Google DeepMind 的 Logan Kilpatrick：为何模型会吞噬缰绳

所以我们可以通过剪辑，让这段看起来像是

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

在这里。

here.

明白吗？

Okay?

对。

Yeah.

没错。

Yeah.

我希望把这段放在我们之前在镜头外聊天的那个位置。

I I want this where where we were talking off camera.

开场白就该这么处理，我觉得这样能让整个内容更有感染力。

Like, we should do that for the intro because I think it just, like, makes all this stuff more capable.

我看过一些例子，那种极其微妙的细节，让我感受到仿佛世界本身在自然演绎。

I've seen these examples of, like, such subtle nuance that make me appreciate that it's like the world understanding playing out.

我正在台上演讲，旁边是我朋友图尔西，她负责带领模型团队。

I was giving a talk and was on stage with my friend, Tulsi, who leads the model team.

我跟台下一个人提了句，让帮忙剪辑视频，结果他们真的当场用Omni拍了张图，实时编辑，一只狗居然就出现在舞台上。

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.

在剪辑后的版本里，其他嘉宾低头一看，发现了那只狗。

In the edited version, the other guests looked down and see the dog.

他们忍不住笑了出来。

They chuckle a little bit.

而我当时正聊着AI跟笑话之间的关系。

This is while I'm opining about whatever AI is about your jokes.

是的。

Yeah.

那根本不是我的笑话。

That it was not my jokes.

大家笑的是那只突然冒出来的狗。

They they laugh at the dog coming up.

它直接跳到了我腿上。

It jumps onto my lap.

我稍微冲它点了点头。

I sort of, like, acknowledge the dog.

我接着继续讲我的话。

I keep talking.

我就顺手摸了摸它，之类的。

I'm, like, petting it or whatever.

要拿捏好那种微妙的分寸其实特别难，但这个模型做得太到位了。

And just, like, there's, like, so much subtle subtlety in getting that right, and the model crushed it.

这事儿太有意思了，我还在慢慢消化，想想这对我们做内容的方式还有其他方方面面意味着什么。

And it's just it it very interesting and, like, still trying to, like, absorb and digest, like, what that means for, you know, the way we make content and all these other things.

这真的太有意思了。

That's so interesting.

很高兴能请到洛根来参加节目。

I'm delighted to have Logan on the show.

Logan 负责 Google AI Studio 和 Gemini API。

Logan runs Google AI Studio and the Gemini API.

你花很多时间思考和打造下一代开发者的工具。

You spend a lot of your time thinking and building for the next generation of builders.

是的。

Yes.

所以我特别想和你聊聊从 AgenTik AI 到 AI 编程、世界模型这些话题，正好赶在 Google IO 之后。

So I'm excited to talk to you about everything from AgenTik AI to AI coding, world models, more today, and right off the heels of Google IO.

还有什么比这更合适的时机呢？

So what better timing?

对。

Yeah.

我特别兴奋。

I'm super excited.

谢谢你邀请我。

Thank you for having me.

太好了。

Wonderful.

我们先聊聊 AgenTik AI 吧。

Let's start with AgenTik AI.

Sundar 在 IO 大会上开场就说了，现在进入了 AgenTik Gemini 时代。

So Sundar opened at IO by calling this the AgenTik Gemini era.

对 Google 来说，AgenTik AI 意味着什么？

What does AgenTik AI mean for Google?

对。

Yeah.

这是个好问题。

It's a good question.

我觉得，如果你一直关注的话，其实我们在 Gemini 2.0 的时候就提到过一些相关想法，只是那时候可能有点早。

I think and we were, we sort of if you if you followed closely, we did sort of mention some of these things back with, like, Gemini two point o, which I think was, like, a a little bit early.

所以现在到了 Gemini 3.5 这个阶段，才真正开始落地，我们确实进入了以智能体为核心的产品时代，尤其是在 Gemini 方面。

And so I think this era, this, like, Gemini 3.5 era, feels like it's actually becoming true now, we're in the era of agentic coating, or agentic products and everything agents as far as Gemini goes.

对我们来说，这个智能代理层——其实我们在IO大会上已经宣布过，它是基于Antigravity Agent框架构建的——现在成了贯穿谷歌所有产品的主线，所有产品都以此为基础了。

I think for us, this agentic layer, and I think we announced this actually at IO, sort of being powered by the Antigravity Agent harness, is this additional through line for Google that sort of connects all of our products that they're sort of based on now.

在Gemini之前，我们那大约五十款产品其实并没有一条清晰的主线。

Historically, prior to Gemini, there actually wasn't a through line for the probably sub 100 number of Google products that we have, the 50 Google products we have.

那时候真的没有一条主线。

There wasn't a through line.

我们有了Gemini。

We had Gemini.

Gemini这才成了那条主线。

It became this through line.

现在所有产品都在以某种方式使用Gemini。

Everything is now sort of using Gemini in some way.

现在Antigravity也在走同样的路，所有产品都在重新调整，变成原生的智能代理产品，能主动为用户做事，帮他们完成任务。

That's now becoming true for Antigravity as sort of all of the products rebase to become sort of like agentic native products and actually taking action on behalf of users and helping them get things done.

你看到一条新的主线正在浮现，我觉得这真的特别有意思。

You see this new through line emerging, which I think is actually really, really interesting.

不好意思，能帮我确认一下，Antigravity 是那个 IDE 吗？

I'm sorry, help me with Antigravity is the IDE, right?

或者说是的，不是

Or the Yeah, non

Antigravity 其实涵盖了很多东西，我觉得这对我们来说反而是个机会。

antigravity is a lot of things, which I think is sort of, again, is an opportunity for us.

你有核心的 IDE。

You have sort of a core IDE.

