本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
欢迎各位收看《The Information》的TITV节目。
Welcome everyone to The Information's TITV.
我是阿卡什·帕什里卡。
My name is Akash Pasricha.
今天是6月23日，星期二。
It is Tuesday, June 23.
今天节目一开始，我们来聊聊企业正在大幅增加对Anthropic和OpenAI的AI预算。
We are starting off the show today focusing on the ballooning AI budgets companies are spending on Anthropic and OpenAI.
我们的AI记者劳拉·布拉顿加入我们，解读她最新的报道：企业如何寻找更便宜的替代方案。
Our AI reporter Laura Bratton joins us to break down her latest story on how businesses are hunting for cheaper alternatives.
接下来，我们请来了Superhuman的首席执行官，这家公司以前叫Grammarly，来谈谈他们最新的收购案。
We've then got the CEO of Superhuman, formerly known as Grammarly, coming on the show to talk about their latest acquisition.
我们还会邀请报道马斯克的记者，聊聊特斯拉和SpaceX的股价是否真的有关联。
We'll bring on our Elon Musk reporter to talk about whether Tesla and SpaceX's share prices are at all interrelated.
最后，我们会深入探讨AI代理循环，以及它对您日常工作的实际影响。
And we're going to close out the show with a deep dive into AI agent loops and what they actually mean for your daily workflow.
这期节目一定会很精彩，咱们马上开始。
It's going to be a great show, so let's get right on into it.
AI模型依然很贵，有些甚至越来越贵，这让客户别无选择，只能寻找更便宜的替代方案。
AI models remain pricey and in some cases are getting more expensive, which has left customers no choice but to seek cheaper alternatives.
我的同事劳拉·布拉顿和凯瑟琳·佩洛夫今天发表了一篇文章，讲为什么这并不总是那么难妥协。
My colleagues Laura Bratton and Catherine Perloff published a story today about why that isn't always such a difficult trade off.
我想请劳拉来分享一下她的发现。
I want to bring on Laura to share more about what she found.
劳拉，欢迎再次做客节目。
Laura, welcome back to the show.
很高兴你来。
It's great to have you here.
很高兴能来。
Good to be here.
现在模型成本上涨，对企业的困扰有多大？
How big an issue is it right now that model costs are going up for companies?
这个问题正变得越来越严重。
It's becoming more and more of an issue.
我接触的那些公司说，他们从Anthropic和OpenAI收到的账单不断上涨，已经开始影响利润空间。
The companies that I spoke to said that the rising bills that they're seeing from Anthropic and OpenAI have begun to affect their profit margins.
因此，他们正在寻找更便宜的替代方案。
And because of that, they're looking to use cheaper alternatives.
如果能以更经济的方式使用AI来推动业务增长——这本来就是使用这些工具的初衷——那他们当然会这么做。
Because if they can use AI in a cost efficient way to grow their business, which is the intention of using the tools in the first place, that's obviously what they wanna do.
明白了。
K.
那他们是怎么降低成本的呢？
So how are they lowering their costs?
他们正在从Anthropic和OpenAI的高端前沿模型，转向更便宜的替代方案。
So they're looking to use cheaper alternatives from Anthropic and OpenAI to the, you know, latest and greatest frontier models.
比如，有时候他们会用OpenAI的GPT-Mini或者Anthropic的Haiku模型，而不是Opus或Sonnet。
So in some cases, that might be OpenAI's GPT mini or Anthropic's haiku models rather than Opus or Sonnet.
在某些情况下，他们还会考虑使用开源模型，或者自己开发小型定制模型，也可能选择像Cohere这样的公司提供的较小或中等规模的模型。
And then in some cases, they're looking to use open source models or develop their own custom small models, you know, or using a smaller or medium sized model from a company like Coke here.
这正是他们降低开支的主要做法。
And that's really what they're doing to try and cut down on some of the costs.
