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。
OpenAI 正在接近一笔高达1000亿美元的融资。
OpenAI is closing in on a massive $100,000,000,000 fundraise.
OpenClaw 被收购，代理热潮迅速升温。
OpenClaw is acquired as agent hype goes into overdrive.
那么，人工智能真的让我们更高效了吗？
And is AI making us more productive, actually?
接下来，就在本节目周五特别版中，Box 首席执行官亚伦·利维将与我们深入探讨这一话题。
That's coming up on a big technology podcast Friday edition with Box CEO Aaron Levy right after this.
我是迈克尔·刘易斯。
Michael Lewis here.
我的畅销书《大空头》讲述了2008年美国房地产市场泡沫形成与破裂的故事。
My best selling book, the big short, tells the story of the buildup and burst of The US housing market back in 2008.
十年前，《大空头》被拍成了获得奥斯卡奖的电影，现在我首次将其以有声书形式呈现，由尤瑞斯·特鲁利倾情朗读。
A decade ago, The Big Short was made into an Academy Award winning movie, and now I'm bringing it to you for the first time as an audiobook narrated by Eurus Truly.
《大空头》的故事、做空市场的意义，以及谁真正为失控的金融体系买单，如今比以往任何时候都更具有现实意义。
The Big Short story, what it means to bet against the market, and who really pays for an unchecked financial system is as relevant today as it's ever been.
现在就去 pushkin.fm/audiobooks 或任何有声书销售平台获取《大空头》。
Get the Big Short now at pushkin.fm/audiobooks or wherever audiobooks are sold.
欢迎收听《大科技播客》周五版，我们将以一贯冷静而细致的方式解析新闻。
Welcome to Big Technology Podcast Friday edition where we break down the news in our traditional cool headed and nuanced format.
今天为大家准备了一场精彩的节目。
We have a great show for you today.
我们将讨论 OpenAI 即将进行的约一千亿美元融资，软银、亚马逊、英伟达，可能还有微软预计将参与其中。
We're gonna talk about the forthcoming $100,000,000,000 or thereabouts funding raise for OpenAI where SoftBank, Amazon, Nvidia, and maybe Microsoft are expected to participate.
我们还将讨论 OpenAI 收购 OpenClaw 的情况，以及一些关于人工智能是否真正提升了我们生产力的新研究。
We're also gonna talk about the acquisition of OpenClaw, also by OpenAI, and some new studies about whether AI is actually helping us be more productive.
今天 Ranjan Roy 不在，我们请到了一位完美的嘉宾。
Ranjan Roy is out today, and we are joined by the perfect guest.
回归冠军 Aaron Levy 现在与我们同在。
Returning champion, Aaron Levy, is here with us.
Aaron，Box 公司的首席执行官，欢迎再次做客我们的节目。
Aaron, Box CEO, welcome back to the show.
谢谢。
Thank you.
很高兴，很高兴来到这里。
Good to, good to be here.
AI领域总是精彩纷呈。
Never a dull moment in AI land.
确实如此。
Seriously.
所以这周我们有模型发布，还有潜在的融资公告，真不知道该从哪儿说起。
So this week we have model releases, we have potential funding announcements, it's hard to figure out where to start.
我们就从大新闻开始吧。
Let's just go with the big story.
几周前，我们曾暗示过OpenAI可能正在进行一轮500亿美元的融资。
A couple weeks ago, we foreshadowed this idea that OpenAI, might be on the way to a $50,000,000,000 fundraise.
猜怎么着？
Guess what?
现在翻倍了。
It's doubled now.
看起来可能达到1000亿美元，其中软银占300亿。
It looks like it might be $100,000,000,000, SoftBank with 30,000,000,000 of that.
亚马逊最终可能投资高达500亿美元，考虑到他们与Anthropic的关系，这太疯狂了。然后，我不知道，这些数字甚至让我觉得至少是1100亿，因为英伟达可能...
Amazon might end up investing as much as 50,000,000,000 which is wild given their connections, to Anthropic and then, I don't know, the numbers are even making it look like at least 110 to me because Nvidia might
我记得我们A轮和B轮融资时就是这类数字，所以这一切都是常规操作。
I remember these kind of numbers from our like series A and B days, so this is all this is par for the course.
没错，这里的背景是，英伟达可能会拿出300亿美元，所以这些数字基本上都会超过史上最大IPO的融资总额。
Right, context here is that any so Nvidia could put up, $30,000,000,000 so this this is all of these numbers would basically be larger than the entire amount raised by the biggest IPO, in history.
那么让我直接问你，围绕OpenAI的叙事一直是‘红色警报’输给谷歌、产品同质化、被Anthropic打得落花流水。
So let me just ask you this, the narrative around OpenAI has been Code Red losing to Google, commoditized, getting its ass kicked by Anthropic.
现在钱只是数字，只是钱，但如此大规模的融资是否反驳了其中一些批评？如果其中一些批评可能属实，你认为这些公司为何还会对OpenAI下如此大的赌注？
Now money is just numbers, it's just money, but does this size of a fundraise rebut some of that, and why do you think these companies would be making such a big bet on OpenAI if some of those criticisms might be true?
嗯，我的意思是，我对这件事持相当务实的看法，那就是，可能OpenAI在达到1000亿美元市值后，每一轮超过10亿美元的融资，都会被问到同样一套问题。
Well, I mean, I I just take a pretty pragmatic view to this, which is, you know, probably every fundraise after 1,000,000,000 of, you know, the 1,000,000,000 market cap from OpenAI, the same set of questions would have would have been asked.
我确信，当他们达到100亿、500亿、1000亿乃至几千亿市值时，问题始终是：这个市场的潜力究竟能有多大？
I'm sure when they were 10,000,000,000 and 50,000,000,000 and 100,000,000,000 and and, you know, couple 100,000,000,000, The question was always how big could this market possibly be?
竞争将会异常激烈。
It's going to be hypercompetitive.
谷歌总有一天会醒悟过来。
Google's going to wake up someday.
还有其他竞争对手存在。
There's other competition.
这些模型难道不会变得商品化吗？
Aren't these models going to get commoditized?
你几乎可以想象，这总会是讨论的常态。
You have to almost imagine that's always going to be the state of conversation.
这在每个阶段都会发生，正如我们过去所见，我认为未来也会如此。
That will happen at at every kind of, you know, juncture, you know, as we saw in the past and and I think as we we'll see going forward.
然而与此同时，几乎从每个指标来看，至少OpenAI产品的使用量都在持续增长。
And yet, at the same time, almost by every metric, the usage of of at least OpenAI's products keep growing.
当然，还有Anthropic、Gemini以及其他相关领域的参与者。
Certainly, Anthropix and Gemini's and and other players in the space.
这些模型的能力水平正在不断提升，因此它们正在承担更多的工作。
The capability level of these models is only increasing, so these models are doing more work.
我们仍然处于智能在组织和企业内部产生真正连锁反应的最初阶段。
We are still only in the earliest innings of of of the actual ripple of intelligence across organizations and across the enterprise.
所以，你刚才提到的所有指标都很相关，但它们就像是云计算早期阶段的指标——比如2010年、2011年或2012年，那时你会想，哇，谷歌刚刚入场，Azure正在提升市场份额，你在观察亚马逊，心里想着：在如此激烈的竞争下，这个市场到底能有多大？
So so I think all of the metrics you just cited are relevant, but they're kind of the metrics that you would look at in the early days of cloud computing, and you're like in 2010 or '11 or '12, and you're and you're like, wow, you know, Google just now got into the game, and and Azure is building up market share, and and you're looking at Amazon, you're saying, well, you know, how how big now could this possibly get given how much competition there is?
我认为在人工智能领域，我们正经历着同样的情况：如果你拉长视野，从十年的角度来看这个市场，我们目前所看到的，只是未来将要发生的巨大变革中极其微小的一部分。
And I think in AI, we're we're kind of experiencing the same thing, which is which is if you actually zoom out and you look at maybe the ten year view of this market, we are looking at a a a really, really small percentage of the total change that is going to happen as a result of this.
所以我们还处于最初的阶段。
So we're in the earliest innings.
当你谈到一笔1000亿美元的融资时，想到这一点确实很疯狂，我知道这中间可能存在某种认知失调。
It's crazy to think that when you're talking about a $100,000,000,000 raise, like, I'm I'm, you know, I'm aware of of the the cognitive dissonance that that might exist from that.
但当你谈论的是像人工智能这样，可能是未来一个世纪经济最基础、最核心的支柱之一时，看到如此激烈的竞争，以及一些公司在这个领域逐步逼近万亿美元的估值，就完全合理了。
But when you're talking about just intel like, one of the most kind of fundamental kind of, you know, core fabrics of the economy in the next century, I I it's just like entirely reasonable that you would both see that level of competition and you might have companies that are now approaching a trillion dollars in this in this category.
好的。
Okay.
但这是反对意见会提出的观点。
But here's what the pushback would be.
过去，这些问题曾经出现过，比如Anthropic会怎么做？
It would be that in the past, these questions have come up, you know, what is Anthropic gonna do?
谷歌能整合好吗？
Is Google gonna get it together?
是的。
Yeah.
那些都是假设。
Those were ifs.
现在，谷歌已经整合好了。
Now, Google has gotten it together.
Gemini，我认为这周推出了Gemini 3.1的新模型，它的价格是其他领先模型的一半，但性能差不多。
Gemini, I think we have a new model from Gemini 3.1 that came out this week that is, you know, half the price of the other leading models and has about the same performance.
这场竞争已经实实在在地加剧了。
This is a competition that has tightened in a real way.
是的。
Yeah.
Anthropic已不再是想象中的幻影了。
Anthropic isn't just a figment of the imagination anymore.
完全正确。
100%.
它正在主导企业市场。
It is dominating an enterprise.
Cloud Code 的表现令人惊叹。
Cloud Code is crazy.
但是，你必须，你必须在这件事上稍微换个算法。
But but you just have to you have to kind of do a slightly different math on this.
你刚才说的一切都是真的，但还没有，还没有影响到估值或融资的问题。
You have to the the everything you just said is true and doesn't doesn't yet doesn't impact the the the valuation or funding question.
我们谈论的是一个领域，其市场规模将由万亿美金来衡量，这些规模将由人工智能创造。
We're we're talking about a category where, you know, it'll be measured in the tens of trillions of dollars, the market caps that will will be generated by AI.
其中一部分会流向芯片供应商，一部分会流向芯片供应商的供应链，一部分会流向AI模型提供商，还有一部分会流向应用和部署层。
Some of that will go to the chip providers, some of that will go to the supply chain of the chip providers, some of that will go to the AI model providers, and some of that will go to the kind of application and deployed layer.
所以，如果你谈论的是一个价值数万亿美元的领域，那我们讨论的只是些小规模的摩擦罢了。
So if you're talking about a category that will be worth tens of trillions of dollars, you know, we're we're talking about little skirmishes Right.
在通往成为这家领域中5万亿美元、2万亿美元、5000亿美元或1000亿美元公司的道路上。
On the path to to you know, who's gonna be a $5,000,000,000,000 company in this space, or a $2,000,000,000,000 company in this space, or a $500,000,000,000 company in this space, or a $100,000,000,000 company in this space.