如果你希望在网页上使用，也可以选择以智能代理为核心的体验。

You have sort of the agent first experience if you want it on the web.

你还有命令行工具。

You have a CLI.

你也有软件开发工具包。

You have an SDK.

我觉得，虽然我们之前可能没这么强调过，但其实它已经形成了一个完整的产品生态，目的就是让开发者无论在哪儿都能用得上。

So I actually think, and I don't know how much we've framed it this way, but it really is an ecosystem of stuff that we built, and it's designed to sort of, like, meet developers wherever they are.

所以如果你想要一个托管的智能代理，又不想操心底层基础设施，可以直接通过 Gemini API 来使用。

So you could use it through the Gemini API if you want to and you want a managed agent that you don't have to do any of the the sort of infrastructure work for.

更有趣的是，这不仅仅局限于 Antigravity 这套工具生态。

And then the most interesting bit is, like, it's not just the ecosystem of Antigravity stuff.

实际上，同一套底层框架还在支撑着谷歌其他所有产品。

It's also powering, like, literally, it's the the same harness is actually powering all the other Google products.

Antigravity 将为搜索、Gemini 应用、云平台和 AI Studio 等多个产品提供智能代理支持，这真的让人兴奋。

So Antigravity will be powering a bunch of agent stuff in search, in the Gemini app, across, like, cloud and and AI Studio, really is exciting.

我明白了。

I see.

以前是靠 Gemini API。

So it used to be the Gemini API.

也就是说，语言模型才是贯穿所有谷歌产品的核心线索，让 AI 能融入每一个产品中。

So, like, the language model was the through line in terms of how AI gets baked into every Google product.

对。

Yeah.

现在不只是API了，连编码框架也在用。

And now it's not only the API, it's the the coding harness

没错。

Exactly.

这些产品都在用这个框架，所以实际上是代理本身在推动产品中更多代理功能的实现。

That's that's being used in each of these products, and, therefore, it's like coating agent itself that's driving more agentic properties inside the products.

我觉得这个说法挺准确的？

And I think Fair description?

很准确，确实如此。

Fair fair description.

我觉得更广义地说，它就是代理框架。

I think more generically too, it's just, like, it is the agent harness.

我觉得‘涂层’只是代理框架的一个具体应用，虽然很强大，但编程其实已经证明，它不仅是适合编程，更是通用的代理框架。

I think, like, coating as sort of, like, a specialized use case of the agent harness, I think, is is obviously powerful, but it is, like, coding has proved to be the general purpose agent harness in addition to also working really well for coding.

代理框架和编程框架是同一个东西吗？

Are agent harness and coding harness synonymous or not?

这中间确实有细微差别。

There's definitely nuance.

我觉得，通过专门化还能进一步优化，实际上你也能看到，AI Studio用的代理框架就稍微针对了编程这个场景，而Gemini应用用的代理框架则更偏向于面向消费者的24/7全天候代理。

I think there's, like, optimization that you can squeeze out of, like, specializing, and actually, you see this where technically the agent harness that gets used for the way that AI Studio uses it is a little bit specialized for the vibe coding use case, and the way that the Gemini app is using the agent harness is a little bit specialized for the consumer, always on 20 fourseven agent.

所以我认为，底层框架大概有80%是共通的，然后根据编程或者其他具体用途再做定制。

So I think you have that base harness that probably has 80% of the same stuff, and then you specialize for coding or for whatever the use case is.

挺有意思的。

Interesting.

你如何看待现有业务被侵蚀的问题？尤其是现在你们正大力推动代理功能的时候。

How do you think about the cannibalization of the existing business, especially now that you are going much more aggressively into agentic properties?

因为如果只是做搜索或者摘要，这种侵蚀的担忧其实没那么大。

Because I could see, for example, if all you're doing is search or summarization, there's not as much of a cannibalization fear.

但如果你直接替我处理邮件、帮我回信，那我还有必要自己看邮件吗？

Whereas if you're actually going through my emails, replying to them for me, am I even going through my email anymore?

所以我觉得，随着代理功能越来越强，用户在你们产品上花的人工浏览时间可能会越来越少。

And so I could imagine that there's actually just fewer human eyeball hours on your products as a result of having more agentic capabilities.

这么说公平吗？你是怎么看待这种自我冲击的？

Is that fair, or how do you think about the cannibalization?

是的，这确实挺有意思。

Yeah, it's interesting.

我觉得有一点值得注意，一开始，我觉得 Sundar 说得特别好，当时大家普遍认为，AI 能直接回答问题，对搜索业务肯定是坏事。

I think one observation I have is that at the beginning and I think Sundar has done a great job of talking through this is at the beginning of the current AI era, everyone assumed that AI being able to answer questions for you was going to be negative some for search.

但实际发生的情况恰恰相反，AI 反而让搜索变得更有活力了。

And actually, what's ended up happening is it's an incredibly positive sum for search.

人们搜索得更多了。

People are searching more.

做的事情也更频繁了。

People are doing more.

而且

And

代理也在主动搜索。

agents are searching too.

是的。

Yeah.

实际上，同样在这个时期，还兴起了一整个市场，代理在增多，人类的搜索行为也在增加。

Actually, again, there's this whole market that spawned at the same time that agents are doing more, at the same time that humans are also searching more.

当然，人类的时间是有限的，但根据我早期对这些趋势的观察，整体来看，这种生态系统的价值创造确实是积极的。

Obviously, there's a finite amount of 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 from an ecosystem value creation.

人类行为会如何演变，未来一两年可能比较清晰，但三到五年后就很难说了，那时候技术更先进，产品形态也可能和现在不太一样。

How the human behavior aspect of it turns out I think is somewhat clear in the next one to two years, much less clear three to five years from now when the technology has improved and the products probably look a little bit different than the way that they do.

但归根结底，这才是产品的成功。

But ultimately, that is the success of product.

