谷歌研究主管尤西·马西亚斯：AI助力癌症研究、量子计算进展与科研人员未来

谷歌研究部门负责人将与我们探讨人工智能在癌症研究、量子计算中的应用，以及产品与研究是否正变得过于紧密。这场在谷歌山景城总部现场观众（包括研究人员和媒体）面前的对话即将开始。事实上，人工智能安全就是身份安全。AI代理不仅是代码片段，更是数字生态系统中的一等公民，理应获得相应对待。

The head of Google Research joins us to talk about AI for cancer research, quantum, and whether product and research are getting too close together. That conversation in front of a live audience of researchers and media at Google's Mountain View headquarters is coming up right after this. The truth is AI security is identity security. An AI agent isn't just a piece of code. It's a first class citizen in your digital ecosystem, and it needs to be treated like one.

这就是Okta率先保护这些AI代理的原因。解锁这层新防护的关键在于身份安全架构。企业需要统一全面的方案，通过一致的政策和监督机制保护所有身份——无论是人类还是机器。别等到安全事故发生才意识到AI代理是巨大的安全盲区。了解Okta身份安全架构如何助您守护包括AI代理在内的新一代数字身份。

That's why Okta is taking the lead to secure these AI agents. The key to unlocking this new layer of protection, an identity security fabric. Organizations need a unified comprehensive approach that protects every identity, human or machine, with consistent policies and oversight. Don't wait for a security incident to realize your AI agents are a massive blind spot. Learn how Okta's identity security fabric can help you secure the next generation of identities, including your AI agents.

请访问okta.com（拼写为0-k-t-a）。第一资本的科技团队不仅讨论多智能体AI，更已实际部署了名为「聊天礼宾」的系统，正在简化购车流程。

Visit okta.com. That's 0kta.com. Capital One's tech team isn't just talking about multiegenthic AI. They already deployed one. It's called chat concierge, and it's simplifying car shopping.

通过自我反思和实时API校验的分层推理，它不仅帮助买家找到心仪车辆，还能安排试驾、获取贷款预批及评估旧车置换价值。先进、直观且已投入使用——这就是第一资本的科技实力。

Using self reflection and layered reasoning with live API checks, It doesn't just help buyers find a car they love. It helps schedule a test drive, get preapproved for financing, and estimate trade in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One.

各位好，我是Alex Kantrowitz，Big Technology播客主持人。非常荣幸能与谷歌研究负责人Yossi Matias探讨研究的未来及其与产品的交集。Yossi，很高兴见到你。

Everyone, I'm Alex Kantrowitz. I'm the host of Big Technology Podcast, and I'm thrilled to be here for a conversation with the head of Google Research, Yossi Matias, about the future of research and how it intersects with product. Yossi, great to see you.

感谢你的到来，Alex。

Well, thanks for being here, Alex.

近期AI领域热议不断。但谷歌最近通过大语言模型提出了关于癌细胞行为的假设，并在活细胞中得到验证。能否谈谈这项发现的意义和来龙去脉？这是否意味着生成式AI可能开启治愈癌症的新纪元，还是说这次只是幸运的偶然？

So there's been a lot of noise in the AI world recently. A lot of noise. But recently, Google has come up with a hypothesis about cancer cell behavior with an LLM that was then proven out in a living cell. So can you talk a little bit about the significance of this and how it came about? Is this the beginning of generative AI being used to potentially cure cancer or was it lucky?

我们应该如何看待这件事？

What should we think about it?

是的，首先我认为人工智能的进步显然是变革性的，而医疗健康可能是AI能产生最大影响的领域之一，因为医疗本质上是一门基于信息的科学。当我们将不同学科结合起来时，自然会解锁新的机遇。借助AI模型和生成式AI，我们现在能更好地理解各种模式。这无疑是一系列研究工作中的一环，而合作往往能创造奇迹。比如这个'细胞到句子'项目，就是耶鲁大学研究者与谷歌研究院、DeepMind的合作成果，旨在探索如何结合基础模型与细胞数据。我认为这是朝着解决医疗领域最大挑战迈出的一步，当然还有更多工作要做。

Yeah, first I think that obviously we see the progress on AI is transformative and one of the areas that AI can probably do more impact than anything is healthcare because healthcare is really about information based kind of science. Now, when you bring together disciplines, then obviously you unlock new opportunities, and with AI models, generative AI, we now have better understanding to understand patterns, and by all means this is one is in a sequence of a lot of research work and you know, a lot of magic happens with collaborations. So this one, for example, on the cell to sentence is a collaboration with Yale researchers and researchers from Google Research and Google DeepMind looking into how to leverage foundation models in combination with the data that we have on cells. So I think that it's a step towards obviously some of the biggest challenges that we have on healthcare. There's a lot of more work to do.

