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。
Anthropic的IPO数据泄露了。
Anthropic's IPO numbers leak.
这些数据够不够好，能不能拯救经济？
Are they good enough to save the economy?
OpenAI推出了一款名为Dots的常驻智能体。
OpenAI has a new always-on agent called Dots.
我们是不是已经通过了视觉图灵测试？
And have we passed the Visual Turing Test?
这些内容将在《大科技播客》周五版中呈现，就在广告之后。
That's coming up on a Big Technology Podcast, Friday edition, right after this.
本期节目由Genesys赞助播出。
This episode is brought to you by Genesys.
要让智能体AI在整个企业中真正落地，到底需要做些什么？
What does it actually take to put agentic AI to work across an entire enterprise?
我最近和Genesys的董事长兼CEO托尼·贝茨聊了聊，探讨了为什么编排正在成为客户体验领域的竞争优势，以及企业在AI时代需要做好哪些事情。
I recently spoke with Genesys Chairman and CEO Tony Bates about why orchestration is emerging as a competitive advantage in customer experience, and what companies need to get right in the AI era.
然后我和Voya Financial的企业业务解决方案主管Adam Mitchell坐下来聊了聊，深入了解了一个大规模的企业级客户体验转型，包括其中的商业决策、取舍和一路上的经验教训。
Then I sat down with Adam Mitchell, head of Enterprise Business Solutions at Voya Financial, to go inside a large-scale agentic CX transformation, including the business decisions, trade-offs, and lessons along the way.
今天你可以在我的YouTube频道上看到这两段对话。
You can watch both conversations now on my YouTube channel today.
欢迎收听Big Technology播客周五版，我们一如既往地用冷静、细致的方式解读新闻。
Welcome to Big Technology Podcast, Friday edition, where we break down the news in our traditional cool-headed and nuanced format.
今天这期节目非常精彩。
We have a great show for you today.
大家期待已久的Anthropic上市前财务数据终于出来了，是路透社泄露的一份报告。
The long-awaited Anthropic pre-IPO numbers are out, courtesy of a leaked report in Reuters.
我们会仔细看看这些数字，讨论它们是否足以支撑当前的经济运行。
We're going to go over the numbers and talk about whether they're good enough to keep this economy running.
另外，OpenAI推出了一款名为Dots的全天候智能代理。
We also have a new always-on agent from OpenAI called Dots.
我已经试过了。
I've tried it.
我会分享我的看法，以及为什么我已经给它派了一大堆活儿。
I'm going to give my reflections and why I'm sending it a ton of work already.
最后，我们要聊聊TAVIS和视觉图灵测试。
And then finally, we're going to talk about TAVIS and the Visual Turing Test.
是的，现在有些机器人能在视频通话中完全骗过你，让你以为它们是真人。
Yes, there are some bots now that can completely fool you into thinking they're human on a video call.
和往常一样，周五和我们一起讨论的是Margins的Ranjan Roy。
Joining us as always on Fridays to do it is Ranjan Roy of Margins.
Ranjan，很高兴见到你。
Ranjan, great to see you.
很高兴见到你。
Good to see you.
我们离真正的S-1越来越近了。
We're getting so much closer to the real S-1.
我们不断得到这些零碎信息，接下来要深入聊聊，但你有没有看到他们要...
And we're getting these tidbits, and we're going to dig into it, but did you see they're going to...
我觉得他们宣布了投资者日定在10月14号，那就意味着Anthropic的IPO会在感恩节前如期进行。
I think they announced Investor Day is supposed to be October 14th, and that means we're on track for a pre-Thanksgiving Anthropic IPO.
没错，是的。
Exactly, yeah.
就在11月。
It's coming in November.
我觉得毫无疑问，今年11月就会上市。
I think there's no doubt it comes this November.
而且就像你提到的，我们拿到了泄露的消息。
And as you mentioned, we got a leak.
路透社基本上把整个S-1文件都独家报道了。
So Reuters did scoop basically the entire S-1.
他们没有把文件打印给我们，但把Anthropic带给华尔街的所有数据都给了我们，我们终于有了确切的数字，不是ARR。
They didn't print it out for us, but they gave us all the numbers that Anthropic is taking to Wall Street, and we finally have some binding numbers, not ARR.
我其实很好奇你怎么解读这些数字，因为里面有些数字很好看，但有些数字让人直摇头，忍不住问：这到底怎么运作？
And I'm actually curious how you read them, because there are some good numbers in there, and there are some numbers in there that make you just kind of shake your head and ask, "How is this going to work?"
那如果你愿意的话，我先给你说几个数字。
So if you will, let me give you a couple of them to start off with.
好消息是，如果你是Anthropic的话，公司收入在指数级增长，对吧？
So the good news is the company's revenue is, if you're Anthropic, the good news is the company's revenue is growing exponentially, right?
2024年4亿，2025年总共46亿。
400 million in 2024, 4.6 billion total in 2025.
到年底的时候，他们说年化收入是100亿，但如果你1月份是10亿，12月份是100亿，那平均下来差不多就是46亿。
Now, at the end of the year, they said they had a 10 billion ARR, but if you have a 1 billion ARR in January, and 10 in December, 4.6 is kind of where you would imagine things net out.
我觉得最疯狂的数字是，他们报告的46亿收入，却告诉投资者未来十年计划在计算和基础设施上花5180亿，而且其中80%是绑定的。
The crazy number to me is on that 4.6 billion in revenue that they reported, they said they told investors they plan to spend 518 billion over the next decade on computing and infrastructure, and of that, 80% of that spend, 80% of the spend is binding.
所以你看，这家公司最近一个报告年度才赚了40亿，却已经承诺了4000亿，而且根本没法脱身。
So you're looking at a company that just basically made, in the most recent year that it can report, $4 billion, that has committed $400 billion and has no way of getting out of it.
就是这些数字。
These are the numbers.
你怎么看？
Your reaction.
我的反应是有点失望，不是对数字失望，这个我们待会儿会聊，而是因为泄露的文件里没有2026年上半年的财务数据。
My reaction is I was a little disappointed, not in the numbers, which we will get into, is that the leak doesn't have the H1 2026, the first half of the year financials.
到时候他们肯定需要提交并公开一些2026年的财务数据、营收和亏损数字，我猜他们很可能会这么做。
When they're definitely going to need to be filing and actually publicly stating some amount of 2026 financials and revenue numbers and loss numbers, which I'm sure they'll likely be.
不过这样我们倒是能好好看看Anthropic在2025年底是什么样子。
But so we have a good look at what Anthropic looked like at the end of 2025.
再说一遍，营收是46亿。
And again, 4.6 billion in revenue.
我觉得净亏损大概是多少呢，参考一下，大家一直在说的那个420亿净亏损数字，我这么说吧，我可能会对他们很多数字持怀疑态度或者深挖一下，但其实不是420亿，其中有340亿是一个奇怪的会计处理问题，跟公司价值大幅增长有关。
I think what was the net loss was around, for reference, there's this $42 billion number that's being thrown around as a net loss, and I say this as someone who probably is going to be skeptical or dig into a lot of their numbers, but it's not actually 42 billion, $34 billion of it is a weird accounting treatment issue related to the company actually growing in value a lot.
但净亏损仍然在80亿左右。
But the net loss is still around $8 billion.
他们每赚1美元，实际上要亏掉大约2美元。
They're still losing around $2 to actually make every single dollar that they're making.
所以对我来说，2025年他们亏了很多钱。
So to me, in 2025, they were losing a lot of money.
2026年会是什么样子呢？
What does it look like in 2026?
有各种小道消息说他们在五月和六月就已经盈利了。
There's all these little leaks that they were supposedly profitable back in May and June.
我不太确定。
I don't know.
我很想看看实际情况。
I want to see that.
我觉得我们在文件里看不到这些，对吧？
I doubt we're going to see that in the documents, right?
还有，记得之前有提到1000亿美元的年化收入，或者说2026年收入达到1000亿美元。
And also, remember, there was talk of $100 billion run rate, or $100 billion in 2026, like revenue.
那就是那些ARR和类似报告数字流传的地方，想想看，要达到2万亿美元的估值，你可能确实需要看到一些接近那个方向的数字。
That's where that ARR and all these kind of reported numbers were going around, and when you think about it, to get to that $2 trillion valuation, you would think maybe you do start to need to see something up in that direction.
所以，我想我们一会儿再聊计算承诺，但我很想知道你认为他们的收入会达到什么水平？
So, I guess let's get into the compute commitments in just a bit, but I want to know where do you think their revenue is going to end up?
你觉得他们会拿出什么样的数字来展示当前业务状况，而不是2025年的？
What kind of numbers do you think they're going to show about the current state of the business as opposed to 2025?
因为说真的，谁还记得2025年啊？
Because come on, who remembers 2025?
太久了。
So long ago.
不，这很重要，因为如果那个数字看起来不像是还在指数增长，那很多大数字就会变得很吓人。
No, it's very important because if that number doesn't look like it's continuing to go exponential, a lot of these big numbers become very scary.
比如说，最近Anthropic，我记得是Anthropic，跟《纽约时报》说了，当然我们说的是接近消息源的那种，对吧？
So for instance, Anthropic recently, I think it was Anthropic, told "The New York Times," we obviously it's like sources close to the matter, right?
是Anthropic吗？
Was it Anthropic?
还是他们的某个投资人？
Was it one of their investors?
可能是其中一个吧。
Probably one or the other.
他们告诉《纽约时报》，到今年年底，年化收入将达到1000亿美元。
Told "The New York Times" that it was going to be at 100 billion annualized revenue at the end of the year, by the end of the year.
所以你会认为，今年他们最终会达到500亿到1000亿之间的某个数字。
So you would assume that this year they're going to hit somewhere between 50 and 100 billion when all is said and done.
如果他们能做到，那么花5000亿就没那么吓人了，比他们只有……10亿的时候要好。
If they can do that, then spending 500 billion is less daunting than if they're at ... 10.
但我们不知道，我觉得你说得对。
But we don't know, and I think you're right.
当我们最终看到这些数字时，那会非常重要。
When we end up seeing those numbers, it's going to matter a lot.
记住，今年大部分时间，他们的年化收入大概在450亿到600亿之间。
For most of the year, remember, they've been at something like 45 to 60 billion ARR.
至少我们被告知的是这样。
At least that's what we've been told.
所以我觉得他们肯定在增长，但增长的速度会非常关键。
So I think they're definitely growing, but the rate that they're growing is going to matter a lot.
这个问题达里奥也提到过。
And this is something that Dario's kind of addressed.
比如他和德瓦克什聊过这个。
He's talked about this with Dwarkesh, for instance.
他也和安德鲁·罗斯·索金谈过。
He talked about it with Andrew Ross Sorkin.
我觉得去年我在那里的时候，他和我也有过一些交流，就是公司必须提前几年预测需要多少算力。
I think he and I had a bit of a conversation about it when I was there last year, which is that the company has to make these projections of how much compute it's going to need years out.
有一个风险是，如果你买得太少，就无法满足应用需求，而你的竞争对手拥有大量算力，这就是最近流传的说法，因为OpenAI似乎拥有更多算力，并且像达里奥说的那样，正在全力冲刺，能够满足需求。
And there's one risk, which is that you buy too little and you can't serve the use case, and your competitor has a lot, which is the story that's been told recently because OpenAI seems to have more compute and is going YOLO, in the words of Dario, and able to meet that need.
但另一方面，如果你的收入无法增长到足以履行承诺，无论你多谨慎，如果无法履行这些承诺，而其中80%是不可取消的，你的公司就会破产。
But the other thing is, if your revenue isn't able to grow to meet your commitments, careful or not, if you cannot meet these commitments and 80% of them are not cancelable, your company goes bankrupt.