我们经常做各种演示和讨论，核心理念就是：开发技术的目的，是为了让它能替你做事。

I think we have a bunch of conversations with demos all the time, and it's like the point of building the technology is so that it can go and do stuff for you.

对谷歌来说，成功可能不是体现在用户在我们产品上花的时间有多长。

Success for Google probably doesn't look like maximizing eyeball time in front of our products.

而是体现在帮用户高效达成目标，让他们能去生活、去做自己想做的事。所以我相信，我们会更倾向于以用户成果为导向，而不是以屏幕停留时间为导向。

It's like maximizing outcome for customers to do the thing that they wanna 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。

Yeah.

我脑子里一直有个词，叫‘代理驱动增长’。

I have this term stuck in my head, agent led growth.

我自己平时用很多编码代理，干脆让代理帮我全权决定基础设施的选型。

It seems to me So I'm using coding agents a lot in my personal time, and I just let the agent make all the infrastructure choices for me.

我不在乎用什么数据库，你告诉我就行。

I'm like, I don't care which database, you tell me.

我问这个是因为，现在写代码确实就是这样。

And the reason I ask is, it's true in coding today.

我觉得以后很多事可能都会这样，比如将来购物也一样。

I would imagine it's maybe gonna be generally true for a lot of things, let's say shopping down the line.

你觉得这会怎么改变广告的运作方式，还有平台的价值捕获方式？

How do you think that's gonna change how advertising works, how value capture works for the aggregators?

感觉这跟之前的趋势很像。

It feels like it's a very similar trend.

这并不完全准确，但很多东西其实都是彼此的替代指标。

This isn't perfectly true, but a lot of these things are just proxies of each other.

我觉得SEO的运作方式，和另一种机制直接相关，只是我一时想不起那叫什么了。

The way that SEO works, I think, is directly correlated with the way that, I forgot what the term now for it.

就像GEO，其实是生成式引擎优化，或者别的什么名字。

It's like GEO is the generative engine optimization or whatever it's called.

所以这些事情之间确实存在很强的关联性。

And so it does feel like there's a lot of correlation between the things.

我觉得，这种变化看起来没我们眼下以为的那么颠覆，因为这些趋势是层层叠加的。

My guess is it looks like much less of a radical shift than I think maybe what we assume right now, just because these things compound on top of each other.

如果用‘爬、走、跑’来衡量代理能力的程度，谷歌的产品系列现在处于哪个阶段？

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?

这是个好问题。

Yeah, that's a great question.

现在肯定还只是在‘爬’的阶段，而且我觉得这背后是谷歌产品本身固有的矛盾。

It's definitely crawl right now, and I think some of this is all of the inherent product tension for Google.

你有超过一千三百亿个用户产品，所以我反而觉得，有些实验性功能可能已经接近走路甚至跑步了，但总的来说，目前大部分产品的体验还是处在起步阶段，毕竟我们肩负着为海量用户打造产品的责任。

You have, what, 13,000,000,000 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 building a product that's being used by lots of people.

我觉得长尾用户还没准备好让AI完全替他们做主。

I don't think the long tail of customers are ready to have AI running and just doing all the things.

他们希望掌握主动权。

They wanna be in the driver's seat.

他们正在小心翼翼地迈出第一步，而谷歌搜索团队可能是最典型的例子。

They're cautiously taking the first step, and I think Google team And Search is maybe the most quintessential example of this.

我觉得他们有责任以一种温和的方式推进，别一下子彻底改变人们使用互联网、与产品互动的方式。

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.

对。

Yeah.

你觉得哪些产品最接近‘走路’阶段？

Which products do you think are closest to the Walk?

这是个好问题。

That's a good question.

我觉得 Gemini 应用是最接近‘走路’级别的。

I think Gemini app is definitely closest to Walk.

至于 Spark，我觉得让一个 24/7 的智能代理随时帮你执行各种操作，这绝对是前沿场景之一；我觉得 Antigravity 也是个例子——它能让你的自主编码代理帮你重建操作系统，处理数十亿个标记，还替你花掉上千美元。

And so for Spark, I think having a 20 fourseven always agent, 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 antigravity is another one where it's like you could have autonomous coding agents rebuilding operating systems and doing billions of tokens and spending thousands of dollars on your behalf.

这些其实也是 GDM 的一部分，从另一个角度看，GDM 在这方面走得非常靠前，而谷歌其他产品则更倾向于逐步推进，我觉得这样其实很合理。

And think those are, again, more And actually, they're in GDM as well as another angle of this, so I think GDM is taking very much a frontier look at this, where I think the rest of Google's products, I think, are more incrementally getting there, which, again, makes reasonable sense to me.

对。

Yeah.

你觉得谷歌最终会形成一两个、三五个 AI 使用界面，还是会成千上万种？

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

这很难说。

It's tough.

我觉得这其实根植于人类使用产品的习惯，我个人觉得，产品之间保持分工和专精是有好处的——如果一个产品啥都替你干了，用起来反而更费劲。

I think a lot of this is actually baked in just 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.

也许将来会有人做出真正神奇的体验，打破这个规律，但大多数普通用户，用通用型产品时，得花更多心思和时间才能让它完成自己真正想做的事；相比之下，我点开日历应用，它就直接显示我的日程，这种体验反而很舒服。

Think would be the default say, 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.

我就不用操心别的事了。

Like, I don't need to worry and deal with anything else.

这就是我对PPT为什么能存在这么久的看法——你想要的信息，就得稳稳地待在同一个地方。

This is my hot take for why slide decks have existed for so long of just like, you know, the thing, the piece of information, you want it to be exactly in the same place.

