为什么谷歌在人工智能竞赛中落后了

NPR。

NPR.

这是《星球金钱》的《指标》节目。

This is The Indicator from Planet Money.

我是特里米·伍兹。

I'm Terrieme Woods.

我是韦伦·王。

And I'm Waylon Wong.

直到最近，谷歌还是互联网的霸主。

Until recently, Google was the king of the Internet.

是的。

Yeah.

人们会用谷歌来查信息。

You Google information.

这已经成了一个动词。

That is the verb.

你不会说‘Yahoo一下’网页。

You don't Yahoo the web.

你也不会说‘Alta Vista一下’网页。

You don't Alta Vista the web.

不会的。

No.

从来没人这么说过。

No one's ever done that.

然而，随着人工智能的兴起，谷歌在这一领域投入了大量资源，但还是没能保住它的霸主地位。

Yet with the rise of AI, Google has invested a ton in the technology, but that hasn't kept its crown.

也就是说，现在人们提到聊天机器人，不会想到谷歌。

Like, Google isn't synonymous with chatbots.

那是指ChatGPT。

That is ChatGPT.

没错。

That's right.

而且它也不是程序员首选的工具。

And it's not the go to tool for coding either.

那是Claude Code。

That is Claude code.

在最近的一次演示中，谷歌好像把所有AI功能都扔出去，看哪个能留下来。

And in a recent presentation, it kinda appeared like Google was throwing everything AI at the wall to see what sticks.

你可以

You can

比如上传你自己的视频，这张自拍，就能以非常有趣的方式改变现实。

give it your own videos, for example, this selfie, and change reality in a really fun way.

戴密斯·哈萨比斯是谷歌旗下AI团队DeepMind的首席执行官。

Demis Hassabis is the CEO of Google's in house AI group, DeepMind.

戴密斯展示了一个看似古怪的视频生成器，能给你换衣服或者换背景。

And Demis was showcasing a seemingly quirky video generator that could give you a new outfit or background.

但紧接着，他又谈起了极其宏大的技术愿景。

But then the next minute, he was speaking with extraordinary technological ambition.

我们的使命是重新构想药物研发

Our mission is to reimagine the drug

然而，尽管目标如此远大，谷歌却始终没能领跑人工智能领域。

And yet, despite these stratospheric goals, Google never seems to quite get to the front of the AI race.

这是记者塞巴斯蒂安·马利比的观点。

That's according to journalist Sebastian Mallaby.

谷歌就像是个特别擅长拿第二名的实验室。

Google looks like the lab that is just fantastic at coming second.

今天在节目中，我们采访了塞巴斯蒂安，聊聊德米斯·哈萨比斯在谷歌取得的惊人突破，以及为什么这家公司总是在第二名打转。

Today on the show, we talk to Sebastian about how Demis Hassabis made incredible breakthroughs at Google, yet the company is only fantastic at coming second.

我们还深入了解了所谓的‘三重创新者困境’。

And we learn about what's been called a triple innovator's dilemma.

拉姆·普拉特纳已经退出了。

Ram Plattner is out.

在最新一期的NPR政治播客中，我们剖析了一位曾被视为民主党旗帜人物的候选人为何走向衰落。

On the latest NPR politics podcast, we are unpacking the downfall of a candidate once seen as a standard bearer for Democrats.

另外，普拉特纳退出主要参议员竞选，可能对国会权力格局产生什么影响。

Plus, what Plattner's exit from the main senate race could mean for the balance of power in congress.

每天下午收听NPR政治播客，无论你在哪个平台听播客，都能率先掌握最新政治动态。

Listen to the NPR Politics Podcast every afternoon to stay a step ahead of all the latest political news wherever you get your podcasts.

我是坦娅·莫西利，《新鲜空气》节目的联合主持人。

This is Tanya Moseley, cohost of fresh air.

从邮寄选票到种族不公正的选区划分，美国公正自由的选举正面临重重考验。我采访了记者阿里·伯曼，了解他所说的《投票权法案》所遭受的最后一击。

From mail in ballots to racial gerrymandering, a lot is at stake for America's free and fair elections, I spoke with journalist Ari Berman about what he calls a final blow to the Voting Rights Act.

我担心，南方正重回吉姆·克劳时代的政治氛围。

My fear is that we are returning to a politics of Jim Crow in the South.

请在NPR应用或你常用的播客平台收听《新鲜空气》。

Listen to fresh air on the NPR app or wherever you get your podcasts.

塞巴斯蒂安·马利比撰写了多部关于行业巨擘的著作，比如前美联储主席艾伦·格林斯潘和投资人乔治·索罗斯。

Sebastian Mallaby writes books featuring titans like former Fed chair Alan Greenspan and investor George Soros.

