从大语言模型到通用人工智能：谷歌 DeepMind CEO 谈进展、局限与突破

谷歌 DeepMind CEO Demis Hassabis 加入我们，讨论从当前到 AGI 的路径、谷歌 AI 眼镜何时推出，以及 AI 进步的速度能否保持。欢迎收听来自达沃斯的《大科技播客》特别版。我是 Alex Kantrowitz，今天的特邀嘉宾是谷歌 DeepMind CEO Demis Hassabis。Demis，欢迎再次来到节目。

Google DeepMind CEO Demis Hassabis joins us to talk about the path from here to AGI. When Google's AI glasses are coming, and whether the pace of AI progress can keep up at this rate. That's coming up right after this. Welcome to a special edition of Big Technology Podcast from Davos. I'm Alex Kantrowitz and I'm joined today by a special guest, Demis Hassabis, the CEO of Google DeepMind. Demis, welcome back to the show.

很高兴来到这里。

It's great to be here.

一年前，人们确实在质疑 AI 进步是否在放缓。当时流行问 LLM 是否会撞墙，这些问题似乎已经尘埃落定。过去一年取得了巨大进展。你能具体说说是什么让 AI 行业从去年的质疑时刻走到了今天吗？

A year ago, there were real questions about whether AI progress was tailing off. It was in fashion to ask whether LLMs were going to hit a wall, and those questions seem like they've been settled. There's been a tremendous amount of progress over the past year. Could you tell us what specifically has happened that's gotten the AI industry from that moment of question last year to the point that it is today?

对我们内部来说，我们从未质疑过这一点。明确地说，我们一直看到巨大的进步。所以我们有点困惑为什么会有这种疑问。部分原因是人们担心数据耗尽。确实，所有数据都已被使用，我们能否创建有用的合成数据来学习？但实际上，我们可以从现有架构和数据中榨取更多价值。所以我认为还有很大空间，我们在预训练、后训练、思维范式以及它们如何结合方面都看到了这一点。因此，仅凭我们已经知道的技术以及在此基础上进行调整和创新，仍有很大的发展空间。

Well, for us internally, we were never questioning that. Just to be clear, I think we've always been seeing great improvements. So we were a bit puzzled by why there was this question in the air. I mean some of it was to do with people being worried about data running out. And there is some truth in that all the data had been used. Can we create synthetic data that's going to be useful to learn from? But actually it turns out you can ring more juice out of the existing architectures and data. So there was plenty of room, I think, and we're still seeing that in both the pre-training, the post-training, and the thinking paradigms and also the way that they all kind of fit together. So I think there's still plenty of headroom there just with the techniques we already know about and tweaking and innovating on top of that.

好吧，怀疑论者会这么说：LLM 之上加了很多技巧。我经常听到脚手架、编排，以及 AI 能用工具搜索网络，但它不会记住学到的东西。一旦关闭会话，它就忘了。这只是大语言模型范式的局限吗？

All right, here's what a skeptic would say. That there have been a lot of tricks that have been put on top of LLMs. I hear often about scaffolding and orchestration and AI that can use a tool to search the web, but it won't remember what it learns. As soon as you close that session, it forgets. Is that just a limitation of the large language model paradigm?

嗯，我认为确实如此，我绝对认同我们可能需要一两个重大突破才能达到 AGI。我认为这些突破涉及持续学习、更好的记忆、更长的上下文窗口，或者更高效地使用上下文窗口——不存储所有内容，只存储重要的东西，这样会更高效，大脑就是这样做的。还有更好的长期推理和规划。目前还不清楚仅靠扩展现有想法和技术是否足够，还是我们需要一两个真正深刻的创新。如果非要选，我可能属于后者。但无论你属于哪一派，我认为大型基础模型都是最终 AGI 系统的关键组成部分，这一点我确信。所以我不认同 Yann LeCun 等人认为它们只是死胡同的观点。我唯一的争论是它们是关键组成部分还是唯一组成部分。我认为就在这两个选项之间。对我们来说，优势在于我们有深厚而丰富的研究基础，可以全力推进这两方面：既扩展当前的范式（扩展也涉及创新，尤其是预训练，我们在这方面很强），又探索全新的架构和想法，就像过去 10 年谷歌和 DeepMind 发明的 Transformer 等。

Well, look, I think there is, and I'm definitely a subscriber to the idea that maybe we need one or two more big breakthroughs before we'll get to AGI. And I think they're along the lines of things like continual learning, better memory, longer context windows or perhaps more efficient context windows would be the right way to say it. So don't store everything, just store the important things. That would be a lot more efficient. That's what the brain does. And better long-term reasoning and planning. Now, it remains to be seen whether just sort of scaling up existing ideas and technologies will be enough to do that. Or we need one or two more really big insightful innovations. I'm probably if you were to push me I would be in the latter camp. But I think no matter what camp you're in, we're going to need large foundation models as the key component of the final AGI systems, of that I'm sure. So I'm not a subscriber to someone like Yann LeCun who thinks that they're just some kind of dead end. I think the only debate in my mind is are they a key component or the only component. So I think it's between those two options and for me, we have the advantage of having such a deep and rich research bench. We can go after both of those things with maximum force both scaling up the current paradigms and ideas—and when I say scaling up that also involves innovation by the way, pre-training especially I think we're very strong on—and then really new blue sky ideas for new architectures and things, the kinds of things we've invented over the last 10 years as Google and DeepMind, of course including transformers.