而且……

And

我觉得人类本来就习惯这种模式，相比之下，生成式界面听着很酷，但我们的大脑真的能适应吗？这难道不是增加了更多认知负担吗？

I think we as humans are just actually very used to that as The opposed to idea of a generative interface sounds so cool to me, but it's like, do our brains really Isn't that just more cognitive overhead for us?

在某些情况下确实如此。

It definitely is in certain cases.

我觉得肯定有人会——世界上聪明人太多了，也许有人能做出一种更自然的体验。

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.

但就我目前来看，一万种可能太极端了。

But to me right now, maybe not 10,000 is the extreme version.

我猜更可能是各种产品分别瞄准不同需求，至于谷歌会怎么走，我也说不准。

I'm guessing it looks more like more products going after different And maybe the other answer is, I don't know what it looks like for Google.

我觉得，整个生态会涌现出更多产品。

For the ecosystem, it looks like a lot more products, I think.

这真的让人很兴奋。

And that's really exciting.

谷歌最终会怎么战略决策——用户是希望我们提供一万个产品，还是只保留三个？

I think how Google will end up strategically deciding, do our customers want to deal with us having 10,000 products, or would it be better to only have three?

这最终得由我们自己来做战略取舍。

Will come down to a strategic decision for us.

这完全说得通。

That totally makes sense.

我跟一些企业客户聊的时候，他们说，人人都在谈智能体AI，但真正见到智能体真正起作用的，只有代码生成类的工具。

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.

你同意还是不同意这个观点？

Do you agree or disagree with that take?

是的。

Yeah.

我觉得这取决于你对‘有效’怎么定义，这其实正是问题的关键所在。

I think it depends what your bar for working is, which I think is a lot of the nuance of this.

如果真想让模型去处理那些它还没达到质量门槛的复杂任务，那确实没错。

Like, I think if truly trying to offload very complicated tasks for domains in which the models haven't actually crossed the threshold of quality, then I think that's definitely true.

它解决不了根本问题，但我特别希望我们能有个方法来衡量这一点。

It's not gonna solve the problem, but this is something that I wish we could measure.

比如OpenRouter就在统计总的token消耗量，这样就能看出，现在世界上的智能总量相比一年前增长了多少。

A good example is OpenRouter, for example, is measuring the total token consumption that's happening, and so you can sort of see these trends play out over time of how much more intelligence is in the world now versus a year ago.

同时，我真正想追踪的是，平均一个智能体任务或一个完整工作流程到底要花多长时间。

In parallel, the thing that I'm actually really interested to measure is how long is the average agent run or the average task actually taking place?

我不知道他们有没有公开这些数据，但我猜他们手里肯定有，其他公司应该也一样。

I don't think it's something that they publish, but I feel like they probably have interesting data, I'm sure there's others.

因为现在新模型一发布，能力一提升，使用量就会明显跳升。

Because I do think you're seeing these new model capability lands or new model drop, and it's spiking up.

虽然目前整体曲线还很低，但已经能看到一些早期迹象了——任务时长在变长，各大模型实验室都在说，我们刚发布的这个新模型，能连续自主运行三天之类的话。

And maybe the curve is still very low right now, but you're seeing those early signs of it spiking upward to long running tasks and all the model labs are talking about, We released this new model and it did three days of autonomous work, or whatever it is.

这是极端情况，但现实中，这种影响正迅速蔓延开来，真的很有趣。

That's the extreme, but I think, in practice, you're seeing that trickling up pretty quickly, which is really interesting.

即使企业还没在编程之外感受到，今年其他那些应用场景也会明显提升。

So even if the enterprises haven't felt it outside of coding, they are going to this year as sort of a bunch of those other use cases get much better as well.

从DeepMind的角度来看，你觉得长周期智能体是一个重要的指标吗？

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

它是不是最关键的那个指标？

Is it the KPI that matters?

这确实很重要。

It definitely matters.

我觉得DeepMind在做很多事，后面我们可以再详细聊。

I think for DeepMind, we're doing lots of things, which we can talk more about later.

目前我们正在推进大量不同的方向。

There's a huge portfolio of different bets that are taking place.

长周期智能体显然非常关键。

Long running agents obviously matters a lot.

我觉得编码代理在这方面也特别重要。

I think also specifically coding agents in that matters a lot.

如果你有一个出色的编码模型，它显然能加速你业务的方方面面。

It clearly is an accelerant of every other part of your business if you have a great coding model.

所以确保我们拥有这一点，我觉得是当前最优先的事。

And so making sure we have that, I think, is super top of mind.

明白了。

Got it.

我想换个话题，聊聊编码方面。

I'd love to shift gears a little bit and talk about coding.

对。

Yeah.

好的。

Okay.

我要问个难题了。

I'm gonna ask a hard question.

我很多做开发的朋友早就开始用Cloud了。

A lot of my developer friends were using Cloud for a long time.

OpenAI注意到这一点，立刻拉响了警报。

OpenAI saw that, declared code red.

现在Codex已经非常强大了。

Codex is now really good.

我觉得我朋友们现在用Cloud和Codex的差不多各占一半。

I would say my friends are maybe split fifty fifty now in using Cloud and using Codex.

我很少听说有人用Gemini，这我一直有点想不通。

I don't hear a ton of them using Gemini, which has always kind of puzzled me.

这背后是啥原因呢？

What's going on with that?

是的。

Yeah.

这是个好问题。

It's a great question.

我觉得还有一个关键点可以补充，这让整个故事变得更有趣了——去年十二月，大家普遍觉得谷歌赢了。

I think there's one there's one part of the story that I'll add, which is which which is makes it even more interesting, which is, December, the narrative was that Google had won.

当我们推出 Gemini 3 的时候，我觉得从模型能力上看，它简直是质的飞跃。

And when we landed Gemini three, I think it was such a profound improvement from a model capability perspective.