这是探索旅程的一部分。我们研究生成式AI模型的应用已有数年——如何调整模型、辅助诊断、通过AI协研员赋能研究者。仔细想想，这其实就是利用AI代理来筛选信息，完成过去只有顶尖专家才能完成的工作。现在我们能解锁这些可能性，让研究者敢于提出更宏大的问题。

It's part of a journey, I mean we're looking into how to use generative AI models actually for a few years now, how to adapt them to models, how to help them with diagnostics, how to actually empower researchers with the like of AI coscientists, which if you think about it is really using AI agents to help out sift through the information and do the kind of work that in the past only very sophisticated people could do, and now we can actually unlock these opportunities and empower the researchers to do, to ask even bigger questions.

没错。我查阅了资料，这个模型的突破在于发现了一种从未用于治疗癌细胞的物质，能促使癌细胞向免疫系统'举手投降'，这非常了不起。

Right. I did the reading. It seems like what happened with this model is that it found a substance that hadn't been used to treat cancer cells. Basically get them to raise their hands to the immune system, which is pretty amazing.

确实，细想之下还有大量未解信息。很多时候我们甚至不知道未知的存在，这正是科学假设探索的意义所在。顺便说，这也是EcoScientist项目的核心理念——辅助生成假设。像'细胞到句子'这样的项目，本质上是探索如何利用AI解析细胞信息，发现潜在隐藏的模式。

Yeah, if you think about it, there's so much information there that we have yet to unlock. Actually, in many cases, we don't know what we don't know. That's why the scientific process of looking for hypothesis. And again, by the way, this is the basic for EcoScientist, which is about how to help out with generating this hypothesis. But when you think about projects and efforts such as the cell to sentence, it's really about how do we actually leverage an AI on the cell information in this case to actually identify the kind of patterns that may be hidden out there.

重申一下，基于'线索无处不在'的假设，我们在科学进程中的任务就是发现、识别这些线索，并进行测试验证。这需要大量时间和精力，而AI正在赋能并加速这一研究过程。

And again, under the assumption that there are hints all over the place, one of our effort on the scientific process is to uncover, identify these hints, test them, validate them. This all takes a lot of effort and time and AR is really empowering that research and accelerates it.

好的，我们简要谈谈量子计算。谷歌本周取得量子突破，其量子芯片完成某算法的速度比传统超算快13,000倍。这类量子领域的头条新闻屡见不鲜，对公众而言可能比实际研究进展显得更频繁。

Okay, so let's talk about quantum briefly. Google this week had a quantum breakthrough where the quantum chip was able to do complete an algorithm 13,000 times faster than a traditional supercomputer. It's one of those headlines that we see all the time about quantum. Maybe it's to the public. It seems more frequent than it does when you're actually doing the research.

但我们经常看到这些关于量子技术的突破性头条。然后当你问，我们离量子计算还有多远时，答案总是五年、十年，或许更久。你能解释这种脱节现象吗？以及我们应该如何看待当今量子技术的真实发展程度？

But we see like these breakthrough headlines about quantum frequently. And then when you ask, well, how far away are we from quantum computing? It's always five, ten years, maybe longer. So can you explain that disconnect and how real we should think quantum is today?

首先，量子计算是一项非常长期的探索。你看一些基础研究，很多都可以追溯到八十年代。事实上，我们最近非常激动，我们自己的米歇尔·德沃雷特与同事约翰·克拉克和约翰·马丁内斯因其八十年代的工作获得了诺贝尔奖。米歇尔和同事们正在我们出色的AI量子实验室里，基于那些早期科学突破，构建我们认为将成为实用量子计算机的装置。当然，这是一个长期努力，不像许多研究项目可能只需数月或几年，这项研究可以追溯到很久以前。但在2012年，我们实际上决定是时候投资于此，我们按照非常可衡量的时间表和明确的里程碑取得了稳步进展。

So first, quantum computing is a very long term quest. Right? I mean, if you look into some of the basic research, a lot of that goes back to the eighties. In fact, we're very thrilled just recently to have our very own Michel de Voret recognized with his colleagues, John Clark and John Martinez, being a Nobel laureate for their work from the 80s, and Michel and colleagues are actually working in our fabulous AI quantum lab in actually building on some of those early scientific breakthroughs and building what we believe is going to be a practical quantum computing. Now, of course, it's a long term effort, unlike many of the research efforts that sometimes will take months or a few years, this one really goes back, but back in 2012, we actually started, we actually decided that this is time to invest in that, and we have a very steady progress on very measurable timeline and very clear milestones.

当然所有成果都经过验证。昨天《自然》杂志发表的论文，首次展示了量子计算机相对于经典计算机可验证的实际应用优势。仔细想想，这为更好理解分子结构打开了潜在机遇，将在众多应用领域创造未来机会。所以我们看到稳定进展，显然还有大量工作要做。

And of course everything is validated. This announcement of yesterday is a paper in Nature that actually shows the first verifiable practical application advantage of a quantum computer over classical computer. And if you think about it, this unlocks potential opportunities, future opportunities on better understanding of molecules in so many different applications. So we see a steady state. Obviously, there's a lot of more work to be done.