带有大量硬编码的东西能被视为 AGI 吗？

Can something with a lot of hard-coded stuff ever be considered AGI?

不，我认为这取决于你所说的“大量”是什么意思。我对混合系统非常感兴趣，有时人们称之为神经符号系统。AlphaFold、AlphaGo 就是例子。我们一些最重要的工作将神经网络和深度学习与蒙特卡洛树搜索等结合起来。所以我认为这是可能的，我们正在做一些非常有趣的工作，使用 LLM 结合进化方法（如 AlphaEvolve）去发现新知识。你可能需要超越现有方法的东西，但我认为学习是 AGI 的关键部分，实际上几乎是定义性特征。当我们说通用时，我们指的是通用学习。它能否学习新知识，能否跨任何领域学习？这就是通用部分。所以对我来说，学习一直等同于智能。

No, I think it depends what you mean by a lot. I think I'm very interested in hybrid systems, what I would call them, or neurosymbolic sometimes people call them. You know, AlphaFold, AlphaGo are examples of that. So some of our most important work combines neural networks and deep learning with things like Monte Carlo tree search. So I think that could be possible and there's some very interesting work we're doing using LLMs with things like evolutionary methods, AlphaEvolve, to actually go and discover new knowledge. You may need something beyond what the existing methods do, but I think learning is a critical part of AGI. It's actually almost the defining feature. When we say general, we mean general learning. Can it learn new knowledge and can it learn across any domain? That's the general part. So for me, learning is synonymous with intelligence and always has been.

好的。如果学习等同于智能，而这些模型仍然没有持续学习的能力。没有。就像我之前说的，它有金鱼大脑。它可以搜索互联网，然后说“我搞明白了”，但它不会改变模型。会话结束后就忘了。你对如何解决持续学习问题有理论吗？愿意和我们分享吗？

Okay. So if learning is synonymous with intelligence, and these models still don't have the ability to continually learn. No. Like I said earlier, it has goldfish brain. It can search the internet and it can be like I figured this out, but it doesn't change the model. It just will forget it after the session. Do you have a theory as to how the continual learning problem can be solved and do you want to share it with us all?

我可以给你一些线索。我们正在非常努力地研究这个问题。过去我们在 AlphaZero 等系统上做过一些工作，它从零开始学习，还有 AlphaGo 的版本。AlphaGo Zero 也在已有知识的基础上学习。所以我们在更窄的领域做到了。游戏显然比混乱的现实世界容易得多。所以这些技术能否真正扩展到并泛化到现实世界和实际问题，还有待观察。但至少我们知道的方法可以做一些相当令人印象深刻的事情。所以现在的问题是，至少在我看来，我们能否将其与这些大型基础模型融合。当然，基础模型在训练期间学习，但我们希望它们在现实世界中学习，包括个性化。我认为这将会实现。我觉得这是构建优秀助手的关键部分：它理解你，为你工作，是为你服务的技术。我们上周刚刚发布了第一个版本。个人智能是迈向那一步的最初尝试。但要做到这一点，你需要的不仅仅是把数据放在上下文窗口中。你需要比这更深入的东西，正如你所说，实际上随着时间的推移改变模型。那才是理想的情况。

I can give you some clues. We are working very hard on it. We've done some work on this in the past with things like AlphaZero that learned from scratch, versions of AlphaGo. AlphaGo Zero also learned on top of the knowledge it already had. So we've done it in much narrower domains. Games are obviously a lot easier than the messy real world. So it remains to be seen if those kinds of techniques will really scale and generalize to the real world and actual real world problems. But at least the methods we know can do some pretty impressive things. And so now the question is can we blend that, at least in my mind, with these big foundation models. And so of course the foundation models are learning during training, but we would love them to learn out in the wild, including things like personalization. I think that's going to happen. And I feel like that's a critical part of building a great assistant is that it understands you and it works for you as technology that works for you. And we've released our first versions of that just last week. Personal intelligence is the sort of first baby steps towards that. But I think to have it, you want to do it more than just having your data in the context window. That's you want to have something a bit deeper than that which is, as you say, actually changes the model over time. That's what ideally you would have.

我们几次提到 AGI。年底我和山姆·奥特曼聊过，我问他：你似乎说了两件事。我们还没到 AGI，但每次他谈论 GPT 模型能做什么时，又好像符合他的定义。他说 AGI 定义不明确，他希望大家都同意我们已经飞速越过 AGI，正迈向超级智能。你同意吗？

We've brought up AGI a couple times. I was speaking with Sam Altman towards the end of the year and I asked him: you seem to be saying two things. We're not at AGI yet, but every time he talks about what GPT models can do, it seems like it fits his definition. And he said that AGI is underdefined, and what he wishes everybody could agree to was that we've sort of whooshed by AGI and we move towards superintelligence. Do you agree with that?