当时很多说法都认为，谷歌已经实现了重大突破，彻底扭转了局面。

I think a lot of the narrative was Google has taken a huge leap forward and made that happen.

但作为生态系统的参与者，真正让我惊讶的不是这个风向转变得多快，而是紧接着假期过后，进入一月及之后，又掀起了一波关于智能编码代理的热潮，而那其实没过多久。

And I think what was interesting to see as an ecosystem participant is not how quickly that narrative shifted, but just the next wind of the narrative obviously was all the agentic coding stuff that happened over the holidays and then into January and beyond, and that was not that long ago.

所以现在的情况是

So it is a

我觉得自从那以后，一切都像在超光速前进。

I feel like we've been in warp speed ever since.

确实如此，但这恰恰提醒我们，事情变化能有多快。

Yeah, for sure, but it's a matter of reminder of just how fast things can change.

我觉得这个观察并不过分。

I think the observation is not unreasonable.

我觉得我们背后真正努力的方向，是尽可能快地突破编程领域的边界，而反重力项目正是其中关键的一环。

I do think what's happening behind the scenes for us is trying to push the frontier as fast as possible on coding, and so I think antigravity actually is an important part of that.

我觉得其中一个重要的体会是，如果你没有一个真正能处理长周期、复杂开发任务的产品，想做出一个出色的编程模型几乎不可能，而谷歌正是意识到了这一点。

I think one of the takeaways is that it's actually really hard to make a great coding model for this developer use case of really long running, sweet work if you don't actually have a product that does that, and so I think Google realized that.

所以这才有了Windsurf的收购。

That's why the Windsurf deal happened.

也正因如此，那批人才被招进来，最终打造了反重力项目。我们内部一直在用，而且Sundar在IO大会上还展示过谷歌内部令牌消耗量的增长曲线。

It's why those folks came over and then ultimately built Antigravity, and we've been using internally, actually, and Sundar showed this at IO, just the graph of growth of token consumption inside of Google.

你得让这个引擎转起来，而说到底，这个引擎现在确实在高速运转。

So you need that engine to spin, and the meta comment, again, is the engine is spinning.

模型的进步需要时间，但我对此充满信心。

It takes time in order to actually make model progress, but I'm super confident.

我觉得负责编程的那支团队，内部都管他们叫AI界的复仇者联盟，真的是谷歌最顶尖的一批人，全力以赴地推动这项工作，认真对待每一个细节。我觉得，尽管有人对价格有议论，但Three Flash的推出，确实是这些努力开始结出成果的重要一步。

I think the group of folks who we have working on code is like I describe it as the avengers of AI internally, and so it really is some of the best people inside of Google trying to push the rock up the hill on this stuff and taking it super seriously and trying to push, and I think three flash, notwithstanding some of the conversation about the price and stuff like that, is sort of a step towards actually starting to bring a lot of these capabilities and the fruits of that labor paying off.

这个Flash模型在编程能力上已经超越了我们过去发布的任何专业模型，而之前的那些专业模型本来就已经非常出色了。

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.

还有另一个方面，大家常常忘了还有预训练阶段。

So there's another thread of this also, which is everyone forgets that there's pre training windows.

我觉得真该有人在网上追踪一下这个，挺有意思的。

I wonder, somebody should track this online, which would be interesting to see.

你是说那些大规模训练，到底用过哪些集群？

Meaning the big run, what clusters have been available?

没错。

Exactly.

大规模训练其实是这个故事里很重要的一环，从外面看，可能觉得你们落后了，但实际上你们完全忽略了那些大规模训练到底在哪里进行、预训练资源在哪。

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

所以我觉得，预训练历来是DeepMind的巨大优势。

So I think that also Obviously, pre training has historically been a massive strength for DeepMind.

我们有一群全球顶尖的人才，我特别期待看到他们的成果和这些年的积累。

We have some of the best people in the world, I'm so excited to see the fruits of that labor and everything else that's happened.

3.5 Flash的提升完全是后训练阶段的成果，这真的很棒。

3.5 flash was all post training gains, which is really cool.

这充分证明了那个团队的努力，他们仅靠后训练就实现了如此大的提升，甚至超越了之前的专业模型，真是太棒了。

So a huge testament to the work that that team did to actually make the level of gains and surpass the previous pro model literally just with post training, which is awesome.

你们在内部有多严格地使用自家产品？

How religious are you all about dogfooding internally?

比如，DeepMind的员工还能不能用其他公司的模型？还是说现在大家都必须用Gemini系统，得把产品做得特别特别好？

For example, are DeepMind folks still allowed to use other models, or is it like you guys are using the Gemini harness now, we have to make this really, really good?

是的。

Yeah.

我觉得多用其他模型特别好，因为如果你不接触外面的模型，真的很难理解整个生态到底在发生什么。

I think people, it's so healthy to be using other models just because it's sometimes hard to actually grok what's happening in the ecosystem if you're not.

我什么模型都用。

So I use all the models.

我什么产品都用。

I use all the products.

我觉得DeepMind其他同事也都是这么做的。

I think folks across the rest of DeepMind are doing the same thing.

不过，你确实得用 Gemini 模型。

You definitely have to use the Gemini models, though.

从反馈循环的角度来看，这效果特别好，也是模型进步的关键。

It's just great from a feedback flywheel perspective, and it's part of how they get better.

DeepMind 和谷歌整体有十多万顶尖工程师在使用这些模型并提供反馈，这本该是谷歌的竞争优势——毕竟我们有这么庞大的工程团队、深厚的人才储备，还能做 A/B 测试和线上实验这些事。

DeepMind has, and Google more broadly, has 100,000 plus incredible engineers who are using the models and giving feedback, and it should be a competitive advantage for Google because we have that scale of engineering resources and the depth of the talent and can run AB tests and live experiments and all that stuff.