所以我只是从整个市场的规模以及这个蛋糕可能如何分配的角度来看，即便谷歌的规模扩大到今天的两倍，并占据消费者流量50%的市场份额，这依然能支持OpenAI、Anthropic或其他一两个参与者获得非常庞大的数字，仅仅因为我们在谈论的市场体量实在太大了。
So so I I just I look at it as just like the total size of the market and how that pie will likely be be divided, and and you can still have, you know, Google become two times bigger than they are today and have 50% of the market share from consumer traffic, and that would still support, you know, very large numbers from OpenAI or Anthropic or one or two other players in the space just because of the the sheer size and scale of of the market we're talking about.
现在我看着这些数字的规模，其中一个让我思考的问题是，我是否应该说，
Now I'm looking at the size of these numbers, and one of the questions that has come up for me is do I mean,
这里，就当是随便说说吧。
here here here here just just for fun.
就当是随便说说。
Just for fun.
你觉得，如果非要我让你说说看，你觉得摩根大通的市值是多少？
Do you think what do you think if if want me to put you on the spot, what do you think the market cap of JPMorgan is?
大概是10万亿美元，或者2万亿美元吧。
Let's say a $100,100,000,000,000, 200,000,000,000.
8400亿美元。
$840,000,000,000.
天啊。
Oh, man.
我真不好意思。
I'm embarrassed.
差得太远了。
Way off.
好吧。
Okay.
所以摩根大通的市值是8400亿美元。
So so the market cap of JPMorgan is $840,000,000,000.
我并不是在说这是公平的市值或不公平的市值。
And I'm not I'm not saying that that's a fair market cap or not a fair market cap.
所以，我对这个市值没有任何看法。
So so no no opinion on on on that market cap.
但你和我可以列出15家摩根大通的竞争对手，而我甚至不确定自己是否了解它们中的任何一家。
But you and I could list 15 competitors to JPMorgan, all of which I I don't even know if I do anything.
我对摩根大通没有任何投资。
I I I don't have any JPMorgan thing.
我想我可能有一笔通过摩根大通办理的汽车贷款之类的。
I I think I have maybe like a car loan or something that's through JPMorgan.
但我日常生活中根本不用摩根大通的服务，而它的市值却高达8400亿美元。
But, like, I don't use JPMorgan in my daily life, and they're worth $840,000,000,000.
如果你把其他所有银行都算上，你知道，你很快就会发现，仅仅在一个小类别里，总价值就已经达到万亿美元级别。
And if you take all of the other banks that that, you know, you you just are in the, you know, you're in the trillions of dollars very, very quickly across just one one little category.
所以，如果你在谈论整个经济体系中的智能配置，你可以在相当合理的情况下，迅速得出非常庞大的数字。
Now and so the the this is the like, if you're talking about intelligence across the entire economy, you can get to, you know, pretty large numbers in a in a pretty pretty reasonable way.
好的。
Okay.
你正好完美地引出了我正要问的问题。
You're setting up the question I was about to ask perfectly.
哦，也许我并不想这样。
Oh, maybe I didn't want to.
不。
No.
我觉得是的。
I think it is.
这真是个完美的铺垫。
It's a great setup.
你刚刚说明了我即将要问的关于这些数字规模的问题。
You've just illustrated what I'm gonna ask about the size of these numbers.
所以这些数字很大。
So the numbers are big.
是的。
Yeah.
我的问题是，投资者是否认为这一切都会是叠加性的？
And the question I have is, are the investors thinking that this is all gonna be additive?
或者可能的情况是，OpenAI之所以变得如此庞大，是因为它能够从摩根大通那里分走一部分市值。
Or maybe what happens is that, OpenAI is getting this big because it's able to take some of that, a little bit of market cap from a JPMorgan.
要知道，摩根大通业务的一大部分是为客户提供投资决策建议。
Know, a big part of JP Morgan's business is advising clients on making investment decisions.
你知道，如果我有一个Chat GPT的投资实例，是不是突然之间，部分市值就会流入OpenAI的市场缺口，更贴近我们自身的情况。
You know, if I have a chat GPT investment instance, know, is that is that all of a sudden some of that market cap is going into into the OpenAI market gap closer to home.
我们正处在SaaS行业的寒冬之中，对吧？
We're in the middle of the SaaS pocalypse, right?
现在有一种观点认为，AI将会吞噬掉许多软件公司目前正在做的事情，而市场在今年年初对软件公司确实非常不友好。
Where there's this belief that AI is going to just ingest lots of what software companies are doing right now and the market has really been unkind to software companies at the start of this year.
非常不友好。
Very unkind.
太不公平了。
Very unfair.
我觉得特朗普，真的很苛刻。
I feel like Trump, like, very unkind.
太不公平了。
So unfair.
说到这个，你能描述一下，如果这是叠加效应，会发生什么？如果这真是一种会吞噬经济大片领域的技术，又会发生什么？
You know, on that note, like, do so can you just sort of describe what you might think as what happens if this is additive versus what happens if this actually is a technology that will just gobble up big swaths of the economy?
我倾向于把它看作是经济的乘数，或者你可以把它看作是一种增效器，它会对经济征税，或者它会通过某种劳动力套利式的定价，从经济中抽取一定比例。
Well, I kind of think about it as a multiplier on the economy or a or, you know, kind of a maybe a maybe it you could either think about it as a force multiplier, and it it gets a it gets a tax on that, or it's a it takes a percentage of of the economy, you know, through through some sort of, you know, kind of, you know, labor arbitrage type type pricing.
但在我看来，整个经济中，有数万亿美元花在了知识型工作者身上。
But to me, I kinda look at it as, you know, tens of trillions of dollars are spent on on knowledge workers across the economy.
如果你能让所有知识型工作的生产率提升30%或50%，那么主要的实验室和围绕它们的应用程序，能否从中收取5%、10%左右的费用？
And and if you could, you know, add a 30 or 50% increase in productivity across all of knowledge work, could the major labs and the applications around that take up a 5%, you you know, 10% sort of fee on on on that?
这样，你就能理解为什么收入能达到数百亿甚至低万亿级别了，从数学上讲，这并非完全不合理。
That that that sort of like I think how you get to to to the math where revenue can get to the hundreds of billions or trillions, low low trillions, and it's not like entirely unreasonable just mathematically.
而你基本上就是在说，好吧。
And that's and you just are are basically saying, okay.
OpenAI 会分得其中一部分。
Well, OpenAI will will take part of that.
Anthropic 也会分得其中一部分。
Anthropic takes part of that.
谷歌也会分得其中一部分。
Google takes part of that.
一些应用层也会分得其中一部分。
Some of the application layer takes part of that.
但我认为，你可以通过多种方式实现这一点，比如广告实际上也可能帮你达成目标。
But I think that you can you know, there's a lot of ways you can get there, including actually just like advertising could probably get you there.
根本没有任何理由让你的AI服务无法通过更精准的广告投放，每年产生500亿到1000亿美元的收入，这本身就是另一种商业模式。
Like, there's just no reason that that that your AI service is not generating 50 to $100,000,000,000 just due to better performing hyper targeted advertising as another business model.
所以我认为，OpenAI 拥有多种商业模式叠加在一起，这些模式将随着时间推移不断创造更多机会。
So I think I think OpenAI has kind of these multiple business models stacked up that all that all will create, you know, more and more opportunity over time.
与此同时，你知道，五年后，他们在推理方面的规模将会扩大一百倍。
And at the same time, you know, in five years from now, both they will be, you know, a 100 times bigger in inference.
Anthropic在推理方面的规模也会扩大一百倍。
Anthropic will be a 100 times bigger in inference.
Gemini在推理方面的规模将扩大一百倍，依此类推。
Gemini will be a 100 times bigger in inference, and and so on.
而且推理的利润更高，这在一定程度上开始回答了一些问题。
And that inference is more profitable, which sort of starts to answer some of these questions.
推理最终会变得更有利可图。
The inference eventually gets more profitable.
我认为你现在正处于一种模式，我知道这听起来可能有点疯狂和泡沫化，而且，你知道，有一定概率我可能完全是在盲目乐观。
I think you're in a a mode right now, and I I I I know it it sort of is it'll it'll sound kinda crazy and bubbly, and, you know, there's a some percentage chance that I'm totally just drinking the Kool Aid.
但我认为，你现在正处于一个阶段，就是在进行基础设施建设，向世界普及人工智能。
But I think I think you're in a period right now where you're just in the infrastructure build out, teach the world about AI.
这在一定程度上值得补贴很多这些用例，因为这是最快找到真正价值所在的方法。
It's sort of worth subsidizing a lot of these use cases, because because it it's the it's the fastest path to figuring out where the actual value is going to be.
所以，虽然在某些情况下，初创公司或实验室会补贴编码等任务的令牌使用，但这确实是一种具有竞争力的明智之举，有助于获取市场份额、收集数据、构建飞轮、建立护城河。
And and so while, you know, there are some scenarios where you have a startup or a lab subsidizing tokens for coding or whatnot, It it is there it it it is like competitively a good move for, you know, gaining market share, getting getting data, building a flywheel, creating a moat.
这些在现阶段都是战略性举措。
Like, those are all strategic things to do at this stage.
就像优步当年不得不在各个地区以亏损的方式抢占市场，但随着时间推移，如今它已成为一个利润丰厚的业务，因为它已经建立了强大的网络效应，并牢牢占据了这些市场。
Similar to how Uber, you know, had to buy their way into to many markets unprofitably on a region basis, and then over time, you know, it's now a wildly profitable business because they now have obviously a very strong network effect, and they're kinda locked in to these markets.
我认为，这些前期资本支出巨大、现金消耗严重的业务，有时候本质上就要求这么做。
And I I I think I think some of these very kinda CapEx or or, you know, cash heavy businesses upfront, you know, sometimes just fundamentally require that.
对。
Right.
在继续之前，关于广告，我想补充一点。
One one note on the ads before we move on.
你提到广告可能成为一个年收入超过千亿美元的业务，而且
You were talking about how ads could be a hundred hundred plus billion dollar annual business and
我需要加一个巨大的星号，因为我还没深入研究过这一点。
Giant asterisk that I've I've not studied at once.
仅从Facebook和谷歌的业务规模来看，没有任何理由认为能够回答你任何问题的消费级智能产品，不能同样实现这种商业模式。
Just going off of the size of Facebook's and Google's businesses and saying there's just no reason that consumer grade intelligence, you know, that's answering any question for you wouldn't also deliver that type of business model as well.
是的。
Yeah.
所以OpenAI已经获得了大量关注，顺便说一句，Facebook上个季度的收入达到了600亿美元。
So OpenAI has gotten a lot of I mean, so Facebook, by the way, did $60,000,000,000 in the last quarter.
所以你所看到的数字，大致是那个数字的一半。
So this would basically the numbers you're looking at is like half half of that.