我肯定他希望如此，但绝对不行。我认为 AGI 不应该变成营销术语或用于商业利益。它一直有科学定义。我的定义是：一个系统能展现人类所有认知能力——我是说所有。包括最高水平的人类创造力，比如我们钦佩的科学家和艺术家。不只是解数学方程或猜想，而是提出突破性猜想——那难得多。不只是解决物理问题或化学问题如 AlphaFold，而是提出新的物理理论，就像爱因斯坦的广义相对论。系统能做到吗？因为历史上最聪明的人类用他们的大脑架构做到了。在艺术方面，不只是创作已知风格的拼凑品，而是像毕加索或莫扎特那样开创全新的艺术流派。今天的系统远未达到。不管你解决了多少厄尔多斯问题。那很好，但离真正的发明——比如拉马努金会做的事——还差得远。你需要一个系统能在所有领域都做到这些。除此之外，我还要加上物理智能——我们能运动、控制身体到惊人水平，就像达沃斯这里的精英运动员。我们在机器人方面还差得远。所以我认为 AGI 系统必须能做所有这些事，才能真正实现 AI 领域的原始目标。我认为我们离那还有 5 到 10 年。

I'm sure he does wish that, but absolutely not. I don't think AGI should be turned into a marketing term or for commercial gain. There has always been a scientific definition. My definition is a system that can exhibit all the cognitive capabilities humans can—and I mean all. That includes the highest levels of human creativity, like the scientists and artists we admire. Not just solving a math equation or a conjecture, but coming up with a breakthrough conjecture—much harder. Not just solving a physics problem or a chemistry problem like AlphaFold, but actually coming up with a new theory of physics, like Einstein did with general relativity. Can a system do that? Because the smartest humans with their brain architectures have been able to do that in history. On the art side, not just create a pastiche of what's known, but actually be Picasso or Mozart and create a completely new genre of art. Today's systems are nowhere near that. It doesn't matter how many Erdős problems you solve. That's good, but it's far from true invention like what a Ramanujan would have done. You need a system that can potentially do that across all domains. And on top of that, I'd add physical intelligence—we can play sports and control our bodies to amazing levels, like elite sportspeople here in Davos. We're still way off on robotics. So I think an AGI system would have to do all those things to fulfill the original goal of the AI field. I think we're 5 to 10 years away from that.

我认为论点会是：如果某物能做所有这些事，那就会被视为超级智能。但你认为 AGI 是个好术语吗？

I think the argument would be that if something can do all those things, it would be considered superintelligence. But you think AGI is a good term?

不，当然不是。因为个体人类能提出新理论。爱因斯坦做到了，费曼做到了，所有伟人——我的科学英雄——他们都做到了。这很罕见，但用人类大脑架构是可能的。所以超级智能是另一个值得讨论的概念。那将是真正超越人类智能能力的东西。我们不能在 14 维空间思考，也不能把气象卫星接入大脑——至少现在还不能。那些是真正超越人类或超人类的。那是另一场辩论。但一旦我们达到 AGI……

No, of course not. Because individual humans can come up with new theories. Einstein did, Feynman did, all the greats—my scientific heroes—they were able to do that. It's rare, but possible with the human brain architecture. So superintelligence is another concept worth talking about. That would be things that can really go beyond what human intelligence can do. We can't think in 14 dimensions or plug weather satellites into our brains—not yet anyway. Those are truly beyond human or superhuman. That's a whole other debate. But once we get to AGI...

我最近听你说话，有件事让我很惊讶。你在 Google DeepMind 播客上被问到：今天有没有接近 AGI 的系统？我以为会是 Gemini 3。你提到了 Nano Banana？那个图像生成器。什么？

I was listening to you recently and something you said really surprised me. You were asked on the Google DeepMind podcast, which is a great listen, if you have a system today that is close to AGI. I thought it might be Gemini 3. You named Nano Banana? The image generator. What?

嗯，有时候你得取些有趣的名字，玩一玩。但图像生成器怎么接近 AGI？我们来看看图像生成器，还有我们的视频生成器 Veo，它是视频生成领域的顶尖水平。从 AGI 的角度看，我觉得那更有趣。你可以想象一个视频模型，能生成 10 秒、20 秒的真实场景。它有点像物理世界的模型——我们有时称之为直觉物理。它直观地理解液体和物体在世界上如何表现。展示理解的一种方式就是能够生成它，至少在人眼看来足够准确、令人满意。显然，从物理角度看它并不完全准确，我们会改进。但这是迈向世界模型的一步——一个能理解世界及其机制和因果关系的系统。这对 AGI 至关重要，因为它能让这些系统在现实世界中长期规划，跨越很长时间跨度，就像我们人类能做到的。我会花四年拿学位，这样十年后就能有更好的工作。这些都是我们毫不费力就能做的长期计划。目前，没有这些系统，我们还不知道如何做到。我们只能做单一时间尺度上的短期计划。但我认为你需要这类世界模型。在机器人领域，这正是你想要的：机器人在现实世界中规划，能够从当前情况想象出许多轨迹来完成任务。最后，从我们的角度看，这就是为什么我们从一开始就让 Gemini 多模态——能够处理视频、图像，并最终将所有内容融合到一个模型中。那是我们的计划。它也会对通用助手非常有用。

Well, sometimes you have to have fun names and have fun with those. But how is the image generator close to AGI? Let's take image generators, but also our video generator Veo, which is state-of-the-art in video generation. I think that's even more interesting from an AGI perspective. You can think of a video model that can generate 10 seconds, 20 seconds of a realistic scene. It's sort of a model of the physical world—intuitive physics, we'd sometimes call it. It intuitively understands how liquids and objects behave in the world. One way to exhibit understanding is to be able to generate it, at least to the human eye being accurate enough to be satisfying. Obviously, it's not completely accurate from a physics point of view, and we're going to improve that. But it's steps towards having a world model—a system that can understand the world and the mechanics and causality of the world. That would be essential for AGI because it would allow these systems to plan long-term in the real world over very long time horizons, which we as humans can do. I'll spend four years getting a degree so that in 10 years I'll have a better job. These are long-term plans we do effortlessly. At the moment, without these systems, we still don't know how to do that. We can do short-term plans over one time scale. But I think you need these kind of world models. In robotics, that's exactly what you want: robots planning in the real world, being able to imagine many trajectories from their current situation to complete a task. And finally, from our point of view, that's why we've worked with Gemini as being multimodal from the beginning—able to deal with video, image, and eventually converge that all into one model. That's our plan. It'll be very useful for a universal assistant as well.