所以我觉得所有模型都得用，但对大多数人来说，Gemini 已经是日常主力了，这挺好。

So I think you have to use all the models, but I think for the majority of folks, it's like Gemini as the daily driver, which is great.

你相信这种说法吗？就是一旦有了足够好的智能编码模型，研究进展就会加速，形成一个自我强化的循环。

Do you believe in this narrative around a soft takeoff of once you have a good enough agentic coding model, then it accelerates the pace of research progress, and it's a self reinforcing cycle.

这看起来明显是对的，但也许是我被洗脑太深了，才觉得真这么回事。

It seems obvious that that's true, but maybe I'm drinking too much Kool Aid that that's the case.

你现在已经看到迹象了吗？

Are you seeing the signs of it yet?

是的。

Yeah.

你确实能看出来一些迹象了。

You definitely see some signs of this.

从模型的角度来看，这些迹象还处于早期阶段，而其中一部分原因是，像这类大规模训练所需的资源投入实在太大了，所以人类依然掌握着决策权——你不可能随便就动用上万个TPU去跑一个其实没什么意义的任务。但从产品角度看，这种变化确实非常明显。

I think the signs that are still early is doing this from a model perspective, and I think part of the context of that is the resource allocation for some of these larger training runs is just significant, and so you definitely still have a human in the driver's seat of making those decisions because you're not gonna accidentally take 10,000 TPUs to go kick off some job that actually doesn't make that much sense, But from a product perspective, you for sure see it.

我们团队已经能明显感受到这一点了。

I think we're seeing this on our team.

我们用Antigravity开发了移动应用，上线速度比谷歌内任何团队过去开发移动应用都要快得多。

We've built mobile apps using Antigravity, and we'll launch them to the world faster than I think any team at Google has ever built a mobile app.

Josh的团队用同样的方式开发了Gemini的macOS应用，端到端地快速交付了一个应用，速度之快是谷歌历史上前所未有的，这完全得益于智能编码系统，所以从产品角度来说，效果太棒了。

Josh's team did this with the Gemini macOS app and sort of like end to end delivered an app sort of faster than any team had ever delivered a Mac app at Google, and it's because of agentic coding, and so it's great from a product perspective.

你以前说过，如果真有一个系统能用代码构建任何东西，人类在这一点上就完全没法跟它竞争了，这其实就是狭义的超级智能。

I think 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.

你觉得我们现在已经达到这个阶段了吗？

Do you think we've reached that point?

这确实挺有意思的。

-It is interesting.

我觉得这个狭义超智能的例子挺值得一看的。

I think this narrow superintelligence example is interesting to see.

现在写代码这块确实有这种感觉，代码生成得太厉害了，简直就像狭义超智能。

Obviously, it kind of feels that way for coding right now, where coding is just so good that it does kind of feel like narrow superintelligence.

我也不确定。

I don't know.

这得看你怎么具体衡量，但关键是，就像你之前说的，它在编程上真的特别出色。

It depends how you actually end up the details of quantifying this, but I think the important thing is, to your point earlier, it works incredibly well for code.

如果它还能干别的事当然更好，但光是能在编程上这么厉害，就已经足够有影响力了。

And so it would be great if it did a bunch of other things, but it's actually just so impactful that it can be great at code.

所以我经常花时间去慢慢消化这个事实——虽然造通用人工智能很重要，也很吸引人，但如果因为太关注这个长远目标，反而忽略了眼下技术已经能做到的事，那可能得不偿失。所以我一直在心里同时提醒自己两件事：我们确实需要发展通用技术，但同时，眼下这项能力带来的影响也实在太巨大了。

And so I spend a lot of time just letting that fact sort of just wash over me because I think it's Obviously, building AGI is super important and very interesting, but building AGI, if it sort of takes away from the story of the current present capability of the technology, I think is actually kind of a bad sort of trade off, and so I'm trying to always hold these two things in my head equal at the same time, which is we need to build general purpose technology, but obviously, it's so impactful to have this thing.

而且我觉得它并没有削弱人类开发者，反而成了一个特别积极的结果。

And it feels like it hasn't taken away sort of It's been one of the best positive outcomes is that I feel like it hasn't taken away from human developers.

它真的像一个加速器，让我这个程序员感觉自己在世界上更有掌控力了。

It really does feel like an accelerant of what human I, as a human developer, feel like I have more agency in the world.

我觉得自己现在能应对各种挑战了，这是我的亲身感受。

Feel like I can tackle This is my personal experience.

我觉得自己能接手更雄心勃勃的项目了。

I feel like I can tackle more ambitious problems.

以前我总在脑子里转悠一些想法，觉得它们有点遥不可及，就只好叹口气：要是能做成就好了。

I feel like I used to kick around ideas, they were slightly out of reach, and I would just be like, Ah, wouldn't it be nice?

现在反过来，我脑子里蹦出一个点子，第一反应居然是：这事儿还能再干得更大一点。

And now I have the opposite problem, which is I'm kicking around an idea, and I'm like, I could probably make this even more ambitious.

这带来了一种新的责任，或者说新的压力——我不能再只做个最小可行版本了。

It adds a different layer of sort of responsibility or a different layer of burden, actually, because I'm like, Oh, I can't just do the sort of MVP of this.

既然技术能支持，我就得往前多走十步，我觉得这种野心的重新定位，我自己也想了很多。

I actually need to go 10 steps further because the technology enables me, and resetting my level of ambition, I think, is something that I've also spent a bunch of time thinking about.

我觉得在这些垂直领域的超级智能上，这种情况会大量出现，挺有意思的，而且在通用人工智能真正成型之前，我们可能先迎来一堆这样的技术，说白了，我们最后得到的可能是种参差不齐的超级智能。

But I think that will happen in these vertical superintelligence domains, which will be interesting, and it feels like we're gonna get a bunch of those before we've solved like, it's almost like jagged superintelligence, I think, is what we'll end up with.