Cohere 这家公司挺有意思的。
Cohere is kind of interesting.
他们是加拿大公司。
They're a Canadian company.
这跟Fable的禁令有关系吗？
So is this at all related to the Fable ban?
这是个很有意思的观点。
That's a really interesting point.
Cohere 特别告诉我，自从Anthropic受到出口管制以来，他们收到了大量客户和政府机构的咨询，大家都想分散模型供应商，不再只依赖某一家。
So Cohere in particular told me that they have actually, since the Anthropic export controls, have seen a massive increase in inbound interest from customers and governments looking to kind of diversify their model usage and not just be locked into any one vent vendor.
这直接促使他们将2027财年的年经常性收入预期提高了三倍。
And that actually prompted them to triple their annual recurring revenue projection for the calendar year 2027.
他们没透露具体数字，但我认为这清楚地表明，从顶尖的前沿模型转向更多模型选择，已经为像 Cohere、Hugging Face、Together AI 这样的公司带来了明显助力。
They wouldn't disclose what the figure was, but I thought that it was a really stark example of how this shift from the top performing frontier models towards more model optionality is already creating tailwinds for companies like Cohere, Hugging Face, Together AI, stuff like that.
我们再聊聊这些闭源模型吧，毕竟 Anthropic 和 OpenAI 的模型目前最受欢迎，都是闭源的。
And let's stick with these closed source models for a minute here because Anthropic and OpenAI, they are largely closed source models, the ones that are most popular.
有没有办法在不完全换掉其他公司模型的情况下，降低这些模型的使用成本？
Are there ways to make those models less expensive without changing to another company's fleet entirely?
当然有。
Definitely.
其实有很多办法可以尝试。
I mean, there's something you can do.
我得说明一下，跟我聊过的那些开始使用开源模型或更便宜模型的公司，并不是全都彻底放弃了 Anthropic 和 OpenAI。
And I should say that the companies I spoke with that are beginning to use open source or cheaper models, not all of them are switching completely away from Anthropic and OpenAI.
他们只是在慢慢探索更多选择。
They're just beginning to explore more options.
当然，也有一些小公司非常干脆，直接从 Anthropic 的 Opus 或 Sonnet 模型彻底转向了阿里巴巴的 Qwen。
Of course, there are some small companies that are really stark examples of saying, know, hey, I'm gonna switch completely from Anthropic Opus or Sonnet models to something like Alibaba's Quinn.
不过有一件事，就像你提到的，公司可以使用框架，这基本上是在 AI 模型之上的一层软件，而且基本上可以编辑他们的
But but one thing, you know, as you mentioned, companies can use harnesses, which is basically a software layer that sits on top of an AI model and also, you know, basically edit their
提示词，让运行 AI 模型稍微更节省成本，并且
prompts to make running the AI model a little bit more cost efficient and a
稍微便宜一些。
little bit cheaper.
所以，除了使用更便宜的模型之外，还有其他方法可以降低使用 AI 的成本。
So there are ways other than just using a cheaper model to make using AI less expensive.
而且有很多初创公司正在致力于帮助企业解决这方面的问题。
And there's a ton of startups working on helping companies with this.
我采访过其中几家，比如 CodeStrap 和 Martian AI。
I spoke to a couple like CodeStrap and Martian AI.
这些公司都在尝试构建一种能够统一调度提示词的中间层。
And those are a couple companies that are trying to help with this sort of harness prompt layer.
模型路由器会把不同的 AI 任务分配给不同的服务商，我想这也就涉及到了选择不同模型的那一层。
Model routers will sort of route different AI tasks to different providers that, I guess, gets into the layer of choosing different models as well.
但是
But
对的
yeah.
关于开源这部分，我得说，没人真觉得公司会完全转向开源。
So, on the open source piece here, and I have to say, I don't think anyone is suggesting that companies are moving to open source entirely.