这周我和一位广告高管交谈时，发现了一个有趣的点：OpenAI的广告现在受到了很多批评，也许是有道理的，但其中一个有趣之处在于它非常注重个性化服务，因此他们收取高达60美元的千次展示费用，这简直不可思议。
And the one interesting I was speaking with an ad executive, this week, and one of the interesting things about OpenAI's advertising, now they've taken a lot of flack for it, maybe with good reason, one of the interesting things about it is it's so high touch and that's why they're charging like a $60 CPM, which is insane.
它非常注重个性化，真的会引导你完成整个流程，体验起来感觉非常好。
It's so high touch, it really guides you through a process, it feels seems like it feels good to go through.
如果你在考虑住宿地点，这种服务会很有帮助。
It's helpful if you're thinking about, like, staying somewhere.
长期以来，广告业面临的难题是，这种定制化且高接触的服务很难实现规模化。
The difficult thing with advertising over time is something that custom and that high touch has been really difficult to scale.
但有了人工智能，这种规模化的机会就出现了，于是你所说的那些数字也就不再荒谬了。
But with AI, that opportunity to scale it presents itself, and then all of a sudden these numbers that you're talking about aren't crazy.
是的。
Yeah.
不过，我和很多人在这个问题上的立场是相反的。
Well, I I'm on the other camp than versus a lot of people on this.
我认为广告在人工智能产品中可以非常有效。
I I think ads can be incredibly powerful in AI products.
我觉得，作为用户，你最终必须做出选择：你是想看到那些被SEO操纵的产品，还是想看到那些被经济模式操纵的产品？
I think that, you know, you you just kinda like you you sort of have to eventually decide as a user, do you want to see products that are kind of SEO hacked, or do you wanna see products that are kinda like marketplace economically hacked?
而能够最好地为你做广告的产品，可能反而是更好的产品，因为它们有明确的财务动机——只有当产品真正优质、运行良好时，才会吸引你访问他们的网站，否则你就会离开。
And and there's there's, you know, many reasons why the the products that can best advertise for you to you might be the better product because they have the they they they they have a very clear financial incentive only to get the you to their site if it's a good product and it works well, or else you're just gonna bail.
相比之下，SEO只是在大量网站上堆砌关键词，发布大量Reddit帖子。
And so versus, you know, SEO, we can just load a bunch of keywords across a whole bunch of sites and create lots of Reddit posts.
现在你看到的全是这些。
That's that's all you're seeing right now.
当你提出需求时，你看到的是某种公司正竭尽全力确保自己能出现在那个算法中。
When you ask for something, you're you're you're seeing some form of a company, you know you know, doing whatever it can to ensure that it's showing up inside that that algorithm.
因此，对我来说，这种商业模式并不会带来更差的结果，并不明显。
And so it's not obvious to me that that the marketplace model of that is is gonna, you know, give you worse results.
事实上，我认为没有任何实验室会因为广告而改变它给出的答案。
And I'm actually very you know, I I think I don't I don't think any lab would ever change the answer that it's giving based on advertising.
我认为它会直接给你答案，然后根据竞价系统推荐相关的内容。
I think it's gonna give you the answer, and then it's gonna give you related and recommended things from from, you know, from from the the bidding system.
我们希望如此。
We hope.
在我看来，这完全说得通。
To me, that kinda makes total sense.
这就像互联网过去二十五年运作的方式。
That's just like how the Internet has worked for twenty five years.
它为互联网上的大量产品创造了惊人的消费者剩余。
It's funded incredible consumer surplus of of products on the Internet.
这就是为什么我们有免费的搜索、免费的邮箱和免费的地图，同样地，没有任何理由认为这种模式不适用于消费级智能产品。
It's why we have free search and free email and free maps and like, there's just no reason that that would not apply to a consumer grade intelligence product as well.
确实如此。
Definitely.
不。
No.
我觉得这是一种非常有趣的方式来思考这个问题。
I think it could it's a very interesting way of thinking about it.
你说得对。
And you're right.
在这些系统中，你 anyway 都会看到推荐的产品。
You're gonna get recommended products anyway in these things.
所以，也许这是一个不错的信号。
So, you know, maybe maybe that's a that's a good signal.
人们希望相信这些系统中存在某种了不起的、真正的仲裁者，但实际上并没有。
People wanna believe that there's some kind of, like, you know, amazing truth arbitra like, arbiter in these in in these systems, and and they they're not.
我的意思是，它们和以前的搜索算法一样，完全受制于同样的机制。
I mean, it it's they are at the exact same mercy of a prior search algorithm would have been.
它只是从各种来源获取信号。
It's just taking signal from a variety of sources.
它正在尽力找出真正的答案。
It's doing its best to figure out what the what the real answer is.
如果在上面再叠加一个市场平台，我觉得这根本不是世界末日，而且你会在过程中获得很多不错的推荐。
And if you also have a marketplace layered layered on top of that, it's just not I just don't think it's the end of the world, and I think you'll actually get a lot of good recommendations along the way.
人们会愿意付费来屏蔽广告，这会带来更多的收入。
And and people will then pay to not see the ads, and that'll be even more revenue.
所以，如果你是一家达到这种规模的AI公司，这简直就是一种非常好的赚钱方式。
So there's just like it's like a very good way to make money if you're an AI company at that scale.
我的意思是，我觉得这只有对两到三家公司才相关，但OpenAI就是其中之一。
Not I mean, I I only think it's relevant for two or three companies, but OpenAI is one of those.
是的。
Yeah.
他们将拥有十亿用户，或者现在可能已经拥有十亿了。
They'll have a billion users or they might already have a billion now.
好的。
Okay.
在我们继续谈融资之前，有一件事一直让我困惑，我想听听你的看法。
So before we move on from the fundraising thing, there's one thing that has puzzled me throughout, and I need to ask you what your thoughts are here.
所以，OpenAI 和英伟达宣布了一项高达一千亿美元的融资计划，由英伟达分批注资，每次一百亿美元，但后来似乎杰ensen 在退缩。
So, OpenAI and Nvidia announced this $100,000,000,000 funding that was gonna come in from Nvidia to OpenAI, $10,000,000,000 at a time and, and then, it seems like Jensen was backing away from that.
《华尔街日报》曾报道这笔交易已搁置，而本周《金融时报》的报道显示，英伟达确实会投资 OpenAI，但金额将从一千亿美元下调至三千亿美元。
There was a Wall Street Journal article saying that the deal was on ice, and we found out this week from reporting from the Financial Times that NVIDIA is going to invest in OpenAI, but it's going to be $30,000,000,000 and not a $100,000,000,000.
此前有报道称，Jensen 对 OpenAI 的发展轨迹感到不满，而且他在谈及此事时的语气与最初的新闻稿截然不同——新闻稿最初说的是‘我们希望他们邀请我们投资’，而现在却变成了‘我们打算投资’，这两种说法差异很大。
Now there were these reports that Jensen was not happy with OpenAI's trajectory and and all of that, and and he seemed like when he was talking about it very different from the original press releases saying we hope they'll invite us to invest as opposed to we intend to invest those two very different ways of talking about it.
所以，我想弄清楚，Aaron，我该怎么理解这件事？
So I'm trying to figure out Aaron, how do I think about this?
因为一方面，如果这笔新交易取代了原计划，那就意味着少了七千亿美元。
Because on one hand they are so this is if this deal replaces it, that's $70,000,000,000 less.
我的意思是，如果你比预期少了七千亿，那确实不好，但据报道他们仍然投入了三千亿，这已经是笔巨款了。
I mean if you if you get $70,000,000,000 less than you anticipated that's bad however they're still putting in $30,000,000,000 reportedly that's a lot of money.
你觉得双方的关系现在处于什么状态？我们应该如何解读这个数字，以及它对最初百亿投资的替代？
Where do you think, where do you think the relationship stands and how are we how should we read the number and the replacement of the initial 100?
哦，我的意思是，这简直就像在搞占星术一样。
Oh, I mean, this is, this is like full astrology on
是的。
Yes.
这是占星术。
Is astrology.
是的。
Yes.
我们简直是在为人工智能行业算手相。
And we're doing palm reading for for the AI industry.
首先，他们有没有明确说过，他们打算在下一轮投资，还是只是打算在某个不确定的时间点投资一千亿美元？
I, you know, I I I first of all, I did did they say that they intended to invest in the very next round, or they intend to invest a 100,000,000,000 at at in some arbitrary point in time
随着时间的推移？
over time?
这本来就是分阶段进行的。
It was over time.
从来就不是一次性完成的一轮投资。
It was never one round.
所以我不确定。
So I I I don't know.
我只是像其他人一样，基于所有已知事实来判断，但我就是对这件事的戏剧性一面提不起兴趣。
I'm gonna just I like I I'm I'm taking all the facts that in the same way everybody else is, and but I just don't have the impulse for the drama side of this.
显然，英伟达希望与OpenAI建立稳固的公司合作关系。
It's, you know, Nvidia obviously wants a very strong corporate relationship with OpenAI.
OpenAI显然也希望优先获得芯片供应。
OpenAI obviously wants to be able to be first in line for for chips.
双方都有很强的动力让彼此都取得巨大成功。
They have a lot of incentive to both make each make each other very successful.
这对双方来说都是个好事，因为整个领域如果持续增长的话。
It's it's like a it's a it's a boon for both of them if if the whole the whole space, you know, keeps growing.
同时，可能还有很多配置上的动态因素，比如英伟达需要考虑投资多少，而百模（Bo Binai）在考虑其整体股权结构和各公司持股比例时也需要权衡。
And at the same time, there's probably a lot of configuration dynamics that, you know, that both NVIDIA has to consider on how much to invest, and that Bo Binai has to consider when they think about, you know, their total cap table and what companies own what percentage of them.
所以，我不得不说，这个答案很无聊，只是因为我感觉，看贾森在街头采访中的 viral 视频还挺有趣的。
So I I you know, it's a very boring answer only because I I think it's like it's it's it's like fun to kinda watch the viral video and of, you know, Jensen in the street interview.
但我其实不太在意这些事。
But like, I I just might like, I I I kind of don't worry about it too much.
我只是觉得，这个领域变化太快了，我完全可以想象，很多配置结果可能和六个月前的初衷，或者律师在新闻稿中设定的条款有所不同。
I just think, like, this space is is changing so quickly that I can imagine many different reasons why some configuration might end up different from, you know, where its intent was six months ago or where the lawyers decided to, you know, kind of put put certain terms in the in the press release.
是的。
Yeah.
我的观点是，我认为贾森确实希望 OpenAI 能成功。
My my hot take here is that this is all I think Jensen does want OpenAI to succeed.
很明显，这是他们和谷歌之间的竞争。
Obviously, it's them versus Google.
我认为这件事本质上是他给他们的一个信号：你们必须表现更好，不要再有红色警报，只要保持领先就行。
I think this whole thing was basically a signal from him to them, you better perform and no more Code Reds and just just stay ahead.
我唯一不同的看法是，我觉得OpenAI在融资方面根本没什么困难。
I I you know, the only the only my only counter take to that is I just don't think that OpenAI has a challenge raising money.
所以我不确定是否真的存在来自资本结构方面的压力能施加在他们身上。
So I don't know that I I don't know that there's sort of some kind of pressure that can be exerted on them from from the cap table side.