你觉得接下来会在哪些领域率先出现超级智能？

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

这是个好问题。

That's a great question.

最近我花太多时间在编程上，得想想别的领域。

I do spend a lot of my time too much time probably thinking about coding these days, so I'll think for a second of, like, the other the other domains.

那些更容易验证的领域，比如金融，显然会更快看到成效。

I think part of this is things that have better verifiability, obviously, are the ones where you'll see the gains happen more quickly, so things with and finance.

科学其实会是个非常有意思的领域。

Actually, science could be a really interesting one.

要是能看到一些有明确验证标准的领域真正爆发出来，那会很酷，这也关系到AI对世界影响的更大图景。

It would be fascinating to see some of these domains where there's some level of verifiability actually really start to take off, which would be cool, and I also think an important thing in this broader narrative about just what impact AI is having on the world.

你其实希望事情就是这样。

You almost want that to be the case.

在技术落地的顺序上，你希望那些真正有益、积极的影响能尽早发生，好让大家看到这项技术的正面潜力。

In the sequencing of things that work, you want a lot of these really, really good, impactful, positive things for the world to happen as early on as humanly possible so that folks understand what the potential positive impact of the technology is.

所以我觉得科学会是个非常有意思的领域。

So I think science could be a really interesting one.

现在数学证明这一块儿发生了好多事，但我不是数学家，有些东西实在有点超纲了，不过……

Obviously, there's all this stuff happening right now with math proofs and stuff like that, which I'm not a mathematician, so it's somewhat over my head, but

前几天我看到一条特别有意思的推文。

I saw a great tweet the other day.

为什么AirDose会出这么多问题？

Why did AirDose have so many problems?

没错。

Exactly.

这问题提得好。

That's a good one.

我喜欢这个说法。

I like that.

这简直可以当T恤标语了。

That's a good like T shirt.

太搞笑了。

So funny.

好吧。

Okay.

不过说到推特，我之前翻过你的推特，再给你念一条。

I but sweet speaking of Twitter, I went through your Twitter before this, so I'm gonna read back another tweet at you.

推特的好处是，你所有的预测都留着公开记录。

The the good thing about Twitter is there's a public record of all your predictions.

那你得开个自动删推的功能才行。

So need to turn on that auto tweet deleting feature or whatever it is.

去年十月，你发过一条推，说到了2025年底，人人都能用 vibe 编程做游戏。

Last October, you tweeted, everyone is going to be able to vibe code video games by the end of twenty twenty five.

对。

Yeah.

这事儿最后成真了吗？

Did that end up being true?

感觉快了。

It feels close.

我觉得肯定不是那种大型3A游戏。

And I think there's I mean, obviously not AAA games.

你还没法直接做出下一个《使命召唤》或者《GTA》。

Like, you're not building, you know, the next Call of Duty or GTA yet.

但感觉比以往任何时候都更接近了。

But I think it feels closer than it's ever been.

其实很多有趣的地方在于，做游戏还得搭很多其他东西。

Think a lot actually, a lot of the interesting bit about video games is you actually need to end up building a lot of this other stuff.

比如模型，我们之前私下聊过，Three.js就是个好例子。

Models, and we were talking off camera before this, Three.

Three.js让很多以前做不到的事变得可能了，

Js is a great example of this.

但还有很多粗糙的地方，光靠代码代理解决不了。

Three.

你需要生成精灵图，而现在的模型本身并不擅长这个，所以还得靠额外的协调工具和系统来实现。

Js makes a lot of things possible that weren't before, but there's 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, and so you need some orchestration layer and tooling in order to make that happen.

还有很多其他东西，同样是游戏体验的核心，需要极高的稳定性，我觉得离实现已经不远了，但真要做出可复用、可重玩、有深度的作品，还得花大量精力做产品基建，而且还得有点品位。

There's a bunch of other things like that that are core to the gaming video game experience that need to have a high degree of reliability that I think it feels like it's within reach, but actually requires a lot of product scaffolding work in order to create experiences that are reusable and replayable and sort of have the level of depth and requires a little bit of taste in there.

你有没有看到很多人在AI工作室和其他开发者平台上做游戏？

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

有的。

Yeah.

这其实是基于我们当时看早期数据发现的，那时候在AI工作室里，用户做的应用里差不多有20%都是游戏。

And so this was actually based on us looking at the early data, and there was something like, in AI Studio at the time, it was like 20% of all apps that folks were making were actually games.

很多人都在尝试做游戏。

People were trying to build games.

数量不少。

Lot them

这是最受欢迎的类别吗？

Is that the most popular category?

现在不是最热门的类别了，因为生态和用户群体都变了，但做游戏的人依然很多。

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.

现在最热门的类别是什么？

What is the most popular category?

我觉得大概20%是和金融相关的，还有20%

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

哇，原来这么多人喜欢算钱啊。

Wow, people like counting their money that much.

我觉得主要是加密货币，很多人拿它来做金融类应用，还有不少个人效率工具，以及生成式媒体类的东西，毕竟谷歌的生成式媒体工具做得确实不错。

I think it's something around crypto, actually, I think is what people are doing a lot of stuff with finance, a lot of personal productivity things and a lot of Genmedia stuff, actually, because obviously the Google suite of Genmedia stuff has done a great job.

不过我也觉得，GDM团队对游戏特别重视，德米斯当初就是从游戏入手做AI的，所以我觉得这方面会有不少突破。

But I also think GDM has sort of like a Obviously, Demis cares a ton about games and sort of started his career in doing AI stuff because of games, and so I think we'll have some interesting swings at this.