这就是Anthropic目前的状况。
And that is the state of Anthropic.
我今天在《Big Technology》上写了这件事，我称之为Anthropic带着所有人进行的一次信仰之跃。
And I wrote about it today in Big Technology, and I called it a leap of faith that Anthropic is taking everybody on.
这确实是一场信仰之跃，因为如今的经济依赖于Anthropic和OpenAI履行这些承诺。
And it really is this leap of faith because the economy today is dependent on Anthropic and OpenAI meeting these commitments.
如果他们做不到，不仅对Anthropic来说是坏消息——如果无法履行这些义务，它会破产——还会引发连锁反应，因为他们向Alphabet、亚马逊、微软，更不用说英伟达，承诺的金额巨大，这些都会影响这些公司的预期。
And if they're not able to, not only is this going to be bad news for Anthropic, which will go bankrupt if it can't meet these obligations, but there is going to be a cascade because the amount of money that they've pledged to companies like Alphabet, like Amazon, and like Microsoft, not to mention Nvidia, is massive, and factors into these companies' projections.
好吧，行。
Well, okay.
我确实同意这些承诺非常巨大，但他们能付得起吗？
I do agree these are massive commitments, and will they be able to pay them?
这对他们的财务未来到底意味着什么？
And what does that actually mean for their financial future?
但我不认为这意味着他们会破产。
But I don't think it means they go bankrupt.
实际上，如果我们推演一下这个情况，即使再说一遍，那80%的计算承诺是不可取消的。
Actually, if we outline how this plays out, even again, 80% of that overall compute commit is non-cancelable.
但现实中，如果这开始产生某种连锁效应，你不觉得律师会介入，然后大家都会意识到吗？
But in reality, if this starts to have some kind of cascading effect, do you not think lawyers are going to get involved and everyone's going to realize?
而且Anthropic，这又是那种“大到不能倒”的情况，跟博通、Alphabet、亚马逊、微软这些公司有关。
And Anthropic, this is, again, too-big-to-fail type stuff relative to the Broadcom, the Alphabet, the Amazon, the Microsoft.
那些跟他们有承诺关系的公司，其实也有很强的利益关系，不希望这一切都崩溃。
The people who they have the commitment to also have a very strong vested interest that this does not all collapse.
所以你觉得他们会因为原始合同的性质，就把Anthropic逼到破产吗？
So do you think they would drive Anthropic into bankruptcy because that's the nature of the original contract?
嗯，好吧。
Well, okay.
那是达里奥的原话。
So that was Dario's words.
可能我有点——
Maybe I was a bit-
等等，他说了YOLO吗？
Wait, did he say YOLO?
他说了YOLO，意思是OpenAI在赌一把。
He said YOLO as in OpenAI is YOLOing.
有些玩家在冒险，把风险调得太高了，我非常担心。
There are some players who are YOLOing, who pull the risk dial too far, and I'm very concerned.
谁在冒险？
Who is YOLOing?
这个问题我就不回答了。
So that's a question I'm not going to answer.
他说：“有些人在冒险，而我们试图保持谨慎。”
He said, "There are others that are YOLOing, and we're trying to be careful."
但他基本上用了这些说法，是的，如果我们不能履行，我们就会破产，因为我们在签署具有约束力的义务。
But he used basically these terms that, yeah, we're going to go bankrupt if we can't because we're signing binding obligations.
你看，当然，如果你是亚马逊，你对Anthropic有大量投资。
Look, of course, if you're Amazon, which you have a big investment in Anthropic.
顺便说一句，这又回到了我们一直在讨论的那些话题。
This is, by the way, it goes back to these discussions we've been having the entire time.
循环融资。
The circular financing.
那些投资。
The investments.
那些算力交易。
The compute deals.
它们都是环环相扣的。
It's all tied in together.
所以如果Anthropic出了问题，我觉得亚马逊不会让它破产的。
So if Anthropic goes bad, yeah, I don't think Amazon's going to let them go bankrupt.
但这对亚马逊和Anthropic来说都会是毁灭性的打击。
But it will be irrevocably bad for Amazon and for Anthropic.
所以听听这些数字。
So listen to these numbers.
Anthropic对Alphabet——也就是它的股东之一——的支出承诺是未来十年1110亿美元。
The spending commitments that Anthropic has to Alphabet, which is a shareholder, is $111 billion over the next decade.
对亚马逊是1100亿美元，对微软是314亿美元。
To Amazon, it's $110 billion, and to Microsoft, it's $31.4 billion.
你还记得看到这些云业务的盈利数字吗？
You remember seeing these earnings numbers of these cloud businesses?
当然，这些数字是分摊到不同时间的，但我们看到这些云业务的盈利数字，其中一部分是直接卖给客户的。
Now, of course, that's spread over time, but we see these earnings numbers for these cloud businesses, and a part of it is selling direct to their customers.
但我们看到的很多增长，其实来自于它们和OpenAI、Anthropic这些公司的合作，以及这些公司的健康状况。
But a lot of the growth that we're seeing is coming from their dealings with the OpenAIs and the Anthropics and the health of these companies.
比如在AI之前，AWS每年稳定增长17%，在这么大的基数上已经不错了，但公司没法提高增长率。
Like pre-AI, for instance, AWS was growing at 17% consistently a year, which on a big base is nice, but the company could not increase its growth rate.
现在它暴涨了。
That has boomed.
增长率在30%左右，考虑到它的规模，这简直不可思议。
It's in the 30% range, which is just crazy given its scale.
所以当然，如果亚马逊看到Anthropic表现不佳，它会想办法达成交易来拯救这家公司，保住它继续销售技术的能力，保住它对公司的投资。
So of course, if Amazon saw Anthropic doing poorly, it would try to make a deal to save the company, save its ability to continue to sell its technology, save its investment in the company.
Alphabet也是一样。
Same with Alphabet.
我不怀疑这一点。
I don't doubt it.
所以，是的，破产可能不会发生，但也许不需要破产就能在我们的经济中引发连锁问题。
So yeah, bankruptcy might not happen, but it might not take a bankruptcy to have a cascading issue in our economy.
这正是我非常想和你聊的，Ranjan，因为你在交易大厅待过，见过经济崩溃。
And this is what I really want to speak with you, Ranjan, because you've been on the trading floor, you've seen economic collapses.
我不认为我们正走向经济崩溃，但我觉得我们必须稍微谈谈整个市场对Anthropic IPO、OpenAI IPO以及更广泛公司财务表现的依赖。
I don't think we're heading towards an economic collapse, but I think we have to talk a little bit about the dependence of the broader market on the financial performance of the Anthropic IPO, of the OpenAI IPO, and the companies more broadly.
所以今天我在《Big Technology》上写了这篇文章，标题是《Anthropic的IPO和我们集体的信仰之跃》。
So I wrote about this in Big Technology today, "Anthropic's IPO and our collective leap of faith."
当今的公开市场以及大部分经济都依赖于AI的持续增长。
Today's public markets and much of the economy depend on AI's continued growth.
标普500指数目前正经历自互联网泡沫以来最差的广度，中位股票比其52周高点低16%，尽管该指数今年上涨了约13%。
The S&P 500 is currently experiencing its worst breadth since the dot-com bubble, with the median stock trading 16% below its 52-week high, even as the index is up around 13% this year.
英伟达一直在支撑着市场，在2026年增加了超过一万亿美元的市值，而根据高盛的数据，今年标普500每股收益增长的一半来自AI投资。
Nvidia has been holding it up, adding more than a trillion in market cap in 2026, and half of the S&P 500's earning per share growth this year is coming from AI investment, according to Goldman Sachs.
史蒂夫·拉特纳放了一张图表来讨论这个问题。
Steve Ratner put a chart up talking about this.
他说，标普500指数接近历史高点，但如果剔除AI公司，自8月底以来已经下跌了5%。
The S&P 500 is near a record, he says, but if you strip out the AI companies, it's down 5% since late August.
所以，这就是问题的关键所在。
So this is kind of what it comes down to.
Anthropic做出了这些重大承诺，我们终于通过这份S-1文件看到了具体数字。它必须兑现这些承诺，因为我们的市场广度非常不正常，标普500中的许多公司股价都远低于它们的高点。
Anthropic has made these big commitments, and we finally got a chance to see the number with this S-1 And it has to make good on them because we have this crazy breadth in our market where many of the companies in the S&P 500 are down significantly off their highs.
你看不到它们有任何增长。
You're not seeing any growth from them.
增长再次集中在大型科技公司，主要是AI交易，比如英伟达和这些大型科技公司。
The growth is concentrated, again, in big tech, mostly in the AI trade in companies like Nvidia and in these big tech companies.
如果Anthropic的故事，或者OpenAI的故事，开始出现动摇，你可能会看到连锁反应。
And should the Anthropic story or the OpenAI story, for that matter, begin to wobble, you could see a cascade here.
好的。
Okay.
那我们来看看这会怎么发展，因为我觉得我们其实都同意，这些对亚马逊、Meta和超大规模云服务商的算力承诺，对它们来说风险和对Anthropic一样大。实际上，如果Anthropic付不起钱，那对Meta、谷歌，尤其是博通、英伟达和其他公司来说，真的很糟糕。
So let's think about how this plays out, because again, I think we're actually agreed that the magnitude of these compute commits to the Amazons, Metas, and hyperscalers actually present as much risk to the hyperscalers as they do to Anthropic in the way that it's actually if Anthropic can't pay, that's really bad for Meta, Google, especially Broadcom, Nvidia, and these others.
所以我觉得他们需要担心这件事。
So I think they need to be worried about this.
他们得开始试着去理解。
They need to kind of start trying to understand.
我不知道，这个交易背后有太多能量了。
I don't know, there is so much energy behind this trade.
有太多资本利益牵扯其中，每个人都有利益，而且围绕AI营销和推广AI有太多热情。我说这话是因为我坚信AI最终会改变一切，但这一切都必须由这一小群集中的公司来完成，才能维持这个交易。
There's so much vested interest in capital that everyone has this interest, and there's so much of the energy around marketing AI and pushing AI, and I say this as someone who strongly believes that it's going to transform everything over time, but all of it has to be done by this especially kind of concentrated group of companies in order to keep this trade propped up.
我觉得我们应该讨论一下，会不会有连锁反应？
I think we should get into are there kind of cascading effects?
会不会真的出现大规模的经济动荡？
Is there actual kind of large-scale economic turmoil?
但我觉得首先，对Anthropic来说，这实际上会是什么样子？
But I think to start, what does it actually look like in terms of Anthropic?
你觉得如果他们的S-1文件没能让市场兴奋起来，这会不会就是导火索？
Do we think them having an S-1 that does not excite the market is the trigger?
你觉得会不会是六个月后，增长放缓，甚至停滞，然后大家开始质疑算力承诺，接着又担心亚马逊的云业务收入等等？
Do we think it's six months from now, growth slows, growth is flat, everyone starts to question compute commits, which then has a follow-on worry around Amazon's cloud earnings, et cetera?
你觉得这会怎么发展，假设Anthropic没有拿出那种好到离谱、能拯救所有人、让大家都没事的数字？
How do you see it playing out, assuming Anthropic does not have just ridiculously good numbers that just saves everyone and we're all okay?
我觉得有可能。
I think it could be.
问题是这样的。
Here's the thing.
SpaceX开了先例。
SpaceX set the precedent.
当然，SpaceX是完全不同的公司。
Now, of course, SpaceX is a very different company.
