智能体 AI 时代：谷歌反重力引擎驱动下一代产品

所以我们可以编辑这个场景，让它看起来像我们在……

So we could edit this set so it looks like we're on...

是的。对对对，我们想要这个——我们之前私下聊的时候说过，应该把那个做成开场，因为我觉得它能让所有这些都更强大。我见过一些例子，那种微妙的细节让我意识到，这就像是世界理解在展开。我之前做演讲，和我的朋友 Tulsi 一起上台，他领导模型团队。我跟观众里某人说让他们编辑视频，他们真的当场拍了照片，用 Omni 实时编辑，然后一只狗出现在编辑后的舞台上。其他嘉宾低头看到狗，笑了笑。而那时我正在高谈阔论什么 AI 废话。

Yes. Yeah, yeah, we I want this where we were talking off camera like we should do that for the intro because I think it just makes all this stuff more capable. I've seen these examples of such subtle nuance that make me appreciate that it's like the world understanding playing out. I was giving a talk and was on stage with my friend Tulsi who leads the model team. I had mentioned to someone in the crowd to edit the video and they had literally taken the picture, edited it with Omni in real time and this dog came on the stage in the edited version. The other guests sort of looked down and saw the dog. They chuckled a little bit. This is while I'm opining about whatever AI nonsense.

笑话。

Jokes.

对，不是我的笑话。他们笑的是狗跑上来。它跳到我腿上，我稍微回应了一下狗，继续说话，一边摸它什么的。要把那个做对需要很多微妙的细节，而模型完美做到了。这非常有趣，我还在努力吸收和理解这对我们制作内容的方式以及其他一切意味着什么。

Yeah, it was not my jokes. They laughed at the dog coming up. It jumped onto my lap. I sort of acknowledged the dog, kept talking, petting it or whatever, and there's so much subtlety in getting that right and the model crushed it. It's very interesting and I'm still trying to absorb and digest what that means for the way we make content and all these other things.

真有意思。

That's so interesting.

我很高兴邀请到 Logan 上节目。Logan 负责 Google AI Studio 和 Gemini API。你花了很多时间思考和构建下一代开发者。所以我很兴奋今天能和你聊从智能体式 AI 到 AI 编程、世界模型等等，而且正好在 Google IO 之后。时机再好不过了。

I'm delighted to have Logan on the show. Logan runs Google AI Studio and the Gemini API. You spend a lot of your time thinking and building for the next generation of builders. So I'm excited to talk to you about everything from agentic AI to AI coding, world models and more today, right off the heels of Google IO. So what better timing?

是的，我超级兴奋。谢谢你邀请我。

Yeah, I'm super excited. Thank you for having me.

太好了。我们从智能体式 AI 开始吧。Sundar 在 IO 上称这是智能体式 Gemini 时代。智能体式 AI 对 Google 意味着什么？

Wonderful. Let's start with agentic AI. So Sundar opened IO by calling this the agentic Gemini era. What does agentic AI mean for Google?

嗯，这是个好问题。如果你密切关注，我们其实在 Gemini 2.0 时就提到过一些，我觉得那有点早。所以我认为这个时代，Gemini 3.5 时代，感觉现在真的成真了，我们进入了智能体式编程或智能体式产品的时代，就 Gemini 而言，一切都是智能体。对我们来说，这个智能体层——我们在 IO 上宣布了——由 anti-gravity 智能体框架驱动，是 Google 的一条新主线，连接了我们所有产品，现在它们都基于它。历史上，在 Gemini 之前，我们大概不到 100 个 Google 产品，50 个左右，没有一条主线。有了 Gemini，它成了主线，现在所有东西都以某种方式使用 Gemini。现在 anti-gravity 也成了这样，所有产品重新架构，成为智能体式原生产品，真正代表用户采取行动，帮助他们完成任务。你看到这条新主线正在出现，我觉得非常有趣。

Yeah, it's a good question. I think if you followed closely, we sort of mentioned some of these things back with Gemini 2.0, which I think was a little bit early. So I think this era, the Gemini 3.5 era, feels like it's actually becoming true now and we're in the era of agentic coding or agentic products and everything agents as far as Gemini goes. I think for us this agentic layer, and I think we announced this at IO, sort of being powered by the anti-gravity agent harness, is this additional through line for Google that sort of connects all of our products that they're based on now. Historically, prior to Gemini, there wasn't a through line for the probably sub-100 number of Google products, the 50 Google products we have. There wasn't a through line. We had Gemini, it became this through line, everything is now using Gemini in some way. That's now becoming true for anti-gravity as all of the products rebase to become sort of agentic native products and actually taking action on behalf of users and helping them get things done. You see this new through line emerging, which I think is really interesting.

抱歉，帮我确认一下：anti-gravity 是 IDE 对吧？还是非 IDE？

And sorry, help me with anti-gravity: is it the IDE, right? Or the non-IDE?

嗯，anti-gravity 包含很多东西，我认为这对我们来说是个机会。你有一个核心 IDE，有网页上的智能体优先体验，有 CLI，有 SDK。所以实际上，我不知道我们有没有这样表述过，但它真的是我们构建的一个生态系统。它旨在满足开发者在任何地方的需求。所以你可以通过 Gemini API 使用它，如果你想要一个托管智能体，不需要自己做任何基础设施工作。然后最有趣的是，它不仅仅是 anti-gravity 的生态系统，它还实际上驱动着所有其他 Google 产品。所以 anti-gravity 将为搜索、Gemini 应用、Cloud 和 ASU 交易中的一堆智能体功能提供动力，这非常令人兴奋。

Yeah, anti-gravity is a lot of things, which I think is an opportunity for us. You have a core IDE, you have the agent-first experience if you want it on the web, you have a CLI, you have an SDK. So I actually think, and I don't know how much we framed it this way, but it really is an ecosystem of stuff that we built. It's designed to meet developers wherever they are. So you could use it through the Gemini API if you want a managed agent that you don't have to do any of the infrastructure work for. And then the most interesting bit is it's not just the ecosystem of anti-gravity stuff, it's also powering literally all the other Google products. So anti-gravity will be powering a bunch of agent stuff in search, in the Gemini app, across Cloud and ASU deal, which is really exciting.

我明白了。所以以前是 Gemini API，语言模型是 AI 如何融入每个 Google 产品的主线。

I see. So it used to be the Gemini API, so the language model was the through line in terms of how AI gets baked into every Google product.

对。

Yeah.

而现在不仅仅是 API，还有编程框架被用于这些产品，因此它本身就是一个编程智能体，驱动产品中更多的智能体特性。

And now it's not only the API, it's the coding harness that's being used into these products and therefore it's a coding agent itself that's driving more agentic properties in products.

对，我认为这个描述很公平。更一般地说，它只是智能体框架。我认为编程是智能体框架的一个专门用例。它显然很强大，但编程已经证明了自己是通用智能体框架，同时也在编程方面表现出色。

Yeah, and I think that's a fair description. I think more generically too, it's just the agent harness. I think coding is a specialized use case of the agent harness. It's obviously powerful, but coding has proved to be the general purpose agent harness in addition to also working really well for coding.

智能体框架和编程框架是同义词吗？

Are agent harness and coding harness synonymous or not?

肯定有细微差别。我认为通过专门化可以挤出优化空间。你可以看到，AI Studio 使用的智能体框架有点专门针对 vibe 编程用例，而 Gemini 应用使用智能体框架的方式则有点专门针对消费者全天候智能体。所以我认为有一个基础框架，大概 80% 的内容相同，然后你针对编程或任何用例进行专门化。

There's definitely nuance. I think there's optimization you can squeeze out of specializing. You see this where technically the agent harness used for AI Studio is a little bit specialized for the vibe coding use case, and the way the Gemini app uses the agent harness is a little bit specialized for the consumer always-on 24/7 agent. So I think you have that base harness that probably has 80% of the same stuff, and then you specialize for coding or whatever the use case is.

有意思。你怎么看待对现有业务的蚕食，尤其是现在你们更积极地进入智能体特性？比如，如果你只是做搜索或摘要，蚕食的担忧没那么大。但如果你真的在处理我的邮件，替我回复，我还会自己看邮件吗？所以我可以想象，由于有了更多智能体能力，人类在你们产品上的眼球时间实际上会减少。这样说公平吗？或者你怎么看待这种蚕食？

Interesting. How do you think about the cannibalization of the existing business, especially now that you are going much more aggressively into agentic properties? Because I could see, for example, if all you're doing is search or summarization, there's not as much of a cannibalization fear. Whereas if you're actually going through my emails, replying to them for me, am I even going through my email anymore? So I could imagine that there's actually fewer human eyeball hours on your products as a result of having more agentic capabilities. Is that fair or how do you think about the cannibalization?