关键是要确保我们实现这些里程碑。我相当乐观地认为，我们将在约五年的框架内看到这些实际应用。

The important thing is actually to make sure that we're having these milestones, and I'm quite optimistic that we are going to see these real life applications in the in the framework of about five years.

好的。我能简短地问一下，如果成功，量子技术将如何改变世界？

Right. Can I ask you briefly, how does quantum change the world if it works?

事实上，我们将能够提出目前实际上无法触及的各类信息问题并获得答案，这将带来实质性改变——因为能更好地理解分子材料，同时也会加速AI本身发展。想想看，当今AI建立在通过计算积累的知识基础上，然后我们据此构建模型。现在想象一下，你将有能力创造对世界的新认知，这些认知可以输入AI并得到放大。所以我认为这将带来实质性的改变（双关语无意）。这个研究领域令人兴奋之处还在于，许多即将发生的重要事情我们甚至还未意识到——因为一旦发现机遇，突然就会产生你可能未曾预料到的事物。想想AI，今天我们能做到的事情对许多人来说几年前还像是科幻小说，而且发展速度还在加快。

Well, the fact that we are going to be able to ask questions and get answers on the kind of information that is practically out of reach today, that's going to be material change because it's better understanding of the materials of molecules and it's also going to accelerate AI itself because suddenly we're actually going to have more, if you think about it, AI today is built on knowledge that we accumulate and build with computation, and then we take it and build the models based on that. Now just imagine that now you're going to have the capability to create new insights into the world that can then be fed and amplified with AI. So I think it's going to be material change, no pun intended, and exciting thing about research and about this domain as well is that a lot of the important things which are going to happen we're not even aware of because once you uncover opportunity, suddenly it creates the kind of thing that perhaps you did not anticipate. I mean, think about AI and what we can do today that for many of us seemed like science fiction just a few years ago and it's just accelerating.

因此量子技术将开启更多可能性。想象这样一个世界：将有更多聪明人投身其中。这将带来新见解、新创意、新创新，我确信会产生新的全球性影响。

So quantum is going to open up more and think about the world where we're going to have many more smart people actually working on that. That's going to open up new insights, new novelty, new innovation, and I'm sure new worldwide world impact.

所以你相信，如果将产品和研究更紧密地结合，实际上能更快地取得更多研究突破。

So you're of the belief that if you bring product and research closer together, you actually end up getting more research breakthroughs faster.

首先，我对深度研究和知识好奇心以及科学研究都感到兴奋。我在谷歌负责产品，实际上在搜索领导岗位上工作了十多年，领导搜索自动完成功能、体育体验和趋势等项目。另一方面，当然，尤其是今天，研究一直是我们所做一切的驱动力，但今天比以往任何时候都更重要，因为当你想到创新，很多都建立在解决研究问题并解锁能力的基础上。这让我非常兴奋，我称之为研究的魔力循环，这是我一直热衷的。

One thing that so first, I'm I'm kind of both excited about deep research and intellectual curiosity and scientific research as well. I'm a product guy in the Google. Was actually over a decade on search leadership working, actually leading auto complete in search and sports experience and trends and so forth. So on the other hand, of course, today, especially today, it was always the case that research is a driver for everything that we do, but today it's more than ever because when you think about innovation, a lot of it is built on unlocking capabilities that we should actually solve the research problem and then it goes back. This goes to what I'm really excited about, which I call the magic cycle of research, something I always was excited about.

事实上，早在我职业生涯初期在贝尔实验室的黄金时期，我最理论性的研究也是由现实世界的例子所驱动，然后将结果应用回现实，这对我来说是最迷人的部分。今天我们一直在这样做，因为所有的研究项目和努力都是由现实世界的问题所驱动，如果我们解决了这些问题，实际上就能解锁机会。有些是长期的，有些需要数年，许多实际上在几个月内就能实现。这个魔力循环是关于如何通过现实世界的问题驱动突破性研究，然后解决问题，研究问题，经常发表，这就是为什么验证、同行评审等如此重要。

In fact, even as early on my career when I was in Bell Labs in their heydays, my most theoretical research was motivated by real world examples and then actually taking the results and applying them back was to me the most fascinating aspect. Today that's what we do all the time because all of our research projects and efforts are motivated by problems in the real world that if we solve it, it would actually unlock opportunities. Some of them longer term, some of them would take years, many of them are actually within months. Now this magic cycle is about how to drive breakthrough research motivated by real world problems, then solving the problem, the research problem, quite often publishing it, that's why it's so important to actually have the validation, the peer reviewed and everything.

那很好。

That's good.

然后将其应用回现实世界的产品、业务、科学和社会中，这又产生了新的问题。这个循环的一个神奇之处在于，谷歌研究实际上在整个循环中工作，经常是同一个团队在取得突破性研究后，又与产品团队和其他合作伙伴一起将其变为现实，并回到下一个大问题，加速这一过程。

And then taking it back to applying it back to real world applications on products, on businesses, on science and society, and this generates the next questions. Now this cycle, one of the magical things about Google research is that we are actually working through the entire cycle, and the same team quite often that actually had the breakthrough research is the team that would actually then bring it together with product teams and others, partners, to actually reality and go back to the next big questions and accelerate that.