它不是纯AI公司。
It's not pure play AI.
但它的上市表现非常强劲，当然，之后也遇到了一些麻烦，尽管今天它的股价仍然高于IPO价格，也没有导致某种崩盘。
But it went out very strong, of course, and has kind of had some trouble since, even though it is today still trading above its IPO price and hasn't led to some sort of collapse.
但Anthropic存在一些结构性风险，投资者对SpaceX表现出的热情，我不确定你是否会在Anthropic身上看到同样的热情，尤其是考虑到我们已经知道的Anthropic面临的其他风险，基本上就是它的前沿模型正在与表现相当不错的标准模型竞争，公司开始说，‘我要走模型路线，我要离开前沿，选择更便宜的方案。’
But Anthropic has some structural risk, and the enthusiasm that investors showed around SpaceX, I don't know if you're going to see the same enthusiasm around Anthropic, especially given all the other risks that we know exist for Anthropic, which is basically its frontier models are competing with standard models that are doing a fairly good job, and companies are starting to say, "I'm going to model route, and I'm going to go off the frontier and be cheaper."
它目前的客户群正试图比过去减少在token上的支出，而且还有所有这些政府和世界末日的逆风，有些人可能会说这是营销，但它们是一个变量，而你在上市时不想有变量。
It has a customer base right now that is trying to spend less on tokens than they have in the past, and there's all these government and the world-ending headwinds that some might say are marketing, but they're a variable, and you don't want a variable when you go out.
但你看，在所有事情中，我认为投资者会关注的基本点，他们必须做出选择。
But look, the fundamental thing here among everything that I think investors are going to look at, and they'll have to make an option for them.
他们必须做出决定，基本上就是，你是想要一家可能创造出改变世界技术的公司的股份，还是看看现实？
They'll have to make a decision, which is basically, do you want a stake in a company that may end up creating some world-changing technology, or do you look at the picture?
而现实显示，巨大的亏损和巨额支出即将到来。
And the picture shows big losses, big spending coming up.
你有被锁定的支出。
You have spending that you're locked into.
再说一次，这5.18亿美元中的80%是不可取消的承诺，不管那意味着什么，对吧？
Again, 80% of that $518 million is commitments that are uncancellable, whatever that means, right?
数十亿。
Billions.
抱歉，是数十亿的不可取消承诺。
Sorry, billion in uncancellable commitments.
虽然你不能取消，但你的客户可以。
And while you can't cancel, your customers can.
所以这是风险因素中提到的另外两件事。
So these are two other things that came up in the risk factors.
去年，Anthropic近四分之一的收入来自两个客户。
Nearly a quarter of Anthropic's revenue came from two customers last year.
是的。
Yes.
去年，四分之一的收入来自两个客户。
One quarter of the revenue came from two customers last year.
在一个似乎每个人都在开发AI模型的世界里，Anthropic的客户会被锁定吗？
And are Anthropic's customers locked in, in a world where everybody seems to be developing AI models?
不是的。同样从风险因素来看，公司很多大客户并没有签订长期合同，基本上可以随时削减或停止支出。
No. Again, from the risk factors, many of the company's largest clients were not locked into long-term contracts and could cut or stop spending with them basically at will.
再次感谢路透社和Echo Eng独家报道了这些内容。
Again, shout out to Reuters and Echo Eng for scooping all this.
这些数字真是令人着迷。
These numbers are fascinating.
但是，一方面自己被迫花钱，另一方面客户却可以随时走人，而且收入还这么集中在少数几个客户身上，我不知道投资者会不会愿意投这样一家公司。
But this contrast of being locked into spending while your customers are not and having such concentration in a couple of customers, I don't know if investors are going to run towards a company that has those dynamics.
从那个层面看，确实不太好看。
It doesn't look good at that level.
当然，过去几年的增长率在人类历史上基本上是前所未有的，就收入增长水平而言。
Obviously, growth rate over the last couple of years is unprecedented in human history, basically, at that kind of level of growth in terms of revenue.
但我同意，这些明显且现实的威胁确实很严重。
But I agree, there's all these clear and present risks that are dramatic.
再说，这种客户集中度，放在任何其他公司身上都会让人害怕。
Again, that concentration of customers in any other kind of company would be terrifying.
有意思的是，SpaceX，这得归功于马斯克，他们收入达到187亿美元，虽然亏损，但规模不如Anthropic，不过还是在180亿的基础上亏了40亿，后来市值超过一万亿，一度达到两万亿，所以从这个角度看，也许不是不合理，但我觉得，是的，今天十月初的市场，大家的心理状态和几个月前相比，已经大不相同了。
And it's funny because SpaceX, to Elon's credit, they were at $18.7 billion in revenue and losing money, not at the scale of Anthropic, but still losing $4 billion on the 18, and they came out at over a trillion and got to $2 trillion for a bit, so maybe it's not unreasonable from that perspective, but I think, yeah, the market, everyone's mentality is in a very different place today in early October versus where it was a few months ago.
所以我觉得大家会仔细看这份S-1文件，但我想，对你来说，最危险的红旗是什么？
So I think people will look carefully at this S-1, but I guess what would be the most dangerous red flag for you?
计算承诺，我觉得我和你的看法不同。
The compute commits, I think I differ from you.
我不认为那是个那么刺眼、那么令人震惊的问题，相对于业务的核心实际状况来说。
I don't think that is as glaring and shocking an issue relative to the actual core state of the business.
所以客户亏损的集中度，如果我们看到增长放缓，那才是最大的风险。
So the concentration of customers' losses, if we see slowing growth, that is actually the greatest risk of all.
所以那些才是最大的风险。
So those are the biggest risks.
计算承诺，那是在更长的时间尺度上才会显现的。
Compute commits, that plays out on such a longer timescale.
我觉得那最终会变成公司律师们去争论，然后大家都会搞清楚的。
I feel that ends up in kind of corporate lawyers hashing it out and everyone figures that out.
你觉得最大的风险是什么？
What do you think is the biggest risk?
你不觉得现在市场很脆弱吗？
Don't you think the market is really tenuous right now?
当然，我们看到债券市场出现了一些令人担忧的迹象，而且所有投资者都集中在AI和AI相关的公司上。
Of course, we've seen the bond market show some worrying signs, and you just have all these investors concentrated in AI and these AI-adjacent companies.
我会回答你的问题，但让我先说，我一直在我的X时间线上看到这些推文，让我想，嗯，这也许是个问题。
I'm going to answer your question, but let me just say, I've continued to see these tweets show up on my X timeline that make me think, huh, maybe that's a problem.
这是来自Citadel的消息。
This is something from Citadel.
仅微软、英伟达、苹果和Meta在第三季度就为标普指数贡献了约300点，超过了该指数总涨幅的200%。
Microsoft, Nvidia, Apple, and Meta alone added about 300 points to the S&P in Q3, more than 200% of the index's total gain.
其他大约499只股票合计下跌了约150点。
The other 499-ish stocks together subtracted about 150 points.
这是来自彼得·希夫的消息。
This is from Peter Schiff.
标普500指数离高点只差0.7%，但有430只股票比它们的高点低了27.1%。
The S&P 500 is 0.7% below its high, yet 430 of the stocks are 27.1% below their highs.
这意味着，平均来看，86%的股票都处于熊市。
That means, on average, 86% of stocks are in a bear market.
市场广度这么差的情况只出现过两次，一次是1973年1月，另一次是1999年到2000年。
Breadth has only been this bad twice, in January 1973 and in 1999 to 2000.
那两次之后，标普指数都暴跌了将近50%。
On both occasions, the S&P then crashed nearly 50%.
我这么想是不是错了？这可能有点夸张，但我还是想说出来。
Am I wrong in thinking that the-- this might be a hyperbole, but I'm going to throw it out there.
市场基本上就是靠AI这个故事撑着的。
The market is basically being held up by this AI story.
这就是个叙事，对吧？
It's a narrative, right?
这个叙事，我觉得挺有说服力的，就是说像Anthropic和OpenAI这样的公司，他们确实发现了什么大东西。
It's a narrative that, and I think it's a strong narrative, that companies like Anthropic and OpenAI, they're just onto something.
他们发现了这项新技术。
They have found this new technology.
它将改变世界。
It is going to change the world.
它会很快改变世界，而且可能在不久的将来就会非常赚钱。
It's going to change the world fast, and it's going to be very profitable in maybe not too long of a time horizon.
如果这个故事崩塌了，那可能就是你所说的那种情况，我觉得这理由再充分不过了，对吧？
If that story crumbles, it might be what you said, which I think that's as good of a reason as any, right?
这周我听了些投资者的看法，他们说：“听着，说25%的收入来自两个客户，我们知道今年早些时候有传言说一个客户一天就花了，多少，5亿——”
Where I was listening to some investors this week, and they're like, "Listen, this idea that 25% is two customers, we know there was that rumor earlier this year that one customer spent, what, 500-
一天就花光了所有token之类的。
Token maxing in a day or whatever it was
……一天就花了5亿美元之类的，然后我们就想，哦，我们得设些限制。
... 500 million in a day or something like that, and then we're like, oh, we should put controls.
我们知道各地的企业都在设限制。
We know businesses everywhere are putting controls on.
他们公布的数字，我们可能看不到增长放缓。
We probably won't see a momentum slow in whatever numbers they present.
那会很糟糕。
That would be bad.
但任何一点波动都可能开始让股市——我不是说会让股市崩盘，那太夸张了——但可能会让股市回到我认为它一直想去的那个位置。
But any little bit of wobbliness could start to, I'm not going to say bring the stock market down because that's ridiculous, but could make the stock market sort of revert to where I think it's been trying to point.
也就是说，标普500今年涨了13%，全靠这些公司。
Which is like S&P is up 13% this year because of these companies.
你可能会看到它把这些涨幅吐回去，跌个5%或10%，因为人们不再相信这个故事了。
You might end up seeing it give up those gains and go down 5 or 10% because people stop believing in the story.
5%或10%？
5 or 10%?
我觉得如果真这样发展，我不是说会跌50%——
I think if this plays out, I'm not saying 50%-
是全年来看。
On the year.
哦，抱歉。
Oh, sorry.
按年度来看。
On the year.
按年度来看。
On the year.
所以那会比现在下跌20%。
So that would be a 20% drop from where they are now.
我知道你不喜欢听这个，但我必须得说。
I know you don't like this, but I have to say it.
这不是投资建议。
It's not investment advice.
我们只是在这里聊聊天。
We're just having a conversation here.
但你说吧，Ranjan。
But go ahead, Ranjan.
不是投资建议。
Not investment advice.
我不会根据这个来交易。
I'm not trading off of this.
我也不会根据这个来交易，因为市场的非理性已经持续够久了，而且，正如你所说，这里确实有真实的故事。
I'm not trading off of this either because the irrationality of the market has gone on long enough, and also, as you said, there's a real story here.
但对我来说，把这一点和实际的短期叙事区分开来很重要。
But separating that out from the actual short-term narrative is important to me.
而且，实际上，那个数据说标普500中86%的股票处于熊市，它们实际上是市场的绝大多数。
And actually, that stat that 86% of the stocks in the S&P 500 are in a bear market, that they're actually the vast majority of the market.
如果你从实际市场机制的角度想想这意味着什么，一直在发生的事情——我相信很多听众都有这种感觉，我自己也有——如果你投资于这个集中交易之外，无论是实际的半导体领域还是实际的科技公司本身，你不仅错过了机会，而且在那些股票上表现得很差。
And if you think about what that actually means from an actual market mechanics standpoint, what has been happening, and I'm sure many of our listeners feel this, and I have felt this myself, of if you are invested outside of this concentrated trade, whether it's on the actual semi side or whether it's the actual tech companies themselves, you're not just missing out, you're doing badly on any of those stocks.