没错。
Right.
我认为这更像是一个动态的市场，人们只是在审视自己的资本配置决策、估值，以及是否有其他获取资金的途径等等。
I think it's I think it's a bit more of a fluid market, and and it's just people looking at their capital allocation decisions, looking at valuations, looking at, you know, do you have other sources of ways of of getting the capital, etcetera.
你从NVIDIA的角度想一想，他们其实并不需要持有OpenAI的股份。
Like, you think about it from NVIDIA's standpoint for one second, like, they don't they don't need to own a percentage of of OpenAI.
对他们来说，真正需要的是向OpenAI销售芯片。
Like, that's like, they need to sell chips to OpenAI.
所以，他们真正需要做的，就是确保拥有一个非常稳固、有力的关系，以支持人工智能领域的整体利好趋势。
And so and so really, they they just need to ensure that they've got a very strong, you know, relationship that is is sort of very sturdy and and supporting the broad tailwinds of AI.
我不知道有没有一个具体的数字，比如如果软银想要获得更多配额，我这完全是瞎猜的。
And I don't know that there's a number that like, if it turns out SoftBank wants to take more of the allocation I'm making all of this up.
但如果软银确实想获得更多配额，我不认为这会在战略上对他们产生重大影响。
But if it turns out SoftBank wants to take more of the allocation, I don't know that they're, like, strategically impacted by that in in in a meaningful way.
因为即使他们持有更多OpenAI的股份，我认为这种股权结构并不会过度影响OpenAI的基础设施决策。
Because if they if they own more of OpenAI, I don't think that that position in the cap table is going to overly sway the infrastructure decisions of OpenAI.
OpenAI的基础设施决策必须基于芯片的供应端、成本端，以及他们在哪里拥有数据中心容量等因素。
In OpenAI will have to make their infrastructure decisions based on just, like like like, the supply side of of chips, the the cost side, you know, where where do they have data center capacity?
这些因素的重要性会超过谁在他们的公司结构中持有多少百分比的股份。
Those things are gonna matter more than who owns a certain percentage of their of their, you know, corporate structure.
我对此的反驳是，面对如此巨大的数字，他们能筹集的资金已经所剩无几，而 NVIDIA 拥有四万亿美元的市值，并且有可观的收入，正是这样的潜在资金来源之一。
My counter to that would be with numbers this big, there's only a certain amount of money left for them to raise, and NVIDIA at $4,000,000,000,000 with, you know, sizable revenues is one of those Yeah.
潜在资金来源。
Potential sources.
我不确定。
I don't know.
有些国家拥有大量
There's countries with lots
资金。
of money.
是的。
Yes.
拥有。
Have.
我们即将看到他们参与进来。
We're we're about to see them get involved.
所以
So
而这些地方希望将资金投入到未来的经济活动中。
And those those places want want to deploy money in future economic, you know, activities.
所以
So
是的。
Yes.
好吧，我们这一轮将是科技巨头轮，然后是海湾国家轮第一轮、海湾国家轮第二轮，接着是IPO。
Well, we have we definitely have we'll have this round, which is gonna be the Tech Giants round, then we'll have the Gulf State Round number one, the Gulf State Round number two, and then IPO.
这很可能就是事情的发展方式。
It's probably what the way it will play out.
但愿如此，老天保佑。
From your lips and God's ears.
说到其他国家，整个AI行业本周都来到了印度，参加印度AI峰会，并且从那里传出了一些非常大胆的言论。
So, speaking of other countries, the entire AI industry made their way to India this week, for the India AI Summit and some really bold statements coming out of there.
所以，我们来玩一个我们这个节目偶尔会玩的游戏，叫‘炒作还是真实’。
So let's play a game, that we play on this show every now and again called, hype or true.
这些声明是炒作还是真实？
Is this are these statements hype or are these statements true?
我们有一个来自萨姆·阿尔特曼的。
We got one from Sam Altman.
按照我们目前的发展轨迹，我们相信可能只需几年时间就能实现早期版本的真正超级智能。
On our current trajectory, we believe we may only be a couple of years away from early versions of true superintelligence.
如果我们是对的，到2028年，世界上大部分智力能力可能会存在于数据中心内部，而非外部？
If we're right, by the 2028, most of the world's intellectual capacity could reside inside of data centers than outside of them?
你怎么看？
What do you think?
要知道，你接下来要说的每件事可能都会受到条件限制，对吧。
You know, all probably every one of the things you're about to say are going to be conditioned on, you know Right.
首先得定义一下我们讨论的到底是什么东西。
One definition of what is the thing that is being talked about.
但我认为，考虑到我们目前的发展轨迹，这似乎完全合理。
But I think that that there's kinda that that seems to be totally reasonable based on the trajectory that we're on.
而且我敢打赌，山姆对于'智能'或类似概念的定义标准，甚至比我还要高得多。
And I would bet that that Sam has an even a far higher bar for for what his definition of, you know, intellectual or whatever the the the term was than even I would.
比如，我认为，即使是现有的最新一轮模型，只要配上合适的AI工具套件，加上恰当的框架和合适的人员参与，我们就能从这些系统中榨取出相当大一部分有价值的工作。
Like, think like I I think already with things like the latest round of models with the right kind of AI harness, we we could squeeze out a significant portion of of valuable work from these systems with the right scaffolding and the right kind of people being involved.
所以我认为，基于他所说的话，这个说法非常合理。
So I I think that that is a very reasonable statement based on what he's saying.
这可能和扬·莱库恩对智能的定义不同，他可能会把智能定义为：一个系统能否在仅十分钟的训练后就能开车。
That might be different than what, like, Jan Lecun would say is the definition of of intelligence, where where he would probably define it as can the thing drive a car, you know, with only ten minutes of training.
我并没有那种更偏向生物学的智能定义。
And I just don't I don't have that same kind of more biological definition of intelligence.
所以这就是为什么我认为萨姆的说法非常合理。
I like like, you know, so so that's why I think Sam's statement is very reasonable.
这是达里奥。
Here's Dario.
过去十年，人工智能一直呈指数级发展。
AI has been exponential for the last ten years.
AI模型在大多数方面超越人类认知能力的剩余时间已经不多了。
There are only a small number of years left for AI models surpassing the cognitive capabilities of most humans for most things.
我想这算是一个类似的说法。
I guess that's a similar statement.
是的，我觉得确实如此。
Yeah, I think so true.
同样的回答。
Same answer.
对。
Yeah.
好。
Yeah.
在印度峰会上发生了一个有趣的时刻。
Interesting moment happened at this India summit.
我相信你已经看到了。
I'm sure you've seen it.
他们让所有首席执行官们上台，要求拍照时手拉手、举起手臂。
They have all the CEOs up there on stage and instructed for a photo to lock hands and raise their arms.
而萨姆和达里奥，他们似乎不太喜欢彼此，却反而
And Sam and Dario, who don't seem to like each other very much, instead
等等，我不是...我看了好几遍那个视频。
of Didn't it didn't I I've watched the video a couple times.
你不觉得那可能有点即兴发挥吗？
Didn't it feel like maybe it was a little impromptu?
或者，或者你认为那是被指示的？有报道说是被指示的吗？
Or or do you think that was instruct is it reported that it was instructed?
我不...我不太确定，我只是根据当时的协调情况做了一个假设。
I don't I don't so I was making a assumption on the coordination of it.
也许那是即兴的。
Maybe it was impromptu.
也许中间的莫迪只是那样做了，然后大家都跟着做了。
Maybe Modi at the middle was just like, and then everybody followed it.
看到一些视频，感觉好像没人真的知道该做什么。
Saw some videos where it kinda felt like nobody really knew what to do.
确实如此。
And and That's true.
是的。
Yeah.
而且他们似乎都在摸索着应对，因为当时有个瞬间，亚历克斯得去抓我的手。
And and and they were kinda like just all figuring out because you have this moment where like Alex had to grab I'm gonna hand.
是的。
Yeah.
而且他们因为看起来并不是每个人都清楚该如何协调这件事。
And and they because and it seemed like like not everybody quite knew how to coordinate this.
所以，也许我们只是暂时出现了混乱，等反应过来的时候，已经太晚了，根本没法再手拉手了。
So so so so you might have maybe we just maybe they just malfunctioned for a minute, and and then by the time it was too late, it was just like, I we can't we can't hold each other's hands.
所以谁知道呢？
So who who knows?
我的意思是，对。
I mean, yeah.
听好了，关键是，我们或许可以在未来的某一集中回放这段视频，然后
Look, the point is, the point we could we we could we should we can, maybe in a future episode, play the, video back and go
好的。
Okay.
进行实时解说。
Do the play by play.
但关键是，所有人都似乎明白了，只有萨姆和达里奥没明白。
But the point is everybody seemed to figure it out except for Sam and Dario.
明白了吗？
Alright?
他们俩双手高举，紧紧攥着
They They had their hands in the air clenched with one
像另一只手一样。
like the other.
龙虾手。
Lobster hands.
好了。
Alright.
对。
Right.
是的。
Yeah.
他们用Photoshop把钳子手P到他身上了。
They did photoshop the claw hands on onto him.
是的。
Yeah.
我有个问题想问你。
Question for you about this.
这两个家伙连怎么尊重彼此的差异都搞不定，如果他们连拍张照握手都做不到，我们能信任他们来处理AI对齐问题吗？
Can can these two guys who can't figure out a way I respect their differences, but if they can't figure out a way to hold hands for a picture, should we trust them to handle AI alignment?
这确实是一个非常非常深刻的元问题。
It's a it's a it's a that's a very it's a very great meta question on that.
有人写过这篇文章了吗？
Has anybody written that piece yet?
不。
No.
我的意思是，这本应该是本周最大的科技新闻。
I mean, I that really should have been the big technology story this week.
我的意思是，写一下这篇文章。
I mean, write write that piece.
我觉得这是我们面临的一个很棒的难题，这是一个更广泛问题的小小缩影。
I I I think it's a it's a it's a great conundrum that we face that is this great little micro, you know, microcosm of of of a broader issue.
但确实如此。
But yeah.
我的意思是，我花了很多钱来了解他们两人对牵手这件事的看法。
I I don't I mean, you know, I pay a lot of money to get both of their takes on on the hand thing.
你知道，有时候你会和对手陷入激烈的争执，人们在公开场合说了太多话，就像你到了这样一个地步：关系太过戏剧化，需要某种中立的平台来让一切重新和解。
You know, sometimes you get into these heated battles with a rival where people are just saying too many things in public and and and it's just like, know, you get to this point where it's just the the relationship is is too dramatic and there needs to be some kind of, you know, kind of neutral ground that brings everything back together.
也许有人会认为印度本该这么做。
Maybe one would have thought India would have would have done that.
但我还是相当有信心，我们会度过这个手部问题的难关，他们总能有办法修复彼此的关系。
But I I I have kind of full faith that we will get through, you know, hand hand issues and and and they can repair the the relationship somehow.
是的。
Yeah.
我也希望如此。
I I I hope so.
