps for us as an industry to work towards this, and regulation will end up playing a role as we go through all of that as well.

好，问你一个技术问题。

Okay, technology question for you.

谷歌有点特别，你们研究语言或者说大语言模型这一块，这始于Transformer架构，同时你们也在研究世界模型。

Google is somewhat unique in that you research the language or the LLM side of the equation, which started with the transformer language, and then you also pursue world models.

有些AI研究机构，比如OpenAI，基本上认定大语言模型就是通往AGI的路径，他们实际上把部分世界模型研究——也就是试图理解物理世界的AI模型——给缩减了。

And there are some AI research efforts out there, like OpenAI, that have basically determined that the LLM is the path to AGI, and they've actually subset part of their world model research, which is the AI model that tries to understand the physical world.

但你们两条路都在走。

But you're pursuing both.

那跟我们聊聊，为什么你觉得世界模型这一块很重要？

So talk a little bit about why you find the world model side to be important.

作为人类，你显然能感知周围的世界，接收所有这些信息，确保这些模型能理解物理定律，高保真地模拟世界。

As humans, you obviously perceive the world around you, and you're taking all those inputs, making sure these models are able to understand the laws of physics, simulate the world in high fidelity.

我觉得这些都是智能的关键基础。

All that I think are key enabling blocks for intelligence.

而且我认为，在我们探索所有可能的应用时，比如机器人领域，你看Argon，在具身推理基准测试上，Argon是最先进的，这可不是偶然的，对吧？

And I think as we are working across all possible applications of this, take an area like robotics, it's not an accident, Argon, if you look at embodied reasoning benchmarks, Argon is state-of-the-art, right?

我觉得在这些基准测试上确实表现很好，这都来自于我们构建世界模型的经验。

I think it literally sold on many of these benchmarks, and that all comes from our experience building world models as well.

所以本质上，我预计未来会有架构将所有这些整合在一起。

So inherently, I would expect there will be architectures which will converge all of this in the future.

但对我们来说，我们认为今天在这个智能的崎岖前沿，我们想要站在所有可能方法的前沿。

But I think for us, we view this as today, in this jagged frontier that is intelligence, we want to be at the frontier of all the possible approaches.

我认为这将帮助我们推动更好的研究、更好的见解，并开发出更好的AGI模型，这就是我们的想法。

And I think that will help us both drive better research, better insights, and develop a much better AGI model, is how we think about it.

是的，Argon发布后我看了你的推特，全是实际应用。

Yeah, I looked at your Twitter after the Argon announcement, and it's all just real-world applications.

没错。

That's right.

不是文本或生成文本的模型。

It's not text or a model generating text.

好的，我想和你聊聊你如何在云端和AI研究之间分配计算资源，这是你最宝贵的资源。

Okay, I want to talk to you about the way that you balance allocating compute, which is your most precious resource, between cloud and AI research.

这是关于谷歌的一个大话题。

This is a big topic about Google.

有人跟我建议说，谷歌应该退出AI研究，把所有的云计算资源和芯片都拿去卖。

I've had people suggest to me that Google should get out of the AI research game and take all of its cloud computing resources and its chips and just sell that.

你看看现在的商业世界，英伟达就是这么做的，市值差不多五万亿美元。

And you look at the world of business right now, and Nvidia, which is just doing that, effectively a $5 trillion company.

这里有个矛盾：云业务想高价卖计算资源，而AI研究需要一台非常大的计算机来做出最好的AI。

There's a push and pull between cloud that wants to sell this computing resource at a high markup, and AI research, that needs a very big computer to make the best AI.

你说过，你的工作很大一部分就是决定怎么分配计算资源。

Now, you've said a big part of your job is figuring out how to allocate compute.

你是怎么做的？又是怎么平衡这两者的？

How do you do this, and how do you balance these two?

听着，我觉得对我们来说，最根本的是确保我们投资于站在AI前沿所需的计算能力。

Look, I think the ground zero for us is making sure we are investing in the compute that's needed to be at the AI frontier.