所以每个人都开始更多地转向这些股票，所以它们背后有更多的动量，这确实一直在发生，如果你不投资英伟达，你为什么投资？
So everyone starts rotating into them more, so there's more momentum behind them, and that has definitely been happening, that if you're not in Nvidia, why are you investing?
如果你不投资博通或其他这些股票，而不是试图寻找市场其他领域甚至科技领域内部的好公司，并试图真正理解它们是不是好企业。
If you're not in Broadcom or any of these other stocks, and versus trying to find what are good companies in other sectors of the market or even within technology itself and trying to actually understand are they good businesses.
再说一次，市场现在的走势并不是这样的，一旦这种心态发生转变，我觉得会非常剧烈。
And again, that is not the way the market's been trading, and any kind of switch in that mentality, I think will be pretty dramatic.
而且我确实觉得，S-1 文件，我尽量不把事情想得太悲观，因为我还是不认为这会导致经济灾难或崩溃。
And I do think the S-1 is like, I'm trying not to go full doomer because I still don't think this is economic calamity or collapse.
但我确实认为，那个简单的叙事会破裂，突然之间，一小撮公司极高的估值开始受到质疑，然后人们真的会开始撤资，或者把钱转回那些在过去一年里被重创的其他公司。
But I do think that simple narrative breaks, suddenly rich valuations, incredibly rich valuations for a very small group of companies, starts to be questioned, and then people actually start to either pull out or move it back to other companies that have been getting hammered over the last year.
所以我觉得 S-1 文件最好表现得好一点。
So I think the S-1, it better be good.
必须得表现好。
It better be.
我们刚刚看到了。
We've just seen it.
这是爱国行为。
It's patriotic.
给出那样的增长率。
Give that growth rate.
我觉得不会再有更多了。
I don't think there's going to be much more.
不，我们还没看到。
No, we haven't.
顺便说一句——
By the way-
我们还没看到。
We haven't seen it.
我们还没看到。
We haven't seen it.
你觉得他们会在上市前加上——
You think they're going to add before they go public-
哦，他们会的。
Oh, they will
……他们会加上这些上半年的数据。
... that they're going to add these first half numbers.
他们必须得这么做。
They will have to.
你根本不可能去——
There's no way you can go-
我们很快就会知道了。
We're going to find out soon
——上市。
... to the market.
再说，这是第一件事，我一直——一般来说，任何S-1文件，你都得展示一些相对近期的数据，虽然你不应该做前瞻性预测，但你应该准确反映公司目前的状况。
Again, this is the first thing, and I kept-- Generally, any kind of S-1, you're going to have to show some relatively recent, again, you're not supposed to make forward-looking projections, but you should give an accurate reflection of where the business is today.
所以如果你要争取2万亿美元的估值，你不能把45亿美元作为你业务经济状况的最后反映，然后还指望市场……再说，通常就是这样运作的，我不知道Anthropic会被允许做什么。
So if you're going for a $2 trillion valuation, you can't throw out $4.5 billion as the last reflection of what the economics of your business are, and then expect that the market... Again, that's how things normally work I don't know what Anthropic is going to be allowed to do.
也许他们会，但他们肯定会拿出更多数据。
Maybe they will, but they're going to come out with more.
会多得多。
It's going to be a lot more.
那个东西出来的时候，我们得赶紧录一期紧急播客，因为——
We're going to need an emergency pod when that thing comes out because-
好的
Okay
——内容肯定很劲爆。
... it's going to be juicy.
我们这边有人提到，我想你那边也是，说：“嘿，路透社那篇报道出来的时候，赶紧录个紧急播客吧。”
We had people in my mentions, I think yours too, saying, "Hey, do the emergency pod when the Reuters story came out."
但我觉得还是等正式文件出来吧，到时候我们可以在YouTube上直播，带大家逐条读一遍。
But I think let's wait till the actual document comes out, and we can read through it for everybody live on YouTube.
顺便说一句，你刚才也提到了，这显然就是为什么特朗普今年在华盛顿跟所有AI公司的CEO待在一起。
And by the way, you hinted at it, but this is obviously why Trump was with all the AI CEOs in Washington this year.
他基本上就是——
He's basically just like-
他参与了AI相关的会议
He's on the meetings in the AI bet
……如果你想想这个方式——
... if you think about the way-
嗯，他是总统。
Well, he's the president.
他希望股市表现好，他基本上就是说：“如果你们再搞这种悲观论调或者放慢速度的事，整个市场都会下跌，那对我的总统任期影响不好。”
He wants the stock market to do well, and he's basically just like, "If you guys do any more of this doomerism or slowdown stuff, the whole market goes down, and that reflects poorly on my presidency.
黄仁勋，你坐我旁边。
Jensen, you're going to sit next to me.
达里奥，你坐得离桌子那么远，我甚至都看不见你。”
Dario, you're going to sit so far down the table, I'm not even going to be able to see you."
别再悲观了。
No more doomerism.
我们要继续建设，因为如果我们停下来，再说清楚一点，这不仅仅是OpenAI或Anthropic的事，更是整个市场的事。
We're going to keep building because if we stop, again, to put it clearly, it's not just an OpenAI or Anthropic story, it's a market story more broadly.
确实是市场的事。
It is a market story.
这肯定是一个更大的市场故事。
It definitely is going to be a larger market story.
你觉得我们得等OpenAI吗？
Do you think we have to wait for OpenAI?
我相信你看到了，他们现在又在筹集300亿美元，所以他们短期内肯定不会消失。
I'm sure you saw they're raising another 30 billion now, so they're definitely not going out in the immediate vicinity.
但你觉得Anthropic会给我们明确的方向，告诉我们这一切会走向哪里吗？
But do you think Anthropic is going to give us clear direction on where this is all going?
是的。
Yes.
我一直认为，Anthropic的走向就是AI交易和AI本身的走向。
I've always thought as Anthropic goes, so goes the AI trade and the AI.
但因为他们如此坚定地押注于纯技术，即使他们现在有一些产品，我一直认为Anthropic的走向就是AI的走向。
But because they have bet so strongly on pure-play technology, even though they have some products now, I've always thought as Anthropic goes, so goes AI.
所以我认为，我们会在那里得到很多清晰的信息，关于市场是否想要这些东西，是否相信它，以及整个叙事能否继续维持，势头能否保持。
So that is, I think, where we're going to get a lot of clarity around whether, A, the market wants this stuff, whether it believes in it, whether this entire narrative can continue to be held up, and the momentum can be held up or not.
所以如果你是马克·扎克伯格——或者扮演马克·扎克伯格——
So if you are Mark Zuckerberg- Or be Mark Zuckerberg-
……就一会儿。
... for a moment.
在总统面前开玩笑——
Cracking jokes with the president-
……趁达里奥讲话的时候在背后搞小动作。
... behind Dario's back as he tries to speak.
回到水翼艇上。
Get back on the hydrofoil.
在新闻发布会上。
In the press conference.
来吧。
Come on.
就带着美国国旗上水翼艇吧。
Just get on the foil with the American flag.
我们都会喜欢的。
We'll all love it.
不，别那样做。
No, just don't do that.
但是——
But-
好的。
Okay.
扔给我吧。
Throw it at me.
好的。
Okay.
所以这里有一种奇怪的竞争关系，因为Anthropic对你做出了支出承诺。
So there's weird competitive dynamics here because Anthropic has spending commits to you.
Anthropic是这个故事的核心。
Anthropic is the basis of the narrative.
再说，Muse 大获成功，Meta 重新回到了游戏里，比我们几个月前甚至一年前能想到的都要厉害。
Again, Muse has been a smash hit, Meta's back in the game more than any of us ever would've given them credit for, even a few months ago, much less a year ago.
但你们是不是想打击竞争对手？
But are you trying to kneecap your competitors?
而且他们在一段时间内一直是开源、开放权重这个话题的代表。
And they are the open source, open weight conversation for a while.
Meta 一直采取这种策略，但如果你非常激进地针对竞争对手，你会不会和 Meta 一起，我们接下来会讨论 Meta 的企业级发布，你会不会大幅降价，就为了打垮 Anthropic 或 OpenAI，同时知道这种后续影响短期内可能会严重冲击你的股价？
Meta's been taking this approach, but by going super aggressively at your competitors, do you go in with Meta, and we're going to talk about Meta's enterprise launch, do you just undercut price massively in an effort to just smash Anthropic and/or OpenAI, knowing that the follow-on effects could actually severely impact your stock price in the short term?
这肯定不是长期商业策略的体现，但你会不会冷酷到采取这种手段？
Definitely not a reflection of longer-term business strategy, but are you cold-blooded enough that you take that approach?
我很感谢这个思想实验，说马克·扎克伯格现在是 AI 胜负的裁决者。
I appreciate the thought experiment that Mark Zuckerberg right now is the arbiter of who is going to win and lose in AI.
但我不这么看。
I just don't see it that way.
不，他是其中之一。
No, he's one of them.
桑达尔也一样。
All of Sundar is, too.
他一直在努力，我们一会儿会深入讨论Muse和Dots的对比。
He's been trying, and we're going to get into Muse versus Dots in a moment.
但扎克伯格从AI竞赛真正开始以来，就一直在试图压低这些公司的价格，但并没有奏效。
But Zuck has been trying since the AI race really began, to undercut these companies, and it hasn't worked.
他们反而一直在继续发展。
They've only continued to do it.
Anthropic，即使以令人失望的估值上市，也会立刻成为比Meta更大的公司。
Anthropic, if it goes out, even at a disappointing valuation, is going to be a bigger company than Meta right off the bat.
所以，是的，我认为扎克伯格会尝试。
So yeah, I think Zuckerberg will try.
记住，他有算力，也有现金。
Remember, he's got compute, he has cash.
所以他会尽一切努力让人们使用他的服务，而不是他们的。
So he'll try to do whatever he can to get people to his service versus theirs.
但区别就在这里，我觉得我们接下来会和Dots聊到这个话题。
But here's the difference, and I think we're going to talk about this with Dots.
说到企业级应用，Muse对消费者来说是个好应用。
When it comes to enterprise, Muse is a good app for consumers.
但在企业级应用上，你真正想要的是最顶尖的智能。
When it comes to enterprise, you really want the best intelligence.
你要的是前沿的智能。
You want the frontier intelligence.
所以这就是为什么我觉得Meta推出的这个企业版Meta，对吧？
So that's why I don't think this Meta for Enterprise thing, which Meta is introducing, right?
这周他们宣布要推出Muse企业版。
So folks this week announced that they're going to be doing this Muse for Enterprise.
我不觉得它能和Anthropic、OpenAI形成很强的竞争，不过我也不怪Meta去尝试。
I don't see it being extremely competitive to Anthropic and OpenAI, though I don't give Meta a hard time for trying.
记住，他们有很好的数据。
Remember, they have good data.
展开剩余字幕（还有 479 条）
他们跟这些大公司都有关系。
They have relationships with all these big companies.
他们了解是什么在推动这些公司的业务。
They understand what drives their business.
所以去争取这些客户是合理的。
And it makes sense to go after them.
那我们要不要聊聊Meta的企业版？
See, should we dig into Meta for Enterprise?
好。
Yes.
行。
Okay.
不过别急，等等。
But no, hold on.
我们应该按这个顺序来。
We should do it in this order.