我的意思是，我觉得如果你现在去问他们中的任何一个，他们都会说，我本该就那样握住那只手，然后……
I mean, I think if you asked either of them right now, they would have just said I would I should have just held the hand and it
本来会
would have
成为整个事件中的那个梗
been as that became the meme out of the whole
事情。
thing.
我不
I don't
我认为他们并不想让那次握手成为峰会的焦点。
think they meant for that to be the takeaway from the summit.
所以，他们各自发表了大约二十分钟的演讲，内容是关于……是的，我不认为握手是预想中的重点。
So they they had they had, like, twenty minute speeches about the and and, yeah, I don't think the hand was meant to be the takeaway.
有趣的是，当你把所有这些AI领域的领导者聚在一起时，有时就只产生了一个绝佳的梗。
It is funny how you get in all these AI leaders together, and sometimes there's just one great meme.
就是这么回事。
There's that.
达里奥和德米斯坐在小沙发上，这是我最喜欢的场景之一。
There's Dario and Demis on the small couch, which is one of my favorites.
还有康奇黑德戴着帽子。
And hats at Conchhead.
所以，实际上在模型前沿方面，这是一个非常有趣的发展。
So very interesting, development actually on the model front.
我们之前已经暗示过了。
We hinted it before.
展开剩余字幕（还有 268 条）
Anthropic 推出了一个新的大型模型 SONNET 4.6，你说这是对最近模型 4.5 的一次重大升级。
Anthropic has a new big model, SONNET 4.6, and you've said that it is a major upgrade over the most recent model 4.5.
我们通常认为这些个位数版本的模型是渐进式更新，但你在复杂工作评估中分享的数据却相当显著，在性能和准确性上，4.5 到 4.6 之间有 15 个百分点的跃升。
We usually expect these single digit models to be incremental updates, but the stats that you shared on your, evaluation for complex work are pretty pretty significant where there's been a 15 percentage point jump in performance and accuracy, you know, between four point five and four point six.
这是你在 Twitter，或者说 X 上发布的。
This is about from you, on Twitter, or X, shall we say.
在公共部门，你看到复杂任务的准确性从 77% 跃升至 88%。
In the public sector, you saw a jump from 77 to 88% in accuracy for complex tasks.
医疗保健领域的复杂任务准确率从60%跃升至78%，法律领域的准确率则从57%提升至69%。
Healthcare saw a jump from 60 to 78%, and legal saw a jump from 57 to 69% accuracy on complex tasks.
这确实是非常、非常大的进步。
That's pretty, pretty big.
看起来这个模型的热度似乎被低估了。
It seems like this model has has almost been underhyped.
你能稍微谈谈这些提升以及它们的重要意义吗？
Can you talk a little bit about these these jumps and what the significance I
我认为，或许主要的启示应该是，过去几年我们在AI编程领域所见证的这些有意义的飞跃式进步。是的。
I think I think probably the the main takeaway should be that that the progress of these meaningful jumps that we've been seeing in AI coding over the past couple of years, where Mhmm.
你知道，大约两年半前，在编程领域，模型最多只能在类型提示格式下生成几行代码，而现在显然人们可以交给模型一个任务，让它为一个完整项目编写数万行代码，我们刚刚见证了这种惊人的进步速度，以及编码能力随时间不断向上提升的进程。
You know, the model at best could do a couple lines of code, in a type ahead type format two, two and a half years ago in coding space, and now obviously people are giving the model a task of write me tens of thousands of lines of code for a full project, and and we've just seen this incredible rate of progress and this march up toward, you know, more and more capability over time with with with encoding.
我认为同样的趋势将会出现在其他知识工作领域，因此SONNET模型从4.5到4.6的这次飞跃，我认为正是一个例证，展示了当这些模型在更多知识工作领域得到训练时会发生什么。
I think that same trend is gonna come to other other now fields of knowledge work, and so so this jump in SONET's model from four or five to four six, think represents an example of what happens when these models just get trained across more areas of knowledge work.
当它们在超越编码的推理能力上变得越来越强时，会发生什么？
What happens when they are getting better and better at reasoning capabilities that go beyond coding?
当它们在工具使用和决策何时使用工具方面变得更强时，会发生什么？这正是我们复杂工作评估所要体现的——它如何思考问题、如何确定自己得到了正确答案、如何检查自己的工作。这些模型在实现这些能力方面正变得越来越出色。
What happens when they get better at using tools and deciding when to use tools, and that's what our complex work eval, you know, is meant to represent is sort of how does it think through a problem, how does it decide it's got the right answer, how does it check its work, and these models are getting much better at at being able to deliver on that.
所以我认为这将是未来几年的趋势，即使在我们自己的评估中，我认为我们正处于知识工作者类型评估的最早期阶段之一。
So I think that'll be the trend for the next couple of years, and even for our own eval, I think we're looking at one of the earliest phases of a knowledge worker type eval.
我认为我们很快将不得不让评估变得越来越难，以更好地体现这些模型的能力。不过，是的，这些性能跃升显然非常引人注目。
Think we're gonna have to make it harder and harder to better represent the capabilities of these models soon, but yeah, these jumps are obviously very eye opening.
你知道，我们稍后回来会深入探讨人工智能在下半年将如何工作，但围绕云服务发生的一件有趣事情是，Anthropic与五角大楼之间就Claude的使用问题出现了一些戏剧性事件。具体来说，是关于五角大楼使用Claude的情况，有报道称五角大楼显然使用了Claude来协调其对委内瑞拉的攻击。
You know, we're gonna get a little bit into how AI will do work in second half when we come back, but one of the interesting things that's been happening around cloud is there's been this drama between Anthropic and the Pentagon about its use of of Claude and this like, the Pentagon's use of Claude and there was this story that came out that apparently the Pentagon used, Claude in its to coordinate its, attack on Venezuela.
这是来自前用户托尼·切夫林的分享。
This is from ex user Tony Chevlin.
克劳德竟被传参与了委内瑞拉总统的直升机撤离行动，却没人质疑‘克劳德怎么帮得上忙？’，这真是对它的极大肯定。
Such a compliment to Claude that amid rumors it was used in a helicopter extraction of the Venezuelan president, nobody's even asking, wait, how can Claude help with that?
当然，它肯定是有用的。
Of course of course it was useful.
你难道会不用克劳德吗？
Like, how would you not have used Clive?
其实这挺搞笑的，两年前，这句话听起来简直让人匪夷所思。
It is actually a very funny like, two years ago, that sentence would have been like, excuse me?
你到底会怎么想？这事儿到底该是什么样子？
What what do you how would this have like, what would the thing have been?
而现在，大家觉得理所当然：他们肯定用了某种智能工具来规划或分析事情，或者整合数据，这种能力正越来越被默认融入到更复杂的工作和软件中。
And now it's just like, yeah, I'm sure they use some kind of intelligence to to plan something or figure something out or, you know, correlate data, and that's just sort of priced into, I think, more and more complex work and and and and software.
太疯狂了。
Wild.
好的。
Okay.
所以我们还有很多要聊的。
So we still have so much to talk about.
我们有OpenClaw。
We have OpenClaw.
我们还有这些关于生产力的新研究。
We have these new studies on productivity.
我们稍后回来就做这件事。
Let's do that when we come back right after this.
开始新事物不仅仅是困难。
Starting something new isn't just hard.
它是令人恐惧的。
It's terrifying.
你投入了大量工作在这件你并不完全确定能否成功的事情上，要迈出这信仰的一跃可能很难。
So much work goes into this thing that you're not entirely sure will work out, and it can be hard to make that leap of faith.
当我开始做这个播客时，我不确定是否有人会听。
When I started this podcast, I wasn't sure if anyone would listen.
现在我知道这是个正确的选择。
Now I know it was the right choice.
当你拥有像Shopify这样的合作伙伴在身旁提供帮助时，事情也会变得更容易。
It also helps when you have a partner like Shopify on your side to help.
Shopify是全球数百万企业背后的商业平台，支撑着美国10%的电子商务交易。
Shopify is the commerce platform behind millions of businesses around the world and 10% of all ecommerce in The US.
从像Allbirds和Cotopaxi这样的家喻户晓的品牌，到刚刚起步的品牌。
From household names like Allbirds and Cotopaxi to brands just getting started.
凭借数百个即用型模板，Shopify帮助您打造一个与您品牌风格相匹配的漂亮在线商店。
With hundreds of ready to use templates, Shopify helps you build a beautiful online store that matches your brand style.
您还可以像拥有一个营销团队在背后支持一样进行推广，轻松创建电子邮件和社交媒体营销活动，无论您的客户是在滑动屏幕还是在漫步。
You can also get the word out like you have a marketing team behind you, easily create email and social media campaigns wherever your customers are scrolling or strolling.
是时候在今天，通过Shopify将那些'如果'变为现实了。
It's time to turn those what ifs into with Shopify today.
前往 shopify.com/bigtech 注册每月1美元的试用。
Sign up for your $1 per month trial at shopify.com/bigtech.
访问 shopify.com/bigtech。
Go to shopify.com/bigtech.
就是 shopify.com/bigtech。
That's shopify.com/bigtech.
问题是这样的。
Here is the problem.
你的数据到处都暴露了。
Your data is exposed everywhere.
个人数据散落在数百个网站上，往往未经你的同意，这意味着数据经纪人买卖你的信息——你的地址、电话号码、电子邮件、社会安全号码，这种暴露会带来真实的风险。
Personal data is scattered across hundreds of websites often without your consent, And that means that data brokers buy and sell your information, your address, phone number, email, social security number, and that exposure leads to real risks.
比如身份盗窃、诈骗、骚扰、更高的保险费率。
Things like identity theft, scams, harassment, higher insurance rates.
Incogni 会追踪并从数据经纪人、目录、人物搜索网站和商业数据库中删除你的个人信息。
Incogni tracks down and removes your personal data from data brokers, directories, people search sites, and commercial databases.
以下是它的工作原理。
Here's how it works.
首先，您创建账户并分享定位您个人资料所需的最少信息。
First, you create your account and share minimal information needed to locate your profiles.
其次，您授权Incogni代表您联系数据经纪人。
Second, you authorize Incogni to contact data brokers on your behalf.
第三，Incogni将通过自动处理数百家经纪人以及客户手动请求的方式，移除您的数据。
Third, that Incogni will remove your data both automatically with hundreds of brokers and via customer removals.
我们还提供三十天退款保证。
There's also a thirty day money back guarantee.
用Incogni夺回你的个人数据。
Take back your personal data with Incogni.
访问 incogni.com/bigtechpod 并在结账时使用优惠码 big tech pod。
Go to incogni.com/bigtechpod and use code big tech pod at checkout.
我们的优惠码可让你享受年度计划六折优惠。
Our code will get you 60% off an annual plan.
去了解一下吧。
Go check it out.
我们回到《大科技播客》，今天我们邀请到Box公司的首席执行官亚伦·莱维。
And we're back here on Big Technology Podcast with Box CEO, Aaron Levy.
亚伦，每次你来都太棒了，我相信你会非常喜欢接下来的环节，因为你一直密切关注这件事，我们非常期待听到你的见解。
Aaron, it's always great to have you here, and I think you're really gonna enjoy this next segment because this is something that you've been following very closely, and it's gonna be great to get your perspective on it.