所以，在2015年，我就把公司定位为AI优先。

So that is, in 2015, I set the company to be AI first.

所以AI是我们所有业务中的关键横向要素，从搜索到云到Waymo，再到我们做的每一件事。

So AI is that key horizontal ingredient across everything we do, all the way from Search to Cloud to Waymo to everything we do.

所以我认为这是我们的根本出发点，确保研究和模型开发有足够的计算资源。

So I think that is ground zero for us, making sure both for research and developing the models that we have enough compute needed.

除此之外，我们还有这些独立的业务，每个都有自己的长期规划，他们用这些规划来预测未来，我们为所有这些业务规划计算资源。

Beyond that, look, we have these independent businesses which each have their own long-range plans, and so they're using those long-range plans to project forward, and we plan for compute for all of them.

云团队有自己的业务，有自己的预测，也有需求信号。

And so the cloud team, they have a business, and they have a projection, and they have demand signals.

他们签了合同。

They've committed contracts.

有了这些帮助，我们就能为他们规划长期的计算资源轨迹。

So with that help, we can plan on a long-term compute trajectory for them.

我觉得我们很幸运能有这么多样的机会。

I think we're fortunate to have this wide set of opportunities.

我觉得世界上没有哪家公司不在任何时刻做出权衡取舍。

I think there's no company in the world which isn't making trade-offs at any given moment in time.

但我们运营的首要原则是确保有足够的算力来投资前沿领域，这是不容商量的，其他一切都建立在这个基础之上。

But the first principles we operate on is to making sure we have enough compute to invest in the frontier, and that's non-negotiable, and everything else works on top of that.

我想你在这些对话中看到的是，如果你现在坐下来，记住，我们正在建设自己的算力，自己的数据中心。

I think what you're seeing in these conversations is if you sit now, remember, we are building our own compute, our own data centers.

所以前置时间很长。

So there's long lead times.

我们今天拥有的算力是我在2024年做出的决定的结果。

So the compute we have today is a function of decisions I made in 2024.

即使你回想一年前，也就是去年，智能体编程的爆发还没有发生。

Even if you think back to a year from today, last year, the agentic coding takeoff hadn't happened.

所以今天，显然，当你坐下来看任何事情时，每个人都短缺，不只是我们，其他公司也一样，每个人都希望你有计划的两倍。

So today, obviously, when you sit and look at anything, everyone is short, not just us, other companies, and everyone wishes you had 2X of what you had planned for.

我认为这是不可避免的，对吧？

I think that comes with the territory, right?

而且在任何时刻，我们都在规划未来两到三年需要什么。

And at any given moment in time, we are planning for what we would need two to three years out in the future.

所以我觉得你会发现，有些决定在两年后看可能看起来不是最优的，对吧，这取决于当时的技术水平。

So I think you're going to find some decisions you make may look suboptimal two years out, right, depending on where the state of the art is.

所以这个，我不会花太多时间去想。

So that, I don't spend too much time thinking about it.

我对算力分配的想法是，确保Google DeepMind，无论是研究还是前沿模型开发，在算力获取方面都是最顶尖的，对吧？

The way I think about compute allocation is making sure Google DeepMind, both for research and frontier model development, is cutting-edge in terms of compute access, right?

在那之后，我们为所有业务做规划。

And after that, we plan for all our businesses.

我觉得我们能把算力提供给其他人也很好，就像我们已经做的那样。

I think it's great we can provide our compute to other people, too, like we have done.

这让我们能扩大规模，在尽可能广泛的应用场景中进行真实测试，在现实世界中检验，并确保我们在AI技术栈的每一层都处于技术前沿。

It allows us to scale that up, get real tests for it in the widest application set possible, test it in the real world, and make sure we are at the cutting edge of technology at every layer of our AI stack as well.

简单说，有个播客里Ben Horowitz提到，Andy Jassy说过类似“算力已经卖到2028年”的话。

Briefly, there was a podcast where Ben Horowitz relayed that Andy Jassy had said something like, "Compute is sold out till 2028."

