本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
欢迎收听《谷歌DeepMind》播客节目。
Welcome to Google DeepMind, the podcast.
我是主持人汉娜·弗莱教授。
With me, your host, professor Hannah Fry.
早在2017年我们首次策划这档播客时，DeepMind还只是个规模较小、专注于人工智能研究的实验室。
Now when we first started thinking about making this podcast back in 2017, DeepMind was a relatively small, focused AI research lab.
他们刚被谷歌收购，得以在伦敦的安全距离内自由开展各种奇特的研究项目。
They'd just been bought by Google, and they've been given the freedom to do their own quirky research projects from the safe distance of London.
如今时过境迁。
Well, how things have changed.
因为自上季节目以来，谷歌已重组整体架构，将人工智能和DeepMind团队置于其战略核心位置。
Because since the last season, Google has reconfigured its entire structure, putting AI and the team at DeepMind at the core of its strategy.
谷歌DeepMind持续追求赋予AI人类级别智能的目标——即人工通用智能（AGI）。
Google DeepMind has continued its quest to endow AI with human level intelligence known as artificial general intelligence or AGI.
他们推出了名为Gemini的强大AI模型系列，以及能处理音频、视频、图像和代码的AI智能体Project Astra。
It's introduced a family of powerful new AI models called Gemini as well as an AI agent called Project Astra that can process audio, video, image, and code.
该实验室在将AI应用于众多科学领域方面取得重大突破，包括能预测人体内所有分子（不仅是蛋白质）结构的AlphaFold第三版。
The lab is also making huge leaps forward in applying AI to a host of scientific domains, including a brand new third version of AlphaFold, which can predict the structures of all of the molecules you'll find in the human body, not just proteins.
2021年他们还分拆成立了Isomorphic Labs公司，专注于新药研发业务。
And in 2021, they spun off a new company, Isomorphic Labs, to get down to the business of discovering new drugs to treat diseases.
谷歌DeepMind还在开发能通过强化学习自主完成任务的强大AI智能体，延续了AlphaGo在围棋比赛中战胜人类的传奇。
Google DeepMind is also working on powerful AI agents that can learn to perform tasks by themselves using reinforcement learning and continuing that legacy of AlphaGo's famous victory over a human in the game of Go.
当然，各位都是这档播客的忠实听众。
Now, of course, you'll all have been following this podcast since the beginning.
大家对这所有变革背后的故事都耳熟能详。
You'll all be familiar with the stories behind all of those changes.
但若您是新听众——欢迎您的加入。
But just in case you are coming to us fresh, welcome.
你可以在Google DeepMind的YouTube频道或任何你获取播客的地方找到我们获奖的前几季节目。
You can find our first award winning previous seasons on Google DeepMind's YouTube channel or wherever you get your podcast.
这些剧集还详细探讨了我们将会反复听到的许多主题，比如强化学习、深度学习、大语言模型等等。
They also those episodes go into detail about a lot of the themes that we're gonna hear come up over and over again from the people here, like reinforcement learning, deep learning, large language models, and so on.
所以不妨去听听看。
So have a listen.
它们真的很棒，即使是我们自己这么说。
They are really good, even if we do say so ourselves.
自从上一季以来，AI获得的新关注意味着现在有更多播客可供你选择。
Now all of the newfound attention on AI since the last series does mean that there are quite a few more podcasts out there for you to choose from.
但在这个播客中，和以往一样，我们想为你提供一些不同的东西。
But on this podcast, in just the same way as we always have, we wanna offer you something a little bit different.
我们想带你直抵这些想法的发源地，向你介绍那些引领我们集体未来设计的人们。
We want to take you right to the heart of where these ideas are coming from, to introduce you to the people who are leading the design of our collective future.
没有炒作，没有渲染，只有引人入胜的讨论和宏大的科学抱负。
No hype, no spin, just compelling discussions and grand scientific ambition.
考虑到所有这些，我现在和DeepMind联合创始人、现任Google DeepMind CEO Demis Hosabis在一起。
So with all of that in mind, I am here with the DeepMind cofounder and now CEO of Google DeepMind, Demis Hosabis.
我现在必须称呼您为Demis爵士了吗？
Do I have to call you sir Demis now?
你不知道。
You don't know.
绝对不用。
Absolutely not.
好的。
Okay.
那么，Demis，欢迎来到这个播客。
Well, Demis, welcome to the podcast.
谢谢。
Thank you.
非常感谢你能来。
Thank you very much for being here.
好的。
Okay.
我想知道，随着公众兴趣的爆发性增长，你现在的工作是变得更轻松还是更困难了？
I wanted to know, is your job easier or harder now that there has been this explosion in public interest?
我认为这是把双刃剑。
I think it's double edged.
对吧？
Right?
我觉得更困难了，因为现在有太多关注和审视，整个领域其实充斥着很多噪音。
I think it's harder because there's just so much scrutiny focus and actually quite a lot of noise in the whole field.
说实话，我更喜欢以前人少的时候，那时可能更专注于科研本身。
I I actually preferred it when it was less people and maybe a little bit more focused on the science.
但这也是好事，因为这表明技术已经准备好以多种方式影响现实世界，并以积极的方式改变人们的日常生活。
But it's also good because it shows that the technology is ready to impact the real world in many different, you know, ways and impact people's everyday lives in positive ways.
所以我觉得这也很令人兴奋。
So I think it's exciting too.
嘿。
Hey.
你对这项技术如此迅速抓住公众想象力感到惊讶吗？
Have you been surprised by how quickly this has caught the public's imagination?
我的意思是，我猜你最终会预料到这种情况
I mean, I I guess you would have expected that eventually
是的。
Yes.
人们本来会接受的。
People would have got on board.
是的。
Yes.
确实如此。
Exactly.
所以到了某个时刻，我们这些多年来——甚至几十年来——一直致力于此的人，你懂的。
So it's we at some point, you know, those of us who've been working on it for like us for for many years now, you know, even decades.
所以我猜想，公众终将意识到这一事实，实际上每个人都已经开始认识到人工智能的重要性。
So I guess at some point, the general public would wake up to that fact and effectively everyone's really starting to realize how important AI is going to be.
但亲眼见证这一切成为现实，仍然感觉相当超现实。
But it's been quite surreal still to see that actually come to fruition and for that to happen.
我认为这要归功于聊天机器人和语言模型的出现，毕竟语言是人人都会使用的工具。
And I guess it is the advent of the chatbots and language models because everyone, of course, uses language.
人人都能理解语言。
Everyone can understand language.
因此这是让公众理解和衡量人工智能发展水平的便捷方式。
So it's an easy way for the general public to understand and maybe measure where AI has got to.
我听到你用'超乎想象的强大'来形容这些聊天机器人，这个说法我很喜欢。
I heard you describe these chatbots as though they were unreasonably effective, which I really like.
实际上在本期播客后续内容中，我们将讨论Transformer模型——这个重大突破可以说为我们提供了这些工具。
And, actually, later in the podcast, we are gonna be discussing transformers, which was the big, breakthrough, I guess, the big advance that that gave us those tools.
不过先告诉我，你说的'超乎想象的强大'具体指什么？
But tell me first, what what do you mean by unreasonably effective?
我的意思是，如果回溯五到十年前，当时人们会说我们需要构建Transformer这类精妙架构，然后进行扩展，而不是专门攻克概念或抽象等具体问题。
What I mean by it is I suppose if one were to wind back five, ten years ago, and you were to say, what we're gonna the way we're gonna go about this is, you know, build these amazing architectures like transformer architectures, of course, and then scale from there and not necessarily crack specific things like concepts or abstractions.
这就是我们五到十年前经常争论的焦点——有人认为需要特殊的抽象处理方式。
So these are a lot of debates we would have five, ten years ago is you need a special way of doing abstractions.
大脑确实似乎能做到这一点。
The brain certainly seems to do that.
但不知何故，这些系统只要获得足够数据，它们似乎就能从中学习并泛化，不仅仅是死记硬背，而是某种程度上真正理解所处理的内容。
But yet somehow the the systems, if you give them enough data, then they do seem to learn this and generalize from those examples, not just rote memorized, but actually somewhat understand what they're processing.
这在某种程度上有点出乎意料地有效——比如五年前没人能想到它能达到现在这样的效果。
And it's sort of a little bit unreasonably effective in the sense that, like, I don't think anyone would have thought that it would work as well as it has done, say, five years ago.
是啊。
Yeah.
我想概念性理解和抽象能力的涌现确实是个意外
I suppose it is a surprise that that that things like conceptual understanding and and abstraction have emerged rather
而非刻意设计出来的。是的。
than been been Yes.
我们上次讨论过概念和基础这类话题。
And and we would have been probably we discussed last time things like concepts and grounding.
对。
Yeah.
你知道，将语言扎根于现实世界经验中——无论是通过模拟还是机器人具身智能——这对真正理解周遭世界是否必要？
You know, grounding language in real world experience, maybe in simulations or as robots embodied intelligence, would it be necessary to really understand the world around us?
当然，这些系统目前还达不到这种程度。
And, of course, these systems are not there yet.
它们会犯很多错误。
They make lots of mistakes.
它们并没有真正建立完整的世界模型，但仅通过语言学习就达到了远超预期的水平。
They don't really have a model of the world, a proper model of the world, but they've got a lot further than one might expect just by learning from language.
我想我们或许该解释下什么是'基础'，毕竟不是所有人都听过第一季和第二季。当然。
I guess we probably should actually say what grounding is for those who haven't listened to series one and series Sure.
因为这曾是个重要议题。
Because this was a big thing.
我们之前讨论过很多关于学习的问题，你想在这里简单总结一下吗？
Mean, we were talking about this a lot about how you need so do you wanna just give us a note here of learning?
接地问题是指，你知道，这是上世纪八九十年代构建的系统（比如MIT开发的经典AI系统）面临的问题之一，它们都是庞大的逻辑体系。
So grounding is when, you know, that that was one of the reasons the the systems that were built in the eighties and nineties, the classical AI systems built in places like MIT, they were big logic systems.
你可以想象那里有庞大的词汇数据库，每个词都与其他词相关联。
So you could imagine there's huge databases of words connected to other words.
问题是你可以陈述一个事实，比如'狗有腿'，对吧？
And the problem was you could say something, a dog has legs, right?
这个陈述会被记录在数据库中。
And that would be in the database.
但问题在于，当你展示一张狗的照片时，系统完全无法理解那些像素集合指向的就是那个符号。
But the problem was as soon as you showed a picture of a dog, it had no idea that collection of pixels was referring to that symbol.
这就是接地问题。
And that's the grounding problem.
你拥有这种符号化的抽象表征，但它在真实混乱的世界中究竟意味着什么？
So you have this symbolic representation, this abstract representation, but what does it really mean in the real world, in the messy real world?