当OpenClaw被OpenAI收购时，我就说我们必须请亚伦来节目中谈谈他的看法。
When when OpenClaw sold to OpenAI, I said we gotta get Aaron on the show for his perspective on this.
这是来自CNBC的报道：OpenClaw的创始人彼得·斯坦伯格加入OpenAI。
So this is from CNBC, Open OpenClaw creator Peter Steinberger joins, OpenAI.
病毒式AI代理OpenClaw的创始人将加入OpenAI，该服务将以开源项目的形式隶属于一个基金会，由OpenAI继续提供支持，萨姆·阿尔特曼表示。
The creator of the viral AI agent OpenClaw is joining OpenAI, and the service will live in a foundation as an open source project that OpenAI will continue to support, Sam Altman said.
他表示，斯坦伯格将加入OpenAI，推动下一代个人代理的发展。
He said that it's that Steinberger is is gonna join OpenAI to drive the next generation of personal agents.
所以我们很想听听你的看法，简单介绍一下OpenClaw是什么，因为经常回顾一下总是有帮助的。
So we'd love to get your perspective here just on, you know, a little bit about very briefly what OpenClaw is because it's always good to sort of refresh there.
那么，OpenAI收购它或聘请Steinberger有什么重要意义呢？
And then why is it significant that OpenAI either acquired it or brought Steinberger aboard?
是的。
Yeah.
所以我认为，Steinberger在OpenClaw上带来的创新在于，过去几年里虽然有过各种尝试，但真正实现可能只在最近几个月模型能力提升之后才成为可能。
So so I think the the innovation that that Steinberger kinda created with OpenCLW was and there's been various attempts at this, you know, obviously, over the past couple of years, but but I think it was only really possible in probably the last couple of months of of model capability.
但关键的突破是，我们现在有了这些能够代表我们行动的智能体，我们控制并引导它们为我们完成任务。
But the the big jump is, you know, we we have these agents that effectively act on behalf of us, and we are controlling it and steering it to go do tasks for us.
比如Claude Code，你可以在终端里输入指令，让它生成一些代码，它就会去执行，完成后回来等待你的下一个任务。
So Claude code, you you kinda, you know, type in your terminal, you tell it to generate some some code, and it goes off and does work, and comes back, and and it's waiting for its next task for for you to give it.
或者在UI界面中，你告诉它去生成代码，比如在Dev和Factory这类智能体中。
Or codecs, you're you're in a UI telling it to to go and and generate some code for you, dev and factory, you know, all these kind of agents.
在过去大约一年左右的时间里，这基本上一直是智能体技术的前沿状态。
And that's that's basically been the state of the art of of agents for for the past, you know, year or so, plus plus or minus.
而OpenClaw则借鉴了许多相同的原则，但提出一个问题：如果这个智能体能够自主运行，拥有访问你的电脑、浏览器以及你使用的所有服务的权限，并且持续不断地运行，会怎样？
And OpenClaw kinda took, you know, many of the same principles, but said, well, what if that agent is sort of running on its own, and it had access to your computer and your browser and all the services that you use, and and it's it's just literally running on an ongoing basis.
你可以和它聊天，向它下达任务，它也可以主动向你推送相关信息。
And it and you you you chat with it, and you can ask it to do things, but it can also ping you as as a sort of relevant.
这是一种非常新颖的代理思维模式，虽然我们之前也见过一些例子，但还没有任何东西像OpenClaw那样取得如此广泛的普及。
And that was that was this this is sort of a a very new kind of way to think about agents that, again, we've seen examples of, but nothing obviously that has taken off at the at the level that OpenClaw did.
这让你稍微窥见了未来的可能性：你不再只是在需要时才启动或关闭这些代理，而是拥有一个真正持续运行、始终为你执行任务的代理。
And and it gives you a little bit of a peek into what the future, you know, could be, where you don't you don't have these agents that you only sort of spin up and spin down as you need them to do work for you, but you have a actually an agent that's sort of always always on, kinda working working for you and executing tasks for you.
这就是为什么人们正在为这些代理设置独立的电脑。
That's why people are setting up, you know, their own separate computers for these agents.
它们可以在自己的环境中持续运行。
They can just keep running off in their own environment.
很难确切知道如何完整地打包这种系统，以及如何让它以一种真正简单、安全的方式呈现给用户，完全安全可靠。
And and and you know, hard hard to know exactly how you fully would package that up and how it could manifest in a way that would be really, really simple for people to use and and fully secure full fully yeah.
没错。
Exactly.
对于那些不太熟悉这些系统的人来说，要确保安全可靠。
Safe and secure for people that don't kinda know their way around all these systems.
那里还有很多需要弄清楚的地方，但其实和我所想的主版本更新或范式更新并没有太大不同。
Lots to figure out there, but but not that different from, you know, what I what I think about it is like a, you know, a principal update or a a paradigm update.
我记得德文的那个病毒式视频，应该是两年前的事了，我不太记得所有细节了，他们是发了Slack消息还是在界面里操作，但你只要告诉德文去干活，就能看到它在行动。
You know, I remember the viral video of of of Devin, must be two years ago now, and, you know, I don't remember exactly all the details if they if they did a Slack message or if they were in the UI, but you you kinda told Devin to go off and do work, you could just see it.
它正在生成自己的代码。
It's producing its code.
它还有一个环境，你可以看到它在构建什么，你知道，他们当时让很多人觉得，哦，这根本不可能成功。
It it had another environment where you could see what what it was building, And, you know, they got they got you know, I think there are a lot of people that were like, oh, this will never work.
这怎么可能实现呢？
How could this possibly work?
它实际上并没有在做那些事。
It's not actually doing that.
当时有很多非信徒发布的病毒式反驳视频。
And there were these viral takedowns from nonbelievers.
但对一些深入人工智能领域的人来说，我们当时心想：糟了。
And but but for for for some people who were deep in the AI space, we were like, oh, shoot.
这完全是另一种思考如何与智能体协作的方式。
Like, that is a very different way to think about, you know, working with an agent.
你并不在IDE里。
You're not in an IDE.
你也不是和它一起写代码。
You're not coding alongside it.
你只是下达一个任务，然后它会为你完成大量工作。
You're just sending off a task, and it's gonna go and do a bunch of work for you.
现在很明显，这正是我们将要进入的主导范式。
And now, obviously, it's very clear that that's the dominant paradigm that we're going to be in.
Codex已经证明了这一点，Cloud Code也证明了，Devin和Factory同样证明了。
Codex has proven it, you know, Cloud Code has proven it, Devin and Factory have proven it.
我想Cursor在智能体方向上押注更多。
You know, I I assume Cursor is betting even more on agents.
你可以明显看到他们更侧重于智能体这一端的用户体验，而不是IDE端。
You can kinda see them them pushing more on the agent side of the user experience as opposed to the IDE side.
所以，这是几年前我们得到的一个更新。
So so that was an update that we got a couple of years ago.
我认为我们现在也会在其他知识工作领域看到同样的情况，开放的，你知道，OpenClaw 引入了一种有趣的范式，这种范式可能会持续应用于越来越多的工作领域。
And I think we're gonna see the same thing now in other areas of knowledge work and and open and and, you know, OpenClaw introduces an interesting kind of paradigm that that could that that could persist across, you know, more and more areas of work.
对。
Right.
现在作为一名软件公司的首席执行官，我真的很想听听你对这意味什么的看法。
And now as a software CEO, I really would love to hear your perspective on what this means for software.
我先提供一些背景信息。
I'll just give some context here.
你知道，过去大概一周半的时间，我一直埋头在 Cloud Code 里。
You know, I've spent the past, I guess, week and a half now just like with my nose in Cloud Code.
我一开始只是想，你能帮我做一个类似电子表格的软件吗？就是当我填完一个字段时，它能自动发邮件。
I've just been going crazy with it and, you know, initially it was like, can you build me like a a basically a software version of a spreadsheet that like sends an email when I complete a field.
但后来我想，为什么不把这个接入 YouTube 的 API 呢？
But then it was like, well, why don't you plug that into YouTube's API?
你为什么不把这接入到，你知道的，我正在找公寓。
Why don't you plug that into, you know, I'd be right, I'm looking for an apartment.
你能接入StreetEasy和Zillow吗？
Can you plug into StreetEasy and and and Zillow?
突然间，这就变成了我不再自己上互联网，而是让AI帮我筛选互联网上的信息。
And all of a sudden, it's like, oh, it goes from basically me going to the Internet to the AI, you know, sorting through the Internet for me.
你确实关于OpenClaw的情况发过推文。
And you actually tweeted about this with the OpenClaw situation.
你说，在OpenClaw、Codex、Cloud Code、Cowork、Manus（已被Meta收购）以及其他代理系统的世界里，软件的未来必须以API为先，同时也要支持人类进行验证、与代理和他人协作，并处理输出结果。
You said, in a world of OpenClaw, Codex, Cloud Code, Cowork, Manus, which Meta acquired and other agentic systems, it's becoming clear that the future of software has to be API first, but also enable human interaction for verification, collaboration with agents and people, and working on the output.
如果软件行业转向以API为先，这意味著什么？
So what does it mean for the software industry if if it becomes API first?
因为一方面，通过这种方式与你互动，你的客户能获得巨大的实用性。
Because, you know, on one hand you're enabling your customers to get a tremendous amount of utility if they're interacting with you this way.
但另一方面，你知道，Zillow可能从我访问它那里获得了一些价值。
On the other hand, you know, Zillow, probably got some value in me going there.
YouTube 可能希望我留在 YouTube 上。
YouTube probably wants me on YouTube.
现在所有这些都发生在我用 Cloud Code 构建的仪表板里。
Now it's all happening in my like, you know, my dashboards that I've built with Cloud Code.
是的。
Yeah.
所以也许我们可以稍微区分一下市场，因为你最后提到了很多消费类产品。
So maybe we'll separate the markets a little bit because you you threw in a lot of consumer products at the end of of that.
说实话，很难说有多少消费互联网会简化为 API 调用，而普通消费者仍然只想去 YouTube 看推荐内容，他们不会……
You know, I hard to say how much how much of the consumer Internet kinda gets collapsed into API calls versus versus, you know, the average consumer just still wants to go to YouTube and see the feed and and they're not gonna
不会那样做。
do that.
对我来说，YouTube 完全是在后台运行的。
For me, YouTube is that's strictly on, like, the back end.
所以这是 YouTube 的创作者端。
So that's, the the creator side of YouTube.
就像是，是的。
Like Yeah.
是的。
Yeah.
我用它来排序缩略图，然后根据点击率进行排名，还能告诉我们用户在视频上停留了多长时间。
I've used it to sort, like, thumbnails and then rank them by, you know, click through rate and then also tell us how how long people are staying on the videos.
但是。
But I
不过你说得对，从消费者角度来看，人们大概不会去你的Claude机器人那里看YouTube。
but point taken on the consumer side, not gonna wanna go to your Claude bot to watch YouTube probably.
是的。
Yeah.
所以这就是我稍微做了一些区分的原因。
And so so that's why I kinda separated a little bit.