嗯，这很有意思。我有一个观察：在开始时，我认为 Sundar 很好地阐述了这一点，在当前 AI 时代之初，每个人都认为 AI 能为你回答问题对搜索来说是负和博弈。但实际上最终发生的是，它对搜索来说是非常正和的。比如人们搜索更多，做得更多。

Yeah, it's interesting. I think one observation I have is that at the beginning, and I think Sundar's done a great job of talking through this, at the beginning of the current AI era, everyone assumed that AI being able to answer questions for you was going to be negative sum for search. And actually what ended up happening is it's been incredibly positive sum for search. Like people are searching more, people are doing more.

是的，智能体实际上又催生了整个市场，在智能体做更多事情的同时，人类也在进行更多搜索。所以我认为，显然，世界上的人类时间是有限的。但根据我对这一切如何发展的初步感受，从生态系统的价值创造以及人类行为方面的结果来看，这确实是非常正面的。我认为未来 1 到 2 年的情况相对清晰，但 3 到 5 年之后，当技术改进、产品形态可能有所不同时，就不那么明确了。但最终，这就是产品的成功。我们经常与 Demis 讨论，构建技术的意义在于让它为你做事。对 Google 来说，成功可能不是最大化用户在我们产品前的眼球时间，而是最大化客户的结果，让他们完成想做的事，从而去生活、去做自己想做的事。所以我认为你会看到我们走向最大化客户结果的道路，而不是最大化眼球时间。

Yeah, and agent actually again, there's like this whole market that spawned at the same time that agents are doing more, at the same time that humans are also searching more. And so I think it will be obviously, there's a finite amount of like human time in the world. But from my early feelings of how a lot of this is playing out, it does feel like it's very positive sum from an ecosystem value creation, like how the human behavior aspect of it turns out. I think it's somewhat clear in the next 1 to 2 years, much less clear 3 to 5 years from now when the technology has improved and the products probably look a little bit different than the way that they do. But ultimately, that is the success of product. I think we have a bunch of conversations with Demis all the time and it's like the point of building the technology is so that it can go and do stuff for you. Like that point, success for Google probably doesn't look like maximizing eyeball time in front of our products. It's like maximizing outcome for customers to do the thing that they want to do so that they can go and live their life and do what they want. And so I feel like you'll probably see us go down the route of maximizing outcomes for customers and not maximizing eyeballs.

是的。我脑子里一直有个词，智能体主导的增长。在我看来，我在个人时间里大量使用编程智能体，我只是让智能体为我做所有基础设施选择。我说，我不在乎你告诉我用什么数据库。

Yeah. I have this term stuck in my head, agent-led growth. Like it seems to me I'm using coding agents a lot in my personal time and, you know, I just let the agent make all the infrastructure choices for me. I'm like, I don't care what database you tell me.

是的。

Yeah.

所以我问的原因是，你知道，这在今天的编程中已经成立。我想可能对很多事情都会普遍成立，比如未来的购物。你认为这将如何改变广告的运作方式，以及聚合商的价值捕获方式？

And so the reason I ask is, you know, it's true in coding today. I would imagine it's maybe going to be generally true for a lot of things, let's say shopping down the line. How do you think that's going to change how advertising works, how value capture works for the aggregators?

感觉这是一个非常相似的趋势。这不完全正确，但很多事情只是彼此的代理，比如 SEO 的工作方式，我认为与所谓的生成式引擎优化（GEO）直接相关。所以感觉这些事情之间有很多相关性。我的猜测是，它看起来不会像我们现在假设的那样剧烈变化，因为这些事物是相互叠加的。

It feels like it's a very similar trend. This isn't perfectly true, but a lot of these things are just like proxies of each other, like the way that SEO works, I think is directly correlated with the way that, um, I forgot what the term was, it's like GEO is like the generative engine optimization or whatever it's called. And so it does feel like there's a lot of correlation between the things. My guess is it looks like much less of a radical shift than what I think maybe we assume right now, just because these things compound on top of each other.

如果按爬行、行走、奔跑来划分智能体化的程度，Google 产品套件的智能体化程度目前处于哪个阶段？

If you were to grade the scale of agenticness in terms of crawl, walk, run, where are we in terms of how agentic the Google suite of products is?

是的，这是个好问题。目前绝对是爬行阶段。我认为部分原因是 Google 固有的产品张力——我们有超过 130 亿用户的产品。所以实际上我认为我们有一些更像实验室的体验，可能更接近奔跑或行走。但我认为今天大多数产品体验绝对更接近爬行。我认为这只是我们构建被大量人群使用的产品所承担的管理责任。我不认为长尾客户已经准备好让 AI 运行并做所有事情。他们可能想坐在驾驶座上，谨慎地迈出第一步。我认为 Google 团队和搜索可能是最典型的例子。我认为他们有很大的责任以引导人们的方式去做，而不是彻底改变他们与互联网互动以及关联产品的方式。

Yeah, that's a great question. It's definitely like crawl right now. And I think some of this is like all of the inherent product tension for Google is like you have what, 13 billion plus user products. And so I actually think we have some more labs-like experiences where you're probably closer to running or walking. But I think most of the product experience today is definitely closer to crawling. And I think that's just the stewardship responsibility we have sort of building a product that's being used by lots of people. Like I don't think the long tail of customers are ready to have AI running and just doing all the things. Like they probably want to be in the driver's seat. They're cautiously taking the first step. And I think the Google team and like search is maybe the most quintessential example of this. Like I think they have a lot of responsibility to actually do that in a way that it brings people along and doesn't just change everything of how they interact with the internet and the way they associate with products and stuff like that.

是的。你认为哪些产品最接近行走阶段？

Yeah. Which products do you think are closest to the walk?

这是个好问题。我认为 Gemini 应用绝对最接近行走。对于 Spark 来说，拥有一个 24/7 全天候在线的智能体，实际上代表你执行一系列操作，绝对是前沿用例之一。我认为你会看到，比如 anti-gravity 是另一个例子，你可以拥有自主编程智能体，重建操作系统，处理数十亿 token，并代表你花费数千美元。我认为这些更前沿，实际上它们也在 GDM 中，这是另一个角度。所以我认为 GDM 以非常前沿的方式看待这一点，而 Google 的其他产品则更渐进地达到那里，这对我来说是合理的。

That's a good question. I think Gemini app is definitely closest to walk. And so for Spark, I think having a 24/7 always on agent like literally going and potentially doing a bunch of actions on your behalf is definitely one of the frontier use cases. And I think you'll see, I think like anti-gravity is another one where it's like you can have autonomous coding agents, you know, rebuilding operating systems and doing billions of tokens and spending thousands of dollars on your behalf. And I think those are again more and actually like they're in GDM as well as another angle of this. So, I think like GDM is taking very much like a frontier look at this where I think the rest of Google's products are more incrementally getting there, which again makes reasonable sense to me.

是的。你认为 Google 最终会有一两个、三个还是数千个使用 AI 的产品界面？

Yeah. Do you think that Google ends up with one, two, three product surfaces for using AI or thousands?

这很难说。我认为很多因素实际上已经融入了人类消费产品的方式。我的感觉是，产品的细分和专业化有好处——如果你最终得到一个为你做所有事情的产品，使用那个版本的产品本质上需要更多工作。我认为这会是默认状态。也许有人会打造出真正神奇的体验，使这不再成立，但我认为长尾用户最终需要花费更多脑力和时间才能让通用产品做他们真正想做的事，相比之下，我点击日历应用，它只显示我的日历，这很好。我不需要担心和处理其他任何事情。

It's tough. I think a lot of this is actually baked in just like how humans consume products. And my sense is that there's something nice about having this compartmentalization and this specialization of products where it becomes, if you end up with a product that is doing everything for you, inherently there's more work involved in using that version of the product. I think that would be the default state. I think maybe somebody will spin together the truly magic experience that doesn't make that true, but I think the long tail of folks end up having to spend more mental energy and more time to actually get the general purpose product to do the thing that they actually want to do versus there's something nice about I click my calendar app, it just shows me my calendar. Like I don't need to worry and deal with anything else.

这是我对为什么幻灯片存在这么久的热门观点——你想要的信息正好在同一个地方。我认为我们人类实际上非常习惯这一点，而生成式界面的想法听起来很酷，但我们的大脑真的能适应吗？这难道不是给我们带来更多认知负担吗？

This is my hot take for why slide decks have existed for so long, of just like the piece of information you want to be exactly in the same place. And I think we as humans are just actually very used to that as opposed to the idea of a generative interface sounds so cool to me, but it's like, are our brains really, isn't it just more cognitive overhead for us?