当然他们后来尝试解决这个问题，但始终无法完全处理好。
And and then, of course, they try to fix that, but you never get that quite right.
而现在的系统则完全不同，它们直接从数据中学习。
And instead of that, of course, today's systems, they're they're directly learning from the data.
某种程度上，它们从一开始就在建立这种关联。
So in a way, they're forming that connection from the beginning.
但有趣的是，如果仅从语言中学习，理论上它们应该会缺失很多必要的现实关联。
But the interesting thing was is that, you know, if you learn just from language, in theory, they should be missing a lot of the grounding that you need.
但事实证明其中很多内容居然可以通过某种方式推断出来。
But it turns out that a lot of it is inferable somehow.
为什么理论上会这样呢？
Why in theory?
那么，这种基础认知是从哪里来的呢？
Well, because where is that grounding coming from?
至少最初这类大型语言模型并不真实存在于现实世界中。
These sis the at least the first kind of large language models Don't exist in the real exist in the real world.
它们没有连接到模拟器。
They're not connected to simulators.
它们没有连接到机器人。
They're not connected to robots.
它们甚至从一开始就不具备多模态访问能力。
They don't have any access to even they weren't multimodal to begin with either.
它们无法获取视觉信息或其他感官数据。
They don't have access to the to to the visuals or anything else.
它们纯粹生活在语言空间中。
It's just purely they live in language space.
因此它们是在抽象领域中进行学习。
So they're living in a they're learning in an abstract domain.
所以它们能从中推断出现实世界的某些信息就相当令人惊讶了。
So it's pretty surprising they can then infer some things about the real world from that.
这就说得通了——基础认知是通过人们与系统的互动形成的，比如人们会说'这个答案很糟糕'。
Which makes sense that the grounding gets in by people interacting with this system and saying, that's a rubbish answer.
这是个好答案。
That's a good answer.
当然，部分原因是如果它们回答错误——早期版本出错就是因为缺乏基础认知，比如不知道现实中的狗是怎么叫的，回答就会出错，而这种反馈会纠正它。
So for sure, part of that is if the question that they're getting wrong, the early versions of this was due to grounding missing, you know, actually, the real world dogs bark in this way or whatever it is, and it's answering it incorrectly, then that feedback will correct it.
部分反馈就来自我们自身的基础认知。
And part of that feedback is from our own grounded knowledge.
嗯。
Mhmm.
所以确实有些基础概念正在这样逐渐渗透进来。
So some grounding is seeping in like that for sure.
我记得看过一个很好的例子，关于横渡英吉利海峡和步行穿越
I remember seeing a really nice example about crossing the English Channel versus walking across the
英吉利海峡。
English Channel.
对。
Right.
没错。
Exactly.
就是这类事情。
Those kinds of things.
如果它回答错了，你会告诉它错了，然后它就得稍微琢磨明白，你知道，人是不能步行穿越海峡的。
And if it answered wrong, you would tell it it's wrong, and then it would have to sort of slightly figure out that, you you know, you can't walk across the channel.
好的。
Okay.
那么这些涌现出来的特性中，有些是预期之外的，我想稍微问问你关于炒作的问题。
So some of these properties that have that have emerged that weren't necessarily expected to be, I kind of wanna ask you a little bit about hype.
你认为我们现在的处境，此刻的情况如何？是的。
Do you think that where we are right now, how how things are at this moment Yes.
是被过度炒作还是低估了？是的。
Is overhyped or underhyped Yes.
或者可能只是炒作方向错了？
Or is it just hyped perhaps in the wrong direction?
是的。
Yeah.
我认为更多是后者。
I think it's more the latter.
所以我认为短期内它的热度被夸大了。
So I would say that in the near term, it's hyped too much.
我觉得人们现在声称它能做各种事情。
So I think people are claiming can do all sorts of things.
其实做不到。
Can't.
现在有各种初创公司和风投资金在追逐那些疯狂的想法，但你知道，这些想法根本不成熟。
There's all sorts of, you know, startups and VC money chasing crazy ideas that don't you know, they're just not ready.
但另一方面，我认为它仍然
On the other hand, I think it's still
这话居然从你口中说出来，德米斯。
Coming from you, Demis.
是的。
Yes.
我知道。
I know.
我知道。
I know.
知道。
Know.
2010年的人工智能。
AI in 2010.
确实如此。
Exactly.
确实如此。
Exactly.
但是，你知道，我认为它仍然被低估了，甚至现在人们都还没充分认识到当我们实现AGI和后AGI时代会发生什么。
But but, you know, I think it's still underhyped or perhaps underappreciated still even now what's going to happen when we get to AGI and post AGI.
我仍然觉得人们还没有完全理解那件事将会有多么重大，以及随之而来的责任。
I still don't feel like that's that's people are quite understood how enormous that's going to be and therefore the sort of responsibility of that.
所以实际上两者兼而有之。
So it's sort of both really.
我认为目前短期内它有点被过度炒作了。
I think it's it's a little bit overhyped in the near term at the moment.
我们正在经历那个周期。
We're kinda going through that cycle.
我想，好吧。
I guess, though okay.
就所有这些潜在初创公司而言...是的。
So in terms of all of these potential start ups Yeah.
关于风投融资这些事，像你这样几十年来一直沉浸其中的人...是的。
And VC funding and so on, you, who have lived and breathed this stuff for for, as you say, decades Yeah.
你完全有能力判断哪些目标是现实的，哪些不是。
Are are very well placed to spot which ones are are realistic goals and which ones aren't.
但对于其他人来说，他们该如何区分真假呢？
But for for other people, how can they distinguish between what's real and and what isn't?
是的。
Yeah.
嗯，听着，我认为显然你需要做技术尽职调查，对技术和最新趋势有所了解。
Well, look, I think you need to look at obviously, you've gotta do your technical due diligence, have some understanding of the technology and the and the latest sort of trends.
我觉得还可以看看那些提出主张的人的背景，他们的技术功底如何。
I think also look at perhaps the, you know, the background of the people saying it, how technical they are.
他们是去年才从其他领域转来做AI的吗？
Have they just arrived in AI like last year from somewhere else?
我不知道。
I don't know.
他们去年在搞加密货币。
They were doing crypto last year.
你看，这些迹象可能表明他们只是在跟风。
You know, these might be some clues that that perhaps, you know, they're jumping on a bandwagon.
当然这并不意味着他们没有好主意，实际上很多人确实会有。
And it doesn't mean to say, of course, they could still have some good ideas and they many will do.
但可以说这更像是在买彩票。
But it's a bit more lottery ticket like, shall we say.
我认为当一个地方突然受到大量关注时，这种情况总是会发生。
And I think that's always happens when there's a ton of attention suddenly on a place.
显然资金就会随之涌入，每个人都害怕错过机会。
And that's obviously then the money follows that and everyone feels like they're missing out.
这就形成了一种投机性的环境，与我们这些数十年来深耕深度科技、深度科学的人所秉持的理念有些相悖，而我认为在接近AGI的过程中，后者才是理想的发展方向。
And that creates a kind of opportunistic, shall we say, environment, which is a little bit opposite to the people those of us who've been in for decades in a kind of deep technology, deep science way, which is ideally the way I think we need to carry on going as we get closer to AGI.
是啊。
Yeah.
我想我们这个系列要讨论的重点之一就是Gemini项目。
And I guess one of the big things we're gonna talk about in this series is Gemini Yes.
我认为它确实源自那种深度科学的研究方法。
Which really comes from that very deep science approach, I guess.
没错。
Yes.
Gemini与其他实验室发布的大型语言模型有哪些不同？
In what ways is Gemini different from from the other large language models that are released by other labs?
从Gemini项目伊始，我们就希望它从一开始就是多模态的。
So from the beginning with Gemini, we wanted it to be multimodal from the start.
这样它不仅能处理语言，还能处理音频、视频、图像、代码等各种模态的数据。
So it could, you know, process not just language, but also audio, video, image, code, any modality, really.
我们之所以想这样做，首先是因为我们认为这是让这些系统真正理解周围世界并构建更好世界模型的途径。
And the reason we wanted to do that was, firstly, we think that's the way to get these systems to actually understand the world around them and build better world models.
实际上还是回到我们之前的基础性问题，仍在构建基础，但这次是依托于语言来实现。
So actually still going back to our grounding question earlier, still building grounding in, but in the but piggybacking on top of language this time.
因此这很重要。
And so that's important.
我们最终还怀有打造一个通用助手的愿景。
And we also had this vision in the end of having a universal assistant.
我们原型开发了一个名为Astra的系统——相信我们会讨论到它——它不仅能理解你输入的内容，还能感知你所在的上下文环境。
And and we prototyped something called Astra, which I'm sure we'll talk about, which understands not just what you're typing, but actually the context you're in.
试想一个个人助手或数字助理，如果它能理解更多关于你请求的上下文或你所处的情境，它的实用性将大幅提升。
And if you think about something like a personal assistant or digital assistant, it will be much more useful if the more context it understood about what you're asking it for or the situation that you're in.
所以我们始终认为这才是更有价值的系统类型，因此从一开始就构建了多模态能力。
So we always thought that would be a a much more useful type of system, and so we built multimodality in from the start.
这是第一点：原生多模态。
So that was one thing, natively multimodal.
当时这是唯一具备该特性的模型。
And then at the time, that was the only model doing that.
现在其他模型正在努力追赶。
So now the other models are trying to catch up.
我们在记忆方面的另一项重大创新是长上下文处理。
And then the other big innovations we had on memory, so like long context.
实际上现在能记住约100万甚至200万token——你可以大致将其理解为记忆中的词汇量。
So actually holding in mind, you know, a million or 2,000,000 now tokens, you can think of them as more or less like words in mind.
因此你可以给它《战争与和平》这样的长篇著作，或者由于多模态特性，现在甚至可以输入整部电影、讲座视频，然后让它回答问题或在视频流中查找内容。
So you can, you know, give it war of peace or or even a whole because multimodal, a whole video now, a whole film, or lecture, and then get it to answer questions or find you things within that video stream.
好的。
Okay.
Astra项目，就是那个新型通用人工智能助手，能够处理视频和音频数据。
Project Astra, that's the the new universal AI agent, the one that can take in video and and audio data.
在Google I/O大会上，我记得你们举过例子，比如Astra能帮你回忆眼镜放在哪里了。
Google IO, I think you used the example of how Astra could help you remember where you left your glasses, for instance.
不过我在想这些东西的技术渊源。
I wonder, though, about the lineage of this stuff.
因为这难道不就是那些老款谷歌眼镜的高级豪华版吗？
Because is this just a a kind of fancy advanced version of those old Google Glasses?