他意识到，德米斯·哈萨比斯正是当今最激动人心的行业——人工智能领域里另一位极具影响力的人物。

And he figured out that Demis Hassabis was another powerful person in one of the most exciting industries today, AI.

于是塞巴斯蒂安向他提了这个想法。

So Sebastian pitched him.

有人要给你写本书了。

There is going to be a book about you.

你彻底被写进去了。

You're cooked.

这事就这么定了。

It's done.

德米斯同意让塞巴斯蒂安来写这本第一本书。

Demis acquiesced to having Sebastian be the one to write that first book.

塞巴斯蒂安花了三十多个小时采访他，大部分时间都在伦敦的一家酒吧里。

Sebastian spent over thirty hours interviewing him mostly in a London pub.

有个疯了似的科学家半夜两点盯着我，说：塞巴斯蒂安，我正在读一篇科学论文。

There was this possessed scientist staring at me and saying, at two in the morning, Sebastian, I'm reading a scientific paper.

我能感觉到现实正盯着我，呼唤我，等着我去发现它。

I can feel reality staring at me, calling out to me, waiting to be discovered.

所以我才在做人工智能。

That's why I'm building AI.

我们得弄明白这些东西。

We need to understand these things.

然后他会拍桌子问：为什么这张桌子是实心的？

And then he would bang the table and say, why is this table solid?

其实它就是一堆原子在空隙之间跳来跳去。

It's bunch of atoms jumping around with gaps in between.

那台笔记本电脑呢，Sebastian？

Why is that laptop, Sebastian?

它怎么就能思考？

Why can it think?

这些谜题我们非得搞清楚不可。

These are mysteries that we have to understand.

我得在死前弄明白这些。

I need to understand before I croak.

桌子为什么是实心的，达里安？

Why is the table solid, Darian?

这是在威胁我吗？

Is that a threat?

凌晨两点，啥事听着都像威胁，我觉得。

At 2AM, everything sounds like a threat, I think.

以前从没想过，喝着啤酒的对话会这么生死攸关。

Never before has a a conversation over beer has seemed like at such high stakes.

现在的情况是，丹尼斯想弄明白宇宙的运作原理。

So what's happening here is Dennis wants to understand the universe.

所以他才在开发人工智能。

So he's building AI.

为了做AI，他先拿了认知神经科学的博士学位，然后在2010年于伦敦联合创立了DeepMind。

And to build AI, he studied for a PhD in cognitive neuroscience, and then he co founded DeepMind in 2010 in London.

到2013年，德米斯已经成了各大科技巨头争相挖角的对象，都想收购DeepMind。

By 2013, Demis was being courted by various tech leaders who wanted to acquire DeepMind.

马克·扎克伯格、埃隆·马斯克、谷歌的拉里·佩奇，每个人都有各自的优缺点。

Mark Zuckerberg, Elon Musk, Google's Larry Page, there were pros and cons with each.

最后，他同意了谷歌的收购。

Eventually, he agreed to a purchase by Google.

在德米斯·哈萨比斯的领导下，谷歌在人工智能研究上取得了不少惊人的突破。

Under Demis Hassabis' leadership, Google has made some incredible breakthroughs in AI research.

谷歌DeepMind成功预测了氨基酸如何形成三维结构，也就是所谓的蛋白质折叠这一基本过程。

Google DeepMind successfully predicted how amino acids form three-dimensional shapes, which is a fundamental process known as protein folding.

这项成就让德米斯拿下了诺贝尔化学奖。

That got Demis a Nobel Prize in chemistry.

他告诉我，拿一个诺贝尔奖算是个不错的开始，但他还真想再拿一个。

He tells me that that one Nobel Prize was a nice start, but he'd quite like another one.

对。

Yeah.

听说，第一次拿诺奖的兴奋感消退得特别快。

The buzz of the first Nobel Prize wears off pretty quickly, I hear.

是的。

Yes.

我太理解这种感觉了。

I know the feeling well.

但2022年11月发生的一件事，暴露了身为谷歌这样的巨头可能带来的问题。

But an incident in November 2022 revealed a potential downside to being part of a behemoth like Google.

尽管谷歌早就掌握了这项技术，但著名的AI聊天机器人ChatGPT却是由OpenAI率先推出的。

Even though Google had the technology ready, the famous AI chatbot ChatGPT was first launched by OpenAI.

人们不禁要问，为什么谷歌没有早点推出自己的语言模型？

The question was kinda why did Google not release a language bot?

因为它明明有这个技术。

Because it had the technology.

它完全有能力做出来。

It could have done it.

那为什么没抢先发布呢？

Why did they not release it first?