在某些情况下确实如此，我认为有人需要——世界上有很多非常聪明的人，所以也许有人会找到让体验更自然的方案。但对我来说，现在，也许不是一万个产品的极端版本。我猜测它看起来更像是更多产品追求不同的方向，也许另一个答案是，我不知道 Google 在生态系统中的样子。我认为它看起来会有更多产品。这真的很令人兴奋。

It definitely is in certain cases and I think somebody needs to, again, there's a lot of incredibly smart people in the world and so maybe somebody will find the experience that makes it feel more natural, but to me right now, I'm maybe not 10,000 is the extreme version. I'm guessing it looks more like more products going after sort of different and maybe the other answer is like I don't know what it looks like for Google for the ecosystem. It looks like a lot more products, I think. And that's really exciting.

我认为 Google 最终会如何战略性地决定——比如，我们的客户是希望我们拥有 10,000 个产品，还是只有三个更好？——这将取决于我们的战略决策。

I think how Google will end up strategically deciding—like, do our customers want to deal with us having 10,000 products, or would it be better to only have three?—will come down to a strategic decision for us.

这完全说得通。当我与企业公司交谈时，他们说每个人都在谈论智能体式 AI，但他们唯一看到智能体真正发挥作用的地方就是编码智能体。你同意还是不同意这种看法？

That totally makes sense. When I talk to companies in the enterprise, they say everyone's talking about agentic AI, but the only place they've seen agents really working is coding agents. Do you agree or disagree with that take?

是的，我认为这取决于你对“工作”的标准是什么，这其中有很多细微差别。如果你真的试图为模型尚未达到质量门槛的领域卸载非常复杂的任务，那么我认为这确实是真的。它不会解决问题。但这是我希望能衡量的东西。一个很好的例子是 OpenRouter，它在衡量总 token 消耗量。你可以看到这些趋势随时间推移，现在世界上的智能比一年前多了多少。与此同时，我真正感兴趣的是衡量平均智能体运行时间或平均任务实际耗时。我不认为他们公开了这些数据，但我感觉他们可能有有趣的数据。我相信还有其他来源。因为我确实看到这些新的模型能力落地或新模型发布，并且它在上升。也许曲线现在仍然很低，但你看到了长期任务上升的早期迹象。所有模型实验室都在谈论发布一个能自主工作三天的新模型等等。那是极端情况，但我认为在实践中你看到它正在相当快地上升，这非常有趣。所以即使企业除了编码之外还没有感受到，他们今年也会感受到，因为许多其他用例也会变得更好。

Yeah, I think it depends on what your bar for working is, which is a lot of the nuance here. If you're truly trying to offload very complicated tasks for domains where the models haven't actually crossed the threshold of quality, then I think that's definitely true. It's not going to solve the problem. But this is something I wish we could measure. A good example is OpenRouter, which is measuring total token consumption. You can see these trends play out over time of how much more intelligence is in the world now versus a year ago. In parallel, the thing I'm really interested to measure is how long the average agent run or the average task actually takes. I don't think they publish that, but I feel like they probably have interesting data. I'm sure there are others. Because I do think you're seeing these new model capabilities land or new model drops, and it's spiking up. Maybe the curve is still very low right now, but you're seeing those early signs of it spiking up for long-running tasks. All the model labs are talking about releasing a new model that did three days of autonomous work or whatever. That's the extreme, but I think in practice you're seeing that trickling up pretty quickly, which is really interesting. So even if the enterprises haven't felt it outside of coding, they are going to this year as a bunch of those other use cases get much better as well.

从 DeepMind 的角度来看，你认为长周期智能体是一个重要的 KPI 吗？它是那个最重要的 KPI 吗？

From a DeepMind perspective, do you think long-horizon agents is a KPI that matters? Is it the KPI that matters?

这当然重要。对于 DeepMind，我们正在做很多事情，我们稍后可以详细讨论。有一个庞大的不同赌注组合。长周期智能体显然非常重要。而且我认为特别是编码智能体非常重要。如果你有一个出色的编码模型，它显然会加速你业务的每个其他部分。所以确保我们拥有这一点是重中之重。

It definitely matters. For DeepMind, we're doing lots of things, which we can talk more about later. There's a huge portfolio of different bets taking place. Long-horizon agents obviously matters a lot. And I think specifically coding agents matter a lot. It clearly is an accelerant for every other part of your business if you have a great coding model. So making sure we have that is super top of mind.

明白了。我想稍微换个话题，谈谈编码。好的，我要问一个棘手的问题。我的很多开发者朋友长期以来一直在使用 Claude。OpenAI 看到了这一点，并宣布了“代码红色”。Codex 现在真的很好。我想说我的朋友们现在大概在 Claude 和 Codex 之间各占一半。我没怎么听说他们用 Gemini，这总是让我有点困惑。这是怎么回事？

Got it. I'd love to shift gears a little bit and talk about coding. Okay, I'm going to ask a hard question. A lot of my developer friends were using Claude for a long time. OpenAI saw that and declared code red. Codex is now really good. I'd say my friends are maybe split 50/50 now between Claude and Codex. I don't hear a ton of them using Gemini, which has always kind of puzzled me. What's going on with that?

是的，这是个好问题。我想补充故事的一部分，这让它更有趣。在十二月，叙事是 Google 赢了。当我们推出 Gemini 3 时，从模型能力角度来看，这是一个巨大的改进。我认为很多叙事都说 Google 迈出了一大步。作为一个生态系统参与者，有趣的是看到叙事转变的速度有多快。下一波叙事显然是假期期间及之后发生的所有智能体式编码事件。那并不久远。这提醒我们事情变化有多快。我认为这个观察并非不合理。我们在幕后所做的是尽可能快地推动编码前沿。Antigravity 是其中的重要部分。一个要点是，如果你没有一个真正执行长时间工作的产品，就很难为开发者用例做出一个出色的编码模型。Google 意识到了这一点。这就是为什么 Windsurf 交易发生了，那些人过来并最终构建了 Antigravity。我们一直在内部使用它，Sundar 在 IO 上展示了 Google 内部 token 消耗的增长图。你需要那个引擎运转。元评论再次是引擎正在运转。实际取得模型进展需要时间。但我非常有信心。我们从事编码工作的团队就像 AI 界的复仇者联盟。这真的是 Google 内部一些最优秀的人在努力推动这件事，非常认真。而且我认为 Gemini 3 Flash，尽管有一些关于价格等的讨论，是朝着实际带来许多这些能力以及劳动成果得到回报的一步。从编码角度来看，它是一个比我们发布过的任何 Pro 模型都更好的 Flash 模型。而 Pro 模型之前已经非常好了。还有另一条线索：每个人都忘记了预训练窗口。我想知道是否有人应该在线追踪这个，那会很有趣。

Yeah, it's a great question. I think there's one part of the story I'll add, which makes it even more interesting. In December, the narrative was that Google had won. When we landed Gemini 3, it was such a profound improvement from a model capability perspective. I think a lot of the narrative was that Google had taken a huge leap forward. And what was interesting to see as an ecosystem participant is how quickly that narrative shifted. The next wind of the narrative obviously was all the agentic coding stuff that happened over the holidays and into January and beyond. That was not that long ago. It's a reminder of just how fast things can change. I think the observation is not unreasonable. What's happening behind the scenes for us is trying to push the frontiers as fast as possible on coding. Antigravity is an important part of that. One takeaway is that it's really hard to make a great coding model for this developer use case of really long-running work if you don't actually have a product that does that. Google realized that. That's why the Windsurf deal happened, why those folks came over and ultimately built Antigravity. We've been using it internally, and Sundar showed at IO the graph of growth of token consumption inside Google. You need that engine to spin. The meta comment again is that the engine is spinning. It takes time to actually make model progress. But I'm super confident. The group of folks we have working on code is like the Avengers of AI internally. It really is some of the best people inside Google trying to push the rock up the hill on this stuff, taking it super seriously. And I think Gemini 3 Flash, notwithstanding some conversation about price and stuff, is a step towards actually bringing a lot of these capabilities and the fruits of that labor paying off. It's a flash model that's better than any pro model we've ever released from a coding standpoint. And the pro models were really good before. There's another thread: everyone forgets about pre-training windows. I wonder if someone should track that online, which would be interesting to see.

意思是像大型运行，比如哪些集群可用之类的……

Meaning like the big run, like what clusters have been available and like...

正是如此。大型运行是其中的一个有趣线索。从外部看，你可能在某些方面显得非常落后，但实际上你错过了大型运行和大型预训练运行所在的所有背景信息。

Exactly. The big runs are an interesting thread of this. It might look from an external perspective that you're super behind in some way, but actually you miss all the context of where the big runs are and where the large pre-training runs are.