你那边也是这种情况吗？

Are you seeing the same thing?

我觉得这么说挺合理的，如果现在有客户来跟我们说要2027年的算力，那真的很难办到。

I think it's a reasonable, if today a customer is coming out and asking us for compute in '27, that is incredibly difficult to do.

这也是为什么去年我们甚至去租了一些短期算力，就是为了照顾一些客户，帮他们解决问题。

There's part of the reason, even last year, we went out and leased some short-term compute was to accommodate some customers and help them out.

但确实，我觉得大家都在面对类似的时间线。

But yes, yeah, I think we are all dealing with similar timelines.

问个关于泡沫的问题。

One bubble question.

有人说每次技术变革都会出现过度建设。

There's talk of how every technology shift has seen overbuilding.

铁路。

Railroads.

光纤。

Fiber.

你觉得现在这个领域也有过度建设吗？

Do you think there's overbuilding here?

这取决于你看的是哪个时间段。

It depends on the time frames you're looking at.

如果从足够长的时间跨度来看，这种建设规模对我来说是完全合理的。

This level of build, if I take a long enough time frame, it makes complete sense to me.

我们正处于史上最具影响力技术的早期阶段，这不像光纤，它不是闲置的。

We are at early stages of the most consequential technology ever, and this is not like fiber, it's not dark.

我们正在使用它。

We are using it.

我们每天都面临产能紧张。

We are capacity constrained on a daily basis.

应用和底层技术都在快速进步，对吧？

The applications, the underlying technology is improving rapidly, right?

但经营一家公司，我的处理方式是看需求信号。

But running a company, the way I work through the moment is you look at your demand signals.

看那些以合同形式获得的已承诺收入机会。

You look at your committed revenue opportunities that you're getting in the forms of contracts.

你要把这些和底层技术能力、我们内部的评估、我们内部怎么用它、模型的发展轨迹结合起来，基于所有这些，采取一种有纪律的方法，既自信、激进，又有纪律地规划算力。

You balance that with the underlying technical capability, our internal evals, how we are using it internally, the model trajectories, and based on all that, you take a disciplined approach, a confident, aggressive, but a disciplined approach to plan out compute.

但当你谈论整个更广泛的经济时，人们是在各种风险水平下运作的，所以对我来说很难，我专注于我能控制的公司层面，并据此规划。

But when you talk about the broader economy as a whole, you're having people operate with various levels of risk, and so it's tough for me to, I focus on what I can control as a company and plan through it.

如果你问的是，会不会出现过度建设？

If your question is, can there be an overbuild?

会的，因为可能会有一些受限的组件，记住，关键是你需要的一切组件都要齐全才能建设算力。

Yes, because you can have some constrained components, and remember, what matters is a square set of everything you need to build out that compute.

内存、电力、电工。

Memory, electricity, electricians.

土地。

Land.

大概有2000个组件——

There are like 2,000 components-

……或者更多，对吧？

... or more, right?

所以，你可能会在某个时候遇到瓶颈。

And so you could get constrained at some point.

可能会有一些因素导致瓶颈，然后就会出现过度建设。

There could be things which cause bottlenecks, and then you could have overbuilds.

还可能有外部冲击，比如经济方面的，等等。

There could be external shocks, right, to the economy, et cetera.

所以，宏观风险是存在的。

So there are macro risks.

这些在任何时候都有可能发生。

All that is possible at any given time.

但我是否确信，我们规划的算力，只要有足够的时间，就一定能用上？

But am I confident that the compute we are planning for, we will have uses for it given enough time?

绝对确信，是的。

Absolutely, yes.

好的，关于这个再问一个。

Okay, one more on this.

如果你打算两者都做，有一种观点认为，对AI研究人员来说，电脑越大，AI就越好。

If you're going to do both, there's been a line of thinking out there that AI researchers, for them, it's like the bigger the computer, the better the AI.