当然，谷歌在开发眼镜类设备方面历史悠久，可以追溯到2012年左右。
Of course, Google have a long history of developing glass type devices, actually back to, I think, 2012 or something.
所以他们在这方面遥遥领先。
So they're way ahead of the curve.
但可能他们之前就缺这类技术。
But maybe they it was just missing this kind of technology.
这样它就能真正作为智能助手，理解它所看到的内容。
So you could actually understand as a smart agent, a smart assistant that could actually understand what it's seeing.
所以我们非常期待这个数字助手能陪伴左右，理解你周围的世界。
And so we're very excited about that digital assistant to, you know, to go around with you and understand the world around you.
所以当你使用时，感觉这确实是个非常自然的应用场景。
So it seems like a really you know, when you use it, it feels like a really natural use case.
好的。
Okay.
我想稍微回溯到Gemini的起源，因为它来自公司两个不同的部门。
I wanna rewind a tiny bit to sort of the start of Gemini because it came from two separate parts of the organization.
给我讲讲这个故事。
Tell me that story.
没错。
Yeah.
去年我们实际上将Alphabet旗下的两个研究部门合并了。
So we actually, last year, we combined our two research divisions at at at Alphabet.
就是把原来的DeepMind和Google Brain合并成一个部门。
So, obviously, the old DeepMind and then Brain, Google Brain into one.
我们称之为超级单元，汇集了我们公司内所有顶尖人才，将整个谷歌的优秀人才整合到一个统一团队中。
We call it a super unit, bringing all the talent together that we, you know, amazing talent we have across the company, across the whole of Google into one unified unit.
这意味着我们将所有研究中最优秀的知识整合在一起，特别是在语言模型方面。
And what it meant was is that we combined all the best knowledge that we had from all the research we were doing, but especially on language models.
我们有Trinchilla和Gofer等项目，他们开发了Palm、Lambda等早期语言模型。
So we had Trinchilla and Gofer and things like that, and they were building things like Palm and Lambda and early language models.
这些项目各有优缺点，我们将它们整合起来，最终诞生了Gemini这个首个标志性项目。
And they had different strengths and weaknesses, and we pulled them all together into what became Gemini as the first lighthouse project that the combined group would output.
另一个重要方面是我们还整合了所有计算资源，这样我们就能进行大规模的训练运行。
And then the other important thing is, of course, is it was bringing together all the compute as well so that we could, you know, do these really massive training runs and actually pull the compute resources together.
所以我想这次合并非常成功。
So I guess been great.
在很多方面，Google Brain和DeepMind的侧重点确实有所不同。
In in a lot of ways, I mean, the focus of Google Brain and DeepMind was was slightly different.
是的。
Yes.
这么说公平吗？
Is that fair to say?
是的。
Yeah.
我认为是这样的。
So I think it was.
我们显然都专注于AI前沿领域，之前就有很多研究人员层面的合作，但可能在战略层面还不够。
I mean, we were obviously focused on both of us on the frontiers of AI and there was a lot of collaborations already on a kind of individual researcher level, but maybe not on a strategic level.
显然，现在合并后的团队Google DeepMind，我常形容我们是引擎室，但这种结合运作得非常出色。
Obviously, now the combined group, Google DeepMind, I kind of describe it as we're the engine room, but it's it's worked really well.
我认为实际上我们在工作方式上的相似之处远多于差异。
I think there were a lot more similarities actually in the way we were working than there were differences.
我们持续保持并加倍投入在基础研究等领域的优势。
And we've continued to keep and double down our strengths on things like fundamental research.
那么，下一代Transformer架构会从哪里诞生呢？
So, you know, where does the next transformer architecture come from?
我们想要发明它。
We wanna invent that.
显然，Google Brain发明了上一代架构，但我们将其与我们开创的深度强化学习相结合。
Obviously, we you know, Google Brain invented the previous one, but we combine it with deep reinforcement learning that we pioneered.
我仍然认为需要更多创新，我相信我们能像过去十年那样再次实现突破，你知道，这是Google Brain和DeepMind共同的成就。
And I still think more innovations are gonna be needed, And I would back us to do that just as we've done in the past ten years, you know, collectively, both brain and and deep mind.
所以这非常令人振奋。
So it's been exciting.
我想稍后再回到那个合并的话题。
I wanna come back to that that merge in a in a moment.
嗯。
Yeah.
不过我想先聚焦Gemini片刻，它到底有多强？
But I think just sticking on Gemini for a second, how good is it?
与其他模型相比如何？
How does it compare to other models?
是的。
Yeah.
我认为部分基准测试存在问题——整个领域都需要更好的基准测试体系和评估能力。
Well, I think it's you know, some of the benchmarks are not problem is that we need more I think there's one thing the whole field needs is much better benchmarks and benchmarks capabilities.
确实存在一些知名的基准测试，尤其是学术领域的，但现在它们已经趋于饱和，无法真正区分顶尖模型之间的细微差异。
Well, there are some well known benchmarks, academic ones, but they're kind of getting saturated now, they don't really differentiate the nuances between the different top models.
我认为目前处于前沿的模型大致有三种。
I would say there's sort of three models that are kind of at the top of the frontier.
分别是我们的Gemini，当然还有OpenAI的GPT，以及Anthropic的Claude系列模型。
So it's Gemini from us, OpenAI's GPT, of course, and then Anthropic with their Claude models.
显然还有很多其他优秀模型，比如Meta和Mistral等公司开发的。
And then obviously, there's a bunch of other good models too that, you know, people like Meta and Mistral and others built.
它们在不同领域各有所长。
And they're differently good at different things.
这取决于你的需求。
It depends what you want.
比如编程可能是Claude的强项，逻辑推理可能是GPT更擅长。
You know, coding, perhaps that's Claude and reasoning, maybe that's GPT.
而记忆处理、长文本理解和多模态任务则是Gemini的优势所在。
And then memory stuff, long context and multimodal understanding, that would be Gemini.
当然，我们所有人都在持续改进自己的模型。
Of course, we're continuing to all of us are improving our models all the time.
考虑到Gemini这个项目才启动一年，虽然基于我们之前的其他项目，但我认为我们的发展轨迹非常良好。
So, you know, given where we started from, which Gemini is a project only existed for a year, you know, obviously, based on some of our other projects, I think our trajectory is very good.
所以下次我们交谈时，希望我们能够真正站在最前沿。
So, you know, when we talk next time, we should, you know, hopefully be, you know, right at the forefront.
因为还有很长的路要走。
Because there is still a way to go.
我是说，这些模型在某些方面仍然存在不足。
I mean, there are still some things that these models aren't very good at.
是的。
Yes.
确实如此。
For sure.
实际上，这正是当前争论的焦点。
And and actually, that's the big debate right now.
所以最近这一系列技术成果，基本上都源自五六年前发明的那些技术。
So this last set of things kind of emerged from the technologies that were, you know, invented five, six years ago.
问题在于它们仍然存在大量缺陷。
The question is, is they're still missing a ton of things.
比如事实准确性方面，我们都知道它们会产生幻觉。
So they the factuality, you know, they hallucinate as we know.
它们在规划能力上也还有欠缺。
They're also not good at planning yet.
你指的是哪方面的规划？
They're Planning in what sense?
我的意思是，比如长期规划这类。
I mean Well, like kind of long term planning.
它们无法长期持续地解决问题。
So they can't problem solve something long term.
你给它设定一个目标。
You give it an objective.
它们实际上无法替你执行具体行动。
They can't really do actions in the world for you.
所以它们更像是被动的问答系统。
So they're they're very much like passive Q and A systems.
你需要通过提问来输入能量，然后它们才会给出某种回应。
You know, you put the energy in by asking the question, and then they give you some kind of response.
但它们无法真正帮你解决问题。
But they're not able to solve a problem for you.
你不能像数字助理那样说'帮我预订意大利的假期，包括所有餐厅和博物馆之类的'这样的话。
You can't say something like if you wanted as a digital assistant, you might want to say something like, you know, book me that holiday in Italy and all the restaurants and the museums and whatever.
而且它知道你的喜好。
And and, you know, it knows what you like.
然后它会出去为你预订航班等等。
But then it goes out and books the flights and all of that for you.
所以它现在做不到这些。
So it can't do any of that.
但我认为那是下一个时代。
But I think that's the next era.
这类更具代理性质的系统，我们称之为具有代理行为的智能系统。
These sort of more agent based systems, we would call them, or agentic systems that have agent like behavior.
当然，这正是我们的专长所在。
But, of course, that's what we're expert in.
这就是我们过去构建游戏代理（如AlphaGo）和其他项目时所用的技术。
That's what we used to build with all our game agents, AlphaGo and all of the other things we've talked in about in the past.
我们正在做的是将我们著名的研究成果与新型大型多模态模型相结合。
A lot of what we're doing is bringing to kind of marrying that work that we're sort of, I guess, famous for with the new large multimodal models.
我认为这将成为下一代系统。
And I think that's gonna be, you you know, the next generation of systems.
你可以把它想象成将AlphaGo和Gemini结合起来。
You can think of it as combining AlphaGo with Gemini.
是的。
Yeah.
因为我想AlphaGo非常擅长规划。
Because I guess AlphaGo was very, very good at planning.
没错。
Yes.
它在规划方面非常出色。
It was very good at planning.
当然，目前还仅限于游戏领域。
Of course, only in the domain, though, of games.
所以我们需要将其推广到日常工作和语言的通用领域。
And so we need to sort of generalize that into the general domain of everyday workloads and language.
你刚才提到Google DeepMind现在某种程度上成了谷歌的引擎室。
You mentioned a minute ago how Google DeepMind is now sort of the engine room of of Google.
我是说，这和我几年前问起时相比是个很大的转变。
I mean, is quite a big shift since since I was asking the last couple of years ago.
谷歌是不是在你们身上下了很大的赌注？
Is Google taking quite a big gamble on you?
是的。
Yeah.
嗯，我想是的。
Well, I guess so.
我认为谷歌一直都很清楚人工智能的重要性。
I mean, I think Google have always understood the importance of AI.
要知道，桑达尔接任CEO时就说过谷歌是一家AI优先的公司，我们在他任期早期就讨论过这点，他认为AI将是继移动互联网之后的下一个重大范式转变，而且会比这些影响更大。
You know, we've been Sundar, when he took over as CEO, said that Google was an AI first company, you know, and we discussed that very early on in his tenure and he he saw the potential in AI as the next big paradigm shift after mobile and Internet, you know, but bigger than those things.
但我觉得过去一两年里，我们才真正开始践行这一理念，不仅从研究角度，还包括产品和各方面。
But then I think maybe in the last year or two, we've really started living what that means, not just from a research perspective, but also from products and and other things.