那我们聊聊产品吧。我和另外三亿人一起看了纪录片《思考的游戏》。里面发生了一件有趣的事。整部纪录片中，你和一些同事不停地用手机对着东西，问助手发生了什么。我像往常一样对着电脑大喊：这家伙需要眼镜——他需要智能眼镜才能做到。手机是错误的外形。你对 AI 眼镜的愿景是什么？什么时候推出？

So let's talk product a little bit. I watched the documentary 'The Thinking Game' along with 300 million other people. There was something kind of interesting that happened there. Throughout the documentary, yourself and some colleagues kept pointing your phone at things and asking an assistant what was going on. And I was yelling at the computer as I usually do and said this guy needs glasses—he needs smart glasses to be able to do it. The phone is the wrong form factor. What is your vision for AI glasses and when is the rollout happening?

是的，我认为你说得完全正确，那也是我们的结论。

Yeah, I think you're exactly right and that was our conclusion.

当你自己内部试用这些产品时，这一点非常明显。就像你在影片中看到的，你举起手机让它告诉你关于现实世界的信息。这很神奇，也确实有效，但对于很多你想做的事情来说，这显然不是合适的形态——比如做饭、在城市里闲逛时问路或寻求推荐，甚至帮助视障人士。那里有巨大的应用场景。为此，你需要一个免提的设备。对我们这些戴眼镜的人来说，显而易见的选择就是把它放在眼镜上。但也可能有其他设备。我不确定眼镜就是最终的形态，但它绝对是下一个清晰的形态。在谷歌和 Alphabet，我们在眼镜方面有悠久的历史。也许过去我们有点太超前了。我的分析是，与参与那个项目的人交流后，发现形态太笨重，电池寿命也有问题——这些问题现在基本解决了。但缺少的是一个杀手级应用。我认为杀手级应用就是一个通用的数字助手，它随时陪伴你，在日常生活中帮助你，可以在任何界面上使用——你的电脑、浏览器、手机，还有当你走在城市里时戴的眼镜等设备。它需要无缝衔接，并理解每种情境。我认为我们现在已经很接近了，尤其是有了 Gemini 3。我们终于有了足够强大的 AI 来实现这一目标。这是我们正在做的最令人兴奋的项目之一。我个人正在致力于让智能眼镜真正发挥作用。我们与 Warby Parker、Gentle Monster 和三星建立了很好的合作伙伴关系，来制造这些下一代眼镜。你应该会在今年夏天左右开始看到它们。

It's very obvious when you dogfood these things internally. As you saw from the film, you hold up your phone to get it to tell you about the real world. It's amazing, it works, but it's clearly not the right form factor for a lot of things you want to do—like cooking, roaming around the city asking for directions or recommendations, or even helping the partially sighted. There's a huge use case there. For that, you need something hands-free. The obvious thing for those of us who wear glasses is to put it on glasses. But there may be other devices too. I'm not sure glasses is the final form factor, but it's definitely a clear next form factor. At Google and Alphabet, we have a long history with glasses. Maybe we were a bit too early in the past. My analysis, after talking to people working on that project, is that the form factor was too chunky and clunky, and battery life was an issue—these are now more or less solved. But the thing it was missing was a killer app. I think the killer app is a universal digital assistant that's with you, helping you in everyday life, available on any surface—your computer, browser, phone, and also on devices like glasses when you're walking around the city. It needs to be seamless and understand each context. I think we're close now, especially with Gemini 3. We finally have AI powerful enough to make that a reality. It's one of the most exciting projects we're working on. I'm personally working on making smart glasses really work. We've done great partnerships with Warby Parker, Gentle Monster, and Samsung to build these next-generation glasses. You should start seeing that maybe by the summer.

是的。Warby Parker 确实有文件显示这些眼镜今年很快就会推出。

Yeah. Warby Parker did have a filing that said these glasses are coming out pretty soon this year.

是的，原型设计……取决于进展速度，但我认为很快就会实现，而且它将成为一种定义新品类的新技术。

Yeah, the prototype design... it depends how quickly that advances, but I think it's going to happen very soon, and it'll be a new category-defining technology.

考虑到你个人的参与，可以说这是一个非常重要的项目吗？

Given your personal involvement, is it safe to say this is a pretty important initiative?

是的，这是最重要的项目之一。我喜欢把时间花在重要的事情上，但我也喜欢推动最前沿的东西——通常也是最难的。我选择阶段性目标，给予信心，并判断时机是否合适。几十年来，我已经相当擅长这一点。所以我尽量待在那些我能发挥最大作用的最前沿领域。比如眼镜、机器人和世界模型，这些是我目前投入时间的地方。

Yes, it's one of the most important. I like to spend my time on important things, but I also like to push the most cutting-edge things—often the hardest. I pick interim goals, give confidence, and understand if the timing is right. Over the decades, I've gotten quite good at that. So I try to be at the most cutting-edge parts where I can make the most difference. Things like glasses, robotics, and world models are where I'm spending time.