但让我问你，将产品和研究过于接近是否有危险？我的意思是，研究人员可能会被激励进入产品周期。而产品往往是通过季度增长来评估的。你真的希望研究有长期关注。那么你如何看待这一点？

But let me ask you, isn't there a danger of bringing product and research too close? I mean, you could have the researchers motivated to get into the product cycle. And product oftentimes, it's evaluated by growth quarter to quarter. And you really want a long term focus on research. So how do you think about that?

首先，确实在任何开发环境中，重要的一点是要在明天需要做的事情和如何投资未来之间取得平衡。创新周期，对吧？我的意思是，产品开发和业务中的创新困境当然是众所周知的。研究在这方面没有什么不同，我们需要不断管理这些优先级。所以这是一个判断，什么时候是专注于突破的时候？

Well, first, it's true that in any development environment, one of the important things is to have this balance between what you need to do tomorrow and how to invest in the future. The innovation cycle, right? I mean, innovation dilemma in product development and businesses, of course, is well known. Research is no different in the sense that we need to manage those priorities all the time. So it's a judgment call when is the time to actually focus on the breakthrough?

而且往往是长期的。实际上很多时候你并不确切知道它将如何应用。你只是知道这是件重要的事，对吧？你知道，比如如果我能让大语言模型更高效，我知道这很重要。如果我能更好地预测洪水，哦，总会有办法让它变成现实；或者如果我能更好地理解医疗或基因组，总有途径实现它。

And quite often it's for the long term. Quite often actually you don't exactly know how it's going to be applied. You actually know that this is an important thing, right? You know that, well, if I can make LLMs more efficient, I know it's going to be important. If I can actually have better prediction for floods, oh, there's going to be a way for me actually to bring it to reality, or if I'm going to have better understanding of healthcare or genomes, there's a way to do that.

然后当你与产品团队合作时，关键当然是要知道如何高效地实现。顺便说一句，人们常常对将其变为现实如此兴奋，以至于有时需要提醒：嘿，是时候回到下一个问题了。因为产品和研究都如此令人振奋，把握正确时机和判断始终是我们需要做出的决策之一。

Then you when you work with the product teams, one important thing of course is to know how to do that in an effective way. And by the way, quite often people are so excited about actually bringing it to reality that they need sometimes to say, hey, it's time actually to go back to the next question. And because both product and research are so exciting and having the right timing and the right judgment is always one of the decisions we need to do.

我们之前聊过。你提到的一个观点有点反直觉——或者说对我而言并不意外。我们常听到这些术语被随意抛出：发明、创新、研究突破、突破性进展。但你认为真正的突破与创新存在本质区别。能否简要描述一下创新与突破之间的差异？

So we've talked before. And one of the things that you brought up to me was something kind of counterintuitive because we hear or maybe not surprising to me. We hear these terms tossed out invention, innovation, research breakthrough, breakthrough. But you think there's a real difference between an actual breakthrough and what innovation is. So can you just describe a little bit about what the difference between innovation and a breakthrough is?

首先，创新是我们始终在进行的事情。我们应该在产品开发中、在下一代产品的构建中持续创新。我认为随着新能力的出现，全球创新正在加速。而研究突破针对的是那些原则上我们尚不知如何解决的问题，我们需要在某个点上取得突破。当然，有时应用研究就是把已知事物重新组合。

Well, first innovation is something that we're doing all the time. We should do that on product development, on the next generation of what we're going to build. I think that innovation is actually accelerating around the world with new capabilities. When I think about research breakthroughs, this is about problems that currently we don't know how to solve in principle, and we need to somehow make this dent. Now, sometimes some of the applied research is actually to bring together things that are known.

创新不仅应用于产品，也应用于研究本身——因为提出正确问题是任何研究中最重要的环节之一。就像我之前提到的'魔法循环'，我不喜欢'技术转移'这个说法，因为现实从来不是'你构建了某物，然后转移它投入使用'。这永远是个循环过程，永远需要做出判断。

Innovation is something that we apply both on product but also on the research itself because asking the right questions is one of the most important thing in any research. But also I mentioned earlier the magic cycle. When you think about the magic cycle, it's not, you know, I don't like the term technology transfer because life is never, you build something, oh, let's transfer it and make it in use. It's always this cycle. It's always this making the judgment call.

如何将已有成果进行验证测试，通过试点检验后提出新问题？我认为这就是应用于魔法循环本身的创新部分。有些创新真正体现在认识到：如果这项能力通过研究被解锁，将开辟所有这些新机遇。想想对话式AI——早期问题只是'我能实现对话吗？'，接着就是'我该如何应用它？'

How can I take what I've already built and see and test it and have a pilot or test it out and then ask the next question? So I think this is part of the innovation applied to the magic cycle itself, and some of the innovation is really understanding that, oh, if this capability is unlocked with research, this opens up all these new opportunities. I mean, think about conversational AI. Some of it is really about early on it was asking, can I actually have a conversation? And then the next one, how can I actually use it?