现在，你必须稍微体谅一下，或者至少想一想，因为当答案直接在ChatGPT中弹出，或者某种自动化系统直接提供答案时，那些绝对重要的消费类平台流量肯定会大幅下降。
Now now I'm but but you have to be a little bit sympathetic or at least think through because because again, absolutely major consumer properties are gonna see a reduction in traffic when the answer just comes up in ChatGPT or when, you know, some some kind of automated system is just delivering the answer.
所以我认为这是一个整个类别，人们必须认真思考。
So so I I I think that's a whole whole category people have to think through.
在企业软件方面，这显然是我们投入时间的地方。
On the enterprise software side, that's obviously where where we spend our time.
我先以Box为例说一下，然后你可以再扩展一下软件方面的观点。
We I'll speak for Box for a second, and then maybe you can broaden out for software.
在Box，我们对这一点充满热情，因为代理既擅长处理文件，其工作流程也离不开文件。
At Box, we're, like, a 100% excited about this because because the you know, one of the things that agents are both really good at, but also need for their workflows are are your files.
它们需要能够访问信息，以便为你回答问题、生成新信息、存储记忆，以及参与你的工作会话。
They they need to be able to access the information to work with, to answer questions for you, to produce new information, to be able to store off memories, and and it's workings and they're working sessions that you can go and interact with.
它们需要能够阅读规范和文档。
They need to be able to read specifications and documentation.
所有这些最终都会变成文件。
All of that ends up being files.
因此，我们正在构建一个平台层，无论你是个人与数据交互，还是应用程序需要访问数据，亦或是代理需要一个文件系统来交互，我们都希望成为连接这一切的平台层。
So what we are building is a platform layer that whether you're a person interacting with your data, whether you're an application that needs to access data, or whether you're an agent that needs a file system to interact with, we wanna be the platform layer that connects all of that.
嗯。
Mhmm.
而我们Box公司认为，我们之所以处于一种独特的位置，关键在于我们不认为仅仅让智能体拥有自己的沙盒文件系统环境就足够了，同样，仅仅让人们拥有一个独立的环境也是行不通的。
And the the key why we we at Box, we think we're in a a a kind of unique position is we don't think it's enough for the agent just to have its own sort of sandbox environment of of of a file system, nor is it it is it gonna work for just people to have a a separate environment.
你需要一个能够真正将这两个世界连接起来的东西。
You're gonna need something that actually connects those two worlds together.
所以，人们将需要某种形式的终端用户界面，即使那是一个聊天机器人中的界面，他们仍然需要通过某种视觉化的方式与自己的数据进行交互，并且他们最终很可能希望登录某个系统，查看所有内容，并能够管理他们的共享权限以及他们正在与谁合作。
So people are gonna need some form of end user interface, even if that's an end user interface in a in a chatbot, they're still gonna need to kinda interact with their data with with something visual, and they'll likely eventually wanna like log into something and see all their content, and be able to manage their sharing permissions, and who they're working with.
但智能体只需要一套API接口，它们需要能够通过这些API进行工作，并促进它们正在执行的所有任务。
But agents just need a set of APIs, and agents need to be able to work with those APIs, and facilitate all of the work that they're that they're doing.
因此，我们正在投入资源，确保为智能体提供最强大的能力，使其能够在其中工作，并处理您想要提供给它的所有内容。
So what we're investing in is is making sure we've got the most powerful capabilities for agents to be able to work in and, you know, work with all of this content that that you wanna give it.
现在出现了所有这些新的考量，比如如何为智能体提供一个独立的工作空间，以便您与其协作，同时将其影响范围控制在一定的范围内，这样它就不会误删您的所有数据。
Now there's all these new implications, which is how do you give an agent a separate space to work in that you're collaborating with that agent, but it its blast radius is somewhat contained, so it doesn't kinda delete all of your data.
然后突然间，您可能就面临这样的危机，因为您的OpenClaw智能体跑去把所有东西都搞得一团糟。
And and now all of a sudden, have this kinda crisis on your hands because your OpenClaw agent went and and mucked with everything.
顺便说一下，亚马逊就刚发生过这种事。
That just happened to Amazon, by the way.
我的意思是，不打断你，但亚马逊最近有篇报道说，他们因为代理程序决定‘我知道该怎么解决这个问题’而出现了多次中断。
I mean, not to interrupt you, but Amazon, they there was just this story in the Feet that Amazon had lots of had outages because the the agent was like, know what I'm gonna do to fix this problem?
直接删除一切。
Just erase everything.
我可以让问题消失。
I I can make the problem go away.
不再有代码。
No more code.
删除。
Delete.
是的。
Yes.
难道这不是你的
Is there didn't like That was your
一个解决方案。
a solution.
你不喜欢你的文件夹结构吗？
You didn't like your folder structure?
太好了。
Great.
现在没有了。
Now there is none.
所以，你确实需要仔细考虑如何在这些系统之间建立恰当的分界线。但对我们来说，如果你想象未来智能体的数量将是人类的五倍、十倍甚至百倍——考虑到它们将带来的生产力提升，我认为这是一个相对稳妥的假设——所有这些智能体都将处理企业信息。
So so you do have to you have to be thoughtful about about how do you kinda create the right, you know, lines of demarcation between these systems, but again for us, if you imagine that there's five or 10 or a 100 times more agents in the future than people, which is I think a relatively safe assumption given the productivity increase that they're gonna enable, all of those agents are gonna work with enterprise information.
它们需要一个安全的空间来处理这些信息。
They're gonna need a secure space to work with that information.
它们将能够存储这些数据。
They're gonna be able to store that data.
它们将能够基于这些数据进行操作。
They're gonna be able to operate off of it.
它们将能够为最终用户解答问题。
They're gonna be able to answer questions for end users.
它们将需要能够存储自己的数据。
They're gonna be able to be able to, you know, need to be able to store their own data.
这就是我们正在构建的东西。
So that's what we're building.
我们必须确保，再次强调，我们让代理尽可能轻松地去利用它。
And we have to make sure, again, we we make that as easy as possible for agents to go and utilize.
我认为已经有相当数量的现有软件也必须做同样的事情。
I think that there's a meaningful amount of software that already exists that will also have to do the same thing.
它们必须让自己的软件为智能体做好准备。
They will have to make their software ready for agents.
我认为某些形式的软件会被压缩，因为智能体并不需要像人类那样使用它们的工具，这显然会在软件市场的某些领域带来压力。
I think there'll be some forms of software that get kind of compressed where agents don't really need to use their tools in the in the same way that people did, and that's obviously where you're gonna see some pressure in in in the in the software market in some areas.
然后还会出现全新的平台，因为我们没有预见到智能体将会遇到的新型问题，而这正是那些从一开始就以API为先、完全以平台思维构建的公司会涌现的地方。
And then there's gonna be all new platforms that have to exist, because we didn't anticipate the kind of new problems that agents are are gonna run into, and that's where you'll have, again, API first companies get launched from the start thinking, you know, only in terms of platforms.
我认为，对于任何至少参与这一架构的人来说，这都将带来巨大的增长。
And I think this is just gonna be, you know, a tremendous amount of growth for anyone who at least has a a a play in that in that architecture.
好的。
Okay.
你提到了生产力，我认为在节目结束时，这个问题值得深入探讨，因为关于人工智能，确实存在一些讨论。
So you mentioned productivity, and I think this is something that's worth examining as we end the show because I think there is this sort of discussion around AI.
通常人们会说，生产力确实提升了，这一点似乎已被普遍接受，无论现在已有还是将来会出现，但数据其实有些矛盾，我想跟你确认一下，听听你对这些数据的看法。
Oftentimes, it's well, there's productivity increases, and it's sort of accepted, you know, that's there already are or there will be, but the data is a little bit mixed, and I just want to run it by you and get your perspective on what the data is saying.
这是来自《财富》杂志的报道。
So this is from Fortune.
数千名首席执行官承认，人工智能对就业或生产力没有任何影响，这使得经济学家重新审视了四十年前提出的一个悖论。
Thousands of CEOs just admitted AI had no impact on employment or productivity, and it has economists resurrecting a paradox from forty years ago.
文章提到，在20世纪60年代，我们有了晶体管、微处理器和集成电路，但生产力增长率却从之前的2.9%下降到了1973年的1.1%。
So it talks a little bit about how in the 1960s we had transistors, microprocessors, integrated circuits, and productivity growth actually ended up slowing from 2.9% beforehand to 1.1% in 1973.
如今，这些被调查的首席执行官中，有6000人参与了调查，三分之二的高管表示他们使用了人工智能，但平均每周仅使用1.5小时；25%的受访者表示在工作中完全未使用人工智能；近90%的企业表示，过去三年中人工智能对就业或生产力没有任何影响。
And so now you have all the CEOs that have been pulled, and it is, yes, 6,000 CEOs, Two thirds of the executives reported using AI, but it was one out one point five hours a week, 25% of the respondents reported not using it in the workplace at all, and nearly 90% of the firm said AI had no impact on employment or productivity over the last three years.
我的意思是，也许这是去年做的研究，但即便如此，
I mean, maybe this is research done last year, but even still,
说实话，我很好奇，现在也很想知道。
you know I was curious Actually, am curious.
这项研究是什么时候发布的？
When was that published?
或者这项研究是什么时候进行的？
Or when was the research taken?
它发布于2026年2月。
It is published February 2026.
我不太清楚研究具体是在什么时候进行的，
I don't know exactly when the
研究是在什么时间段进行的。
research was conducted over.
是的。
Yeah.
但考虑到受访者的数量，显然这大概是在去年某个时候进行的。
But with with the number of respondents, obviously, that would have been probably, you know, sometime last year.
但，抱歉。
But but sorry.
你继续说。
Keep going.
不。
No.
你请说。
Go ahead.
哦，比如，只是去辩护AI，或者
Oh, like, as in just, like, like, defend defend AI I or
我本来想问一下，你的看法是什么，因为看起来我们确实在某种程度上，我们想稍微检验一下这些假设，比如我们会拥有比工人更多的AI代理，这将会，是的。
was just gonna ask, like, what your perspective is here because it does seem like we're we're, you know, in some ways, and this is sort of we wanna pressure test a little bit about, like, some of these assumptions that we're gonna have more AI agents than we'll have workers, that it will Yeah.
带来生产力的提升，但目前我们看到的数据表明，这一点在最好情况下也仍存疑。
Lead to this increase in productivity, whereas we are still seeing data where that is at the at sort of best when you look at this data up in the air.
是的。
Yep.
对。
Yeah.
我能理解科技驱动型经济与其他经济领域之间可能存在的不协调。
I I I can understand the the dissonance that might be out there be between the tech enabled economy and and the rest of the economy.
因为实际情况是，在科技领域，这些智能体在编码方面非常高效，而且开发者采用智能体进行编码的障碍，远比知识工作者经济中其他领域为获得同等生产力提升而采用类似用例的障碍要少得多。
Because what what's happening is in tech, these agents are are so effective at coding, and and developers have have far fewer barriers to adopt agents for coding than the rest of the knowledge worker economy has for the same level of productivity gain kind of use cases.