所以这非常令人兴奋，但我觉得整合我们所有人才并全力以赴是正确的选择。
So it's very exciting, but I think it's the right bet for us to kind of coordinate all of our talents together and then push as hard as, you know, as possible.
那么反过来看呢？
And then how about the other way around?
因为我想DeepMind原本有很强的基础科研属性，现在成为谷歌的引擎室是否意味着你们必须更关注商业利益而非纯粹的科研？
Because I guess from DeepMind having that very strong research and, like, science focus, does becoming the engine room for Google now mean that you have to care much more about commercial interest rather than the the sort of purer stuff that
你呢？
you do?
确实，我们现在必须更多地关注商业利益，这已明确属于我们的职责范围。
Well, we do definitely have to come worry more about, and it's in our remit now, the commercial interests.
但实际上，关于这方面有几点需要说明。
And but, actually, there's sort of a couple of things saying about that.
首先，我们正在继续推进AlphaFold的科研工作。
First of all, we're continuing on with our science work and AlphaFold.
你们刚刚看到AlphaFold三代的发布，我们正在加倍投入这方面的资源。
And, you know, you just saw AlphaFold three come out, and, you know, we're we're doubling down on our investments there.
我认为这是谷歌DeepMind目前独有的工作方向。
That's, I think, a unique thing that we do at at at Google DeepMind now.
你明白吗？
You you know?
就连竞争对手也承认这些成果是AI带来的普世价值。
And even our competitors point at those things as sort of, you know, universal goods, if you like, that come out of AI.
这方面进展非常顺利。
And that's going really well.
我们还分拆成立了Isomorphic公司专门进行药物研发。
And we spun out isomorphic to to do drug discovery.
这非常令人振奋。
So it's very exciting.
所有工作都进展得很好。
That's all going really well.
所以我们会继续坚持这个方向。
And so we're gonna continue to do that.
那么我们之前关于气候等领域的各项工作呢？
And then what was all our work on climate and all of these things?
不过我们团队规模相当大，所以能同时处理多项任务。
But then we're quite a large team, so we can do more than one things at once.
我们还在构建大型模型，比如Gemini等等。
We're also building our large models, Gemini and etcetera.
此外我们还有产品团队，他们正致力于将这些惊人技术应用到谷歌的所有平台上。
And then we have a product team that we're building out that is going to, you know, bring all this amazing technology to all of the surfaces that Google has.
某种程度上，能有机会将所有成果整合其中是种难得的特权。
So it's an incredible sort of privilege in a way to have that there to plug in all of our stuff.
要知道，我们每发明一样东西，立刻就能造福十亿人。
And, you know, we invent something, it immediately can become useful to a billion people.
这确实非常激励人心。
And so that's really motivating.
实际上另一个变化是：如今产品所需的AI技术与通用人工智能研究之间的界限越来越模糊。
And actually, the other thing is is there's a lot more convergence now between the technology you would need to develop for a product to have AI in it and what you would do for pure AGI research purposes.
不像五年前，你还得专门为产品定制AI解决方案。
So there's not really, you know, five years ago, you'd have had to build some special case AI for a product.
现在可以直接从主研项目分支出产品方案。
Now you can branch off your main research.
当然仍需做些产品定制工作，但可能只占10%的工作量。
And of course, you still need to do some things that product specific, but maybe it's only 10% of the work.
因此AI产品开发与AGI研究之间已不存在那种矛盾了。
So there's actually not that tension anymore between what you would develop for a an AI product and what you would develop for trying to build AGI.
可以说90%的研究路线都是相同的。
It's it's it's 90%, I would say, the same research program.
最后，通过产品落地你能获得大量用户反馈，比如会发现内部指标与用户实际体验存在差异，这些认知对研究非常有帮助。
And then finally, of course, if you do products and you get them out into the world, you learn a lot from that, which and and people using it, and you learn a lot about, oh, your internal metrics don't quite match what people are saying, you know, so then you can update that, and that's really helpful for your research.
确实如此。
Absolutely.
好吧。
Well, okay.
我是说，我们在这期播客中会更多讨论那些将AI应用于科学所带来的突破。
I mean, we are gonna talk a lot more in this podcast about those breakthroughs that have come from applying AI to science.
但我想问问你关于那种紧张感——如何判断向公众发布某件事物的最佳时机。
But I wanna ask you about that tension that there is between knowing when the right moment is to to release something to the public.
因为在DeepMind内部，像大语言模型这样的工具当时主要用于研究，而非被视为潜在的商业化产品。
Because internally at DeepMind, those tools like large language models were being used for research rather than being seen as a potentially commercial thing.
是的。
Yeah.
没错。
That's right.
正如你所知，我们从一开始——早在2010年甚至更早之前——就极其严肃地对待责任和安全问题。
So we we you know, as as you know, we've always taken responsibility incredibly seriously here and safety right from the beginning, you know, way back when we started in 2010 and before that.
后来谷歌将我们的一些伦理准则实质性地纳入了他们的AI原则。
And Google then adopted some of our basically ethics charts effectively into their AI principles.
因此我们始终与谷歌整体保持高度一致，希望作为该领域的领导者之一负责任地部署技术。
So we've always been well aligned with the whole of Google and wanting to be responsible about about deploying this as one of the leaders in this space.
现在开始推出包含生成式AI的真实产品是件很有意思的事。
And so it's been interesting now starting to ship real products with Gen AI in them.
实际上我们正在进行大量快速学习，这很好，因为当前技术的风险相对较低。
You know, actually, there's a lot of learning that is going on and we're learning fast, which is good because we're at relatively low stakes here with the current technology.
对吧？
Right?
它目前还没那么强大。
So it's not that powerful yet.
但随着能力增强，我们必须更加谨慎。
But as it gets more powerful, we have to be more careful.
这仅仅是了解产品团队和其他小组如何测试生成式AI技术的过程。
And that's just learning about the product teams that, you know, and other groups, learning about how to test Gen AI technologies.
它与普通技术不同，因为它不会总是做同样的事情。
It's different from a normal piece of technology because it doesn't always do the same thing.
人们几乎像是在测试一个开放世界游戏。
People can it's almost like testing an open world game.
你可以尝试用它做的事情几乎是无限的。
It's almost infinite what you can try and do with it.
所以如何对它进行红队测试是个有趣的问题。
So it's it's sort of interesting to figure out how do you do the red teaming on it.
那么这里的红队测试是指你们几乎是在与自己竞争吗？
So red teaming in this case being where you're competing against yourselves almost?
是的。
Yeah.
红队测试是指设立一个独立于技术开发团队的专门团队，对其进行压力测试，尝试用各种可能的方式破坏它。
So red teaming is when you set up a specific separate team from the team that's developed the technology to stress test it and try and break it in any way possible.
实际上你需要使用工具来自动化这个过程，因为没有人能完全手动进行红队测试。
You know, you actually need to use tools to automate that because nobody can red team.
即使有成千上万的人在做这件事，与产品发布后数十亿用户的使用相比还是远远不够。
Even if you have thousands of people doing it, that's not enough compared to billions of users when you put it out there.
他们会尝试各种事情。
They're gonna try all sorts of things.
所以将这些经验用于改进我们的流程是很有趣的，这样我们未来的发布就能尽可能顺利。
So it's kind of interesting to take that learning and then improve our processes so that, you know, our future launches will be as smooth as possible.
展开剩余字幕（还有 387 条）
我认为我们需要分阶段进行，比如先有实验阶段，然后是封闭测试，最后才是正式发布。
And I think we got to do it in stages, you know, where there's experimental phase, then a kind of, you know, there's closed beta and then and then launch.
有点像我们过去发布游戏时的做法。
A little bit again like we used to launch our games back in the day.
所以你基本上是在每一步中学习。
So you sort of and learn at each step of the way.
所以有个，你懂的吧？
So there's a you know?
然后我们还需要做的另一件事，我认为需要加强的是利用AI本身在内部进行红队测试，甚至自动发现一些错误或进行分类，这样我们的开发人员和人工测试人员就能专注于那些真正棘手的案例。
And and then the other thing we gotta do, I think we need to do more on, is use AI itself to help us internally with with red teaming and and and actually spotting some errors automatically or triaging that so that the then the human you know, our our kind of developers and human testers can actually focus on those hard hard cases.
这里有个非常有趣的点，就是你完全处在一个更概率化的空间里。
So there's something really interesting there about how you're just in a much more probabilistic space here.
对吧？
Right?
然后如果某件事发生的概率哪怕很小，是的。
And and and then if there's even a very small chance of something happening Yeah.
只要尝试次数足够多，最终总会发生的。
If you have enough tries, eventually Right.
总会出问题的。
Something will go wrong.
而且我猜已经发生过几起错误了，你知道的，公开的错误。
And I I guess there have been a couple of mistakes that you know, public mistakes.
是的。
Yeah.
所以这就是为什么，如我所说，产品团队正在适应这类测试。
So that's why, you know, I think that, as I mentioned, that product teams are just getting used to the sorts of testing.
他们测试过这些东西，但这些测试具有随机性和概率性特征。
They, you know, they tested these things, but they have this stochastic nature, probabilistic nature.
实际上在很多情况下，如果是传统软件，你可以说已经测试了99.999%的情况，然后就能推而广之。
So in fact, a lot of cases where, you know, if it was a normal piece of software, you could say, I've tested 99.999% of things, So then it extrapolates.
是的。
Yeah.
所以这就足够了，因为就像，你知道的，如果它有地板的话，也没有办法暴露出来。
So then it's enough because it's like, you know, there's no way of exposing the floor that it has if it has one.
但这些生成系统并非如此。
But that's not the case with these generative systems.
要知道，它们能做各种有点出格或标新立异的事情，某种程度上与你之前见过的完全不同。
You know, they can do all sorts of things that are a little bit left field or out of the box, out of distribution in a way from what you've seen before.
如果有聪明人或对手决定——这几乎就像黑客决定以某种方式测试推动它。
If someone clever or adversarial decides to it's almost like a hacker decides to test push it in some way.
我的意思是，它甚至可能是组合性极强的。
And it could even be I mean, it's so combinatorial.
它甚至可能与你之前对它说过的所有话有关。
It could even be with all the things that you've happened to have said before to it.
然后你——它就处于某种奇特状态，或者它的记忆被特定事物填满，这就是为什么它会输出某些内容。
And then you you then it's in some kind of peculiar state, which then or it's got its memories filled up with particular thing, and then it that's why I've output something.
所以这里面有很多复杂性，但并非无限。
So there's a lot of complexity there, but it's not infinite.
所以还是有办法应对的，只是比推出常规技术要微妙得多。
So there's there's ways to deal with it, but it's it's just a lot more nuanced than than launching normal technology.