我当时在场，听你说AI会像火和电一样深刻，当时我们都笑了，现在想想，我有点后悔当时笑。

And I was in the room when you said that AI would be as profound as fire and electricity, and we all laughed at the time, and I kind of regret that laugh now.

但如果你真的相信这一点，那么思路就是给研究人员尽可能大的电脑，即使这会损害今天的利润，比如云业务。

But if you're someone that actually believes that, then the thinking would be you give the researchers the biggest possible computer, even if it is to the detriment of today's profits, like the cloud business.

你怎么看？

Your thoughts?

我们都在某种程度上这么做，对吧？

We're all doing some form of that, right?

我认为我们的云业务在任何时候都在放弃短期收入，因为我们正在做出计算上的权衡，以支持前沿AI的发展。

I think our cloud business is at any given moment in time foregoing short-term revenues because we are making compute trade-offs to support frontier AI development.

所以我觉得你说的已经是事实了，对吧？

So I think what you're saying is true already, right?

但你需要深思熟虑，对吧？

But you want to be thoughtful, right?

即使对我们的AI研究人员来说，约束也是有好处的。

Even our AI researchers, constraints are also good.

正是约束激发了创造力。

That's what inspires creativity.

它能帮助你找到更高效的算法突破，比如用更少的内存等等。

That's what helps you come with a more efficient algorithmic breakthrough, maybe to use less memory, et cetera.

所以我们承诺的是，投入必要的资源来保持在最前沿，对吧？

So what we are committed is to invest what's needed to stay at the frontier, right?

我就是这么想的。

That's how I think about it.

我想知道你觉得未来的界面会是什么样子，随着AI开始普及，我们未来使用计算的方式会怎样。

I wonder what you think the future of the interface looks like, the way that we use computing in the future as AI starts to roll out.

几个月前我和OpenAI的总裁格雷格·布罗克曼聊过这个话题。

I was speaking with Greg Brockman, the president of OpenAI, about this a couple of months ago.

他有一句非常有趣的话。

He had a very interesting quote.

他说：“你想要的界面是什么？

He said: "What is the interface that you want?

你想要的的产品是什么？

What is the product that you want?

答案是，你不需要界面。

The answer is that you want no interface.

你不需要产品。

You want no product.

你希望这成为你和我之间的界面吗？”

You want this to be what's the interface between you and me?"

基本上就是说，直接对话。

Basically saying, just talking.

如果AI真的那么聪明，也许就不需要屏幕，也不需要那么多按钮和需要你触碰的东西了。

That if AI is that smart, there shouldn't be maybe a screen or just so many buttons and things that you have to touch.

在AI时代，你觉得界面会是什么样子？

What is your view of what the interface looks like in the age of AI?

我的想法是，我觉得历史上计算一直是人类去适应它。

The way I think about it is, I do feel computing historically has been humans adapting to computing.

有了AI，情况就会反过来。

With AI, it'll be the other way about.

技术会无缝地适应人类的生活方式，适应我们日常的节奏。

The technology will adapt seamlessly to how humans live and go about their day-to-day lives.

所以这里面有一个方面。

So there's an aspect of it.

你希望在你需要的时候，能跟AI说话。

You want that AI to be able to speak to it when you want it to.

它有上下文。

It has context.

它可以一直在后台运行，给你提供连续性。

It can be persistently running somewhere, give you continuity.

你不需要面对很多复杂的东西。

You're not dealing with a lot of complexity.

这很自然。

It's natural.

这些对我来说都讲得通。

All that makes sense to me.

不过，这并不意味着它就没有界面。

That doesn't mean it won't have an interface, however, to me.

你看，就拿现在来说，即使在手机上，Gemini的语音模型已经非常好了，我只要按一下麦克风按钮。

Look, today, for example, nowadays, even in my phone, Gemini voice models are so good, I can just press the microphone button.

我就可以随便说。

I can just ramble.

所以我们才叫它“漫谈者”，对吧？

That's why we call it rambler, right?

我可以随便聊。

I can ramble.