这又引出了核心问题：他们真正需要驱动的是什么能力？在此基础上持续构建，在这个案例中它实际是研究与创新的结合体。

And then it brings back to the question of what is actually the capability that they need to drive here? And building on that, and it's really a combination of both research and innovation in this case.

那么，与当前创新需求脱节的长期研究究竟有多重要呢？

So how important then is the long term research that is detached from the need to innovate right now?

首先，没有任何研究是孤立的。正如我所说，最好的研究要么源于已知需求，要么探索可能性艺术。当思考探索可能性时，动机在于——如果我能解决这个问题，将为业务、产品和能力开启真正有意义的新局面。因此研究始终是相互关联的。回答您的问题：长期研究的重要性比以往任何时候都更加关键。因为我们的职责正是推动突破性研究，这些研究将带来变革，使产品、能力、体验、科学及社会挑战的解决方案远超当前水平。当然，其中部分研究可以较快实现创新。

First, no research is detached. Research, as I mentioned, the best research is research that is motivated by either a need that you already know or by exploring the art of the possible, and when you think about exploring the art of the possible, it's motivated by saying, well, no, if I manage to solve it, that is going to unlock things that are actually going to be meaningful for my business, for my products, for capabilities. So it's always connected. To your question, the importance of long term research is greater than, is more than ever, and here's why we are actually, when you think about our job is really to drive breakthrough research that is going to be transformative, that could enable actually products and capabilities and experience and science and all societal challenges to actually be solved in a way that is materially better than we can do today. Now, some of it is something you can actually innovate and find the kind of the shorter term research.

更多时候是要寻找全新范式。比如谷歌研究院2017年开发的Transformer架构——这种新范式一经问世就重塑了整个行业。我们在基因组学和量子计算领域的工作也是如此。量子计算显然属于超长期研究。实际上，在许多领域都存在这种组合：有些突破性研究能快速催生新算法并立即应用，推测性解码就是绝佳案例。

A lot of it is really to find entirely new paradigms to think about I mean, think about the transformers that were developed by Google Research back in 2017. It was a new paradigm that once done, it actually created a lot of the industry or thinking about some of the work we're doing on genomics or quantum. Quantum of course is very long term as we know. So in many areas actually I can see this combination of things that are we can do that very quickly because with breakthroughs and research and we can have a new algorithm and then apply it very quickly. Speculative decoding is a great example.

一旦获得关键洞见，我们就能迅速应用，继而引发行业级影响——包括行业标准确立和众多变体衍生。但有些领域需要全新架构和能力，比如生成式AI、医疗或地球AI，这些都需要多年研究积累。以地球AI为例：它整合了我们多年研发的地理空间模型（用于解决各类问题），结合其他尖端模型，再叠加生成式AI技术，最终实现用自然语言询问地球相关问题并即时获得答案——这实际上是所有模型协同工作的成果。现在想来，这项长期研究的每个组件本身都是长期研究的结晶。比如我们从2017年开始的洪水预测研究，如今全球模型已服务150个国家20亿人口。

Once we had the right insights, we could very quickly actually apply it and then it got its own kind of impact across the industry and industry standard as well and many variations. And there are things that you need to actually think through new architectures, new capabilities, new ways in which to do generative AI or healthcare or Earth AI, for example, that is built on years of actually research when you think about it. Earth AI is about taking all our geospatial models that we developed over the years to tackle various problems and take those state of the art problems with a lot of other models that we developed over the years, then leverage generative AI on top of that and enable anybody to ask any question about Earth and planet in plain language and suddenly get the result which actually is based on combination of all these models. Now, if you think about it, this is a long term research that is based on various components that each of them was a pretty long term research itself. I mean, our work on flood forecasting started in 2017, now we have a global model serving 2,000,000,000 people, 2,000,000,000 people in 150 countries.

我们经历了多年神奇迭代才达到这个水平，现在它又与风暴预测、短时天气、人口动态等模型协同，加上AI基因层共同开启新机遇。要让企业和组织真正能用它解决问题，整个过程确实漫长，但期间存在诸多里程碑。因此我坚信：对许多看似大胆的愿景，应该采取长期视角，再将其分解为具体里程碑——有些是研究节点，有些是产品节点——就像攀登高峰时的各个营地。

It took us years of magic cycle iterations to get there, and now this comes with other models such as storms, weather now casting, population dynamics, etcetera, along with a genetic layer of AI to actually enable and unlock new opportunities. If you think about this dynamics of this to get to the point that now businesses, organizations can actually use it to solve their problems, it actually was a pretty long cycle, but there were many milestones in between. So I'm a great believer that in many cases you take a very long term vision on something that looks very audacious, but then you actually unpack it into tangible milestones. Some of them are research milestones, some of them are product milestones that actually helps you get into that kind of you know, what you're trying to get into this mountain that you try to climb.