我们在Kaggle的团队，也就是负责GDM各项AI基准测试的那部分，其实和GDM一起打造了一个游戏竞技场，用游戏作为衡量通用人工智能进展的替代指标——这其实深深植根于GDM的历史。

Our team actually in Kaggle, which is sort of a bunch of the AI benchmarking stuff we do in GDM, sort of works with GDM to build this game arena, which is sort of our way of testing progress towards AGI, using games as a proxy, which, again, is very deeply rooted in GDM's history.

你觉得普通人随便有个好点子，就能轻松写出一个好玩又可玩的游戏了吗？

How close do you think we are to Rando off the street, what a good idea, can vibe code a really fun, playable game?

我觉得今年就能做到。模型的能力已经让这件事变得可行了，这也是我最近在产品层面特别兴奋的地方。之前我们私下也聊过，这个生态里的初创公司，感觉真的有戏。

I wanna say this year Actually, I the model capability makes it I think this is where I've gotten excited on the product side, and again, we were also talking off camera about sort of the startups in this ecosystem because it feels like it's possible.

模型本身的质量其实没什么缺口。

It doesn't feel like there's a gap in model quality.

真正缺的是有人懂怎么做出一款好游戏，并且能把整个框架搭对，让这事真正可行。

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.

现在确实有人在这么做，一部分问题是大家根本不知道这事儿能做，另一部分是某些模型能力还差那么一点，再过几周或几个月，等那道坎跨过去，大多数人就能用上了。

I think there are folks who are doing this right now, and so some of it is a discoverability and awareness thing that people just don't even know that they can do that, And some of it is just maybe certain categories of model capabilities are just slightly off and we're weeks or months away from that chasm being crossed and then it just working for most people.

这正好引出我想问的下一个问题——关于世界模型。你觉得靠‘ vibe 编程’做出视频游戏，更可能是靠游戏引擎加编程代理，还是靠世界模型？

And so this is a good segue into, I wanna ask about world models next, but do you think vibe coded video games is more likely going to be game engine plus coding agents based, or do you think it's more likely to be world model based?

对。

Yeah.

我觉得最终世界模型的定义会变得模糊，这点我们真该跟 Omni 一起好好聊聊。

I think the what will end up happening is the definition of world models will blur, which we should which we should talk about with Omni.

编程代理看起来还是会像某种世界模型系统，但要想让世界模型真正派上用场，光有模型还不够，还得有支撑结构。

And it will still I think the coding agent 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.

所以我觉得，现在有不少初创公司在研究，到底该怎么搭建世界模型的支撑体系，把它们从那种开放、模糊的设计，转化成具体、可重复使用的实际应用。

And so I think, again, there's actually a bunch of interesting startups doing work, 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 reoccurring way.

也许有人能搞定世界模型的框架，让游戏成为可能，但就目前来看，世界模型本身的特性其实不太适合做游戏。

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, make it so that it's actually not well suited for games in the current form.

不过进展真是快得惊人，谁知道呢？

But the progress has been crazy, so who knows?

也许两年后版本就能行了，但至少在短期内，我觉得还是编码代理加上某种游戏引擎，能更快出成果。

Maybe in two years, the versions will be able to, but at least in the short term, it's like coding agent plus some sort of game engine, I think, is where you'll see way more alpha from a game's perspective.

这说得通。

That makes sense.

你刚才说世界模型的定义很模糊，能详细说说吗？

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

对，能展开讲讲。

Can we unpack that?

是的。

Yeah.

我觉得 Omni 就是个很好的例子。

I think Omni is an example of this.

我们在IO大会上发布了这个产品。

We launched this at IO.

你可以输入任何内容，生成任何输出，戴米斯当时向外界介绍它时，把它称为世界模型，这很恰当，因为它对世界的理解程度确实很高。

You can sort of take in 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.

从技术架构上看，它和我们过去的做法不太一样，虽然我不是架构专家，但确实有区别，我觉得这是好事，因为它更接近某些可能更具扩展性的实现方式。

I think that technically looks different than And I'm not an architecture expert on the way that we've done world models before, 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.

我明白了。

I see.

过去这种模型根本没法大规模应用。

Historically, it's been super not scalable.

运行传统的在线世界模型成本高得吓人。

It's very, very expensive to run traditional online world models.

是的，比如Gemini就是典型的例子。如果你把传统世界模型理解为一种基于动作条件的视频模型，那现在我们说的‘世界模型’，其实指的是具备某种世界理解能力的模型，而不只是严格意义上的动作条件视频模型。

Yeah, like Gemini being traditional 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 a action conditioned video model.

对，有意思的地方在于，它确实有对世界的理解，同时又能完成很多类似的应用场景，这正是我觉得界限模糊的地方。

Yeah, 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's like it can do a lot of those same use cases.

现在还不是实时的，但它能完成很多你原本用那种世界模型能实现的用途，我觉得这才是最让人感兴趣的地方。

It's not real time right now, but it can do a lot of those same use cases that you would describe or visually could create with that same exact world model, which I think is what's most interesting to me.

我觉得这个世界模型和视频模型的结合，最终会以一种和以前预期不同的方式发展。

So I do feel like this world model, video model thing is gonna change and play out in a different way than was obvious before.

那它的底层原理是怎样的？能分享一下吗？

And how does it work under the hood, like, whatever you're able to share?

是 Gemini 加上视频模型的组合吗？

Is it Gemini plus video models?

还是完全不同的东西？

Is it something different entirely?

它是一个单一模型，我觉得这才是关键。

It is a single model, which I think is the important part.

当初我们最初的设想之一就是，过去你得训练八个不同的模型来分别完成这些任务。

Was actually part of the original desire was, like, you were training, like, eight different models to do all of those things historically.

比如你有一个基于 Gemini 基础的文本模型。

It's like you have a text model with the baseline Gemini model.