我认为预训练历来是 DeepMind 的巨大优势。我们拥有世界上最好的人才，我很期待看到这些努力的成果以及随之而来的其他进展。例如，3.5 Flash 完全是通过后训练取得的成果，这非常酷。这充分证明了团队的能力，他们仅凭后训练就取得了如此大的进步，甚至超越了之前的 Pro 模型，这太棒了。

I think pre-training has historically been a massive strength for DeepMind. We have some of the best people in the world, and I'm excited to see the fruits of that labor and everything else that's happened. For example, 3.5 Flash was all post-training gains, which is really cool. It's a huge testament to the team that they made such gains and surpassed the previous pro model literally just with post-training, which is awesome.

你们内部对“吃自己的狗粮”有多执着？DeepMind 的人还能用其他模型吗，还是说你们现在都用 Gemini 工具链，必须把它做得非常好？

How religious are you about dogfooding internally? Are DeepMind folks still allowed to use other models, or is it like you guys are using the Gemini harness now and have to make it really good?

我认为使用其他模型是健康的，因为如果不这样做，有时很难真正理解生态系统中正在发生的事情。所以我使用所有模型和所有产品。DeepMind 的其他人也一样。不过，你肯定得用 Gemini 模型。从反馈飞轮的角度来看，这很好，也是它们变得更好的方式之一。DeepMind 乃至整个 Google 有超过十万名出色的工程师在使用这些模型并提供反馈。这应该是 Google 的竞争优势，因为我们拥有如此规模的工程资源和人才深度，可以进行 A/B 测试和实时实验。所以我认为你必须使用所有模型，但对大多数人来说，Gemini 是日常主力，这很棒。

I think it's healthy to use other models because it's sometimes hard to actually grok what's happening in the ecosystem if you don't. So I use all the models and all the products. Folks across DeepMind are doing the same. You definitely have to use the Gemini models, though. It's great from a feedback flywheel perspective, and it's part of how they get better. DeepMind and Google more broadly have over a hundred thousand incredible engineers using the models and giving feedback. That should be a competitive advantage for Google because of that scale of engineering resources and depth of talent, allowing us to run A/B tests and live experiments. So I think you have to use all the models, but for the majority of folks, Gemini is the daily driver, which is great.

你相信关于“软起飞”的说法吗？即一旦你有一个足够好的智能体式编码模型，它就会加速研究进展，形成一个自我强化的循环。

Do you believe in the narrative around a soft takeoff, where once you have a good enough agentic coding model, it accelerates the pace of research progress and becomes a self-reinforcing cycle?

这似乎很明显是真的，但也许我喝太多“迷魂汤”了，才会这么认为。

It seems obvious that's true, but maybe I've drunk too much Kool-Aid that that's the case.

你看到迹象了吗？

Are you seeing signs of it yet?

是的，你肯定能看到一些迹象。但从模型的角度来看，这些迹象还处于早期。部分原因在于，较大规模训练的资源分配非常重大，所以仍然需要人类掌舵做决策。你不会不小心用 10,000 个 TPU 去启动一个没什么意义的任务。但从产品角度来看，你肯定能看到。在我们的团队中，我们用 anti-gravity 构建了移动应用，并且发布速度比 Google 历史上任何团队构建移动应用都要快。Josh 的团队用 Gemini Mac OS 应用做到了这一点，端到端交付应用的速度比 Google 任何团队交付 Mac 应用都要快。这都归功于智能体式编码。所以从产品角度来看，这很棒。

Yes, you definitely see some signs. But from a model perspective, the signs are still early. Part of the context is that resource allocation for larger training runs is significant, so you still have a human in the driver's seat making decisions. You're not going to accidentally take 10,000 TPUs to kick off a job that doesn't make sense. But from a product perspective, you for sure see it. On our team, we've built mobile apps using anti-gravity and launched them faster than I think any team at Google has ever built a mobile app. Josh's team did this with the Gemini Mac OS app, delivering an end-to-end app faster than any team had ever delivered a Mac app at Google. It's because of agentic coding. So it's great from a product perspective.

你过去说过，如果你有一个系统可以用代码构建任何东西，人类就无法在同一水平上竞争，那就是狭义超级智能。你认为我们已经达到那个点了吗？

You've said in the past that if you could have a system that could build anything with code, humans can't compete on the same level, and that's narrow superintelligence. Do you think we've reached that point?

这很有趣。这个狭义超级智能的例子在编码领域现在就有这种感觉。编码能力如此之强，确实有点像狭义超级智能。我不知道如何精确量化，但重要的是它在编码方面表现得极其出色。如果它还能做很多其他事情当然很好，但仅仅在编码上出色就已经非常有影响力了。我花了很多时间来消化这个事实，因为构建 AGI 非常重要且有趣，但如果它削弱了当前技术能力的故事，我认为这是一个糟糕的权衡。所以我试图平衡这两件事：我们需要构建通用技术，但拥有这个东西也非常有影响力。感觉它并没有削弱人类开发者；它真的像是一种催化剂。作为一个人类开发者，我感觉自己在世界上有了更多能动性。我可以解决更雄心勃勃的问题。过去我有些想法觉得有点遥不可及，现在相反：我有了一个想法，觉得可以把它做得更宏大。这增加了不同层面的责任或负担，因为我不能只做最小可行产品；我需要再向前走十步，因为技术允许我这样做。重新设定我的抱负水平是我花时间思考的事情。我认为这也会发生在其他垂直超级智能领域，感觉在解决所有问题之前，我们会先得到一堆这样的领域。这几乎就像锯齿状的超级智能。

It's interesting. This narrow superintelligence example feels that way for coding right now. Coding is just so good that it does kind of feel like narrow superintelligence. I don't know how to quantify it exactly, but the important thing is that it works incredibly well for code. It would be great if it did a bunch of other things, but it's so impactful that it can be great at code. I spend a lot of time letting that fact wash over me because building AGI is very important and interesting, but if it takes away from the story of the current present capability of the technology, I think that's a bad trade-off. So I try to hold these two things equal: we need to build general-purpose technology, but it's so impactful to have this thing. It feels like it hasn't taken away from human developers; it really feels like an accelerant. As a human developer, I feel like I have more agency in the world. I can tackle more ambitious problems. I used to kick around ideas that were slightly out of reach, and now I have the opposite problem: I'm kicking around an idea and I think I could probably make it even more ambitious. It adds a different layer of responsibility or burden because I can't just do the MVP; I need to go 10 steps further because the technology enables me. Resetting my level of ambition is something I've spent time thinking about. I think this will happen in other vertical superintelligence domains, and it feels like we'll get a bunch of those before we solve everything. It's almost like jagged superintelligence.

你认为下一个会出现超级智能的垂直领域是什么？

What verticals do you think we'll get superintelligence at next?

这是个好问题。我最近花太多时间思考编码了，所以让我想想其他领域。我认为部分原因是那些可验证性更好的领域，你会更快看到进展。比如数学和金融。实际上，科学可能非常有趣。看到一些具有一定可验证性的领域真正起飞会令人着迷。那会很酷。我也认为，在关于 AI 对世界影响的更广泛叙事中，这种有效事物的顺序很重要。

That's a great question. I spend too much time thinking about coding these days, so let me think of other domains. I think part of it is things with better verifiability, where you'll see gains more quickly. So things like math and finance. Actually, science could be really interesting. It would be fascinating to see some of these domains with some level of verifiability really take off. That would be cool. I also think it's important in the broader narrative about AI's impact on the world that this sequencing of things that work happens.

你希望这些真正好的、有影响力的、对世界积极的事情尽可能早地发生，这样人们才能理解这项技术的潜在积极影响。所以，我认为科学可能是一个非常有趣的领域。显然，现在有很多关于数学证明之类的事情，我不是数学家，所以这有点超出我的理解范围。

You want a lot of these really good, impactful, positive things for the world to happen as early as humanly possible so that folks understand what the potential positive impact of the technology is. So, I think science could be a really interesting one. There's obviously all the stuff happening right now with math proofs and stuff like that, which I'm not a mathematician, so it's somewhat over my head.

我前几天看到一条很棒的推文：“为什么他们会有这么多问题？”

I saw a great tweet the other day. "Why did they have so many problems?"

没错，这条很好。我喜欢。这可以印在 T 恤上。

Exactly, that's a good one. I like that. That's a good t-shirt.

太有趣了。好吧，说到 Twitter，我之前翻过你的 Twitter，所以我要再念一条你的推文。Twitter 的好处是它公开记录了你所有的预测。

So funny. Okay, speaking of Twitter, I went through your Twitter before this, so I'm going to read back another tweet at you. The good thing about Twitter is there's a public record of all your predictions.