广告回来后我们将继续聆听谷歌研究院负责人Yossi Matias的分享。第一资本的技术团队不仅讨论多智能体AI，更已部署名为'聊天管家'的系统来简化购车流程。通过自反思机制、分层推理和实时API校验，它不仅能帮买家找到心仪车型，还能预约试驾、获取贷款预批及评估旧车置换价值。

We'll be back with more from Google research head Yossi Matias right after this. Capital One's tech team isn't just talking about multi agentic AI, They already deployed one. It's called chat concierge, and it's simplifying car shopping. Using self reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get preapproved for financing, and estimate trade in value.

先进、直观且已投入使用——这就是第一资本的科技实力。

Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One.

嘿，科技爱好者们。我是杰森·豪威尔。我是杰夫·贾维斯。在《AI内幕》节目中，我们带着好奇心和一点幽默，拨开AI领域的重重迷雾。

Hey, big technology fans. I'm Jason Howell. And I'm Jeff Jarvis. On AI Inside, we cut through all the AI noise with curiosity. And a bit of humor.

每周我们都会花一小时剖析那些重要的突破性进展，并进行现实检验。与杨立昆这样的行业先驱和艾米丽·本德这样的批评者对话。我们与听众共同学习，让AI的复杂性变得易于理解。想要获取既增长见识又不煽动情绪的AI资讯吗？请在您收听播客的平台订阅《AI内幕》播客。

Every week, we spend an hour unpacking the breakthroughs that matter, and we reality check them. With industry pioneers like Jan Lecun and critics like Emily Bender. We're learning alongside you, making the complexity of AI make sense to all of us. Want AI news that informs and doesn't inflame? Subscribe to the AI Inside Podcast wherever you get your podcasts.

好的，现在让我们从实践层面来探讨。我的意思是，您非常接近生成式AI的发展前沿，能看到最新的突破性研究。下一个突破会来自哪里？

Alright, so let's take it on a practical level now. I mean, you're very close to what's happening in generative AI. You're looking at the latest breakthrough research. Where is the next breakthrough coming from?

研究的美妙之处在于它本质上是探索，很多时候是在探索未知。我们都需要保持谦逊的是，在任何时刻我们都存在认知盲区，而真正令人兴奋的正是去探索那些未知领域。当然这不是随机的，我们不是要碰运气遇到机会，而是要有意识地行动，敢于下注。所以最令人期待的就是那些我们尚未知晓的事物。

Beautiful thing about research is that it's really exploring, in many cases exploring the unknown, and one thing that we all need to be very humbled about is that in any given moment we don't know what we don't know, and the exciting thing is to actually explore that terrain. Of course it's not at random, we just don't try to bump into opportunities. We try to be intentional about it. We try to take some bets. So the most exciting thing are the things that we don't know yet.

显然我们想要研究新架构，获得新见解。我们的很多工作实际上都受到人类大脑、人类和动物行为方式的启发。我们知道存在差距，某些人或动物能比人类更高效地完成某些事情。这实际上证明了可能性。所以在研究中，你通常首先要确认某件事是否可能——说实话我还没见过不可能的事。

Now obviously we want to look into new architectures. We want to do new insights. We want to be inspired by a lot of what we do is really inspired by the human brain and people and animals and how we see behavior, and we know there are gaps, we know that certain people or animals can do things much, much more efficient than we can do as humans. This is actually a proof of existence. So in research quite often what you do, you first want to understand that if something is possible and I've yet to see something that is not, to be honest.

一旦确认可能性，问题就变成了如何实现以及步骤是什么？我认为未来我们将揭开许多目前甚至尚未意识到的事物。

And then if you know it's possible, the question is how do I get there and what are the steps? So I think there's a lot that we're going to unlock to uncover that we're not even aware of.

简单来说，您认为生成式AI的多数进展将来自算法改进还是单纯依靠算力提升？

Briefly, do you think the majority of progress in generative AI is going to come from algorithms or just more compute?

我认为这将是一种组合。显然，我们所取得的许多进展实际上可以追溯到早期——我指的是深度学习的新革命，它吸收了一些已有的理念，但当你投入足够的计算能力和数据时，它就会在实用性和功能上产生质的飞跃。所以这始终是一种组合。想想我们之前讨论的从细胞到句子的例子。大量材料和知识已经存在，但当你拿出一个拥有270亿参数的大模型并在此基础上构建时，它突然就解锁了新的可能性。

I think it's going to be combination. Obviously, lot of the progress that we've done, we've seen actually, even going back to the early days of, I mean, the new revolution of deep learning was taking some ideas that were there before and suddenly when you put enough computing power, enough data, suddenly it has a phase transition in terms of utility and what it can do. So it's always a combination. I mean, think about we discussed earlier about the cell to sentence. So a lot of the material and knowledge is there, but then when you take a big model, you put a 27,000,000,000 parameter model out there and you build on that, suddenly it unlocks new opportunities.