好的，所以眼镜的时机成熟了。我们来谈谈广告。广告的时机成熟了吗？让我铺垫一下。有消息说 Gemini 可能会包含广告，一些竞争对手也可能包含广告。我在社交媒体上看到最好笑的是有人说：‘这些人离 AGI 还差得远呢。如果商业模式是广告，那它就不会是颠覆世界的技术。’你同意吗？

Right, so timing is right for glasses. Let's talk about ads. Is the timing right for ads? Let me set it up. There's been news that Gemini might include ads, and some competitors might include ads. The funniest thing I saw on social media was someone saying, 'These people are nowhere close to AGI. It's not going to be this world-disrupting technology if the business model is advertising.' Do you agree?

嗯，这很有意思。我认为行动胜于言语。回到我们之前与 Sam 等人的对话，他们声称 AGI 即将到来——那为什么还要费心做广告呢？这是一个合理的问题。从我们的角度来看，我们目前没有计划在 Gemini 应用中做广告。我们会非常仔细地观察 ChatGPT 声称要做什么的结果。我认为必须非常谨慎地处理，因为我看到的矛盾是：如果你想要一个为你服务的助手，最重要的是信任——信任、安全和隐私。你可能会与那个助手分享你的生活，你需要确信它是在为你工作，并以你的最佳利益为重。所以你必须小心，不要让广告模式渗透进来，混淆用户对助手推荐内容的判断。这将是一个有趣的挑战。

Well, it's interesting. I think actions speak louder than words. Going back to the original conversation with Sam and others claiming AGI is around the corner—why would you bother with ads then? That's a reasonable question. From our point of view, we have no plans at the moment to do ads in the Gemini app specifically. We're going to watch very carefully what the outcome of what ChatGPT is saying they're going to do. I think it has to be handled very carefully because the dichotomy I see is that if you want an assistant that works for you, the most important thing is trust—trust, security, and privacy. You want to share potentially your life with that assistant, and you want to be confident it's working on your behalf with your best interests. So you have to be careful that the advertising model doesn't bleed into that and confuse the user about what the assistant is recommending. That's going to be an interesting challenge.

这就是不该做的。Sundar 在最近的财报电话会议上说，谷歌内部有一些关于如何正确应对的想法。你如何看待广告？

And that's what not to do. Sundar in a recent earnings call said there are some ideas within Google of the right way to approach this. How do you approach advertising?

嗯，我们还在集思广益。但当你考虑眼镜设备时，也有一些非常有趣的方式——还有其他收入模式存在。所以拭目以待会很有趣。我认为我们还没有得出任何强有力的结论，但这是一个需要非常仔细思考的领域。

Well, we're still brainstorming that. But there are also very interesting ways when you think about glasses devices—there are other revenue models out there. So it's going to be interesting to see. I don't think we've made any strong conclusions on that, but it's an area that needs very careful thought.

只是为了从你这里得到一个明确的答案，我想你已经给出了，但我再问一次。在我们见面之前，我读到谷歌最近几天（从去年开始）告诉广告商，计划在 2026 年将广告引入其 AI 聊天机器人 Gemini。

Just to get a definitive answer from you, I think you've given it, but I'm going to do it one more time. I read before we met that Google has told advertisers in recent days, from last year, that it plans to bring ads to its AI chatbot, Gemini, in 2026.

不，我们目前没有计划。我只能说这么多。

Nope, we have no current plans. That's all I can say.

这相当明确。好吧，我们继续聊聊你的一些竞争对手。Anthropic、Claude Code 和 Claude Co-work 引起了巨大的轰动。看到一些人做到的事情真是令人惊叹。我看到一位前亚马逊高管的帖子，他说他在一个周末——实际上是一天半——建了一个自定义 CRM。我们就说一个周末吧。

That's pretty definitive. All right, let's keep going through some of your competitors. Anthropic, Claude Code, and Claude Co-work have caused a tremendous amount of buzz. It is amazing to see what some people have done. I saw a post from an ex-Amazon executive who said he built a custom CRM in a weekend—or actually a day and a half. Let's call it a weekend.

你怎么看？你打算回应吗？

What do you think about it? And do you plan to have an answer to it?

这非常令人兴奋。向 Anthropic 致敬，他们用 Claude Code 打造了一个非常好的模型。我们对 Gemini 3 当前的编码能力很满意，它在某些方面非常出色，比如前端工作。我在圣诞节期间用它来制作游戏原型，太棒了，让我重新开始编程。我喜欢整个“氛围编码”浪潮。我认为它将为设计师、创意人员、艺术家打开生产力的大门，这些人以前可能需要团队程序员，现在他们自己就能做更多事情。一旦它更广泛地普及，将会创造大量新的创意机会，这太棒了。我们对代码方面的工作很满意，但还有更多要做。我们刚刚发布了我们自己的 IDE——anti-gravity，非常受欢迎，我们甚至无法满足所有需求。我们正在大力推动 Gemini 的编码和工具使用性能。但 Anthropic 完全专注于编码和语言模型，不涉及图像模型、多模态模型或世界模型。他们在这方面非常出色。我们一方面很高兴与之合作，另一方面这也激励我们改进自己的模型。

It's very exciting. Kudos to Anthropic. They built a very good model with Claude Code. We're very happy with Gemini 3's current coding capabilities. It's very good at certain things like front-end work. I've been using it over Christmas to prototype games. It's amazing, getting me back into programming. I love the whole vibe coding wave. I think it will open up productivity for designers, creatives, artists who might have needed teams of programmers. Now they can do a lot more on their own. That will be amazing once it's more widely available, creating new creative opportunities. We're very happy with our work on code, but we have more to do. We just released our own IDE, anti-gravity, which is very popular. We can't serve all the demand. We're pushing hard on coding and tool use performance of Gemini. But Anthropic has fully focused on coding and language models, not image models, multimodal models, or world models. They're very good at that. We're pleased to partner on one hand, and it gives us something to push for to improve our own models.