塞巴斯蒂安认为，谷歌遭遇了他所说的‘三重创新者困境’。

Sebastian believes that Google suffered from what he calls a triple innovator's dilemma.

创新者困境的核心意思是，如果你是一家成功的企业，拥有一个赚钱的好产品，你就不太愿意去支持一个会蚕食现有产品的创新，因为现状实在太诱人了。

Innovator's dilemma basically says, look, if you have a successful incumbent company with a very good product that's making a lot of money, it won't want to back an innovation that will cannibalize the existing product because the status quo is too sweet.

这就是柯达没能抓住数码相机市场的原因，因为它的一切都建立在模拟时代的框架上。

That's why Kodak failed to seize the digital camera market because it was set up for the analog era.

这也是施乐公司没能从个人电脑中获利的原因，尽管他们的研究人员率先开发了这项技术。

And it's also why Xerox couldn't profit from personal computers even though its researchers pioneered the technology.

谷歌从搜索业务中赚了大把的钱。

Google was making a ton of money from search.

既然搜索业务这么赚钱，它为什么要推出像大语言模型这样可能颠覆搜索的竞争对手呢？

Why would it want a rival product like a large language model that would disrupt and destabilize search?

这正是标准的创新者困境。

That's the standard innovator's dilemma.

塞巴斯蒂安说，对谷歌来说，这个困境具体体现在三个方面。

Sebastian says for Google specifically, this innovator's dilemma came in three forms.

首先，它在搜索领域的主导地位依赖于提供可靠信息的声誉。

First of all, you know, its dominance in search depended on a reputation for providing reliable information.

所以它不敢推出会胡编乱造的聊天机器人。

So it couldn't afford to release chatbots that hallucinated.

你还记得它曾经说，可以把胶水涂在披萨上防止奶酪滑下来吗？

Remember when it was saying that you could put glue on pizza to keep the cheese from sliding off?

这方法确实管用。

I mean, it does work.

但代价是什么，达里安？

But at what cost, Darian?

不过这么一来，披萨就没法吃了。

It does make it inedible.

谷歌反对人工智能的第二个原因是，它不清楚如何靠AI赚钱。

And Google's second force working against AI was that it was unclear how to make money from it.

谷歌的收入主要来自在搜索结果旁投放广告，但怎么在聊天对话里插入广告，还不太清楚。

Google's revenues kind of depended on serving ads alongside the search results and it wasn't so clear how you would integrate ads into the chat.

第三，谷歌的市场份额太大了，政界人士、记者，甚至律师都称它为垄断企业。

And then thirdly, Google's market share was so big that politicians, journalists, even lawyers described it as a monopoly.

而这种本身就政治上很敏感的地位，如果谷歌推出像ChatGPT那样的模型——早期频频生成有害内容、胡编乱造，还让一些人觉得它诡异得像有意识——那肯定会惹怒政界、媒体和广告伙伴，根本撑不住。

And so that position, which was kind of politically precarious, would be untenable if Google alienated politicians and journalists and advertising partners by putting out a model like ChatGPT, which in the early days spewed toxic results and, you know, hallucinated and appeared weirdly sentient to some people.

这简直就是自取灭亡的捷径。

I mean, that would have been a shortcut to business suicide.

所以，政治风险确实是第三个障碍。

So, yeah, political risks was barrier number three.

谷歌最终在2023年3月推出了自己的AI聊天机器人。

Google did eventually release an AI chatbot in March 2023.

多亏了德米斯·哈萨比斯和他的团队，这个产品后来演变成了今天的Gemini。

Thanks to Demis Hassabis and his team that's morphed into Gemini today.

它算不上最顶尖的模型，但在一些关键评测中，确实一度站在了前沿。

It's not the leading model, but it has at some points been right at the frontier on various benchmarks.

对。

Yeah.

我们在NPR工作时会用Gemini。

We use Gemini at work at NPR.

对。

Yeah.

当然不是用来写东西，但有时候会用来整理大量文档和研究资料。

Not for writing, obviously, but sometimes for sorting through numerous documents and our research.

谷歌也是NPR的财务支持方之一。

Google's also a financial supporter of NPR.

那这给德米斯·哈萨比斯和谷歌，或者它的母公司Alphabet留下了什么样的局面？

And so where does this leave Demis Hassabis and Google or its parent company, Alphabet?

它会不会最终像柯达或施乐一样，成为创新者困境中的历史教训？

Will it end up like a Kodak or Xerox consigned to the history files as a cautionary innovator's dilemma?

如果你必须押注一家公司，认为它几年后会领先，你会选哪家？

If you had to put your money on one company to be in the lead in few years time, which company would it be?