当你以Medjema为基础加入某些医疗能力或信息时，突然就能解锁未知的新机遇。这其中部分关乎规模，但还有一层我们为AI科研人员设计的推理能力——它不仅关乎外部检索，更在于运用研究者应有的推理方式：形成假设，然后通过测试和验证来推进。再想想我们在辅助建模的实证软件上的工作——科研过程中最大的障碍往往是：你遇到问题需要建模，却要反复测试一堆模型来寻找最优解。而基于AI的实证软件能自动构建并帮你筛选合适模型，这大大加速了整个流程。显然，这种结合不仅需要更强大的模型，更需要具备更优推理和思考能力的智能模型，以及实现这一切的算力——这是一条路径。

When you take Medjema and you put some capabilities or medical information and suddenly you can unlock new opportunities that you don't know. So some of it is about scale, but then there's a layer of reasoning that we have for AI coscientists, for example, it's not only about doing the search out there, it's really about applying the kind of reasoning that typically you'd expect researchers to do, which is to form hypothesis to actually then go through ways of testing them and then measuring them. Or think about our work on empirical software to help model building, when a lot in the scientific process, some of the biggest hurdles is really you have a problem, you want to build a model, you actually have a bunch of models just testing and see what's the best and then trying to get the answers. It's very tedious work with this AI based empirical software that can actually build and help you select the right model for that. It accelerates the entire So obviously this combination of not only stronger models but more intelligent models with better reasoning and thinking as well as the power to do that is one approach.

另一方面，算法创新——任何长期从事研究的人都明白，有些问题会在某个时刻因某个突破性创新迎刃而解。想想Transformer的诞生。未来必将出现更多带来突破的算法创新，其中有些已在研发中。当我看到团队在算法创新上的成果时非常振奋，也对整个生态系统中学术界、研究机构和其他企业的进展感到兴奋。

On the other hand, algorithmic innovation and anybody who's been long enough in research knows that there are some problems that at some point somebody comes with this innovation that is an moment and oh, I can actually solve it in a way that was previously impossible. Think about the transformers. There are going to be more algorithmic innovations that are going to make breakthroughs. Some of them are already in the work and I'm really excited when I look into some of the work our teams are doing on algorithmic innovation. I'm excited about what they see from the ecosystem, the academic community, research communities, and other companies.

但我认为最精彩的还在后头。

And but but I think the best is yet to come.

你作为谷歌研究院的负责人，如何说服研究人员从事与生成式AI无关的工作？

So you're the head of Google Research. How do you convince researchers to work on something that's not generative AI related?

要知道，研究人员的驱动力往往来自解决那些无人能解的难题——这种奥林匹亚式的挑战本身就充满魅力。关键在于找到既有科研趣味性，又能产生实际重大影响的问题，这种交集才是真正的动力源。这就是我所说的研究循环。

You know, when you ask yourself what drives researchers, I would say it's a combination of working on interesting problems that, you know, typically when you have a problem that nobody could solve, that makes it interesting, right? It's real. It's kind of an Olympiad type of challenge. Problems that matter, that could make a difference, and the intersection of finding a problem that is going to be both interesting, exciting from a research point of view, then something that could be applied and have a big impact is really the motivation. This is the research cycle that I was talking about.

这正是顶尖研究者的动机。事实上我们各领域都存在这样的机会——想想今天发布的量子计算、基因组学、地球AI等项目。虽然某些领域可能包含生成式AI成分（它确实是能引发深刻问题的惊人技术），但最终人们渴望的是能施展才华、实现突破的研究——无论是否与生成式AI相关。

This is the motivation for the brightest researchers. The thing is that we have that across the board. I mean, think about just announcements today, quantum, think about genomics, think about Earth AI. Now each of them may have some, some of them may have some strong generative AI component and generative AI is an amazing technology that also brings up some exciting questions. Mentioned research on factuality, I mentioned on efficiency, but there are so many other disciplines that and ultimately people are excited to work on things that matter and can actually apply their brilliance and innovation and have breakthrough research.

因此我们从不缺乏这类重要的问题与机遇。再次，我要向谷歌研究院的优秀团队和杰出的研究人员致敬。当我们汇聚来自不同领域的英才——那些理解语言、健康、气候和量子的人才，将他们凝聚在一起时，奇迹就会发生。令人惊叹的是，人们经常跨学科流动，将洞见从一个领域带到另一个领域。所以我认为，当今在谷歌从事研究的激动之处，还在于我们拥有完整的研究体系——从AI基础设施、卓越模型、世界级研究，到能启发我们并付诸实践的产品。

So we're at no lack of such important problems and opportunities, Again, I'd like to give a shout out to the amazing team at Google Research, brilliant researchers. When we bring together talents looking into the different disciplines, bring people who understand languages and health and climate and quantum, and we bring them all together, then a lot of the magic happens. And it's quite amazing to see how people actually also quite often move between disciplines and bring their insights from one to another. So I think again, exciting part of Google, of being in research today is also the fact that we have the full stack of research. We have AI infrastructures, great models, world class research, products that we can actually be inspired by and then apply to.