所以在编程领域，这些模型具有令人难以置信的特性，因为它们接受了海量的代码训练。
So in coding, you've got these just incredible properties, which is the models are hyper hyper trained on code.
要知道，编程本身是一种纯文本媒介。
They you know, coding itself is a is a text only medium.
达里奥和杜阿尔凯什在他们最新的播客中暗示了一个有趣的观点，即你的代码库包含了大部分你最终需要处理的上下文信息。
You know, Dario and Duarkesh on their latest podcast kind of hinted at at an interesting point, which is your code base contains most of the context that that you end up working with.
它包含了你的文档。
It's got your documentation.
它包含了你所有已完成的工作。
It's got your all of the existing work that you've done.
如果你对比一下，开发者通常更技术化，更关注互联网和最新趋势。
And if you kind of compare and then you developers are just, you know, are are obviously more technical, generally more tapped into the Internet and what's going on and the latest trends.
他们会下载最新的产品并尝试使用。
They pull down the latest new products and and and try them out.
现在你可以将这与其余的知识工作对比，比如快消品公司的市场人员、中型律师事务所的律师，我随便举个例子，某种‘是的’。
Now you can compare that to the rest of knowledge work, you know, the marketer at a CPG company, the the the, you know, lawyer at a mid sized law firm, You know, I'm making up a, you know, kind of some some kind of Yep.
这是对各种职位的夸张描述。
Caricatures of of of, you know, various job functions.
但基本上，他们每天照常工作，并不会想着：‘我该如何构建我的工作流程，以充分利用智能代理，自动化我所有的工作？’
But basically, like, they're going about their day, and they're they're not thinking like, how do I go and construct my workflow to to just fully take advantage of agents and automate everything I'm doing?
对于大多数知识工作者来说，这大概根本不是他们首要考虑的事情。
Like, that that's just like probably not top of mind for, you know, most knowledge workers.
他们会去使用ChatGPT，问一些问题，让AI帮他们写一封邮件。
They're gonna go to ChatGPT, they're gonna ask some questions, they're gonna get an email written for them.
他们会总结一份文档。
They're gonna summarize a document.
他们会制定一个新的战略计划。
They're gonna build a new strategy plan.
然后，你知道，公司会因此逐步多做一点，也许他们的策略会稍作调整，或者财务分析师会得出一些新的见解。
And then, you know, they're gonna be you know, the company will will do incrementally a little bit more as a result of that, and maybe their strategy changes a little bit more, or the financial analyst comes up with some new insights.
我认为，至少在过去几年里，每当有类似这样的调查试图分析时，AI的状态就是这样。
That's, I think, probably been the state of AI for for the past couple of years, at least whenever a survey like this would have would have tried to analyze.
将这与工程领域对比，你知道，我们有产品，我们能构建五倍于以往的功能，这些是工程师根据AI编码给出的实际估算。
Compare that to engineering, where, you know, we have products that we build five you know, these are the estimates from the actual engineer that we will build five times faster because of of AI coding.
因此，我们将能够向客户交付更多功能。
And and we will, as a result of that, be able to ship significantly more capabilities to our customers.
我们将能够为客户解决更多问题。
We will be able to solve significantly more problems for our customers.
在许多情况下，我们甚至可能不会为此功能收取更多费用。
In many cases, we might not even charge more for that functionality.
我们现在可以把这些功能打包进他们现有的许可证中，因为我们已经能做到这一点了。
We are going to pack that into their existing licenses because we now can.
那么，在我们的生产力方面，你到底会衡量哪些指标呢？
So so to some extent, what would you measure in our in our kind of productivity?
现在这已经成为我们理所当然要做的事了，因为我们必须不断提供更多的价值，毕竟科技行业竞争异常激烈，我们希望为客户提供更多功能。
We this is now just a priced in thing that we do because we we have to deliver more and more value, because obviously tech is hyper competitive and and we wanna now add more capability to our customers.
我认为，这种变化还没有波及到其他知识型工作领域，但我相信它迟早会到来。
I think that that has not yet rippled through the rest of knowledge work, and I I think it just will.
它必然会到来，因为工具会变得越来越好，而市场上一旦有竞争对手能利用人工智能来降低成本、降低客户费用，或向客户提供显著更优质的产品，就会产生巨大影响。
It it just it it it will have to because the tools will get better and better, and you'll have one competitor in a in a market that is able to use AI to either lower their costs or lower their fees to the customer, or be able to deliver a substantially higher product to the customer.
随着你看到越来越多这样的例子，这些市场动态就会开始发生转变。
And as you see more and more examples of that, that will just start to to transform these market dynamics.
我想说，同样重要的是，我喜欢根据这样的原则行事——你知道，贝索斯说过一句话：当轶事和数据相矛盾时，你必须关注轶事。
You know, I would say equally that, you know, I I I like to operate off of you know, I think Bezos had this line is when the anecdotes and the data disagree, you have to look at the anecdotes.
所以，你看，两周前普华永道要求他们的审计师降低收费，就是因为人工智能。
And so, you know, look at the the, you know, the equal headline from two weeks ago of KPMG asking their auditor to lower their fees because of AI.
我认为这正是你最初会看到的信号，预示着即将发生的变化：一家公司将表示，对于我们现在知道可以引入自动化的工作，我们应该减少投入，然后将这些资金用于公司内部其他更具生产力、更高效或更具竞争力的方面。
That I think is your in that's your initial signal of of actually what's going to happen, which is a company is gonna say, you know, that kind of work that that that we now know we can we can bring automation to, we should be spending less on, and then using those dollars to do something else in our company that is that is higher productivity or more or that makes us more effective or more competitive.
而当你在一个生态系统中这样操作几十次、几百次、几千次甚至几万次时，你就会开始看到这些市场格局如何重塑。
And you once you do that dozens or hundreds or thousands or tens of thousands of times in an in an ecosystem, that's where you'll start to see kind of this reshaping of of how these markets will play out.
这在科技领域无疑正在发生，现在唯一的问题是，这种变化扩展到其他经济领域的路线图是怎样的？
It's happening in tech unquestionably, and now the only thing is what's the roadmap to that happening across the rest of the economy?
这需要时间。
That's gonna take time.
人们必须改变他们的工作流程。
People have to change their workflows.
人们的数据设置方式尚未为智能体做好准备。
People don't have data set up in a way that is sort of prepared for agents.
智能体本身并不总是具备适合知识工作的正确界面或工具支持。
The agents themselves don't always have the right interfaces or tooling to to be supported in knowledge work.
所以，我对此实际上非常务实，我认为我可以同意你刚才读到的调查结果，同时完全不为所动，甚至更想说，人们应该为此做好准备，因为这将会影响到知识工作的更多领域。
So so I'm actually extremely pragmatic about this, where I think I could agree with the survey that you just read, and equally be completely unfazed, and and more of anything, say people should be probably prepared for this will come for more areas of knowledge work.
我是对这项技术对就业影响最乐观的人，所以我不觉得这有什么可怕的。
I'm the biggest optimist on the jobs impact of that, so I don't see that as a scary thing.
我认为这意味着公司必须为客户提供多得多的服务。
I think it's just gonna mean companies will have to sign up to do way more for their customers.
我认为这将体现在我们与所有供应商打交道时，消费者端会出现供过于求的情况。
I think that that that will be where it shows up is is is we will have a surplus on the consumer side of all of the vendors that we work with.
我们必须不断提供更优质的服务。
We'll just have to deliver better and better service for us.
或者如果你是一家B2B公司，那么你的所有供应商都必须提供更优质的服务。
Or if you're a b to b company, then all of your vendors will have to deliver greater services.
五年或十年后，我们会醒来，发现这一切其实已经变得相当正常了。
And and we will wake up in five or ten years, and it'll actually kinda feel like relatively normal.
不会有什么疯狂的变化，也不会像科幻电影那样。
Like like, it there's not gonna be some kind of crazy it's not gonna be the the sci fi movie.
我们会发现，消费者体验和服务只是在逐步提升，就像你回到四十年前，试着想象一位律师或医疗专业人士的生活，你会说：天啊。
It's going to be that that that we just we just have incrementally better consumer experiences and better services, just as if you went back, you know, forty years ago and tried to imagine life of of a lawyer or a healthcare professional, and you'd be like, wow.
你以前是怎么做你的工作的？
How did you do your job?
就是，没有电脑的时候？
Like, without a computer.
你是怎么在没有互联网搜索的情况下理解法律案例先例的？
Like, how how did you, like, how did you understand the legal case precedence without an Internet search that you could go do?
五年后的工作会是这样：你会惊讶地问，你们以前是怎么做的？居然没有一个代理能瞬间帮你起草整个合同，好让你立刻回应电话那头的客户？
Like, that's gonna be work in five years from now is you'll be like, how did you do that without an agent that drafted your entire contract, you know, for you instantly so you could respond to the client that was on the phone?
我们将来也会有同样的困惑，不明白我们今天是怎么工作的。
That that we will have that same set of questions and be confused how we even work the way we do today.
但那并不会是一种彻底的变革，我们仍然会有人类。
But yet it won't be some kind of, you know, completely transformation of of, you know, we we'll still have people.
他们会一起工作。
They'll be working together.
他们会把任务分配给代理。
They'll deploy tasks to agents.
这些代理会去执行更多任务并开展工作，然后人们再把成果带回当前任务，推动他们的工作或项目向前进展。
Those agents will go off and farm more and and do work, and then people will go and bring it back to to the to the task at hand to move whatever their their sort of, you know, work or project is forward.
没错。
That's right.
是的。
Yeah.
当我看着Claude Code运行时，我会想：等等，人们以前真是这样做的吗？
When I'm watching Claude Code go, I look at it and I say, wait, people did this before?
那些现在被自动化完成的事情，以前居然要花那么多时间，真是不可思议。
That seems like a lot of time to do things that are automated, No.
但是
But
比如说，以前你想要在代码库中做一次库的更改，得花上整整两周，而现在只需要十分钟。
like, literally, like, we you'd used to have to spend, like, two weeks on, like, a a library, you know, change that you wanted to make in your code base, and and that's now a ten minute activity.
但我们的软件开发时间真的变少了吗？
And but but are we spending any less time building software?
不。
No.
这是因为我们现在正在做那些之前没来得及做的事情，因为我们当时花了两周时间进行库更新。
It's because we're just now doing the things that we didn't get to because we were spending the two weeks doing the library update.
对。
Right.
好的，亚伦。
Okay, Aaron.
我们得让你离开了，因为你接下来还有会议要参加，我想。
We have to get you out of here because you have to go to your your next meeting, I think.
但还是想再次表示感谢。
But just wanna say thank you again.
太好了。
Great.
你来做客总是很棒。
Always great having you on the show.
下周三，我们将邀请迈克尔·波伦做客节目。
Next Wednesday, we're gonna have Michael Pollan on.
他是一本关于意识的新书的作者，所以我们会讨论人工智能意识。
He is the author of a new book about consciousness, so we'll talk about AI consciousness.
好的，各位。
Alright, everybody.
请继续关注，我们下次在《大科技播客》再见。
Stay tuned for that, and we'll see you next time on Big Technology Podcast.
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。