得开启那个自动删推功能之类的。

Need to turn on that auto tweet deleting feature or whatever it was.

去年十月，你发推说：“到 2025 年底，每个人都能用 vibe coding 做视频游戏。” 这成真了吗？

Last October, you tweeted, "Everyone is going to be able to vibe code video games by the end of 2025." Did that end up being true?

感觉快了。我的意思是，显然不是 3A 大作。你还做不出下一部《使命召唤》或《GTA》。但我认为它比以往任何时候都更接近。视频游戏的一个有趣之处在于，你实际上需要构建很多其他东西。比如模型，我们之前私下聊过，3.js 就是一个很好的例子。3.js 让很多以前不可能的事情成为可能，但仍然有很多粗糙的边缘，仅靠一个编码智能体无法解决。所以你需要精灵生成，而模型本身并不擅长这个。所以你需要一些编排层和工具来实现它。还有很多其他核心游戏体验的东西需要高度可靠性。我认为这触手可及，但实际上需要大量的产品脚手架工作来创建可复用、可重玩且有深度的体验，这还需要一点品味。

It feels close. I mean, obviously not AAA games. You're not building the next Call of Duty or GTA yet. But I think it feels closer than it's ever been. A lot of the interesting bit about video games is you actually need to build a lot of this other stuff. Like models, and we were talking off camera before this, 3.js is a great example. 3.js makes a lot of things possible that weren't before, but there are still all these rough edges that just a coding agent doesn't solve. So you need sprite generation, and the models aren't very good at doing that natively. So you need some orchestration layer and tooling to make that happen. There are a bunch of other things core to the gaming experience that need a high degree of reliability. I think it feels within reach, but actually requires a lot of product scaffolding work to create experiences that are reusable, replayable, and have depth, and it requires a little bit of taste.

你看到有人在 AI Studio 和你提供的其他开发者平台上制作大量视频游戏吗？

Do you see people making a lot of video games inside AI Studio and the other developer surfaces that you have?

是的，这实际上基于我们查看的早期数据。当时在 AI Studio 中，人们制作的所有应用里大约 20% 实际上是游戏。大家都在尝试做游戏。

Yeah, this was actually based on us looking at the early data. In AI Studio at the time, it was like 20% of all apps that folks were making were actually games. People were trying to build games.

这是最受欢迎的类别吗？

Is that the most popular category?

它不再是最大的类别了，因为我认为生态系统变了，用户群也变了，但游戏仍然很多。

It's not the most popular category anymore, just because I think the ecosystem has shifted and the user base has shifted, but it is a lot of games.

那最受欢迎的类别是什么？

What is the most popular category?

我想大概 20% 是金融相关的东西，20%……

I think it's like 20% finance related stuff, 20%...

看来大家很爱数钱。

Like counting their money that much.

人们喜欢……我觉得实际上跟加密货币有关。很多金融相关的东西。很多个人生产力工具，还有很多生成式媒体内容，因为显然 Google 的生成式媒体套件做得很好。但我也认为 GDM 有点……显然 Demis 非常关心游戏，他的职业生涯也是因为游戏而开始做 AI 的。所以我认为我们会在这方面有一些有趣的尝试。实际上，我们在 Kaggle 的团队——也就是我们在 GDM 做 AI 基准测试的那群人——与 GDM 合作构建了这个游戏竞技场，这是我们测试 AGI 进展的方式，用游戏作为代理，这再次深深植根于 GDM 的历史。

People like... I think it's something around crypto actually. A lot of stuff with finance. A lot of personal productivity things, and a lot of gen media stuff, because obviously the Google suite of gen media stuff has done a great job. But I also think GDM has sort of... obviously Demis cares a ton about games and sort of started his career in doing AI stuff because of games. So I think we'll have some interesting swings at this. Our team actually in Kaggle, which is sort of a bunch of the AI benchmarking stuff we do in GDM, works with GDM to build this game arena, which is our way of testing progress towards AGI, using games as a proxy, which again is very deeply rooted in GDM's history.

你认为我们离一个街上的普通人，只要有个好点子，就能用 vibe coding 做出一个真正好玩可玩的游戏，还有多远？

How close do you think we are to a random person off the street with a good idea being able to vibe code a really fun playable game?

我想说今年。我实际上认为模型能力已经使之成为可能。这就是我在产品方面感到兴奋的地方。我们之前私下也聊过这个生态系统中的初创公司，因为感觉这是可能的。模型质量上似乎没有差距。差距在于，知道如何打造一款好游戏的人，是否以正确的方式把脚手架搭起来，让这件事成为可能。我认为现在有人正在做这件事。所以部分原因是可发现性和认知问题，人们甚至不知道他们可以做到。另一部分原因可能是某些类别的模型能力还略有不足，我们距离跨越那个鸿沟还有几周或几个月，然后它就能对大多数人奏效了。

I want to say this year. I actually think the model capability makes it possible. I think this is where I've gotten excited on the product side. We were again talking off camera about sort of the startups in this ecosystem, because it feels like it's possible. It doesn't feel like there's a gap in model quality. It feels like there's a gap in someone who knows what it takes to build a great game actually putting the scaffolding together in the right way to make that possible. I think there are folks who are doing this right now. So some of it is a discoverability and awareness thing, that people just don't even know they can do that. And some of it is just maybe certain categories of model capabilities are slightly off, and we're weeks or months away from that chasm being crossed, and then it just works for most people.

这正好引出我接下来要问的世界模型问题。但你认为用 vibe coding 做视频游戏，更可能是基于游戏引擎加编码智能体，还是更可能基于世界模型？

This is a good segue into when I'll ask you about world models next. But do you think vibe coding video games is more likely to be game engine plus coding agents based, or more likely to be world model based?

我认为最终会发生的是，世界模型的定义会变得模糊，这一点我们应该和 Omni 讨论一下。而且它仍然……我认为编码智能体看起来会像某种世界模型类型的系统。但你实际上需要让世界模型对真实事物有用；你需要脚手架。所以我认为有很多有趣的初创公司正在做工作，比如弄清楚世界模型的脚手架是什么，这样你就可以把它们从这些非常开放式的固有设计、非常开放的空间中拿出来，以一种具体的方式去做，使其扎根于一个可以重复使用的用例。也许有人会找出世界模型的脚手架来让游戏成为可能，但世界模型目前的固有性质我认为实际上并不适合当前形式的游戏。但进展非常疯狂，谁知道呢，也许两年后的版本就能做到了。但至少在短期内，我认为从游戏角度看，编码智能体加上某种游戏引擎会带来更多的阿尔法。

I think what will end up happening is the definition of world models will blur, which we should talk about with Omni. And it will still... I think the coding agents will look like some sort of world model type system. But you actually do need to make world models useful for real things; you need scaffolding. So I think there are a bunch of interesting startups doing work like figuring out what is the scaffolding for world models, so that you can take them from these very open-ended inherent design of world models, very open-ended spaces, and do it in a tangible way so that it's grounded in a use case that you could use in a recurring way. That could be somebody maybe will figure out the scaffolding for world models to make games possible, but the inherent nature of world models right now I think makes it actually not well suited for games in their current form. But the progress has been crazy, so who knows, maybe in two years the versions will be able to. But at least in the short term, it's coding agent plus some sort of game engine, I think, where you'll see way more alpha from a games perspective.

有道理。好吧，你说世界模型的定义很模糊。

That makes sense. Okay, so you said the definitions of world models are blurry.

我们能展开聊聊吗？

Can we unpack that?

是的，我认为 Omni 就是一个例子。我们在 IO 大会上发布了它，它可以接受任意输入并生成任意输出。我觉得 Demis 向世界介绍它时，恰当地将其称为一个世界模型，因为它对世界有深刻的理解。从技术角度看，它和我们之前构建世界模型的方式不同。我不是架构专家，但它在架构上确实与过去不同，我认为这是积极的，因为它更接近可能更具可扩展性的方式。历史上，世界模型非常不可扩展，运行传统的在线世界模型成本极高。

Yeah, I mean, I think Omni is an example of this. You know, we launched this at IO. You can sort of take any input, create any output. And I think Demis sort of framed it to the world, rightfully so, as a world model because of just the level of understanding that it has of the world. I think that technically looks different than the way we've done world models before. I'm not an architecture expert, but it is different from an architectural standpoint than what's happened in the past, which I think is positive because it's getting closer to some of the ways in which it might actually be more scalable. Historically, it's been super not scalable. It's very, very expensive to run traditional online world models.

是的。

Yeah.