我们应该先聊聊Dots，然后聊聊和Muse的竞争，最后再聊Meta for Enterprise。
We should talk about Dots, then talk about the competition with Muse, and then talk about Meta for Enterprise.
我们来实时过一遍流程吧——
Let's do real-time run of show here-
实时过流程
Real-time run of show
……给听众们。
... for listeners.
所以我们会在休息之后做这件事，但休息之前我得先做两件事。
So we're going to do it after the break, but I need to do two things before the break.
首先，Ranjan，我得跟你说，这周对我来说简直是彻底搞砸了，因为今天我才发现自己被邀请去白宫参加America.gov的揭幕仪式，而我错过了那条消息。
First of all, Ranjan, I got to tell you, this was a absolute fumble the ball week for me, because today I just discovered that I was invited to the revealing of America.gov at the White House, and I missed the message.
那是——
It was-
你本来可以站在那儿，挨着——
You could've been up there next to-
是啊，我本来可以的。
I could.
我本来会——
I would've been-
就在那些——
Next to all the-
……在那一群人里什么的。
... in the gaggle or something.
不不不。
No, no, no.
但是——
But-
你本来会就在达里奥旁边。
You would've been right next to Dario.
就在达里奥旁边。
Right next to Dario.
你看到达里奥怎么——
Did you see how Dario-
哦，对。
Oh, yeah.
哦，对。
Oh, yeah
——躲到大家后面去了，然后特朗普——
... went back and hid behind everybody, and Trump-
——还是把他叫出来了？
... still called him out?
如果我在那附近，我肯定会躲得比达里奥还远，真的。
If I was anywhere close to there I would be hiding even further behind than Dara did, trust me.
好，第二件事，我想承认我之前犯的一个错误。
Okay, second thing I want to say is just a mistake I want to acknowledge on my part.
上周我说Muse被推广给了Meta的三十亿用户。
Last week I talked about how Muse is being promoted to Meta's three billion users.
当然，它并不是面向那三十亿用户的。
Of course, it's not available to the three billion users.
它只对一小部分人开放，主要是在美国。
It's just available to a small segment of the population, mostly in the United States.
所以，看Meta Muse的增长时，我应该以那个为基础，而不是Meta的所有用户。
And so, looking at Meta Muse's growth, I should've used that as the baseline, not all of Meta's users.
我为这个错误道歉。
Regret the error.
Muse增长得不错，但现在它有了真正的竞争对手。
Muse has grown nicely, but now it has real competition.
所以，感谢那些指出这一点的人，包括Podcast Enjoyer和其他一些人。
So thank you to those who've pointed that out, including Podcast Enjoyer and a bunch of others.
我一直很感激听众的反馈。
I always appreciate feedback from listeners.
我们稍后回来，会讨论Muse和OpenAI Dots的竞争。
We will get into the Muse versus OpenAI Dots competition when we come back right after this.
嗨，我是里克·鲁巴克。
Hi, I'm Rick Ruback.
我是罗伊斯·尤德科夫。
And I'm Royce Yudkoff.
我们是哈佛商学院的教授，也是热门播客《想大买小》的联合主持人。
We're professors at Harvard Business School and co-hosts of the chart-topping podcast, "Think Big, Buy Small."
如果你想成为企业家，从零开始创业并不是唯一的选择。
If you want to become an entrepreneur, starting a company from scratch isn't your only option.
你可以收购一家成熟且盈利的小企业，以CEO身份加入，在已有的基础上继续发展。
You can buy an established, profitable small business, step in as CEO, and build on what already works.
这条路被称为收购型创业，我们花了超过15年时间帮助哈佛商学院的学生探索这一领域。
This path is known as acquisition entrepreneurship, and we've spent over 15 years helping Harvard Business School students explore.
在《想大买小》的新一季中，我们通过与企业家、投资者和行业专家的坦诚对话，把这些经验带给你。
On the new season of "Think Big, Buy Small," we bring those lessons to you through candid conversations with entrepreneurs, investors, and industry experts.
我们一起探讨整个旅程的每个阶段，从寻找和融资合适的业务，到完成交易、领导公司，最终出售它。
Together we explore every stage of the journey, from finding and financing the right business to closing the deal, leading the company, and eventually selling it.
无论你是在认真考虑拥有自己的企业，还是单纯好奇创业的另一条路，《Think Big, Buy Small》都能帮你迈出第一步。
Whether you're actively considering business ownership or simply curious about a different route to entrepreneurship, "Think Big, Buy Small" can help you take the first step.
在Apple播客、Spotify或任何你收听播客的平台关注《Think Big, Buy Small》。
Follow "Think Big, Buy Small" on Apple Podcasts, Spotify, or wherever you get your podcasts.
欢迎回到《Big Technology》播客周五版，今天我们请到了Margins的Ranjan Roy。
And we're back here on "Big Technology" podcast Friday edition with Ranjan Roy of Margins.
Ranjan，这周在常驻AI助手方面可是大新闻不断。
Ranjan, big week this week in terms of always-on AI assistants.
它们不能只是助手，必须时刻在线。
They can't just be assistants, they have to always be on.
这就要说到OpenAI推出的Dots了。
And that is with the introduction of Dots from OpenAI.
我来简单介绍一下。
I'll just introduce it.
这是《纽约时报》的报道：一款新的AI代理进入市场，就是OpenAI的Dots。
This is from "The New York Times," "A new AI agent enters the marketplace, OpenAI's Dots.
OpenAI现在有了人工智能代理，它们充当个人数字助理，这是在周二于旧金山举行的公司年度开发者大会上宣布的。
OpenAI now has artificial intelligence agents that act as personal digital assistants, announced at the company's annual developer conference on Tuesday in San Francisco.
这些代理叫做Dots，将与Meta的AI代理Muse竞争，Muse也是本月发布的，还有谷歌的Gemini Spark，它是五月份推出的，以及Instinct，一家AI代理初创公司，在最近一轮融资中筹集了10亿美元。
The agents, called Dots, will compete with Meta's AI agent called Muse, which was also released this month, Google's Gemini Spark, which came out in May, and Instinct, an AI agent startup that raised $1 billion in its latest funding round."
《纽约时报》里怎么没提Grokbot呢？
Where's Grokbot in "New York Times," hmm?
我们也应该说说，我们的讨论里怎么没提Grokbot？
And we should also say, where's Grokbot in our discussions?
好吧，我提过它了，各位。
Well, I've mentioned it, folks.
Grokbot应该出现在那段话里，对吧？
Grokbot belongs in that paragraph, okay?
Dots是OpenAI在打造AI代理产品方面最激进的尝试，这些代理能做预订餐厅或管理财务之类的事情。
"Dots are OpenAI's most aggressive push to create a product out of AI agents, which will do things like booking restaurant reservations or managing finances.
用户可以在ChatGPT、Slack或Microsoft Teams里，或者通过短信，给他们的Dot代理发消息，这个代理他们可以自己命名，还能连接到第三方应用。
Users can message their Dot agent, which they can name, in ChatGPT, Slack, or Microsoft Teams, or through text and connect to third-party apps.
他们还可以通过电话跟这个代理通话。
They can also talk to the agent via phone call.
这些代理从周二开始对ChatGPT专业版的订阅用户开放，价格从每月100美元起。
The agents will be available starting Tuesday to subscribers of ChatGPT's pro plans, which start at $100 a pop."
我觉得Dots的首秀有点不稳。
Shaky debut, I would say, for Dots.
他们的演示没有做好。
They did not nail the demo.
演示出问题的时候，Meta的股价反而上涨了，这说明市场认为这些产品是直接竞争的，如果OpenAI失误，Meta就能捡到便宜。
As they were struggling with the demo, Meta's stock went shooting up, an indication the market believes that these products are directly competitive, and if OpenAI stumbles, it's Meta's bounty.
那你觉得Dots的发布怎么样，Ranjan？
So what do you think about the release of Dots, Ranjan?
我觉得说到失误，你有没有看到Sarah Friar在CNBC上……
I think in terms of the stumble, did you also see Sarah Friar on CNBC-
哦，看到了。
Oh, yes
……不小心把自己的产品叫成了Muse——
... accidentally called her own product Muse-
Muse
Muse
……然后不得不纠正自己，说那是Dots？
... and had to correct herself to saying it's Dots?
好的。
Okay.
我还没试过。
I have not tried it yet.
我相信你试过了。
I believe you have.
对吧？
Is that correct?
是的。
Yes.
那我给你——
So I will give you-
你能告诉我——
Can you tell me-
……我现在对Dots的评价。
... my Dots review now.
说说你对Dots的评价。
Tell me your Dots review.
这东西太棒了。
The thing is amazing.
真的很棒。
It's amazing.
所以我决定，你知道吗？
So I decided, you know what?
我现在要全力投入，让它完全掌控……
I'm going to go all the way in now and give it full control of...
我本不该说这个，但我让它完全控制我的笔记本电脑和浏览器，并在ChatGPT里操作。
I shouldn't even be saying this, but give it full control of my laptop and my browser and have it in ChatGPT.
而且这太不可思议了。
And it's incredible.
你只要不停地给它发消息，它基本上就能明白你想让它做什么，接手你的工作，然后开始安排执行。
You can just keep messaging it, and it will basically figure out what you want it to do and pick up your work and start sequencing it.
尤其是当你有很多任务堆积的时候，我得承认，或者说我得说，这对我来说太合适了。
And especially when you have a lot of tasks piled up, and I'll admit, or I will say, it's perfect for me.
因为特别是我的工作，有太多事情堆积起来，而且我还想腾出时间做内容创作，这需要深度工作，远离收件箱。
Because in my job in particular, I have so many things that pile up, and I also want to make time for content creation, which requires deep work and moving away from the inbox.
所以事情就会堆积到一定程度，我不得不快速处理完。
So things can sort of just build up to the point where I have to just get things done very quickly.
大量的邮件、大量的发票、旅行规划，所有这些。
Lots of emails, lots of invoices, trip planning, all this stuff.
所以昨晚我就坐在那里，告诉它……好吧，再说一次，我不知道该分享多少，但为了透明起见——
And so last night I just sat there, and I told it... Well, again, I don't know how much of this I should be sharing, but for the sake of transparency-
说吧。
Go
……
...
我这就说。
I'm going to do it.
为了听众，快说吧。
For the listeners, come on.
我跟它说：“尽可能多读我最近的邮件和文章，学学我的写作风格。”
I told it, "Read as many of my recent emails and articles as you can, and pick up my writing style."
然后我就开始翻我的邮件。
And then I started going through my email.
你知道那种感觉吧，打开一封邮件，你很清楚自己想说什么，但要用那种邮件的方式表达出来，既直接又不冒犯，这得花掉90%的精力。
And you know when you open an email, and you know exactly what you want to say, but it's phrasing it in that email way where it comes off as direct but not offensive, you know what I'm saying, takes 90% of the energy.
这家伙一下子就抓住了我的风格，开始起草邮件了。
This thing got my voice right away and started drafting email drafts.
所以我本来打算做的就是这个，但它开始生成一堆邮件草稿。
And so the action I had was the intended action, but it starts piling up these email drafts.
我就说，"写一封这样的。
I'm just like, "Write one of these.
写这个。
Write this.
写这个。"
Write this."
然后草稿就出来了，我可以去草稿箱里稍微改一下，然后发送，很快就把收件箱清空了。
And it comes in the drafts, so I can go into my drafts, alter it a bit, and then send, and just very quickly clear out the inbox.
然后我说，"好吧。
Then I said, "All right.
我需要发一张发票。"
I need to send an invoice."
然后我说，"我的发票在云端硬盘里。
And I said, "My invoice is in Drive.