我记得你说过，应该是在我第一次采访你时，你说过一句很有意思的话——实际上必须认为这是一种完全不同的智能，一种不同的计算方式。
I remember you saying, I think it was like in the first first time I interviewed you, you said something really interesting about how actually, you have to think that this is a completely different intelligence, a different way of computing.
某种程度上必须摆脱我们完全理解的那些确定性事物
You kind of have to move away from the sort of the things that we completely understand, the deterministic stuff
嗯。
Mhmm.
转向这种更加混乱的，概率性的——是的。
Into this much more messy, like, probabilistic Yeah.
充满错误的——没错。
Error ridden Yeah.
是啊。
Place Yeah.
还有你们的测试人员。
As well as your testers.
你认为公众是否需要对我们现在从事的计算类型稍微转变下思维方式？
Do you think the public slightly has to shift its mindset on the type of computing that we're doing?
我也这么认为，因为或许这是我们需要考虑的——有意思的是，我们正在考虑在发布产品前先发布一份原则性文件，明确说明对这套系统的预期。
I think so because and maybe that's something we you know, another thing interestingly that that that, you know, we're thinking about is actually putting out a kind of principles document or something before you release something to show show what is the expectation from this system.
你知道它是为什么设计的吗？
You know, what's it designed for?
它能用来做什么？
What's it useful for?
它不能做什么？
What can't it do?
我认为这里需要某种教育引导，比如：如果你这样使用会发现它很有用，但不要尝试用它做那些事，因为根本行不通。
And I think, you know, there is some sort of education there needed of like, you'll be able to find it useful if you do these things with it, but don't try and use it for these other things because it won't work.
我觉得这正是我们作为该领域需要更清晰阐述的，而用户可能也需要更多使用经验才能理解。
And I think that that's something that, you know, we're we, you know, we need to get better at clarifying as a field and then probably users need to get more experienced on.
实际上有趣的是，这可能正是聊天机器人本身（包括ChatGPT）会让人感到突然的原因——甚至OpenAI自己都感到意外。
And actually, this interesting this is probably why chatbots themselves would came a little bit out of the blue by even obviously ChatGPT, but even to OpenAI, it surprised them.
我们有自己的聊天机器人，谷歌也有他们的。
And and we had our own chatbots and Google had theirs.
当时的情况是，我们审视着这些产品，同时也清楚看到它们仍存在的所有缺陷。
And one of the things was is we were looking at them and we were looking at the all the flaws they still had.
对吧？
Right?
现在依然存在这些问题。
And they still do.
就像这样，它会犯这些错误，有时确实会出现幻觉，诸如此类的问题。
And it's like, well, it's getting these things wrong and it sometimes does, you know, hallucinates and blah blah blah.
而且这类情况实在太多了。
And there's so many things.
但我们当时没意识到的是，实际上这仍有很多非常有价值的应用场景。
But then what we didn't realize is, actually, there's still a lot of very good use cases for that.
即便是现在，人们发现总结超长文档、撰写尴尬邮件或填写繁琐表格等功能都极具价值。
Even now that people find very valuable, you know, summarizing documents and really long things or writing, you know, awkward emails or mundane, you know, forms to be filled in.
这些应用场景中，人们其实不介意出现小错误，可以轻松修正，同时节省大量时间。
And there's all these use cases which actually people don't mind if there's some small errors, they can fix them easily and saves a huge amount of time.
我想这才是最令人意外的发现。
And I guess that was the surprising thing.
某种程度上，当把技术交给大众使用时，人们自发发现了这些用途。
They sort of discovered people discovered when you put it in the hands of everyone.
尽管系统存在各种已知缺陷，但确实涌现出这些有价值的应用场景。
There were these they were there were actually these valuable use cases even though this the the systems were flawed in all of these ways we know.
嗯，好吧。
Well, okay.
这正好引出了我想问的关于开源的下一个问题。
So I think that sort of takes me on to the next question I wanna ask, which is about open source.
嗯。
Mhmm.
正如你所说，当技术掌握在人们手中时，确实可能产生非凡的成果。
Because when things are in the hands of people, as you as you mentioned, really extraordinary things can happen.
我知道DeepMind过去开源了许多研究项目，但感觉现在这种趋势正在逐渐改变。
And I know that DeepMind in the past has open sourced lots of its research projects, but it feels like that's slightly changing now as we go forward.
请直接告诉我你们对开源的态度。
Just tell just tell me what your stance is on open source.
是的。
Yeah.
嗯，你看，我们一直是开源和开放科学的坚定支持者。
Well, look, we we're huge supporters of open source and open science.
如你所知，我们几乎公开并发表了我们所做的一切。
As you know, we've we've I mean, we've given away and published almost everything we've done.
要知道，包括像Transformers和AlphaGo这样的成果，我们都在《自然》和《科学》杂志上发表了这些内容。
You know, if you collectively, including like things like Transformers, right, and AlphaGo, we published all these things in Nature and Science.
正如我们上次提到的，AlphaFold是开源的。
AlphaFold was open source as as we covered last time.
这些都是很好的选择。
And these are all good choices.
你说得完全正确。
And you're absolutely right.
这一切之所以有效，是因为通过分享信息，技术和科学才能以最快速度进步。
That's the reason that all works is because that's the way technology and science advances as quickly as possible by sharing information.
因此，这样做几乎总是普遍有益的。
So it almost always that's a universal good to do it like that.
这就是科学的运作方式。
And that's how science works.
唯一的例外是当我们面对具有双重用途的技术时。
The only exception is when we have a dual purpose technology.
对。
Right.
而AGI和强大的人工智能确实属于这一类。
And AGI and AI, powerful AI does fall into this.
于是问题就变成了：你既要支持所有善意的应用场景和真诚的科学家，也要让技术人员能基于这些理念进行建设、批评等等。
And so then the problem is, is that you want to enable all the good use cases and and all the genuine scientists who are acting in good faith and so on, technologists to to build on the ideas, critique the ideas, and so on.
要知道，社会就是这样最快发展的。
That's the way, you know, society advances the quickest.
但问题在于，你如何同时限制那些恶意行为者——他们会利用同样的系统，将其改用于不良目的，比如武器系统之类的，谁知道会用来做什么。
But the problem is how do you restrict access at the same time for bad actors who would take the same systems, repurpose them for bad ends, misuse them, you know, weapon systems, who knows what.
而且，这些通用系统确实可能被这样重新利用。
And, you know, those general purpose systems can be repurposed like that.
目前还好，因为我认为这些系统还没那么强大。
And it's okay today because I don't think the systems are that powerful.
但两三年后，特别是当系统开始具备代理行为时，如果有人滥用，甚至某个流氓国家滥用，可能会造成严重危害。
But in two, three, four years time, especially when you start getting agent like systems or agentic behaviors, then I think, you know, if it's something's misused by someone or perhaps even a rogue nation state, there could be serious harm.
所以我认为，虽然我还没有解决方案，但作为社区，我们需要思考这对开源意味着什么。
So then I think that as a as I don't have a solution to that, but as as a community, we need to think about what does that mean for open source?
或许前沿模型需要更多限制措施。
Perhaps the frontier models need to have more checks on them.
只有在发布一两年后，才能开放源代码。
And then only after they've been out for a year or two years, then they can get open sourced.
这类似于我们Gemini开源模型Gemma采用的模式。
That's sort of the model we're following with because we have our own open models of Gemini called Gemma.
它们规模较小。
They're smaller.
所以不属于前沿模型。
So they're not frontier models.
但对开发者仍然非常有用，因为参数较少，在笔记本电脑上就能轻松运行。
So their capabilities are very useful still to the developer because they're also easy to run on a laptop or because they're small numbers of parameters.
不过它们目前的能力已经被充分掌握。
But but the capabilities they have are well understood at this point.
对吧？
Right?
因为它们不是前沿模型。
Because they're not frontier models.
所以它的能力不如最新版本，比如Gemini 1.5那些模型。
So it's just not as powerful as the latest, say, Gemini, you know, 1.5 models.
因此我认为我们最终可能采取的做法是：我们会开放模型，但这些模型会滞后于前沿技术大概一年左右。这样我们才能真正通过用户公开评估这些模型的能力——那些前沿模型能做到什么。
So I think that's probably the approach that we'll end up taking is we'll have open models, but they'll be lagging, you know, maybe one year behind the the the most cutting edge models just so that those model we can really assess out in the open you know, by by users what those models can do, the frontier ones can do.
这样你才能真正测试那些边界，对吧。
And you can really, I guess, test those those boundaries of the Yeah.
我们会看清这些边界在哪里。
And we'll see what those are.
开源的问题在于一旦出问题就无法撤回。
The problem with open source is if something goes wrong, you can't recall it.
对吧？
Right?
如果是专有模型，当恶意使用者开始滥用时，你完全可以切断访问渠道。
With a proprietary model, if your bad actor starts using it in a bad way, you can just you can just sort of close the tap off.
极端情况下甚至可以直接关闭系统。
You know, in the limit, you could switch it off.
明白吗？
Right?
但开源后就再也无法收回了。
But once you open source something, there's no pulling it back.
所以这是条单行道。
So it's a one way door.
因此在开源前必须万分谨慎。
So you should be very, very sure when you do that.
不过真的有可能将AGI完全限制在某个组织的围墙内吗？
Is it definitely possible to contain an AGI though within the sort of walls of an of an organization?
嗯，那完全是另一个问题。
Well, that's a whole separate question.
我认为目前我们还不清楚如何实现这一点。
I don't think we know how to do that right now.
所以这时候你讨论的就是AGI级别的强大，类似人类水平的人工智能。
So that's that's when you start talking about AGI level powerful, like human level AI.
或者说中间阶段呢？
Or what about intermediary?
关于中间阶段，我认为我们已经有不错的实现思路。
Well, intermediary, I think we have good ideas of how to do that.
比如可以采用安全沙箱这类方案。
So one would be things like secure sandboxing.
所以我希望在游戏环境或非完全连接的互联网版本中测试智能体行为。
So you test that's what I'd wanna test the agent behaviors in is in a game environment or a version of the Internet that's not quite fully connected.
对吧？
Right?
要知道在金融科技等领域，我们已经积累了大量成熟的安全方案。
So there's a lot of security work that's done and known, you know, in this space, in in fintech and other places.
我们会借鉴这些思路来构建系统，并以此测试早期原型。
So we probably borrow those ideas and and then build those kinds of systems, and that's how we would test the early prototype systems.
但我们也清楚这远不足以约束一个可能比我们更聪明的AGI。
But we also know that's not gonna be good enough to contain an AGI, something that's potentially smarter than us.
因此我认为需要更深入理解这些系统，才能设计出适用于AGI的协议。
So I think we gotta understand those systems better so that we can design the protocols for an AGI.