它会根据我的个人背景智能地纠正。

It kind of intelligently corrects with my personal context.

所以即使是我朋友的名字，那些奇怪的名字，它也能大致说对，哪怕一开始猜错了。

So even my friends' names, weird names, it kind of gets it right, even if it first guesses it wrong.

真的很厉害。

It's amazing.

我说完之后都不用回去删改，对吧？

I'm not going and backspace editing once I finish saying it, right?

它进步得真是太明显了。

It's remarkable how good it's gotten.

但有些时候你还是需要界面的。

But there are going to be moments you will need an interface.

你可能想娱乐一下自己。

You may want to entertain yourself.

你可能想看个视频。

You may want to watch a video.

所以你可能让AI帮你做点东西，然后你想看它、消费它，对吧？

So you may ask your AI to create something, and you'll want to watch it, consume it, right?

所以我觉得，理想的状态就是让它尽可能无缝、自然，你不需要费劲去适应某个界面。

So I think the way I think about it is you want it to be as seamless and as natural as possible, and that you are not doing the hard work adapting to whatever the interface is.

技术应该适应你当下的工作方式。

The technology is adapting to how you're working at any given moment in time.

我接下来要聊聊OpenAI的那些大动作。

I'm just going to play the OpenAI hits.

因为去年，你把他们逼到了紧急状态。

Because, last year, you sent them into a code red.

你肯定知道这事。

I'm sure you know.

当时Gemini表现很好，OpenAI就严阵以待了。

Gemini was doing very well, and OpenAI battened down the hatches.

我跟山姆·奥特曼聊过谷歌带来的挑战。

And I was speaking with Sam Altman about the challenge from Google.

他说他觉得OpenAI能挡住你们，因为他们是在原生AI上构建的，我不知道他是不是原话这么说，但他暗示谷歌是在现有产品上硬加AI。

What he said was he thought OpenAI would fend you off because they're building AI native, and I don't know if he said it specifically like this, but he implied Google is bolting on AI to existing products.

你怎么看，一个附加式的做法——我不知道你是否接受这种说法——能不能跟一个原生AI产品竞争？

What is your view of whether a bolt-on, I wonder if you would accept that framing, can compete with an AI-native product?

首先，现在是一个非常动态的时刻。

Look, first of all, it's a very dynamic moment.

我们都在互相推动。

We are all pushing each other.

我觉得OpenAI也把我们公司逼得很紧。

I think OpenAI has pushed us hard as a company too.

所以这不仅仅是我们在对他们做什么，等等。

So this is not just we are doing something to them alone, et cetera.

远非如此。

Far from it.

你们也进入了红色警戒状态。

You went into code red, too.

我知道关于是否真的如此有争议，但是——

I know it's disputed about whether it was, but-

我们都在——

We all-

某种程度上来说

A version of this

……突然之间，AI 时代来临，我们得随时以不同的紧迫程度来应对。

... it's been the AI moment all of a sudden to respond with varying levels of urgency at any given moment in time.

你看，我觉得我们也在不断开发新产品。

Look, I think we are constantly building new products as well.

Gemini 是从零开始打造的——

Gemini was built from the ground up-

……作为产品，我说的是 Gemini 应用。

... as a product, the Gemini app I'm talking about.

或者明天，当我们都在适应智能体时，你是把智能体加到 ChatGPT 上，还是原生构建它们？

Or tomorrow, as all of us are adapting to agents, are you bolting agents onto ChatGPT, or are they building it natively?

你会怎么考虑这个问题，对吧？

How would you think about that, right?

所以你可以随时把这些反问回去。

So you can always flip these questions back.

所以我不这么看。

So I don't think of it that way.

我觉得关键在于打造世界级、原生AI的产品体验，这些产品会胜出。

I think of it as you want to build world-class product experiences that are AI native, and they are winning.

其中一部分是改造现有产品，另一部分是重新构思新的体验。

And some of it is adapting product, some of it is rethinking new experiences.

所以这些都是你应对这个时代的方式。

So they will all be part of how you approach this moment.