Genie 就是这样的例子。所以，如果你把传统世界模型看作一种近乎动作条件的视频模型，那么现在我们说世界模型，实际指的是对世界有一定理解的模型，而不是严格意义上的动作条件视频模型。

Genie being like, yeah, okay. So if you think of traditional world models as being like an action-conditioned video model almost, then right now, when we say world model, what we actually mean is a model that has some understanding of the world, as opposed to being strictly technically an action-conditioned video model.

是的。

Yeah.

有趣的是，它既有对世界的理解，又能做到很多——这让我觉得界限模糊——它能完成很多相同的用例。目前还不是实时的，但它能完成你描述或通过那个世界模型视觉创建的那些用例，这对我来说是最有趣的。所以我觉得世界模型与视频模型之间的区别会发生变化，并以不同于以往明显的方式展开。

And so the interesting thing though is it has understanding of the world, but then it also has that really great—and that's where the line is blurry to me—where it can do a lot of those same use cases. It's not real-time right now, but it can do a lot of those same use cases that you would describe or visually create with that same exact world model, which I think is what's most interesting to me. So I do feel like this world model versus video model thing is going to change and play out in a different way than was obvious before.

它在底层是如何工作的？你能分享多少？是 Gemini 加上视频模型吗？还是完全不同的东西？

And how does it work under the hood? Like whatever you're able to share? Is it Gemini plus video models? Is it something different entirely?

它是一个单一模型，我认为这是关键。这实际上是原始目标的一部分：过去你需要训练大约八个不同的模型来完成所有这些事情。你有文本模型（基础 Gemini 模型）、音频模型、音乐模型（Lyria）、Nano Banana、视频模型，还有一整套音频模型。如果我们和客户能用一个模型完成所有事情，那就太好了。所以这是一个新的设置，使得这成为可能。它不是路由到多个不同的模型——你可能想象我们之前就能做类似的事情，做一个 Gemini Omni 模型——但这是一个真正的 Omni 模型。它从当前效果最好的用例开始，也就是可用的视频编辑功能。技术上，它也能处理其他事情，只是质量还不够完美，不是最先进的，所以我们还没有推出。这也是 Omni 模型的第一次迭代，是 Omni Flash 模型的第一版。未来我们会有更强大、更厉害的版本，那会非常令人兴奋。

It is a single model, which I think is the important part. This was actually part of the original desire: you were training like eight different models to do all of those things historically. You have a text model with the baseline Gemini model, you have audio, you have music models with Lyria, you have Nano Banana, you have video models, you have a whole suite of audio models. It would be great for us and our customers if you just had a single model to do all those things. So it is a new setup that sort of makes that possible. It's not routing to a bunch of different models, which you could have imagined we could have done something like that actually before and done a Gemini Omni model, but this is a true Omni model. It's starting with the use case that works the best right now, which is why it's the one that's available: this video editing capability. Technically, it's functional with the other things, it's just that the quality isn't perfect and is not state-of-the-art. So we haven't rolled that out yet. It's also just the first crank of the model turn on Omni. It's the Omni Flash model, the first iteration. And so we'll have much more capable, powerful versions, which will be exciting to see.

嗯。所以我们可以编辑这个场景，让它看起来像我们……

Mhm. So we could edit this set so it looks like we're...

是的，没错。我们——我想要这个——我们之前在镜头外聊过，我们应该在开场时做这个，因为我觉得这会让所有东西都更强大。我见过一些例子，其中微妙的细节让我意识到这是世界理解在发挥作用。我之前做演讲时，和我的朋友 Tulsi 一起上台，她领导模型团队——我不知道你是否邀请过她，但她很棒，我很喜欢 Tulsi。我当时对观众中的一个人说，让他们编辑视频，他们真的拍了照片，用 Omni 实时编辑，然后一只狗出现在舞台上。在编辑版本中，其他嘉宾低头看到狗，轻轻笑了笑。当时我正在发表观点。他们是在笑你的笑话。是的，不是我的笑话，他们笑的是狗出现。狗跳到我腿上，我注意到它，继续讲话，抚摸它之类的。要完美呈现这些细节非常微妙，而模型做到了。这非常有趣，我还在消化这对我们制作内容的方式意味着什么。

Yes. Yeah, yeah. We—I want this—we were talking off camera, we should do that for the intro because I think it just makes all this stuff more capable. And I've seen these examples of such subtle nuance that make me appreciate that it's the world understanding playing out. I was giving a talk and was on stage with my friend Tulsi, who leads the model team—I don't know if you've ever had her on before, but she's amazing. I love Tulsi. And I had mentioned to someone in the crowd to edit the video, and they literally took the picture, edited it with Omni in real time, and this dog came on the stage. In the edited version, the other guests sort of looked down and saw the dog, they chuckled a little bit. This is while I'm opining about whatever. They were laughing at your jokes. Yeah, it was not my jokes. They laughed at the dog coming up. It jumps onto my lap. I sort of acknowledge the dog, I keep talking, I'm petting it or whatever. And just there's so much subtlety in getting that right, and the model crushed it. And it's very interesting, and I'm still trying to absorb and digest what that means for the way we make content and all these other things.

这太有趣了。

That's so interesting.

是的。

Yeah.

我是生成式媒体及其意义的坚定支持者。我们考虑过，对于播客来说，视觉效果和内容同样重要。这才是最初吸引人们注意力的方式，对吧？所以，我很期待体验 Omni。

I'm the biggest bull on generative media and what it means. I mean, one of the things we've thought about for our podcast is the visuals matter as much as the content. That's how you catch people's attention in the first place, right? And so okay, I'm excited to play with Omni.

我也很兴奋。我想作为内容创作者，你可能也有同感，但我个人历来不使用 AI 来制作我发布的任何内容。所有内容都是我的文字、我的声音、我的形象。我觉得这其中蕴含着巨大的价值和真实性。所以我更希望是我本人，而不是某个 AI 版本的我。我喜欢 Omni 的一点是它不改变我本人。它改变的是其他那些不属于我的部分。我没有选择我们周围的布景或咖啡桌。所以我们的文字可以保持不变，你可以改变那些非个人的部分，用它们做更有趣的事情，我觉得这非常酷。这感觉就像我希望生成式媒体成为的样子：不是一堆 AI 化身，而是……

I'm excited too. And I think you probably feel this way as somebody who makes content, but I've historically been very, for myself personally, I don't use AI to make any of the content I produce. It's all my words, it's always my voice, it's always my image and picture showing up. I feel like there's just so much alpha and authenticity. So I would much rather it be me than some AI version of me. What I like so much about Omni is that it's not changing me. It is changing a bunch of these other bits which are not me. I didn't choose any of the set around us or the coffee table. So our words can stay the same, and you can change these bits that are not personal and do something more interesting with them, which I think is really cool. It feels like the version of what I want generative media to be: not a bunch of AI avatars, it's...

海岛视频？

Island videos?

完全正确。真的。它就是原始内容，是那个人。人的特质还在，只是被改变和放大了。

Exactly. Truly. It really is the original content. It's the person. The personhood is there. It's just different and amplified.

非常有趣。好的，我很期待体验它。

Super interesting. Okay, I'm excited to play with it.

是的，我们应该结束后就发一些提示词，试试看。

Yeah, we should send some prompts right after this and try some things.

不过别管水果视频了。我对两者并存的世界感到满意。在编程方面，你们在 AI Studio 中推出了让人们编写 Android 应用代码的功能。我很想听听目前进展如何，以及你们计划将其发展到什么程度。

Mind the fruit videos though. I'm happy for a world of both. On the coding side, you launched the ability in AI Studio for people to write code Android apps. I'd love to hear how that's going so far and where you plan to take that.

是的，这非常令人兴奋。

Yeah, it's super exciting.

我认为 AI Studio 的战略性目标之一——这其实基于来自生态系统和开发者的大量反馈——就是谷歌有太多产品了。在创业或实现创意的过程中，你通过无数不同的方式接触谷歌。所以我们有一个首要原则：如何将功能引入 AI Studio，让你无需在谷歌的九个不同界面间切换，就能接触到谷歌生态系统的其他部分。Android 就是一个很好的例子，不仅体现了这一点，还让那些原本不会开发 Android 应用的人也能做到。我就在 AI Studio 里开发了我的第一个 Android 应用，这非常酷。

I think one of the strategic things for AI Studio, and this is based on a lot of feedback from the ecosystem and from developers, is that there are so many Google products. There are so many different ways you touch Google through all these different journeys of building a startup or bringing an idea to life. So we have this first-class principle of how do we bring things into AI Studio that make it so that you are exposed to other parts of the Google ecosystem without having to go through nine different UIs across Google. Android is a great example, not only of that but also of enabling people who wouldn't have otherwise built an Android app. I literally built my first Android app in AI Studio. It's very cool to see.