我需要你进入Drive——下载发票，放到Gmail草稿里，发给平时常发的人，然后写上备注。
I need you to go into Drive-" Download the invoice, put it in Gmail in a draft to the people I usually send it to, and write my note.
它完成得还挺不错的。
And it was able to do that fairly effectively.
然后我就想，好吧，我知道我们还在讨论旅行评估的事。
Then I was like, okay, I know we debate the travel eval.
跟你说一声，我加入。
Just so you know, I'm in.
你加入了。
You're in.
我放弃了。
I give up.
我当时就想：“我干脆把所有事情都交给这个Dot好了。”
I was like, "I am just going to delegate everything to the Dot here."
我周末要去萨尔瓦多旅行，我就想：“这些是我们想去的地方。”
And I have a weekend trip to El Salvador coming up, and I was like, "These are the places we want to go.
我的酒店预算是每晚总共100美元，所以你可以根据每晚的情况稍微高一点或低一点。
My hotel budget is $100 a night cumulatively, so you can go a little bit above and a little bit beneath, depending on the night.
帮我找租车，找酒店，订酒店，然后制定一个我们到那里可以遵循的行程。
Find me the rental car, find me the hotels, book the hotels, and then put together an itinerary that we can follow when we're there."
你让它订了？
You let it book?
嗯，它基本上已经进入预订流程，只需要我点最后一步。
Well, it basically got into booking and needed me to just hit the final step.
所以我选了它选的那些酒店。
So I chose the hotels that it chose.
我点了预订来完成这件事。
I hit the book to complete the thing.
然后它意识到……你听这个。
Then it realized... Listen to this.
等等，等等，等一下。
Wait, wait, hold on.
不不不，我想追问你一下。
No, no, I want to push you.
它给了你几个选项？
How many options did it give you?
还是说你就直接锁定了一个，然后就选了它？
Or you literally narrowed it down to one and went with it?
好的。
Okay.
所以，三个里面有两个是我喜欢的。
So, two of the three I liked.
有一个我不喜欢。
One I didn't like.
好的。
Okay.
还是有人在把关的。
There's still human in the loop.
人在回路中。
Human in the loop.
我就在那个回路里。
I got in that loop.
我要完全的人工。
I want a full human.
你想要全自动的。
You want totally automated.
好的。
Okay.
我想让你在萨尔瓦多——
I want you in El Salvador-
……在一家Dot连锁酒店里。
... in a Dot period hotel.
听听这个。
Listen to this.
它发现其中一家酒店在到达之前有一段越野路。
It realized that one of the hotels had some off-roading before you get there.
所以它就说：“我觉得这能行，但我不太确定。”
So it's like, "I think this will work, but I'm not 100% sure."
然后它又说：“我是不是该给租车公司Budget和酒店发邮件，问问那辆车能不能开到那儿？”
And it's like, "Should I just send emails to Budget, the rental car agency, and the hotel seeing if the car can make it there?"
我说可以。
And I said yes.
然后我看了看已发送邮件，发现用我的语气写好的邮件已经发出去了。
And then I just looked in my sent messages, and perfectly drafted in my voice emails have been sent to these folks.
所以我真的全力以赴了，我觉得因为它为企业打造，又用了前沿智能，真的能帮你干很多活，我得说，我挺佩服的。
So I really went ham on it, and I think because it's built for enterprise, because it's built with frontier intelligence, it really can just bang out a lot of work for you, and I have to say I'm impressed.
跟那个比，Muse在消费类产品上确实不错，但我几乎觉得：“天哪，我真该把Muse删了。”
And compared to that and Muse, Muse is really good for some consumer things, but I almost thought, like, "Man, I should just delete Muse."
我自己都不敢相信，但我确实想删掉它，因为我觉得那个体验太好了。
I can't believe I'm saying it, but I felt like deleting it because I thought the experience was so good.
嗯，所以这就是……
Well, so this is where...
再说，我觉得我们应该区分一下，因为你描述的并不是那种意义上的企业级应用。
And again, I think we should differentiate because what you're describing is not enterprise in that sense.
而且，我可以深入探讨一下——
And again, I can dig into-
嗯，发票处理是，邮件也是，但旅行不是。
Well, the invoicing is, the emails are, the travel is not
好的。
Okay.
有道理。
Fair.
对。
Yes.
不过，它仍然是工作。
Still, it's work.
这不是企业级的。
It's not enterprise.
这是工作。
It's work.
在隐私和安全方面，规模不算大。
It's not large scale in terms of privacy, security.
哦，我们得聊聊这个。
Oh, we've got to talk about this.
不不。
No, no.
对，但企业业务是卖给企业客户的，而不是个人消费者。
Yeah, but enterprise business is selling for work clients versus consumer clients.
不是吗？
No?
不不，但等等，当这些公司谈论的时候……再说一次，这是我的领域，我每天都要处理这些。
No, no, but wait, when any of these companies are talking... Again, this is my world, what I have to deal with every day.
当这些公司说企业级的时候，他们指的不是中小企业，我其实觉得Meta应该去争取中小企业，而不是……
When these companies are saying enterprise, they're not talking about SMB, small and medium-sized business across, which I actually think Meta should go after and not...
而且我们可以讨论一下他们真正做好企业级的可能性。
And we can get into the likelihood of them actually nailing the enterprise.
但是大型复杂组织的安全，那个发票，需要谁来批准？
But large complex organization security, that invoice, who does it have to get approved?
那还是企业级的东西，而不是个人，个人对某些工作有完全的管理权，然后责任就到此为止。
That's still the kind of stuff for enterprise versus individual, kind of has full say in management over some set of work that's being done, and then that's the end of the responsibility.
所以我要反驳这一点，因为我真的认为这很重要，关系到Dots以及这类工作和这类代理真正能发挥潜力的地方。
So I'm going to push back on that side because I actually think it is important in terms of where Dots and this kind of work and these kind of agents are actually going to have a lot of potential.
但我确实认为，对我来说，这周最有趣的部分是，我还没试过Dots，但我实际上回去看了……
But I do think, to me, the most interesting part of this week, and I've not tried Dots yet, but I actually went back...
我甚至没意识到Manus，Meta在2024、25年收购的那家。
I didn't even realize Manus, who Meta bought in 2024, '25.
他们是最早为消费者提供某种自然语言驱动的工作流生成的之一。
They're one of the earliest that had some kind of natural language-driven workflow generation for the consumer.
他们不得不取消那个，而Manus推出了自己的个人代理，叫Q，有点像那种小众潮流的。
They had to unwind that, and Manus is out with their own personal agent called Q, which is kind of like the hipster.
我得说，这周我一直在玩那个，试了试。
I'm going to say I was playing with that one this week and trying that out.
他们桌面上其实有个视频编辑器，挺酷的。
They actually have a video editor in line on the desktop, which is kind of cool.
重点是，所有这些已经开始让人觉得都差不多，简直就像每个人都在做同样的东西，但我还是觉得最终取决于你信任谁。
Main point being, all these are already starting to feel interchangeable, and it's literally like everyone is building the same thing in this type of product, but I still think it's going to come down to who do you trust?
我觉得有意思的是，你把车钥匙给了OpenAI，但没给Muse完整的钥匙。
I think it's interesting that you gave OpenAI the keys to the car, but you didn't give Muse the full keys to the car.
我把信用卡给了Instinct。
I gave Instinct my credit card.
我没给Muse。
I didn't give it to Muse.
但后来我给了Muse我的Amex登录，因为我想让它查一下我的白金福利，它做得不错。
But then I gave Muse my Amex login because I was curious for it to check in on my platinum benefits, and it did a good job on that.
所以我觉得，随着这些产品开始趋同，你觉得你会怎么区分它们呢？
So I think, how do you think you're going to differentiate as all these products start converging?
我觉得你说得对。
I think you're right.
归根结底是信任问题，我其实不太信任它们中的任何一个，但有些我愿意容忍，有些我不愿意。
It comes down to trust, and I don't really trust any of them, but there are some that I'm willing to tolerate and some that I'm not.
就像我说的，我甚至不——
And like I said, I don't even-
你是什么意思？
What do you mean?
不过，容忍是什么意思呢？
What's tolerate, though?
容忍基本上就是愿意把我的信息给它们。
Tolerate means basically give my details.
我其实试着把我的信用卡给Dots，但它不让我给，这挺有意思的。
I actually tried to give Dots my credit card, and it didn't let me, which was interesting.
真的吗？
Really?
但其实挺奇怪的。
But actually, it's weird.
比如说，我最近想让Claude Cowork帮我签一份合同。
Some of these things, for instance, I've tried to get Claude Cowork to sign a contract for me recently.
它不肯签。
It wouldn't do it.
然后我就想，那我接下来要去……
And then I was like, so I'm going to be going to...
这个我都不该再说出口。
I shouldn't even say this again out loud.
我要去参加Web Summit。
I'm going to be going to Web Summit.
Web Summit有个演讲者授权书，我都签过一千遍了。
There's a speaker release that you sign for Web Summit that I've signed 1,000 times.
然后我就去了Dot，我说：“去我的邮箱里把那份演讲合同签了。”
And I just went to Dot, and I was like, "Go into my email and sign that speaker contract."
它真的签了。
And it did.
这在法律上有效吗，Web Summit？
Is that legally binding, Web Summit?
不。
No.
如果这次我被Web Summit赶出来，我就在这个节目上大嘴巴。
If I get thrown out of Web Summit after this, I'm going to shoot my mouth on this show.
不，那会是个好故事，值得。
No, it'll make for a good enough story that it'll be worth it then.
其实，这就是头条新闻。
Actually, that's the headline right there.
不是吧，老兄。
Not, man.
我跟我的经纪人签了那个网络发布协议。
"I had my agent sign the web release."
我想去里斯本。
I want to go to Lisbon.
好的。
All right.
对，但是呢。
But yeah.
不，我觉得——
No, I think-
这标题自己就出来了，"我的AI签了合同。"
It would be the headline writes itself, "My Dot signed a contract.
现在我去不了Web Summit了"之类的。
Now I can't go to Web Summit" or something.
但我觉得我最后还是会去的。
But I think I'll end up going.
我对这件事感觉挺好的。
I feel good about it.
不，好吧。
No. Okay.
那你打算怎么……
So where do you...
那到时候你的差异化优势是什么？
What is going to be the differentiator then?
TikTok 可能也会推出一个。
TikTok is probably going to release one.
我一直听到一些传闻，比如……
I've been hearing rumors like-
哦，TikTok 啊。
Oh, TikTok.
哦，我当然想去。
Oh, I definitely want to.
我会把我的出生证明给他们。
I'll give them my birth certificate.
嘿，现在它是美国公司了。
Hey, it's American-owned now.
所以没问题。
It's all okay.
不过，是的。
But yeah.
我大概知道哪些社交媒体公司我不想把这类信息托付给它们。
I generally know social media companies I do not want to trust with this stuff.
另外，我确实有笔账要跟你算。
Also, I do have a bone to pick with you.
我认为前沿智能很重要。
I think frontier intelligence matters.
我觉得模型越聪明，你得到的就越多。
I think you get more if you have a smarter model.
你不这么认为吗？
You don't think so?
哦，等等。
Oh, wait.
抱歉。
Sorry.
等一下。
Hold on.
我并不是反对那个，但关于那一点——
I didn't argue on that, but actually, on that point-
……不，其实我要反驳一下，因为你描述的大部分任务并不需要前沿智能。
... no. Actually, I will push back on that because I think the majority of the type of tasks you describe do not require frontier intelligence.