这一切使我们能够在众多领域开展真正激动人心的研究，从机器学习基础、算法系统到量子科学，再到解决社会问题。

So this altogether enables us to actually get really exciting research on many disciplines, anything from machine learning foundations and algorithms into systems, quantum, into science, into applying to societal problems.

好的。在我们结束前还有最后一个问题。关于癌症研究——如果我没理解错，最酷的一点是模型筛选了所有尚未尝试的潜在治疗方案，并实际发现了一种比人类已找到的更有效的方案。显然，这项生成式AI技术将全面应用于各个研究领域。

Okay. I got one last one for you before we have to go. The cancer research. One of the cool things about that, if I get it right, was that the actual model went through all these different potential treatments that hadn't been tried yet and actually found one that would work better than the ones that humans had uncovered. Obviously, this technology, generative AI technology is gonna be applied in research all across the board.

你预计这会减少对研究人员的需求，还是反而需要更多？

Do you anticipate that it's going to lessen the need for researchers, or are we gonna have more?

事实上我们需要更多各领域的研究人员。想想研究人员的角色是什么？是在现有基础上提出正确问题并推动下一阶段发展。唯一可能减少研究人员的情况，是假设我们已经解答了所有必要问题——我想在场没人会这么认为。

Well, we're going to need many more researchers in all disciplines. I mean, think about what's the role of a researcher? It's really to build on what we can and ask the right questions and build for the next one. Now, the only situation where you need less researchers is if you assume that we practically almost answered all the questions that we need to have. I don't think anybody here in the audience would think that.

我们对世界的认知仅是冰山一角。AI赋能研究者带来的机遇，不仅会催生更多研究人员，更将让他们能提出更宏大的问题、加快研究进程、获得更优成果。就像我同事Demis和John获得诺贝尔奖的AlphaFold——现在研究蛋白质的学者不是变少了，而是更多了，对吧？

We are only understanding tiny bit of what we need to understand. In fact, the opportunity that we have with AI to empower researchers is going to give opportunity not only for more researchers but for each of them to ask bigger question, move faster on the research agenda, have better results. I mean, think about AlphaFold, which my colleagues were recognized with Nobel Prize, Demis and John. I mean, we don't have less researchers working on proteins. We have actually many more, right?

如今他们无需再攻克蛋白质折叠问题，转而研究更重大的课题。想象每位研究生、博士后都拥有AI协研实验室辅助文献检索和假设验证——他们将提出以往只有资深科学家才能应对的问题，从而加速科研进程。医疗、气候、教育领域同样如此。

But now they don't need to work on the protein folding problem. They're actually using it for bigger questions. With AI coscientists, again, think about the fact that every grad student, every postdoc have now their own research lab which can help them with literature search, looking at hypothesis, So now they are going to ask bigger questions. They are going to ask the kind of questions that previously we expected only very senior scientists to do, and we can actually accelerate the kind of scientific process. Similarly in healthcare, similarly in climate, similarly in education.

我的意思是，借助人工智能，教师们有机会更高效地工作，为更多学生提供更优质的教育。而且我们完全有机会让下一代接受更好的教育。事实上，我认为最重要的一点是如何真正赋能下一代，因为创新将来自他们，从而解决许多其他问题。所以我的想法是，我们在理解科学、理解医疗保健、理解世界方面还处于非常早期的阶段，比如在危机中，我们的北极星目标是：没有人应该对即将发生的自然灾害感到意外。通过使用人工智能并让专家们运用它，我们实际上可以更接近这个目标。

I mean, with AI, there's an opportunity for more teachers to be more effective, do more effective work with more students. And again, where no lack of opportunity to actually have the next generation be educated in a better way. In fact, one of the things that are most important in my opinion is how do we actually empower the next generation because the innovation is going to come from them to unlock many of the other problems. So the way I think about it, we're so early on in our ability to understand science, to understand healthcare, to understand the world in a way, for example, in crisis, our North Star is nobody should ever be surprised from a natural disaster coming their way. And by using AI and having the experts using that, we can actually get closer to that.

在医疗保健领域，没有人应该对突如其来的疾病感到措手不及。我们还有大量工作要做，我认为人工智能是人类智慧的放大器。它真正赋能了科学家、医护人员、教师以及我们日常生活中的商业人士。随着人工智能的不断进步，我们就能期待这些专业人士承担更重大的使命，为人类福祉做出更大的贡献。这让我对研究和技术在放大人类智慧方面所扮演的角色感到非常乐观。

On healthcare, there's no reason why anybody should be surprised by a disease that is hitting them. So there's so much more work to do and I think about it as AI as an amplifier of human ingenuity. It really empowers the scientists, the healthcare workers, the teachers, the business people in our everyday life, and the more we're making advancements with AI, then the more we can actually expect all these professionals actually to do, to take on bigger missions, to do bigger progress for the benefit of humanity. Makes me really optimistic about our role at research and in technology in general to actually play a role in actually making this amplification of human ingenuity with AI.

乔斯，非常感谢你。

Josse, thank you so much.

非常感谢你，亚历克斯。

Thank you very much, Alex.