我会选Alphabet和谷歌，因为我觉得，能在别人之后做到最好，其实是一种很聪明的长期生存策略。

I would say Alphabet and Google because I think, you know, being the best at coming second is a very good strategy for, you know, surviving the distance.

你知道，世界上有太多人用过谷歌的产品，不管是地图、搜索、谷歌云盘还是Gmail，谷歌的普及程度是别人根本比不了的。

You know, there's so many people in the world who are touched by some kind of Google product, whether that's maps or search, Google Drive, Gmail, that they are just going to roll it out at a scale that nobody else can match.

从长远来看，能从这些产品上赚钱，才是你在这场资本密集型竞赛中活下去的关键。

And in the long term, being able to make money from that is what you need to stay alive in this very capital intensive race.

从某种角度说，这些关于商业主导、资本和竞争的讨论，似乎和德米斯·哈萨比斯的科学家形象有点不太搭。

There's a sense in which all of this talk of business dominance and capital and competition might seem a little at odds with Demis Hassabis' scientist persona.

他一直就是这样两种身份并存。

He's always had that duality.

对吧？

Right?

他花了不少时间做电子游戏，跟解决科学难题这种宏大目标比起来，游戏显得挺微不足道的。

He spent some of his time building video games, which are kind of trivial compared to the grand ambitions of solving science.

但他脑子里能同时装得下这两样东西。

And he's got space in his head for both.

你根本没法给他贴标签。

Like, you can't pigeonhole him.

这并不削弱他对信息本质、涌现特性，以及我们能否理解现实结构的浓厚兴趣。

It's not detracting from his fascination with what is the nature of information, what is an emergent property, and can we understand the fabric of reality.

这些宏大的目标还包括一个宣言：有朝一日用人工智能攻克所有疾病。

And those lofty ambitions include that pronounce ment to one day use AI to solve all diseases.

目标是有一天彻底解决所有疾病。

The goal of one day solving all disease.

你觉得这听起来完全荒谬吗？

Does that ring as completely absurd to you?

他喜欢做这种宏大的宣称，说要解决所有疾病听起来有点夸张，但我确实相信，如果人工智能是一种强大的工具，能帮我们整理和理解信息，那我们完全有可能通过顶尖的AI技术，把疾病的根源看得一清二楚。

Likes to make these very sweeping claims and all disease feels a bit much, but I do think that if AI is a super tool with which to sort and make sense of information, you should be able to understand the origin of disease quite minutely through very advanced AI.

你被说服了吗，韦隆？

Are you sold, Waylon?

是的。

Yeah.

也许高级人工智能能陪我喝杯啤酒，给我讲讲这事儿。

Maybe advanced AI could explain this to me over a beer.

AI 解释自己还挺在行的。

AI is pretty good at explaining AI.

它对自己了解得很清楚。

It knows itself well.

本期节目由朱莉娅·里奇制作，技术由特拉维斯·哈根负责。

This episode was produced by Julia Ritchie with engineering by Travis Hagen.

内容核实由西耶拉·华雷斯完成。

It was fact checked by Sierra Juarez.

凯特和卡农负责剪辑，《指标》节目由 NPR 制作。

Kate and Cannon edits the show, and The Indicator is a production of NPR.

特朗普总统表示，与伊朗的停火已经结束。

President Trump says the ceasefire with Iran is over.

本周《请思考》节目，一位前美国高级外交官分享了他对下一步局势的见解。

This week on Consider This, a former high ranking US diplomat shares his thoughts on what could happen next.

德黑兰的当权者显然认为，他们的意志比美国更坚定。

The rulers in Tehran clearly believe they have a stronger will than The United States.

我觉得没什么好选择。

I don't think there are any good options.

这就是本周《Consider This》的全部内容。

That's this week on Consider This.

每天下午，你都可以了解到新闻背后的完整故事。

You can get the full story behind the headlines every afternoon.

可以在NPR应用上收听，或者在你常用的播客平台收听。

Listen on the NPR app or wherever you get your podcasts.

你已经决定要多动一动了。

So you've decided you want to be more active.

运动科学家基思·迪亚兹说，先把智能手表收起来。

Exercise scientist, Keith Diaz, says put away the smartwatch.

先留意身体本身的自然信号。

Start by noticing your body's natural cues.

如果我们能养成重新关注身体感受的习惯，这种行为才更可能坚持下去。

For us to build a habit to check back in with our bodies, that behavior is much more likely to stick.

关注身体真正的需求。

Tracking what the body needs.

更多内容请收听《TED电台》播客。

That's on the Ted Radio Hour podcast.

在NPR应用或您常用的播客平台收听。

Listen on the NPR app or wherever you get your podcasts.