不需要。
They don't.
尤其是少量的数据点，比如一小部分信息语料。
Especially small amounts of Dots, like a small corpus of information.
其实我们的收件箱信息量都不大，处理起来挺简单的。
Any of our inboxes are actually relatively small volumes of information to process, work through.
创建一个符合这种需求的语音或技能，对AI来说也是挺简单的任务。
Creating a voice or a skill that is in line with that's a pretty simple task for AI, too.
这大概就是GPT-4、5，或者o3的水平。
That's like GPT-4, 5, maybe o3.
可能我们还在o3的范围内。
Maybe we're still in o3 territory.
但如果前沿智能开始偏向于处理这些任务，那对这些公司来说成本会非常高。
But actually, frontier intelligence, if they start erring on the side of that or kind of biasing towards that, this is going to get very expensive for these companies.
尤其是这种持续在线的工作方式，要不断去挖掘你的收件箱，工作量会很大。
The amount of work that's being done in that to go do, especially this always-on nature of work, to go be mining your inbox.
再说，Instinct，我一直很喜欢你，但你真的不用那么费劲地提醒我某封邮件，比如“嘿，顺便说一下”这种。
Again, Instinct, and I've been loving you for a while, but you don't have to try so hard about reminding me of some email as like, "Oh, hey, by the way."
他们需要用户参与度嘛。
Well, they need the engagement.
这就是关键所在。
That's the thing.
我知道。
I know.
不，不，我知道。
No, no. I know.
而且那可能会变得很奇怪。
And that can get weird.
不，Muse一直在烦我——
No, Muse has been hammering me-
Muse根本停不下来。
Muse does not stop.
哦，我也是。
Oh, me too
……不停地。
... nonstop.
不停歇。
Nonstop.
而且就像，兄弟。
And it's like, dude.
你猜 Muse 这周给我发了什么？
Guess what Muse sent me this week?
你收到了什么？
What do you got?
就是提醒一下，另外说一句，加州总检察长 Rob Bonta 昨天向 OpenAI 发出了调查传票，涉及他们 AI 模型的网络安全风险。
Just a heads-up that, on a separate note , California AG Rob Bonta issued an investigative subpoena to OpenAI yesterday over cybersecurity risks around its AI models.
顺便提一下。
Just as a by the way.
然后我就想，"等等，我让你给我发 AI 新闻了吗？"
And I'm like, "Wait, did I ask you to send me AI news?"
等等，什么？
Wait, what?
然后我就想，"不，这有点越界了。"
And it's like, "Nah, that was an overreach."
我当时就想——
I was like-
等等，你在琢磨什么呢？
Wait, what are you musing about?
你是在琢磨你的新闻吗？
Are you musing your news?
你会琢磨你的新闻吗？
Do you muse your news?
我才不琢磨我的新闻呢。
I don't muse my news.
它肯定是发现我是记者，就开始主动给我发新闻标题了。
It must have figured out that I'm a reporter and started sending me headlines unprompted.
好吧。
Okay.
那倒也是。
That's fair.
我还没收到任何——
I haven't gotten any-
但这是它正在尝试对付的头号竞争对手。
But this is its top competitor that it's trying.
它就像在说：“嘿，顺便提一下——
It's like, "Hey, by the way-
哦，对。
Oh, yeah.
确实。
That's true
……说到信任，OpenAI 有网络安全问题。
... speaking of trust, OpenAI has cybersecurity problems.
加州总检察长。
California AG.
Muse——
Muse-
我是说——
I mean-
……从来不是问题。
... never a problem.
Facebook可能是个问题，但Muse不是Facebook。
Facebook may be a problem, but Muse is not Facebook.
不，我觉得那种渴望肯定会开始成为更大的问题。
No, I think that thirstiness is definitely going to start to be more of an issue.
而且实际上，我在想，Meta一直在高调宣传日活跃用户数，我肯定他们会被追问这种参与度的问题。
And actually, I wonder, again, Meta has been from the rafters shouting daily active user numbers, and I'm sure they're going to be pushed on this level of engagement.
所以这会是用户体验上一个有趣的点，因为要在给人们有价值的东西之间保持一个非常微妙的平衡。
So that's going to be an interesting user experience thing, because it's going to be such a sensitive balance between giving people something valuable.
说实话，我上周就说过，当主动性有价值的时候，那是AI最神奇的体验。
Honestly, and I said this last week, proactivity, when it's valuable, is the most magical experience in AI.
这是我迄今为止遇到过的最神奇的经历之一，就像‘哦，顺便说一下，我刚发现这个东西，我觉得它对你很重要。’
One of the most I've had to date of like, "Oh, by the way, I just realized this thing, and I think this will be really important to you."
当它真的有用时，那就是改变游戏规则的事情。
And when it is, that's a game changer.
当它说‘哦，顺便说一下，你两天前找过晚餐预订’，实际上，这在Muse上就发生过。
When it is, "Oh, by the way, you had looked for a dinner reservation two days ago," and actually, this happened with Muse.
我周五找过晚餐预订，然后周日它去Resy上查了一下，发现现在有空位了，这挺好的。
I'd looked for a dinner reservation on a Friday, and then on Sunday, it went to Resy and looked and found that it was now free, which that's nice.
但那不是我想要的。
That's not what I was looking for.
所以，是的。
So yeah.
当他们有点越界的时候，事情会变得有趣。
As they kind of overreach, it's going to be interesting.
他们得保持冷静。
They got to play it cool.
Instinct 挺酷的。
Instinct's kind of cool.
Instinct 的写作和说话方式都挺随性的。
Instinct has the way it actually writes and speaks is kind of laid back.
我其实看过创始人 Noah Shinn 聊过这个，他说你的智能体不应该表现得太急切之类的。
And I actually saw the founder, Noah Shinn, talk about that, though, that your agent should not be overly eager and stuff.
所以这会是一场有趣的较量。
So it's going to be a fun battle.
我想花一分钟聊聊这个智能水平的问题。
I just want to talk about this intelligence thing for a minute.
你可能会说前沿智能水平不重要，但我发现，尤其是 OpenAI 和 Anthropic 的最新模型，跟其他模型相比，它们的深度和思考细致程度远远超过了它们之前的版本，当然也超过了竞争对手。
You might say that frontier intelligence doesn't matter, but I have found that, especially with the latest models from OpenAI and Anthropic compared to everything else, there's just a level of depth and thoughtfulness that far exceeds even the previous versions of their models and certainly their competitors.
那是什么意思？
What does that mean?
那具体是什么样子的？
What does that look like?
这意味着它们似乎更懂事了。
It means they seem to understand things better.
我觉得这是最好的说法。
I think that's the best way to put it.
比如呢？
Like what?
说到写作，好吧，如果我放一个故事草稿进去，对吧？
And with writing, okay, if I put a story draft in, right?
所以很多时候，我不会让AI替我写。
So oftentimes, I won't have AI write for me.
我反对那样做。
I'm against that.
至少，它不会写我的新闻报道，或者我写的关于重大科技的文章。
At least, it won't write my news stories or anything that I write on big technology.
但很多时候，我会把故事和采访内容扔进去，然后问：“我漏了什么？”
But oftentimes, I'll throw a story and interviews in and be like, "What did I miss?"
或者把一篇报道和不同的文章放进去，然后问：“我漏掉了什么？”
Or a story and different articles in and say, "What did I miss?"
我现在发现，它在诊断故事中发生了什么方面做得更好。
What I found now is that it will do a much better job diagnosing what's going on in the story.
它在处理不同输入方面做得更好，比如来自网络各处的文章，澄清事物的含义，做一些像“哦，那真的很聪明”的事情，而不是“好吧，那有帮助，但还是非常基础”。
It will do a much better job taking in different inputs, like articles from around the web, clarifying what things mean, doing things that like, "Oh, that was really smart," as opposed to, "Okay, that's helpful, but still very basic."
所以我确实认为前沿很重要。
So I do think the frontier matters.
所以至少从我个人层面来看，这就是我所看到的。
So that's at least what I see on my side from a personal level.
但实际上，我有一个问题想问你，从企业的角度来看，前沿能做什么。
But actually, I have a question for you on what the frontier can do from an enterprise standpoint.
那么，当公司处理你定义的企业问题时，这些下一代模型能让它们做到以前做不到的事情吗？
So what are companies able to do with these next-level models when they work on your definition of enterprise problems that they couldn't do previously?
不，但这是我们亲眼所见，而且你在每一份支出报告里都能看到。
No, but this is what we've seen firsthand, and you see it in every ramp spending report that comes out.
企业正在远离前沿模型，因为从成本角度和实际要完成的工作或任务来看，大家都意识到，其实你刚才提到的那些让你觉得Dots很神奇的功能，并不是前沿任务。
The enterprise is moving away from frontier models because from a cost perspective and from the actual jobs to be done or the tasks that need to be done, everyone is recognizing that actually, again, all those things that you started outlining that were magical to you about Dots are not like- frontier tasks.
它们只是有限的信息集，用来做相当……
They're limited sets of information to do fairly...
我不是要贬低它的神奇之处，我是说它们不需要，我们又不是在测序DNA。
I'm not trying to demean the magic of it, I'm saying they're not requiring, we're not sequencing DNA here.
每个人似乎都在Astra或3D模拟环境上构建世界模型。
Everyone seems to be building world models on Astra or 3D-simulated environments.
但是写邮件、理解文本，再说一遍，任何文本理解，只要涉及有限步骤的逻辑和推理，我说的有限是指数量有限，但仍然可能是几十步，甚至几百步。
But email writing, understanding, again, anything that's text understanding that has some finite number of steps of logic and reasoning, and by finite I mean some limited number, but that still can be tens, maybe even hundreds.
去找这么多信息，拉进来，处理它。
Go find this much information, pull it in, process it.
那已经不再是前沿了。
That's not frontier anymore.
所以这就是为什么，对我来说，再说一遍，这些都在所有的ramp数据里，而且要把这一切联系起来，对Anthropic来说是个风险。
So that's why, to me, and again, it's been in all the ramp data and it's going to be an anthropic risk to tie it all back together.
企业不再只是把钱都砸在最前沿的模型上。
Enterprises are no longer just pouring money only into the frontier.
现在资金分散到了开源权重模型，很多公司也在自己开发。
It's getting distributed across OpenWeight models, people are building in-house.
好，我明白了，但我想问的是，什么是前沿任务？
Okay, I get it, but I'm asking you, what is a frontier task?
这就是我想说的。
That's what I'm trying to say.
有哪些任务是前沿智能能处理或者比这些标准任务做得更好的？
What are the tasks that frontier intelligence can tackle or do better on than these standard tasks?
不，但这就是关键，如果你这周末要测序DNA，那你就得用前沿模型。
No, but that's where, if you've got to sequence some DNA this weekend, you've got to go frontier.
再说，我见过的那些让人吹嘘的酷炫用例，记得六个月前，那种YouTube上搞钱的那种AI视频，就是“我自动化了我的生活，第二大脑”那种。
Again, every cool use case I've seen people kind of bragging about, remember, six months ago, the kind of YouTube hustle bro AI type of thing would be, "I just automated my life, second brain."
就是你描述的那些东西。
The kind of things you're describing.
现在，大家，再说一次，我也不确定。
Now, everyone, again, I don't know.
大家好像都在搞虚拟世界和电子游戏，还有——
Everyone seemed to just be building virtual worlds and video games and-
我可没干那个。
I did not do that.