我们来广泛谈谈 AI 行业业务。我有一个关于这一切如何崩溃的理论。这是一个三步过程。第一，大语言模型训练运行产生有限回报。第二，像 Gemini Flash 这样的闪存模型让 AI 计算变得像搜索一样便宜。第三，鉴于这两个因素，大规模基础设施投入变得有些无用，导致连锁崩溃。这是一个合理的担忧吗？

Let's talk broadly about the AI industry business. I have a theory for how this could all fall apart. It's a three-step process. First, large language model training runs produce limited returns. Second, flash models like Gemini Flash make AI computing as cheap as search. Third, the massive infrastructure commitments become somewhat useless given those two factors, leading to a cascading collapse. Is that a legitimate worry?

我认为这是一个可能的情景，但在我看来不是最可能的那种。毫无疑问，AI 已经足够证明了，尤其是在科学、AlphaFold 和药物发现方面。它已经站稳脚跟。我们早已过了怀疑 AI 是否有效的阶段。它将成为人类历史上最具变革性的技术。问题在于时间线：2 年还是 5 年？无论哪种，对于如此变革性的技术来说都非常快。我们仍处于弄清楚如何使用和部署它的时代，因为技术发展太快了。存在巨大的能力过剩——即使是今天的模型也能做我们不完全了解的事情。产品机会巨大。在 Google，我们才刚刚开始将这些技术融入现有产品，更不用说构建像 AI 收件箱这样的新产品了。谁想做邮件管理？那是我工作中的头号痛点。还有很多这样的例子。浏览器中的智能体、帮助 YouTube、驱动搜索。所以机会巨大。至于 AI 泡沫，我的观点不是二元的。AI 行业的部分领域可能确实存在泡沫，比如那些没有产品或研究却获得数百亿种子轮的公司，这似乎不可持续。但像我们这样的企业拥有庞大的基础业务，AI 显然能提高效率。新的 AI 原生产品如聊天机器人和眼镜的货币化效果如何，还有待观察。我认为会有巨大的市场，但尚未得到验证。我在 Google DeepMind 的工作是确保无论 AI 泡沫如何，我们都能赢。Alphabet 在两种情况下都处于非常有利的位置：要么加倍押注现有业务，要么处于前沿。

I think it's a plausible possible scenario, but not the likely one in my opinion. There's no doubt AI has already proven out enough, especially in science, AlphaFold, and drug discovery. It's here to stay. We've blasted past the point of wondering if AI works. It will be the most transformative technology in human history. The question is timelines: 2 years or 5 years? Either way, it's very soon for something this transformative. We're still in the era of figuring out how to use and deploy it because the technology improves so fast. There's a huge capability overhang—even today's models can do things we don't fully know. There are vast product opportunities. At Google, we've only scratched the surface of plugging these into our existing products, let alone building new ones like AI inbox. Who wants to do email admin? That's my number one pain point. There are many such examples. Agents in browsers, helping with YouTube, powering search. So enormous opportunities. As for the AI bubble, my view is it's not binary. Parts of the AI industry probably are in a bubble, like seed rounds of tens of billions for companies with no product or research. That seems unsustainable. But businesses like us have massive underlying businesses where AI obviously increases efficiency. It remains to be seen how popular monetization of new AI-native products like chatbots and glasses will be. I think there will be enormous markets, but they're yet to be proven. My job at Google DeepMind is to ensure that whatever happens with an AI bubble, we win either way. Alphabet is incredibly well positioned in either case: doubling down on existing businesses or being at the frontier.

回到思考游戏，谈到这将如何影响经济，我开始为你的技术的对手感到难过。李·多尔。士气低落。那个玩星际争霸的 Mana 击败了你的机器人，但意识到人类对机器基本上已经结束了。现在随着这些东西进入知识工作，我们都面临这个问题。我们将面临同样的情况：这些模型曾经出色地对抗世界顶尖的星际争霸和围棋选手，现在开始做我们的工作。我们最终会处于同样的位置吗？

Going back to thinking game, speaking of the way this will impact the economy, I started to feel bad for the opponents of your technology. Li Doll. Demoralized. This guy Mana who played Starcraft beat your bot but realized it's basically over for humans versus machines. Now we're all up against this as this stuff makes its way into knowledge work. We're going to have the same situation with knowledge work that these models that performed admirably against the world's best Starcraft and Go players are now starting to do our work. Are we going to end up in the same position?