是什么应用？

What is it?

嗯，我就做了个植物应用，不是加密货币应用。就是个植物类的。我在后院种树来着。

Yeah, I just did a plant app, not a crypto app. Just a plant one. I was planting trees in my backyard.

那挺酷的。

That's cool.

对，所以就是在试水的时候玩了个园艺应用。我还没想出那种突破性的移动应用创意，但我会想点东西出来，看看能不能在 App Store 上竞争。

Yeah, so it was just playing around with a gardening app as I was kicking the tires. I haven't had my breakthrough idea yet for a mobile app, but I'm going to come up with something and see if I can compete on the App Store.

你有没有看到过什么用代码写出来的东西在 App Store 上大获成功？

Have you seen anything coded that's really flying in the App Store yet?

问得好。看看分析数据会很有意思。我确信它加速了 App Store 上的很多事情，但具体程度不清楚。我个人不认识谁做到了。

That's a good question. It would be interesting to see some analysis. I'm sure it's accelerating a lot of things on the App Store, but I don't know how much. I don't know anyone personally who's done that.

嗯。

Yeah.

这很有趣，我还想提一点：今早我最后一次看数据时，自上周以来 AI Studio 里已经构建了 35 万个 Android 应用，这太疯狂了。而且令人兴奋的是，这 35 万个应用可能以前根本没人会去构建。其中很多也是个人应用。所以我认为，虽然可能千禧一代更超前，但你现在为自己解决个人问题而构建软件的想法非常真实。人们正在这样做。这是很多这类产品最常见的用例之一。能够解锁手机的大量原生功能也很有意思，因为你在不同地方拥有那么多上下文。所以我对这个机会感到非常兴奋，Android 感觉正在成为构建者的平台。

It is interesting, and I was going to make the observation too that the last time I checked the numbers this morning, it was like 350,000 Android apps built in AI Studio since last week, which is crazy. And excitingly, it's 350,000 apps that probably no one was going to build before. A lot of these are personal, too. So this is where I think maybe Gen Y is farther out there, but I think the idea of you building software to solve your personal problem is very real right now. People are doing that. It's one of the most common use cases of a lot of these products. Being able to unlock a bunch of the native capabilities of the phone is also really interesting because you have so much context that's in different places. So I'm getting very excited about that opportunity, and Android feels like it's becoming the platform for builders.

现在网络已经如此强大，应用和网页相比还有意义吗？

Does it matter that something is an app versus just the web is so powerful now?

嗯，看到这一点如何发展也很有趣。网络确实强大。但操作系统有一些你无法解锁的东西，比如大量原生丰富性，让体验感觉丰富得多。我其实想到了短信，所有主流操作系统中的短信体验对我来说都比我用过的任何 AI 聊天应用丰富得多。如果我能在我用的任何短信应用里直接和 AI 对话，我会比不得不去另一个应用开心得多。因为我觉得我们也被操作系统所习惯了。

Yeah, that is also very interesting to see play out. Web is definitely powerful. There are certain things that the operating systems have that you just can't unlock, like lots of native richness that actually make experiences feel so much richer. I think about this for text messaging actually, that the text messaging experience in all the main operating systems feels way richer to me than any AI chat app I've ever used. If I could just talk to AI in whatever texting app I use, I would be way happier than having to go to some other app. Because I think we're also just conditioned on the operating systems.

嗯，有道理。好，我想问一下关于模型吃掉工具框架或模型吃掉脚手架的问题。你怎么看？

Yeah, makes sense. Okay, I want to ask about the model eats the harness or the model eats the scaffolding. What are your thoughts?

嗯，我认为这是真的，而且我认为部分原因在于我们历史上认为的模型已经不再是模型了。两年前大语言模型流行时，模型实际上只是一组权重。它是一组权重，关键是如何尽可能简单地输入 token 并输出 token。我认为我们逐步地——我们仍然称之为模型，仍然称之为 Gemini 3.5、GPT 什么的和 Claude 什么的——但实际上它不再只是权重了。它是一个围绕权重构建的不断扩展、蔓延的系统，支持许多下一代体验，从智能体式工具调用到所有托管工具、搜索、代码执行等。模型现在在容器中启动，并且有某种智能体工具框架之类的东西。所以脚手架通常比直接烘焙到模型中的东西领先几步。然后最终发生的是模型吃掉那个脚手架，它成为原生模型系统的一部分。在某些情况下，外部脚手架仍然有价值，比如搜索可能是一个例子。很多人使用不同的搜索提供商和不同的用例。所以当然，也许模型可以原生使用搜索，但你也想要别的东西。代码执行是另一个例子。但确实感觉智能体工具框架现在是典型的例子，每个人都觉得我们必须去构建一个工具框架，而工具框架就是 alpha 所在。我认为这在 12 个月内可能不再成立，至少以我们今天对工具框架的理解来看。我认为模型会消化掉其中很多东西。它会被上游化到模型中，而 alpha 现在会在别处。它不会在于尝试自己构建工具框架，因为模型原生就能做到。

Yeah, I think it's true, and I think part of this is that what we have historically thought of as the model is not the model anymore. Two years ago when LLMs were popular, the model was actually just a set of weights. It was a set of weights, and it was really about how you can, as simply as possible, send tokens in and get tokens out. And I think we've just progressively step by step, we still call it the model, we still call it Gemini 3.5, GPT whatever, and Claude whatever, but it's actually not just the weights anymore. It's an entire expanding, sprawling system built around the weights that enables a lot of these next-generation experiences, from agentic tool calling to all these hosted tools, search, code execution, etc. The models are now being spun up in containers and sort of have an agent harness and all that stuff. So the scaffolding is often a couple of steps ahead of what is baked directly into the model. And then what ends up happening is the model eats that scaffolding and it becomes part of the native model system. There's still value in having external scaffolding in certain cases, like search maybe is an example. There are lots of folks who use different search providers and different use cases. So sure, maybe the model can natively use search, but you also want something else. Code execution is another example. But it does feel like the agent harness is the quintessential example right now, where everyone is like, we have to go build a harness, and the harness is where the alpha is. I think that perhaps won't be true, at least in the way we think of the harness today, in 12 months. I think the models will have sort of digested a bunch of that. It'll be upstreamed into the model, and the alpha will be somewhere else now. It won't be in trying to spin your own harness because the model just does it natively.

但我认为人们构建自己的工具框架的部分原因是，如果你使用任何特定模型提供商的工具框架，你就会被锁定，对吧？所以很多应用公司想要灵活性，这就是他们构建自己工具框架的原因。

But I thought that part of the reason why people are building their own harnesses is because if you use a harness from any given model provider, you're locked in, right? So a lot of the application companies want flexibility, which is why they're building their own harnesses.

嗯，我认为这是脚手架故事的一部分：一开始可能确实如此，但随着模型能力的提升，它会逐渐变得不那么正确。如果一个模型不能使用另一个工具框架，那它就不是一个通用模型。所以重要的是要指出——我几周前在另一次对话中提到过——我们需要类似工具框架基准测试的东西，它实际上衡量所有这些不同模型适应不同工具框架的能力。我觉得这似乎是生态系统应该衡量的合理事物。我很好奇哪些模型实际上是最好的，但我认为随着时间的推移，你会期望它们能够使用每一个工具框架。

Yeah, and I think that's part of the scaffolding story: it starts out perhaps true, but as model capability improves, it becomes less true over time. You don't have a generalized model if it can't use another harness. So it is important to note—I mentioned this in another conversation a few weeks ago—we need something like a harness bench, which actually measures how good all these different models are at adapting to all the different harnesses. I feel like that seems like a reasonable thing we should measure as an ecosystem. I'd be curious to see which models are actually best, but I think over time you'd expect they'd be able to use every harness.

那应用层呢？当模型吞噬了工具链和周边的一切，独立公司还有希望生存下去吗？

What about the application layer? How do you think about where independent companies can have a hope of surviving when the model eats the harness and eats the stuff around it?