不过我确实看到了那些例子。
Although I did see all the examples
……不，但这就是Astra——
... No, but that's where Astra-
……你看我刚做的这个，超级复杂的图形，各种东西到处飞，这就是我看到的真正有意思的东西。
... is, look what I just did, this super complex graphic with stuff going everywhere, and that's all I've seen that looks genuinely interesting.
相比之下，那些用接近前沿模型就能搞定的事，就不算什么了。
Versus something that can't be done with something that's close to a frontier model.
所以我觉得开源权重模型是个巨大的威胁，因为以前开源权重模型总是落后12到18个月，现在只落后三到六个月了。
Which is why I think OpenWeight is such a massive threat in the sense of, in the past, OpenWeight would always be 12 to 18 months behind, and now it's three to six months.
这就让我们回到了Anthropic IPO的问题上。
Which brings us back to this anthropic IPO question.
好的。
Okay.
这显然是我们会争论很久的话题。
And that's obviously something that we're going to be debating for a long time.
没错。
All right.
所以，我们拭目以待吧。
So, we'll see.
这场竞争非常激烈，局势未定，而且越来越有看头。
Very interesting battle, unsettled, continuing to get more competitive.
在结束之前，我想跟你聊聊视觉图灵测试。
Before we go, I want to talk to you about the Visual Turing Test.
有一家公司叫Tavus。
So there's a company called Tavus.
我之前从来没听说过他们，但他们一直在做AI视频头像。
I had never heard of them before, but they have been making AI video avatars.
我就直接读一下这个故事。
And I'll just read the story.
这是来自《The Rundown》的报道：“AI公司Tavus预览了Griffin，一个在视频通话中能做出逼真反应的AI。”
This is from The Rundown: "AI Tavus Previews Griffin, a lifelike AI that reacts on video calls.
这是一个AI模型，能渲染出一个逼真的人，在实时视频通话中能听、能看、能说、能反应。
It's an AI model that renders a lifelike person who can hear, see, speak, and react during a live video call.
这家初创公司表示，在一次简短的公司测试中，近一半的参与者以为自己在和真人对话。”
The startup says nearly half of the participants in a short company test believed they were talking to a human."
我们放一下片段，让大家看看这个AI头像到底长什么样。
Let's play it for a moment just so folks can see what this AI avatar actually looks like.
你有没有再想想我们怎么介绍Griffin？
Have you thought any more about how we're introducing Griffin?
我不知道。
I don't know.
你有什么想法？
What are your ideas?
我觉得我们别再解释了，直接给他们看看效果就行。
I think we should stop trying to explain it and just show them what it looks like.
好的，Ranjan。
All right, Ranjan.
这真的让我大开眼界。
That completely blew me away.
你对此有什么反应？
What was your reaction to this?
当然，这项研究只是一个小型公司测试，所以参与者可能本来就是相信AI的人。
And of course, this study was a small company test, so it might've been true believers anyway.
但是，视频图灵测试，或者说视觉图灵测试，已经通过了吗？
But has the Video Turing, or the Visual Turing Test been passed?
我觉得已经通过了。
I think it has.
我觉得——
I think-
这确实有点疯狂。
It's kind of crazy
……当你观看视频的时候——
... when you watch the video-
那真是太不可思议了。
That was insane.
不，不。
No, no.
当你观看视频时，实际上让我感兴趣的是所有的延迟。
When you watch video, and actually, what was interesting to me is it's all the latency.
实际的像素保真度，我觉得其实还没有完全解决，但肯定已经达到了，你能创造一个能相对快速适应的人类吗？
The actual pixel fidelity, I feel has actually not been fully solved, but has definitely been like, can you create a human that kind of adapts relatively quickly?
实际上需要多长时间才能……
How long does it take to actually...
你能渲染出那个人吗？
Are you able to render that person?
但要做到不是那种效果，因为，我也不确定，你看过那个Tilly Norwood的东西吗？
But doing it where it's not, because even, I don't know, have you ever seen this Tilly Norwood stuff?
我没看过。
I have not.
就是那个AI女演员。
It's like this AI actress.
有一个，她上了皮尔斯·摩根的节目。
There's one, she was on Piers Morgan.
她——
She-
哦，我看过。
Oh, I did.
那个太糟糕了。
That was terrible
……开始表现得怪怪的。
... started acting strangely.
这其实挺好的。
This is actually good.
对，我就是这个意思。
Yeah, that's what I mean.
她开始说中文了，对吧？
She starts speaking Chinese, right?
突然就开始了。
Out of nowhere.
这个玩意儿不小心开始说中文了，有意思的是，我觉得这就像那种老一辈人玩的Meta Facebook上的怪事。
Accidentally, this thing started speaking Chinese, and it's funny because I feel that's one of those weird boomer Meta Facebook things.
所以对，就是那个，有延迟。
So yeah, that, there was a latency.
你能感觉到有很长一段停顿，模型在运行、处理、推理，然后才得出答案。
You could feel there's this long pause while the model is running and processing and inferencing, and they come up with the answer.
而这个就是很自然。
Whereas this one is just natural.
我得试试所有AI的东西，比如谷歌的Frontier模型，我们都没聊过，因为他们还没——
I got to try, with everything in AI, like Google's Frontier model, which we didn't even talk about because they haven't even-
哦，对。
Oh, yeah
……还没发布给我们所有人。
... released it for all of us.
Gemini Pro。
Gemini Pro.
不是。
No.
Argon。
Argon.
Argon。
Argon.
Argon。
Argon.
Argon。
Argon.
好的。
Okay.
谷歌，求你快发布吧。
Just release it soon, Google please.
恭喜各位。
Congrats, guys.
但确实非常令人印象深刻。
But it was very impressive.
真的非常非常厉害，同时也挺吓人的。
It was very, very impressive, and it's terrifying as well.
好吧，关于这个的梗图简直太离谱了。
Okay, the memes on this were just unbelievable.
Autism Capital 简直是在疯狂刷屏。
Autism Capital was just on a rampage.
他们写道：“你爸要把你的全部遗产送人了”，还配了那个人的视频。
They wrote, "Your dad is about to give away your entire inheritance," quoting this video of this person.
然后他们还写道：“这会在 Facebook 上的婴儿潮一代群体里像 80 年代的可卡因一样火爆。”
And then they also wrote, "This is going to hit the Facebook boomer community like crack in the '80s."
这些玩笑确实好笑，但也真的让人好奇，这到底是干嘛用的？
And the jokes are funny, but they really do kind of make you wonder, what is this for?
因为如果我是骗子，我肯定高兴得跳起来了。
Because it seems like if I'm a scammer, I am jumping up and down and thrilled.
等等，什么——
Wait, what-
也许是一些客服的例子，但我不明白这有什么用。
Maybe some customer service examples, but I don't get what it's for.
你觉得这是干嘛的？
What do you think it's for?
那到底是怎么骗人的呢？
What is the scam, though?
我是真心想问这个问题。
This is a genuine question.
人们是不是在和完全陌生的人视频通话，还是不是？
Are people having video calls with- total strangers or not-
你想想看——
If you think about-
我猜是在面试的时候，比如如果——
I guess in an interview, like if-
……那些短信、网络钓鱼，那些“嘿，你好吗，美女？”
... the text message, the catfishing, those "Hey, how are you, beautiful?"
那种你莫名其妙收到的短信。
texts that, I don't know, you get randomly.
不，对。
No. Yeah.
我承认，我时不时也会收到这些。
I'll admit it, I get them every now and again.
就是那些杀猪盘短信。
The pig butchering texts.
我们在节目里聊过这个。
We've talked about it on this show.
老听众都知道，我们之前提过。
For longtime listeners, you know we've talked about it.
如果你能随便发一条这种消息，而且有50%的人——如果你相信那个调查的话——会当真，那杀猪盘是不是就更容易得手了？
Doesn't it get easier to do pig butchering when you can throw one of these things out there, and 50% of people-
……如果你相信那个调查的话，会当真。
... if you are to believe the survey, believe it's real.
这挺吓人的。
It's scary.
对，我觉得这是个问题。
No, I think it's a problem.
这东西就像……
This stuff is like...
我也不知道。
I don't know.
偶尔我会刷一下脸书，上面那些乱七八糟的AI内容已经够吓人的了。
Occasionally, I'll go on Facebook, and the amount of just random AI stuff on there anyway is kind of terrifying already.
虾仁耶稣，宝贝。
Shrimp Jesus, baby.
虾仁耶稣。
Shrimp Jesus.
我们真是……
We're so...
这就是虾仁耶稣的下一个进化版。
This is the next evolution of Shrimp Jesus.
你现在可以跟虾仁耶稣聊天了。
You can talk to Shrimp Jesus now.
但其实，我也不确定。
But actually, I don't know.
你觉得他们会提供有限的头像选择，还是让你自己生成头像？
Do you think they'll have a limited set of avatars, or you'll generate your avatar?
AI领域就是这样发展的。
This is how it goes in AI.
有一家公司做得比较正规，设置了一些防护措施，然后其他公司学着做，但完全不设防。
There's one company that kind of does it properly, and they have some guardrails, and then other companies figure out what they're doing, and they do it without guardrails.
所以，是的。
So yeah.
我觉得迟早会有人做出一个虾耶稣的头像，甚至可能因此开启一个新的宗教。
I think there will be a Shrimp Jesus avatar out there at some point, and it may even begin a new religion.
好吧。
All right.
那么，Tavis，如果你不为广大听众做那个创业项目——一个真正零延迟、能跟人对话、让人以为虾耶稣真实存在的AI头像，甚至可能……
Well, Tavis, if you're not doing that startup idea for listeners everywhere, Shrimp Jesus AI avatar that actually has zero latency and is able to speak to people, and actually make them think Shrimp Jesus is real, and maybe...
这周我们还没聊到Anthropic、天主教和意识呢。
We didn't even get into Anthropic and Catholicism and consciousness this week.
天哪。
Oh my God.
对。
Yes.
但也许下周，全是关于宗教和AI的话题。
But maybe next week, it's all religion and AI.
我觉得我们应该——
I think we should do-
还有虾耶稣。
And Shrimp Jesus
……我们下周再录一期，继续聊这些，因为，天哪，新闻根本停不下来。
... let's do another show next week and keep talking about this stuff because, man, just the news doesn't end.
所以我不知道你怎么样，但我周末会跟Tavis聊，跟这个虚拟形象聊，估计等我们聊完，就准备好录下一期节目了。
So I don't know about you, but I will be spending my weekend talking to Tavis, talking to this avatar, and I guess by the time our conversation is done, we'll be ready to record another episode of the show.
对。
Yep.
周末计划。
Weekend plans.
我打算让它给我读下一份Meter研究报告。
I'm going to have it read the next Meter research report to me.
你能想象那个画面吗？
Can you just imagine that?
就是念80页关于机器人逃跑的内容。
Just reciting 80 pages of bots escaping.
但朗读的是个机器人。
But the narrator is a bot.
是个机器人。
Is a bot.
这就是我们的世界。
That's our world.
对。
Yep.
“真是又酷又正常，各位。
"Really cool and normal, everybody.
真是又酷又正常。”
Really cool and normal."
好的。
All right.
下期节目见。
See you next show.
哇，Ranjan，这期节目过得真快。
Wow, Ranjan, that show went by fast.
我们怎么这么快就结束了？
How have we done already?
确实很快。
That was fast.
我知道。
I know.
天哪。
Goodness.
我知道。
I know.
好的，兄弟。
All right, man.
和往常一样，聊得很开心。
Great talking as always.
谢谢你，Ranjan。
Thank you, Ranjan.
感谢大家的收听和观看，我们下次在《Big Technology Podcast》再见。
Thanks everybody for listening and watching, 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。