我们来看看游戏中发生了什么。自 90 年代以来，国际象棋计算机就比人类更强，深蓝击败了卡斯帕罗夫。但国际象棋比以往任何时候都更受欢迎。没人对看计算机对弈感兴趣。我们感兴趣的是马格努斯·卡尔森与其他顶尖棋手对弈。所以我认为知识工作也会发生同样的情况。AI 将增强人类，而不是取代他们。人们会使用 AI 作为工具来提升自己的能力，就像棋手用计算机分析棋局一样。人的因素仍然是核心。

Let's look at what happened in games. Chess computers have been better than humans since the '90s, with Deep Blue beating Kasparov. But chess is more popular than ever. No one is interested in watching computers play computers. We're interested in Magnus Carlsen playing other top players. So I think the same will happen in knowledge work. AI will augment humans, not replace them. People will use AI as tools to enhance their own abilities, just like chess players use computers to analyze games. The human element remains central.

有趣的是，在围棋领域，世界最强的棋手是一位韩国人。AlphaGo 比赛时他大约 15 岁，现在二十多岁，按 ELO 评分他是史上最强的棋手，因为他从小学习，吸收了 AlphaGo 的知识。他是第一代在知识库中融入 AlphaGo 知识的棋手，可能比当年的 AlphaGo 还要强。我们仍然喜欢玩星际争霸和其他电脑游戏，享受人类的努力。这就像我们仍然热爱百米奥运比赛，尽管有比博尔特快得多的交通工具，但那是另一回事。我认为我们有无限的适应能力，能与技术共同进化。为什么？因为我们是通用智能体。我们是 AGI 系统——当然不是人工的，而是通用的系统。我们有能力发明科学，是制造工具的动物。这使我们与其他动物不同：我们制造工具，包括计算机，而 AI 是计算机的终极表达。这一切都源于我们的人类思维，它原本是为狩猎采集生活方式而进化的。我们能达到现代文明并讨论 AI 和物理学，这显示了我们的通用性。我认为我们会再次适应，但除了经济和就业，还有一个重要问题：目的和意义。我们从工作中获得很多目的感——我本人从科学中获得。当很多工作被自动化后会发生什么？这就是为什么我一直呼吁需要新的伟大哲学家。这将是人类境况的改变，但不一定更糟。就像工业革命，可能强十倍，但我们会适应。我们会找到新的意义。我们现在已经做很多不是为了经济利益的事情：艺术、极限运动、极地探险。未来可能会有更复杂、更精妙的版本。

Interestingly in Go, the best player in the world is a South Korean. He was about 15 when the AlphaGo match happened. He's in his mid-20s now, and he's by far the strongest player ever by ELO ratings because he learned natively, young enough. He's the first generation that learned with AlphaGo knowledge in the knowledge pool, and he may actually be stronger than AlphaGo was back then. I think we all still enjoy StarCraft and other computer games. We enjoy human endeavor. It's similar to how we still love the 100-meter Olympic race even though we have vehicles that can go way faster than Usain Bolt. That's a different thing. I think we have infinite capacity to adapt and evolve with our technologies. Why? Because we are general intelligences. We are AGI systems—obviously not artificial, but general systems. We are capable of inventing science and are tool-making animals. That separates us from other animals: we make tools, including computers, and AI is the ultimate expression of computers. All that comes from our human minds, which evolved for a hunter-gatherer lifestyle. It's amazing how general we are to reach modern civilization and discuss things like AI and physics. I think we'll adapt again, but there is an important question beyond economics and jobs: purpose and meaning. We get a lot of purpose from our jobs—I certainly do from science. What happens when much of that is automated? That's why I've been calling for new great philosophers. It will be a change to the human condition, but not necessarily worse. It's like the industrial revolution, maybe 10 times that, but we'll adapt. We'll find new meaning. We already do things not for economic gain: art, extreme sports, polar exploration. Maybe we'll have more sophisticated versions of those in the future.

还剩两分钟。我有两个问题。先问我最想知道的。在最近一次采访中，你说你有一个理论，认为信息是宇宙最基本的单位。不是能量，不是物质，而是信息。为什么？

Two minutes left. I have two questions. Let me ask the one I want to know most. In a recent interview, you said you have a theory that information is the most fundamental unit of the universe. Not energy, not matter, but information. How?

如果你看能量和物质，很多人认为它们与信息同构，但我认为信息是理解宇宙的正确方式。在生物学中，生命系统是抵抗熵的信息系统，试图在随机性中保持结构和信息。在更大的物理尺度上也能看到这一点。不仅是生物学，山脉、行星、小行星都受到选择压力——不是达尔文进化，而是某种外部压力。它们长期稳定意味着信息是稳定且有意义的。我们可以从信息复杂度的角度看待世界。我想到这些是因为 AlphaGo，尤其是 AlphaFold，我们解决了所有已知的蛋白质结构。怎么做到的？在几乎无限的蛋白质结构中，只有一定数量的结构是稳定的。你必须理解那个信息拓扑并跟随它。然后那些看似棘手的问题——比如大海捞针——就变得容易处理了，如果你理解能量或信息景观。我认为最终我们将通过 AI 帮助我们导航信息景观，从而解决大多数疾病、开发新药、新材料和超导体。

Well, if you look at energy and matter, many people think they are isomorphic with information, but I think information is the right way to understand the universe. In biology, living systems are information systems resisting entropy, trying to retain structure and information against randomness. You can see this at a larger physics scale too. Not just biology, but mountains, planets, asteroids have been subject to selection pressure—not Darwinian evolution, but some external pressure. Their stability over time means that information is stable and meaningful. One could view the world in terms of its information complexity. I think about this because of AlphaGo and especially AlphaFold, where we solved all known protein structures. How? Only a certain number of the almost infinite possibilities of protein structures are stable. You have to understand that information topology and follow it. Then problems that seem intractable—like finding a needle in a haystack—become tractable if you understand the energy or information landscape. That's how I think we'll eventually solve most diseases, develop new drugs, materials, and superconductors with AI helping us navigate that information landscape.