当那个时刻来临时，我们将有更好的方案——可能还会借助AI系统和工具来监控下一代AI。
When that time comes, we'll have better ideas for how to contain that potentially also using AI systems and tools to monitor the next versions of the AI system.
那么首先谈谈安全问题，我知道您深度参与了2023年英国政府在布莱切利园举办的AI安全峰会。
So one, the subject of safety then, because I I know that you were a very big part of the AI Safety Summit at Bletchley Park in 2023, which is, of course, hosted by the UK government.
从外部视角来看，我觉得很多人提到‘监管’这个词时，仿佛它一出台就能解决所有问题。
And and from the outside, I think a lot of people just say the word regulation as though it's just gonna come in and and fix everything.
但你如何看待监管体系应如何构建？
But what is your view on how regulation should be structured?
我认为各国政府能迅速跟进并参与其中是件好事。
Well, I think it's great that governments are getting up to speed on it and involved.
近期兴趣激增带来的积极影响之一，就是政府层面开始高度重视这件事。
I think that's one of the good re things about the the recent explosion of interest is that, of course, governments are paying attention.
我觉得这非常棒。
And I think it's been great.
特别是英国政府——我与他们交流频繁，美国也是，他们的公务员体系中有很多精通技术的聪明人，现在对技术已有相当深度的理解。
UK government, specifically, who I've talked to a lot and US as well, they've got very smart people in the civil service staff that are understand the technology now to to a good degree.
看到英美两国成立AI安全研究所令人振奋，相信更多国家会效仿。
And it's been great to see the AI safety institutes being set up in The UK and US, and I think many other countries are gonna follow.
我认为这些都是很好的先例和规范，值得在风险升级前就确立下来。
So I think these are all good precedents and protocols to settle into, again, before the stakes get really high.
对吧？
Right?
所以这某种程度上也是个验证阶段。
So it's this is a sort of proving stages again as well.
我确实认为需要国际合作，特别是在监管框架、安全护栏和部署规范等方面。
And I do think international cooperation is gonna be needed ideally around things like regulation and guardrails and deployment norms.
因为AI本质是数字技术，很难用国界来限制其发展。
So because AI is a digital technology very much so, it's hard to contain it within national boundaries.
对吧？
Right?
如果英国或欧洲——甚至美国——采取了措施，但中国没有，这真的有助于世界吗？尤其当我们越来越接近通用人工智能时。
So if if The UK or Europe does something, but or even The US, but China doesn't, does that really help the world as oppose you know, when we start getting closer to AGI?
并不完全是这样。
Not really.
因此我认为关键在于，由于技术发展如此迅速，我们必须保持监管的灵活性和敏捷性，这样才能及时适应最新技术的发展方向。
So I think my my view in it is you've gotta be because the technology is changing so fast, we've gotta be very nimble and and light footed with regulation so that it's easy to adapt it to where the latest technology is going.
如果五年前就对人工智能进行监管，那监管对象与今天我们看到的生成式AI完全不同。
If you'd regulated AI five years ago, you regulated something completely different to what we see today, which is Gen AI.
五年后可能又会是另一番景象。
And it might be different again in five years.
届时风险最高的可能是这些基于智能体的系统。
It might be these agent based systems that are the ones that carry the highest risk.
所以目前我会建议先加强现有领域的监管，比如医疗、交通等已有监管框架的行业。
So right now, I would, you know, recommend a sort of beef up existing regulations in in domains that already have them, health, transport, so on.
我认为完全可以像当年为移动互联网更新法规那样，为AI时代更新这些监管条例。
I think, you know, you can update them for AI for an AI world just like they were updated for mobile and Internet.
这或许是我在开展监测工作时首先要做的——确保我们理解并测试这些前沿系统。
That's probably the first thing I do while doing a watching brief you know, and making sure you understand and test this the frontier systems.
待事态发展更加明朗后，再围绕这些新情况制定监管政策。
And then as things become clear and sort of more clearly obvious, then start regulating around that.
可能再过几年实施会更合适。
You know, maybe in a couple of years time would make sense.
我们目前缺失的是能力评估的基准测试——整个行业和领域都想知道：能力发展到什么程度会构成重大风险？
One of the things we're missing is, again, the benchmarks, the right test for capabilities that what we'd all wanna know, including the industry and the field, is at what point are capabilities posing some sort of big risk?
目前除了我刚才提到的智能体能力可能是下一个临界点外，这个问题尚无定论。
And there's no answer to that at the moment, right, beyond what I've just said, which is agent based capabilities is probably a next threshold.
是的。
Yeah.
但对此还没有公认的测试标准。
But there's no agreed upon test for that.
要知道，你可以想象一种能力，比如测试欺骗行为。
You know, one thing you might imagine is, like, testing for deception, for example, as a capability.
你绝对不希望系统具备这种能力，否则你就无法信任它报告的其他任何内容。
You really don't want that in the system because then you can't rely on anything else that it's reporting.
对吧？
Right?
所以我认为这应该是首要测试的新兴能力。
So that would be my number one emergent capability that I think, you know, would be good to test for.
但还有很多方面，比如实现特定目标的能力、复制能力，目前这方面已有大量研究工作在进行。
But there's many, you know, ability to achieve certain goals, ability to replicate, and there's quite a lot of work going on on this now.
我认为安全机构——基本上是政府机构——应该大力推动这方面工作，当然实验室也要贡献我们的所知。
And I think the safety institutes, are basically sort of government agencies, I think it'd be great for them to do a lot, you know, to push on that as well, as well as the labs, of course, contributing what we know.
我好奇，在你描述的世界图景中，机构的定位是什么？
I wonder, in this picture of the world that you're that you're describing, what's the place for institutions in this?
我是说，如果我们发展到AGI支持所有科学研究的阶段...嗯...
I mean, if we get to the stage where we have AGI that's kind of supporting all scientific research Mhmm.
是否还存在...嗯...
Is there still a place Mhmm.
伟大机构的立足之地？
For great institutions?
是的。
Yeah.
我认为有。
I think so.
你看。
Look.
在达到AGI之前还有很长的路要走。
Well, there's there's there's sort of the stage up to AGI.
我认为这必须是民间社会、学术界、政府和工业实验室之间的合作。
I think that's gotta be a cooperation between civil society, academia, government, and and the industrial labs.
所以我真心相信这是我们能够达到最终阶段的唯一途径。
So I think I really believe that's the only way we're gonna get to the sort of final stages of this.
如果你问的是在通用人工智能实现之后——也许这正是你的问题——那么通用人工智能，当然，我一直想构建它的原因之一就是我们可以用它来回答关于现实本质、物理学、意识等一些最重大、最根本的问题。
Now if you're asking after AGI happens, you know, that maybe that is what you're asking, then AGI, of course, one of the reasons I've always wanted to build it is then we can use it to start answering some of the biggest, most fundamental questions about the nature of reality and physics and all of these things and consciousness and so on.
这取决于它采取的形式，是人与AI专家的结合体还是其他方式。
It it depends, you know, what form that takes, whether that will be a human expert combination with AI.
我认为这种情况会持续一段时间。嗯。
I think that will be the case for a while Mhmm.
就发现下一个前沿领域而言。
In terms of discovering the next frontier.
就像现在，这些系统还无法自主提出猜想或假设。
So, like, right now, these systems can't come up with their own conjectures or hypotheses.
它们可以帮助你证明某些东西，我认为我们将能够证明——比如在国际数学奥林匹克竞赛中获得金牌这类事情。
They can help you prove something, and I think we'll be able to prove, you know, gold get gold medals on international mass Olympiad, things like that.
或者甚至可能解决一个著名猜想。
But I I or maybe even solve a famous conjecture.
我认为这现在已经触手可及，但它们还不具备首先提出黎曼假设或广义相对论的能力，对吧？
I think we're that's within reach now, but not they don't have the ability to come up with Riemann hypothesis in the first place, right, or general relativity.
所以这一直是我对真正通用人工智能的测试标准——它要能够做到这些，或是发明围棋。
So that's really was always my test for maybe a true artificial general intelligence is it will be able to do that or invent Go.
你明白吗？
You know?
而我们目前没有任何这样的系统。
And and so we don't have any systems.
我们甚至可能都不知道理论上该如何设计一个能做到这些的系统。
We don't really know how even probably, you know, know how we would design, in theory even, a system that could do that.
你知道计算机科学家斯图尔特·罗素吗？他告诉我他有点担心，一旦我们实现通用人工智能（AGI），我们可能会变得像过去的皇室王子那样——你知道的，就是那些永远不必登基或做任何工作，只需过着放纵奢华生活却毫无人生目标的人。
You know the computer scientist, Stuart Russell, so he told me that he was a bit worried that once we get to AGI, it might be that we all become like the royal princes of the past, you know, the ones who never had to ascend the throne or do any work, but just got to live this life of unbridled luxury and have no purpose.
是啊。
Yeah.
所以这就是那个有趣的问题。
So that's that is the interesting question.
或许这已经超越了AGI的范畴。
Is it maybe it's beyond AGI.
更像是人工超级智能之类的。
It's more like artificial superintelligence or something.
有时人们称之为ASI。
Sometimes people call it ASI.
但到那时我们应该已经实现了，你知道的，极度富足。
But then we should have, you know, radical abundance.
假设我们能确保公平公正地分配这些资源，那么我们就会处于这样的境地：拥有更多选择做什么的自由，而人生意义将成为重大的哲学命题。
And assuming we make sure we distribute that fairly and equitably, then we will be in this position where, you know, we'll have more freedom to choose what to do, and then meaning will be a big philosophical question.
我认为我们需要哲学家，甚至可能需要神学家，以社会科学家的视角开始思考这个问题——他们现在就应该着手研究。
And I think we'll need philosophers, perhaps theologians even, to start thinking as social scientists that they should be thinking about that now.
到底是什么赋予人生意义？
What what brings meaning?
我是说，我当然认为还存在自我实现的需求，我不认为我们都会只是坐在那里冥想，但...但也许我们会沉迷于电子游戏。
I mean, I still think there's, of course, self actualization, and I don't think we'll all just be sitting there meditating, but but but but maybe we'll be playing computer games.
我不知道。
I don't know.
但这是坏事吗？或者其实不是？
But is that a bad thing even or or not?
对吧？
Right?
谁
Who
谁知道呢？
who knows?
我觉得过去的王子们表现得特别好。
Think the princes of the past came off particularly well.
不。
No.
星际旅行。
Traveling the stars.
但还有人们做的极限运动。
But then there's also, you know, extreme sports people do.
他们为什么要做这些？
Why do they do them?
我是说，你知道的，攀登珠穆朗玛峰。
I mean, you know, climb Everest.