聊聊你对未来AI界面商业模式的看法吧。

Talk a little bit about the way that you see the business of this AI interface in the future.

比如，如果我让我的代理去帮我做某件事，它推荐了一个产品，即使有披露，如果我买了，它会不会从中分成？

So, for instance, if I'm asking my agent to go out and do something for me, and it suggests a product, maybe even if disclosed, does it get a cut of the sale if I make the sale?

或者，当你考虑如何从这些产品中赚钱时，你的计划是什么？

Or when you think about the way that you're going to make money from these products, what's the plan?

我觉得，显然，企业这块我先放一边。

I think, obviously, I'm going to put enterprise aside.

我觉得我——

I think I-

好的。

Okay.

对，消费者的问题。

Yeah, consumer question

……所以，如果是消费者的话，你看，我觉得消费者会越来越看重这项技术。

... so, if it's consumer, look, I think consumers are going to increasingly value this technology.

就拿现在的软件开发者来说吧。

Take software developers today.

我觉得开发者愿意为这些编程模型付费，对吧？

I think developers are willing to pay for these coding models, right?

因为它确实能帮他们提高效率。

And because it really helps makes them more productive.

没有人会想回到三年前或者去年这个时候那种编码方式。

Nobody would want to code the way they were coding three years ago, or even same time last year.

所以我觉得，当你提供价值的时候，自然就有价值可图。

So I think when you're delivering value, there's value to be had.

我觉得消费产品的历史表明，如果你做一个消费产品，我从来没见过哪个消费产品能真正扩展到数十亿用户，人们高强度使用，却没有伴随相应的变现模式。

I think history of consumer product shows if you build a consumer product, I've never seen examples where consumer products really scale to billions of users, where people are using it intensely, and hasn't had monetization to go along with it.

如果你给人们带来价值，那本身就蕴含着价值。

If you're providing value to people, there is inherent value to be had.

完全同意。

Totally.

所以它的本质可能会变，对吧？

And so the nature of it may change, right?

我想你问的是一个场景，就是有代理在帮你做事。

I think you're asking a scenario in which agents are doing stuff for you.

归根结底，你会有商业需求。

At the end of the day, you will have commercial needs.

这些代理需要满足那些商业需求。

The agents need to fulfill those commercial needs.

另一边的提供方会看到其中的价值，而且会有某种机制来实现这一点。

The providers on the other side are going to see value in that, and there's going to be some mechanism by which it gets fleshed out.

这就是高效系统的作用。

That's what an efficient system does.

但我觉得重要的是，你在打造那些世界级的消费体验，对吧？

But I think what matters is that you're building those world-class consumer experiences, right?

所以，我回答你的问题了吗？

And so, am I answering your question?

我只是想确认一下。

I just want to make sure.

是的，当然。

Yeah, definitely.

我答应自己只提一个产品功能需求，就一个。

I promised myself one product feature request and one only.

是的。

Yes.

那我开始了。

Here it goes.

有没有可能让Gmail里由Gemini驱动的搜索，跟你在ChatGPT界面里接入Gmail后的搜索一样好用？

Is it possible to make search in Gmail powered by Gemini as good as search in Gmail that you access when you plug Gmail into your ChatGPT interface?

这到底是怎么回事？

What's going on here?

我们最近宣布的一个令人兴奋的事情是，面向企业的统一Gemini代理，我们已经发布了，而且我们正在把它带给普通消费者，对吧？

One of the exciting things we announced recently is in a unified Gemini agent, for enterprise, which we have announced, and we're bringing that to consumers, right?

这些是持久化的代理，在云端全天候运行，就像每次查询都启动一个微型虚拟机一样。

So these are persistent agents which run 24/7 in the cloud, like you're spinning up a micro VM effectively for every query.

从运行时、框架到模型，它都是一个无头服务，对吧？

And from the runtime, the harness, the model, it's all a headless service, right?

而且它会在任何地方都能工作。

And it'll work everywhere.