在我们结束前，我想以此收尾。在思考游戏中，谈到健康和 AI，实验室里曾讨论是否发布 AlphaFold 的结果，你坚定地说：‘发布它。现在就发布。’谈谈那里的教训。

Before we go, I want to wrap with this. In the thinking game, speaking of health and AI, there's a moment in the lab about whether to release AlphaFold results, and you sat there adamantly saying, 'Release it. Release it now.' Talk about the lesson from there.

我们启动 AlphaFold 是为了攻克蛋白质折叠和结构预测这个 50 年的重大挑战。我们之所以研究它，是因为我们认为它是一个根节点问题。如果我们解决了它并发布到世界，它将对人类健康和生物学产生惊人的影响。但我们的团队，无论多么有才华，也只能独自挖掘其中极小一部分潜力。因此，为了最大化对世界的益处，显然正确的做法是将其发布给庞大的科学界，让他们在此基础上发展。看到全球 300 万研究人员在重要研究中使用它，我感到无比欣慰。

We started AlphaFold to crack a 50-year grand challenge of protein folding and structure prediction. We worked on it because we thought it was a root node problem. If we solved it and put it out in the world, it could have amazing impact on human health and biology. But our team, no matter how talented, could only scratch a tiny amount of that potential on our own. So it was obviously the right thing to do to maximize benefit to the world: put it out to the massive scientific community to build on. It's been incredibly gratifying to see 3 million researchers around the world use it in their important research.

我认为未来，几乎所有从今往后发现的药物都可能在其研发过程中的某个环节用到 AlphaFold，这对我们来说太棒了。而这正是我们所有工作的意义所在。

I think in future, almost every single drug that's discovered from now on will probably have used AlphaFold at some point in that process, which is amazing for us. And really, that's what we do all the work we do for.

我还读到那个时刻——如果我说错了你纠正我——有点像一种隐喻：一个充满激情的小型 AI 部门在大公司里呐喊，‘把这个推出去，砍掉繁文缛节。’

I also read that moment, you tell me if I'm wrong, as something of a metaphor: a small, passionate AI division kind of yelling in a big company, 'Get this out, cut the red tape.'

有可能。但你看，我们从一开始就得到了谷歌的大力支持。我们在 2014 年与谷歌联手，原因就在于谷歌本身是一家科学研究、工程、技术公司，一直如此，并且以此为核心理念。这就是为什么我认为我们在所有事情上都秉持科学方法、科学态度，那种深思熟虑、严谨的态度。所以他们当然会喜欢像 AlphaFold 这样的东西。

Yeah, potentially. But look, we've had amazing support from the beginning from Google. And the reason we joined forces with Google back in 2014 is that Google itself is a scientific research, engineering, technical company, always has been, and has that at its core. That's why I think we have the scientific method, the scientific approach, that thoughtful, rigorous approach in everything we do. So of course they're going to love something like AlphaFold.

好的，最后有个大问题。你们构建了 AlphaGo，用人类知识训练计算机下围棋，一旦它掌握了人类水平的下法，你们就用一个叫 AlphaZero 的程序把它释放出来，然后它开始做你从未想象过的事情，以让你惊讶的方式开辟新路径。最终，也许会有那么一天，LLM 或它们的某种版本以同样的方式掌握人类知识。当你把它释放出来，它可能做出和 AlphaZero 一样的事情，那时会发生什么？

Okay, here's the big question at the end. You built AlphaGo, trained the computer to play Go on human knowledge, and once it mastered human-level playing, you kind of let it loose with a program called AlphaZero, and it started doing things that you could never even imagine, making new circuits in ways that surprised you. Eventually, maybe there will come a time where LLMs or some version of them reach a mastery of human knowledge in the same way. What is going to happen when you then let that loose and it does the same potentially as AlphaZero?

我认为那会非常令人兴奋。对我来说，那将是 AGI 的时刻。届时它将发现一种新的超导体——室温超导体，这在物理定律上是可能的，但我们只是还没找到那根大海捞针；或者发现新的能源，构建最优电池的新方法。我认为所有这些都将成为可能，而且不仅仅是可能，我认为一旦我们拥有一个首先具备人类知识水平的系统，然后会有一些技术——也许它需要帮助发明其中一些技术——但就像 AlphaZero 那样，让它能够超越，进入未知的新领域。

I think it'd be very exciting. I mean, that's what to me would be the AGI moment. It will then discover a new superconductor, room temperature superconductor that's possible in the laws of physics but we just haven't found that needle in the haystack, or a new source of energy, a new way to build optimal batteries. I think all of those things will become possible, and indeed not just possible, I think they will happen once we get to a system that's first of all got human-level knowledge, and then there'll be some techniques, maybe it will have to help invent some of those techniques, but kind of like AlphaZero that will allow it to go beyond into new uncharted territory.

那种把它像天气系统一样接入大脑的想法，它将会……

That idea of it like plugging weather system into its brain, it's going to be on that...

正是如此。

Exactly.

好吧，真是激动人心的时代。Demis，感谢你来做客。

All right, so exciting times. Demis, thanks for coming on the show.

谢谢。

Thank you.

谢谢大家。

Thanks everybody.

非常感谢。

Thank you so much.