但我觉得这会非常有趣，而且那个我不清楚。
But I think it's gonna be very interesting, and and that I don't know.
但这正是我之前提到的观点。
But that's that's kind of what I was saying earlier about.
人们低估了即将发生的事，你知道，回到近期炒作与远期前景的对比。
It's underappreciated what's gonna happen, you know, going back to the hype near term versus far term.
所以即便你想称之为炒作，我认为即将发生的变革程度绝对是被低估的。
So if you wanna call that hype even, it's it's definitely underhyped, I think, the amount of transformation that will happen.
我认为从长远来看会非常好。
I think it will be very good in the limit.
我们将治愈许多疾病和所有疾病，你知道，解决我们的能源问题、气候问题。
We'll cure lots of diseases and all diseases, you know, solve our energy problems, climate problems.
但接下来的问题是：存在意义吗？
But then the next question comes is is is there meaning?
所以让我们稍微回归到AGR的话题，而不是总讨论监管者。
So bringing us back, like, slightly closer to to AGR rather than superintendents.
我知道你们的伟大使命是构建造福全人类的人工智能。
I know that your big mission is to build artificial intelligence to benefit everybody.
是的。
Yeah.
但你如何确保它真的能造福所有人？
But how do you make sure that it does benefit everybody?
如何涵盖所有人的偏好，而不仅仅是设计者的？
How do you include all people's preferences rather than just the designers?
是的。
Yeah.
我认为必须...我的意思是，在一个系统中包含所有偏好是不可能的，因为从定义上说，人们的意见本就不同。
I think you've gotta I think what's gonna have to happen is I mean, it's impossible to include all preferences in one system because by definition, people don't agree.
对吧？
Right?
不幸的是，我们从当今世界现状就能看到这点。
We can see that in, unfortunately, in the current state of the world.
国家之间无法达成一致。
Countries don't agree.
政府之间无法达成一致。
Governments don't agree.
我们甚至无法在应对气候问题这样明显的事情上达成共识。
We can't even get agreement on obvious things like dealing with the climate situation.
所以我认为这非常困难。
So I think it's that's very hard.
我设想未来会出现这样的情况：我们将建立一套安全架构体系，个性化AI可以在其基础上构建。
What I imagine that will happen is that, you know, we'll have a set of safe architectures, hopefully, that personalized AIs can be built on top of.
然后各国将根据自身需求决定如何使用和部署这些AI，界定哪些行为是被允许或禁止的。
And then everyone will have, you know, all all different countries will have their own preferences about what they use it for, what they deploy it for, what they, you know, what can and can't be done with them.
这很正常。
And that's fine.
就像现在一样，应该由个人或国家自主决定这些事项。
That's for everyone to individually decide or countries to decide themselves just like they do today.
但作为社会整体，我们需要确保这些架构具有可验证的安全性。
But as a society, we know that there's some provably safe things about those architectures.
对吧？
Right?
之后就可以放心让它们普及应用。
And then you can let them proliferate and and so on.
我认为我们需要经历一个关键转型期——随着接近AGI，国际间需要加强合作，确保以安全架构开发AGI。因为显然存在安全和不安全的开发方式。
So I I I think that we're gonna kind of gotta get through the eye of a needle in a way where as we get closer to AGI, we probably gotta cooperate more ideally, ideally internationally, and then make sure we build AGIs in a safe architecture way, because I'm sure there are unsafe ways, and I'm sure there are safe ways of building AGI.
渡过这个阶段后，就可以重新开放发展，让人们按需拥有个性化便携式AGI。
And then once we get through that, then we can sort of open the funnel again, and everyone can have their own personalized pocket AGIs if they want.
这就是未来的一个版本。
What a version of the future.
嗯。
Yeah.
好的。
Okay.
不过说到安全开发方式，你是指可能会出现某些不良行为特征吗？
But then in terms of the safe way to build it, I mean, are we talking about undesirable behaviors here that might emerge?
是的。
Yes.
不希望出现的行为模式，以及欺骗性能力。
Undesirable emergent behaviors, capabilities that Deception.
欺骗就是你不希望看到的一个例子。
The deception is one example that that you don't want.
价值体系，你知道的，我们必须更好地理解所有这些。
Value systems, you know, we we gotta understand all of these things better.
什么样的防护措施有效且无法被规避？
What kind of guardrails work, not circumventable?
有两种情况需要担心。
And there's two cases to worry about.
一种是坏人或者坏国家滥用技术，也就是人为的误用。
There's bad uses by by bad individuals or or nations, so human misuse.
另一种是AI本身的问题，随着它越来越接近AGI，可能会失控。
And then there's the AI itself, right, as it gets closer to AGI doing going off the rails.
所以我认为这两个问题需要不同的解决方案。
So that and I think you need different solutions for those two problems.
是的，这就是我们在开发这些技术时必须面对的挑战。
And so, yeah, that's that's what we're gonna have to contend with as we get closer to building these technologies.
回到你关于普惠大众的观点，我们正在通过AlphaFold和同构技术等指明方向。
And also just going back to your benefiting everyone point, of course, what what I'm you know, we're showing the way with things like AlphaFold and isomorphic.
我认为如果AI药物设计成功，未来一二十年内我们可能治愈大多数疾病。
I I think we could, you know, cure most diseases within the next decade or or two if, AI drug design works.
还能实现个性化医疗，根据个人病症和代谢特征定制药物，最大限度减少副作用。
And then there could be personalized medicines where it minimizes the side effects on the individual because it's it's mapped to the the person's individual illness and their individual metabolism and so on.
这些都是了不起的突破——清洁能源、可再生能源、核聚变或更高效的太阳能等等。
So these are kind of amazing things, you know, clean energy, renewable energy sources, you know, fusion or better solar power, all of these types of things.
我认为这些都触手可及，届时还能通过海水淡化技术解决全球水资源获取问题。
I think they're all within reach, and then that would sort out water access because you could do desalination everywhere.
所以我坚信这些技术会带来巨大的好处，但我们也必须降低风险。
So I just feel like this enormous good is gonna come from these technologies, but we have to mitigate the risks too.
你提到降低风险的方法之一，就是在某个时刻召集科学界的‘复仇者联盟’。
And one way that you said that you would want to mitigate the risks was that there would be a moment where you would basically do the scientific version of Avengers Assemble.
是的。
Yes.
当然。
Sure.
特伦斯·陶克。
Terence Tauke.
对。
Yes.
合作。
Co.
合作。
Co.
正是如此。
Exactly.
请他过来吧。
Bring him on down.
没错。
Exactly.
这仍是你的计划吗？
Is that still your plan?
是的。
Yeah.
嗯，我想是的。
Well, I think so.
我认为如果我们能获得国际合作，你知道，我希望能建立一个类似国际CERN的组织，专门针对人工智能，汇集全球顶尖的研究人员，你看。
I think if we can get the international cooperation you know, I'd love there to be a kind of international CERN, basically, for AI, where you get the top researchers in the world, you know, look.
让我们专注于这个项目——AGI项目的最后几年，确保每一步都科学、谨慎、深思熟虑，尤其是最后的这些关键步骤。
Let's focus on the final few years of this prod you know, AGI project and get it really right and do it scientifically and carefully and thoughtfully at every every step, the final sort of steps.
我仍然认为这是最佳方式。
I still think that would be the best way.
你怎么知道什么时候该按下
How do you know when is the time to press
答案？
the answer?
这才是核心问题，因为不能过早行动，否则永远无法获得足够的支持来实施。
That's all that's the big question because you you can't do it too early because you would never be able to get the buy in to do that.
很多人会反对。
A lot of people would disagree.
我是说，现在人们对风险存在分歧。
Mean, today, people disagree with the risks.
对吧？
Right?
你会看到一些知名人士声称没有风险，而像杰夫·辛顿这样的人则认为存在大量风险。
You see very famous people saying there's no risks, and then you have people like Jeff Hinton saying there's there's lots of risks.
而我，你知道，我处于中间立场。
And, you know, I'm I'm I'm in the middle of that.
所以我认为必须把握好时机。
So I think you gotta get the timing right.
我认为这些基于代理的系统将成为下一个有趣的发展阶段。
I would say, like, these agent based systems are gonna be an interesting next step.
因此我认为每当出现重大范式转变或涌现新特性时，都需要重新评估并思考这对时间线意味着什么。
So I think every time there's a big paradigm shift or big emergent new property comes along, you've gotta take stock and think about, you know, what does that do to your timelines.
我想和你多聊聊神经科学。
I wanted to talk to you a bit more about neuroscience.
它对你现在的工作还有多少启发？
How much does it still inspire what you're doing?
因为我前几天注意到DeepMind公布了一个配备人工大脑的计算机化老鼠，这改变了我们对大脑控制运动方式的理解。
Because I noticed the other day that DeepMind had unveiled this computerized rat with a with an artificial brain that that helps to change our understanding of of how the brain controls movement.
但在播客第一季时，我记得我们讨论过DeepMind如何直接从生物系统中获得灵感。
But in the first season of the podcast, I remember we talked a lot about how DeepMind takes direct inspiration from biological systems.
这仍然是你们方法的核心吗？
Is that still the core of your your approach?
不是。
No.
现在已经演变了，因为在过去两三年里，我们进入了工程阶段，大规模系统，你知道的，庞大的训练架构。
It's evolved now because I think we've got to a stage in the last, I would say, two, three years, we've gone more into an engineering phase, large scale systems, you know, massive training architectures.
所以我认为神经科学在这方面的影响稍微减弱了。
So I would say that the influence of a of of neuroscience on that is a little bit less.
它可能会重新回来。
It may come back in.
任何需要更多创新的时刻，你都希望获取尽可能多的灵感来源，而神经科学就是这些创意来源之一。
So anytime where you need more invention, then you want to get as many sources as possible, and neuroscience would be one of those sources of of ideas.
但当工程比重更大时，我认为它的优先级就会稍低一些。
But when it's more engineering heavy, then I think that takes a little bit more of a backseat.
所以现在可能更多是将AI应用于神经科学，就像你看到的虚拟老鼠大脑那样。
So maybe more applying AI to neuroscience now, like you saw with the virtual rat brain.
我认为随着我们更接近AGI，我们会用它来理解大脑。
And I think we'll see that as we get closer to AGI using that to understand the brain.
我认为这将是AGI在科学领域最酷的应用场景之一。
I think it'd be one of the coolest use cases for AGI and science.
我猜这些东西大概会经历几个阶段，比如工程阶段、干预阶段。
Guess this stuff kinda goes through phases of, like, the engineering time, intervention time.
这是其中一部分。
Its part.
你知道的，就目前而言，它一直很棒。
It's, you know, for for now, and it's it's been great.