是的，感觉机会非常多。这两件事都成立：一方面，现在比以往任何时候都更有机会去构建东西；另一方面，模型的能力也比以往更强。存在能力过剩的威胁，我认为其中蕴含着巨大的阿尔法。模型公司可能会去解决非常通用的问题，但垂直领域有巨大的价值。如果你有领域专长，了解客户和生态系统，你就能跑赢最顶尖的模型实验室，因为专注是初创公司的超能力。只要你能专注，你就能做成任何事。像谷歌这样的大公司有很多产品和用户，所以他们无法在一个领域专注。初创公司则不同。24 个月前我们还在想，初创公司的机会会不会变少，但事实并非如此。相反，机会甚至更多了。编码工具帮助你缩小与大型公司的差距，因为你可以更快地运行和编写软件。智能体式原语是一个新的产品类别。这其中存在风险，但不同公司的风险偏好不同。如果你愿意承担更多风险，你就能赢得那些同样愿意冒险的用户群体。所以，机会非常多。

Yeah, it feels like there's so much opportunity. Both things feel true: on one hand, there's never been more opportunity to build something; on the other hand, models are doing more than ever. There's the threat of capability overhang, which I think has huge alpha. There's the threat that model companies go after very general problems, but there's so much value in verticalized domains. If you have domain expertise, know the customers and ecosystem, you can run laps around even the best model labs because focus is the superpower of startups. If you can focus, you can do anything. Big companies like Google have many products and users, so they can't focus in one domain. That's not true for startups. 24 months ago we wondered if there would be less opportunity for startups, but that hasn't played out. If anything, there's even more opportunity. Coding tools help close the gap with larger companies because you can run faster and write software quicker. The agentic primitive is a new category to build products around. There's risk involved, but different companies have different risk appetites. If you're willing to take more risk, you can win a user cohort interested in that. So, there's so much opportunity.

我想聊聊 Google DeepMind 的文化。现在在 GDM 内部感觉如何？我们在 AI Sense 上请过 Demis，他非常鼓舞人心。我听说 Sergey 回来了，你们还有 Noam Shazeer 也回来了。给我讲讲现在在 GDM 的感受吧。

I'd love to talk about Google DeepMind's culture. What does it feel like to be inside GDM right now? We had Demis at AI Sense. He was so inspiring. I've heard Sergey's back. You guys have Noam Shazeer back. Walk me through what it's like to be at GDM right now.

这太不可思议了。我努力去感受这一切，因为这是一个重要的时刻。在混乱中，我尽可能多地反思。GDM 的文化很有趣。我有三点观察：第一，回到专注这个话题，我们正在做很多事情。从投资组合的角度看，我们拥有最强的组合之一。但有时你会看到其他实验室在我们投资不足的领域领先。看到我们如何缩小差距很酷。我看过几次《Demis Thinking Game》纪录片。你能看到最初的文化：把一群聪明人聚在一起解决问题。我喜欢这一点。另一个观察：文化从领导者渗透出来。Demis 是一位诺贝尔奖科学家，你在 DeepMind 的文化中能感受到这一点。Sam Altman 是世界上最优秀的商人之一，你在 OpenAI 的文化中能看到这一点。我对 Dario 不太了解，但 Anthropic 看起来很有趣，有点深奥。我喜欢 DeepMind 的科学方法。Demis 开始这个使命是为了解决疾病。很容易迷失在基准测试的竞争中，但我们需要记住，这样做是为了解决人类真正的问题。我最喜欢的硅谷名言来自 Gavin Belson：“我们不能让别人比我们更能让世界变得更好。”这就是此刻的感觉。这不是零和游戏。关于 DeepMind 文化的最后一点：我们是谷歌的引擎室，这现在是 DeepMind Twitter 账号的简介，我很喜欢。

It's incredible. I try to take it all in because it's a moment. I reflect as much as possible in the chaos. GDM's culture is interesting. Three observations: One, back to focus, we're doing a lot of things. From a portfolio perspective, we have one of the strongest portfolios. But you see moments where another lab pulls ahead in an area where we under-invested. It's cool to see how we close that gap. I've watched the Demis Thinking Game documentary a few times. You see the original culture: get a bunch of smart people together and solve the problem. I love that. Another observation: the culture permeates from the leaders. Demis is a Nobel Prize scientist, and you feel that in the DeepMind culture. Sam Altman is one of the world's best businessmen, and you see that in OpenAI's culture. I don't have a strong sense of Dario, but Anthropic seems interesting and somewhat esoteric. I like the scientific approach at DeepMind. Demis started this mission to solve disease. It's easy to get lost in the competitive race of benchmarks, but we need to remember the reason is to solve real human problems. My favorite Silicon Valley quote is from Gavin Belson: 'We can't let other people make the world a better place more than we can.' That's what this moment feels like. It's not zero-sum. The last thing about DeepMind's culture: we're the engine room of Google, which is now the Twitter bio of the DeepMind account, which I love.

所以一方面，你有根深蒂固的实验室文化，另一方面，你有整个 Google 生态系统中的合作伙伴——从我们之前谈到的 Android，到 Google Cloud，到 Gmail，到 Workspace 等等。这是一个有趣的融合。我认为有很多研究工作在进行，但也有大量的应用工作，与一些前沿客户合作。将 Gemini 部署到拥有十亿用户的产品中，是世界上只有两家公司才有的问题，而我们拥有 13 个这样的产品。Google 现在经常经历这些。看到这一切发生，看到为实现这一目标而进行的创新，是一个非常有趣的地方。我觉得只有在 Google 内部才能做到这一点，这真的很酷。

So it's like on one hand you have the deep-rooted lab culture, and on the other hand you have all these partners across the Google ecosystem that we're collaborating with—everybody from Android that we talked about earlier, to Google Cloud, to Gmail, to Workspace, etc. So it's an interesting blend. I think there's lots of research work happening, but there's tons of applied work happening to work with some of the forefront customers. Deploying Gemini to billion-user products is a problem that only two companies in the world have, and we have 13 of those products. Google goes through this all the time now. It's such an interesting place to see that happen and see the innovation that takes place to make it possible. I feel like you can only do that inside Google, which is really cool.

说得太好了。你刚加入时发了很多推文，他们有没有让你很头疼？你需要得到公关部门的批准吗？

Beautifully said. Did they give you a lot of heartburn when you joined and were tweeting a lot? Did you have to get sign-off from comms?

这是个好问题。我在 Google 的经历中，一个亮点就是市场部和公关部的同事们非常棒。他们的工作是保护 Google，确保我们讲述正确的故事，避免坏事发生。所以我非常感激并与他们合作。但能够以真实的方式讲述与开发者产生共鸣的故事，而不必每次都让推文获得批准，这真是一次不可思议的经历。这是一种非常积极的文化。我一直努力不辜负与那些同事建立的信任和善意。但这一切都非常积极，因为归根结底，Google 很难讲述这个真实的故事。这是一家大公司，有很多人和很多意见。你拿走了 Google 的魔力，通过很多人和流程把它稀释了，结果错过了那个美丽的故事：Google 正在做世界上最有意思的技术，帮助用户解决一些最困难的问题。能帮助讲述这个故事是一种荣幸。所以这很有趣，我很享受。

That's a good question. One of the silver linings of my Google experience has been how great the folks across marketing and comms are to work with. Their job is to protect Google, make sure we tell the right story, and make sure bad things don't happen. So I have a ton of appreciation and partnership with them. But it's been an incredible experience to be able to go try to tell the story that resonates with developers in a way that feels authentic, without having to get my tweets approved all the time. It's a very positive culture. I'm always trying to walk the line of not burning the trust and goodwill I've accumulated with those folks. But it's been super positive, because ultimately it's really hard for Google to tell this authentic story. It's a big company with a lot of people and a lot of opinions. You take the magic of Google and water it down through a lot of people and process, and you miss the beautiful story: Google is doing the most interesting technology in the world and helping our users with some of the hardest problems. It's a privilege to help tell that story. So it's a lot of fun. I enjoy it.

我喜欢你正在做的事情，也喜欢 Josh 正在做的事情。你们为这个时代最重要的问题注入了真诚的人性化触感。

I love what you're doing. I love what Josh is doing. You guys have put a really sincere human touch on, as you put it, the most important problem of our time.

谢谢。

Thank you.

太好了，Logan。非常感谢你今天来做客。这是一次非常广泛的对话——从智能体和编码到世界模型、工具和 DeepMind 文化。这里有很多精华。谢谢你今天来。

Well, wonderful Logan. Thank you so much for joining me today. This was a very far-ranging conversation—everything from agents and coding to world models and harnesses and DM culture. Lots of nuggets here. Thank you for joining me today.

这非常有趣。谢谢你邀请我。我很期待看到大家在我们一直坐着的地方——也许就在我们面前——会做出什么。

This was a ton of fun. Thank you for having me. I'm excited to see what the folks cook up where we've been sitting this whole time—maybe in front of us.

也许还会有一只狗。

And maybe they'll be a dog.

一只狗，之类的。

A dog, something.

梦想成真。

Come true.

我喜欢。

I love it.

太棒了。谢谢 Logan。

Awesome. Thanks Logan.

不客气。

Of course.