我们显然仍在密切关注它，并采纳其他任何想法。
And we still obviously keep a close track of it and take any other other ideas too.
你所描绘的所有未来图景仍然相当贴近现实。
All of the pictures of the future that you've painted are still anchored quite in reality.
但我知道你说过，你真心希望AGI能够窥探宇宙的奥秘
But I know that you've said that you really want AGI to be able to peer into the mysteries of the universe
是的。
Yes.
深入到普朗克尺度。
Down at the Planck scale.
没错。
Yes.
比如亚原子层面的...是的。
Like, kind of subatomic Yes.
量子世界。
Quantum worlds.
对。
Yeah.
你认为有些东西甚至超出了我们目前的想象 嗯。
You think that there are things that we have not even yet conceived of Mhmm.
最终可能会实现？
That might end up being possible?
我这里说的是虫洞。
I'm talking wormholes here.
完全同意。
Completely.
是的。
Yes.
我希望虫洞理论能成立。
I love wormholes to be possible.
我认为...可能还存在很多误解，应该说，我们对物理和现实本质仍有许多未解之谜。
I I think we there is a lot of probably misunderstanding, I would say, still things we don't understand about physics and then and the nature of reality.
而且，你知道的，显然量子力学与引力的统一这些问题，标准模型还存在诸多缺陷。
And and, you know, obviously, the quantum mechanics and unifying that with, you know, gravity and all of these things, and there's all these problems with the standard model.
所以我觉得...还有弦理论，我是说，我认为...
So I think there's there's there and string theory, you know, I I mean, I just think
这里面存在巨大的漏洞
There's giant gaping holes in
在物理学领域，真的。
physics, really.
物理学问题无处不在。
Physics all over the place.
我和搞物理的朋友讨论过，很多理论都难以自洽。
And if you're you know, I talk to my physics friends about this, and there's a lot of things that don't fit together.
我不太喜欢多重宇宙的解释。
I don't really like the multiverse explanation.
所以我认为，提出新理论并在大型装置上验证会很有意义，或许可以在太空进行——我痴迷于普朗克尺度的事物，普朗克时间、普朗克空间，正是因为那似乎是现实的分辨率。
So I think that it will be great to come up with new theories and then test those on massive apparatus perhaps out in space at these these tiny you know, the reason I'm obsessed with Planck scale things, Planck time, Planck space, you know, is because that seems to be the resolution of reality.
对吧？
Right?
从某种意义上说，这是你能将任何事物分解到的最小量子级别。
That in a way, that's the kind of smallest quanta you can break anything into.
所以这感觉像是你想要进行实验的那种层级。
So that feels like the kind of level you wanna experiment on.
如果你拥有由AGI和极大丰富资源赋能设计的强大装置，这两者你都需要。
If you had powerful apparatus perhaps designed or enabled by having AGI and radical abundance, you would need both.
这样才能负担得起建造这类实验。
So to be able to afford to build those types of experiments.
现实的解析度。
The resolution of reality.
是啊。
Yeah.
多么精妙的说法。
What a phrase.
没错。
Yeah.
你是指我们当前所处的解析度层级吗？是的。
What so as in, like, the resolution that we're at at the moment Yeah.
某种程度上说，人类层级只是对现实的近似模拟。
Sort of human level is just an approximation of reality.
正是如此。
Yes.
没错。
That's right.
而我们知道还有原子层级，更底层则是普朗克尺度——就目前所知，这是可以讨论事物存在的最小解析度。
And then we know there's the atomic level, and below that's the Planck level, which as far as we know is the smallest resolution one can even talk about things.
因此在我看来，这才是真正理解事物本质需要实验的解析度层级。
And so that, to me, would be the resolution one wants to experiment on to really understand what's going on here.
我在想你是否也预见到，AGI将帮助我们揭示那些超出人类理解极限的事物，而实际上我们根本无法真正理解。
I wonder whether you're also envisaging that there'll be things that are beyond the limits of human understanding, that AGI will help us to to uncover, that actually we're just not really capable of understanding.
然后我在思考，如果有些事情无法解释或无法理解，它们是否还能被证伪？
And then I sort of wonder if if things are are unexplainable or un understandable, are they still falsifiable?
是的。
Yeah.
这些都是很好的问题。
Well, look, I mean, these are great questions.
我认为AGI系统有可能理解比我们更高层次的抽象概念。
I think there will be a potential for an AGI system to understand higher level abstractions than we can.
再回到神经科学，我们知道是前额叶皮层负责这个功能，大概能处理六到七层的间接关系。
So through again, neuro going back to neuroscience, we know that, you know, it's your prefrontal cortex that does that, and there's sort of up to about six or seven layers of of indirection, you know, one could take.
比如这个人这么想，而我在想这个人这么想等等。
You know, this person's thinking this, and I'm thinking this about that person thinking this and so on.
然后我们就跟不上了。
And then we we sort of lose track.
但我觉得AI系统可以拥有一个理论上无限大的前额叶皮层。
But I think an AI system could have an arbitrarily sort of large prefrontal cortex effectively.
所以你可以想象它能发现宇宙中更高层次的抽象模式和规律，这些是我们无法完全理解或一次性掌握的。
So you could imagine higher levels of abstraction and patterns that it will be able to see about the universe that we can't really comprehend or hold in mind at once.
从可解释性的角度来看，我的想法与其他哲学家有些不同——就像在智商方面，我们相对于AGI就像蚂蚁相对于人类。
And then I think the from in terms of explainability point of view, the way I think that is a little bit different to other philosophers who've thought about this, which is like we'll be like to an closer to an ant and then the AGI, right, in terms of IQ.
但我觉得应该这样理解。
But I think that's the way to think of it.
我认为我们是图灵完备的。
I think we are Turing complete.
我们本身就是完全通用的智能体，只是运行速度较慢，因为我们依赖缓慢的生物硬件，而且无法无限扩展自己的大脑。
So we're sort of full general intelligences ourselves, albeit a bit slow because we run on slow machinery, and we can't, you know, infinitely expand our own brains.
但从理论上讲，只要有足够的时间和内存，我们就能理解任何可计算的事物。
But we can, in theory, given enough time and and memory, understand anything that's computable.
所以我认为这更像是加里·卡斯帕罗夫或马格努斯·卡尔森下出一步惊人的棋招。
And so it I think it will be more like, you know, Gary Kasparov or Magnus Carlsen playing an amazing chess move.
我自己想不出来，但他们能向我解释为什么这是一步好棋。
I couldn't have come up with it, but they can explain it to me why it's a good move.
因此我认为通用人工智能系统将能做到这一点。
So I think that's what an AGI system will be able to do.
你说过DeepMind是一个二十年计划。
You said that DeepMind was a twenty year project.
是的。
Yeah.
我们现在进展到哪一步了？
How far through are we?
你们还在按计划推进吗？
Are you are you on track?
我认为我们正在按计划推进。
I think we're on track.
没错。
Yeah.
疯狂的是。
Crazily.
因为通常二十年计划永远都还有二十年距离。
Because usually twenty year projects stay twenty years away.
是啊。
Yeah.
不过现在我们已经取得很大进展，我认为我们
But, yeah, we're a good way in now, and I think we're
二十年就是二三十年后
still Twenty years is twenty thirty for
人工通用智能。
AGI.
是的。
Yeah.
所以我认为，我的说法是，如果它在未来十年内实现，我不会感到惊讶。
So I think I would the way I say is I wouldn't be surprised if it comes in the next decade.
所以我认为我们正按计划推进。
So I think we're track.
这与你上次说的相符。
That matches what you said last time.
你还没有更新你的先验预期。
You haven't updated your prior.
完全正确。
Exactly.
太棒了。
Amazing.
是啊。
Yeah.
丹尼斯，非常感谢你。
Dennis, thank you so much.
谢谢。
Thanks.
真是令人愉快。
Absolute delight.
一如既往地令人愉快。
Absolute delight as always.
和往常一样聊得很开心。
Fun to talk as always as well.
谢谢。
Thank you.
好的。
Okay.
我认为这次谈话得出了一些非常重要的结论，特别是与2022年我们上次与丹尼斯交谈时他所说的内容相比。
I think there are a few really important things that came out of that conversation, especially when you compare it to what Dennis was saying last time we spoke to him in 2022.
因为过去几年确实出现了不少出人意料的发展。
Because there there have definitely been a few surprises in the last couple of years.
这些模型展现出真正概念性理解的能力就是其中之一。
The way that these models have demonstrated a genuine conceptual understanding is one.
这种仅通过语言和人类反馈就实现的现实世界基础认知。
This this real world grounding that came in from language and human feedback alone.
我们原先认为这远远不够。
We did not think that that would be enough.
还有不完美的人工智能对普通人来说是多么有趣和实用。
And then how interesting and useful imperfect AI has been to the everyday person.
德米斯本人也承认他当初没有预见到这一点。
Demis himself there admitted that he had not seen that one coming.
这让我不禁思考：那些我们尚未解决的挑战——比如长期规划、自主性和坚不可摧的安全保障——
And that makes me wonder about the other challenges that we don't yet know how to solve, like long term planning and agency and robust, unbreakable safeguards.
其中有多少（我们将在本期播客详细讨论这些话题）会在几年后回顾时发现比想象中简单？
How many of those, which we're gonna cover in detail in this podcast, by the way, are we gonna come back to in a couple of years and realize that they were easier than we thought?
又有多少会比预期更难？
And how many of them are gonna be harder?
至于他做出的重大预测——比如十年二十年内治愈大多数疾病，本十年末实现AGI，或者我们将进入富足时代——这些听起来是不是有点过于乐观了？
And then as for the big predictions that he made, like cures for most diseases in ten or twenty years or AGI by the end of the decade or how we're about to enter an era of abundance, that sounds a bit like Demis is being overly optimistic, doesn't it?
但话说回来，他到目前为止确实也没错过。
But then again, he hasn't exactly been wrong so far.
您正在收听的是《谷歌深度思维》播客，我是汉娜·弗莱教授。
You've been listening to Google Deep Minds, the podcast with me, professor Hannah Fry.
如果您喜欢本期节目，嘿，
If you have enjoyed this episode, hey.
何不订阅一下呢？
Why not subscribe?
我们还有更多与人工智能前沿人士的精彩对话即将上线，话题涵盖从AI如何加速科学发现到应对这项技术最大风险的方方面面。
We have got plenty more fascinating conversations with the people at the cutting edge of AI coming up on topics ranging from how AI is accelerating the pace of scientific discoveries to addressing some of the biggest risks of this technology.
如果您有任何反馈或想推荐未来嘉宾，请在YouTube上给我们留言。
If you have any feedback or you want to suggest a future guest, then do leave us a comment on YouTube.
下次再见。
Until next time.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。