我的意思是，我们现在看到的趋势是，大家越来越开始考虑开源模型，也更关注成本更低的模型，因为就像我们这档节目里聊过的，有些任务其实不需要顶尖模型，够用就行。
Mean, I think the trend that we're seeing is that people are starting to consider them more and consider cheaper models because as we've talked about on the show, some of these models just need to be good enough.
完成这些任务，不一定非得用最前沿的模型。
They don't need to be the frontier models to get some of these tasks done.
不过你在这篇报道里收集到的一些数据挺有意思，是关于目前开源模型和闭源模型在实际使用中的占比。
But there was some interesting data that you collected in your story about the proportion of model use that is relying on open source versus closed source model now.
我觉得特别有意思，因为现在开源模型的实际使用比例远高于闭源模型，这一点我之前真不知道。
And I thought it was interesting because open source really accounts for a much higher percentage of model use now compared to ClosedHorse, which I didn't know.
对的
Yeah.
说实话，这挺难的，因为数据我是从OpenRouter拿的。
I you know, it's tough because so I got the data from OpenRouter.
OpenRouter会公开很多数据。
OpenRouter shares a lot of publicly available data.
他们之前发布的一些数据就暗示了这个趋势。
Some data they had previously posted suggested this trend.
但他们就是不肯告诉我，具体是谁在用OpenRouter。
But, you know, they wouldn't tell me exactly who's using OpenRouter.
所以我们没法全面了解，到底有哪些公司在用这个工具，也就很难说清这个趋势到底有多普遍。
So we don't have a great, you know, total look into which companies are using this tool to really fully understand how widespread this trend is.
不过我要说，就像你提到的，OpenRouter的数据表明，通过他们平台处理的大部分令牌，都流向了开源模型，而不是Anthropic和OpenAI的顶尖模型。
But I will say that, you know, as you mentioned, open router data showed that the majority of tokens processed through its router are going to open source rather than front tier models from Anthropic and OpenAI.
这些开源模型主要是DeepSeek和Minimax。
And those open source models are primarily DeepSeek and Minimax.
而且OpenRouter每个月处理的令牌量高达数万亿。
And, you know, OpenRoute processes 10s of trillions of tokens monthly.
所以这个趋势真的值得注意，因为去年情况正好相反，大部分计算量都是通过 Anthropic 和 OpenAI 完成的。
So it really is significant to note this trend because it was quite the opposite last year where the majority was processed, you know, via Anthropic and OpenAI.
而且我觉得，这个趋势在其他路由公司身上也能看到，像 Fireworks 这样提供基础设施、帮助企业使用开源模型的公司，也发现通过它们平台处理的计算量越来越多地流向了开源模型。
And I think it's also a trend we see validated through other router companies and you know, companies like fireworks that provide infrastructure and plot a platform for companies to use open source models are really seeing the tokens processed on on open source models through their platforms.
斯科特，为什么现在企业更愿意使用中国产的开源模型，而一年前可能还不太愿意？
Scott, why companies so more willing now to use open source models from China where a year ago they may may may not have been?
确实如此。
Yeah.
这是个很好的问题。
That's a great question.
我觉得首先，成本已经成了一个大问题。
I think that first of all, cost has become, you know, a major issue.
我听说最近三个月尤其明显，有些公司甚至公开提到我们关于优步耗尽 AI 预算的报道。
I really think that what I've heard is that in the last three months in particular, and some companies have shouted out our reporting, you know, on Uber blowing through its AI budget.
今年年初，原本计划全年使用的 AI 预算，才用了几个月就快花完了，这引发了一连串的讨论。
And at the, you know, the first few months of the year that it intended to spend for the whole year really started a slew of conversations.
我听说像ServiceNow、Snowflake，甚至Databricks这样的公司，现在都在认真考虑内部的AI成本，对AI费用不断上涨感到担忧。
And I've heard from companies like ServiceNow, Snowflake, even Databricks that they're thinking about internal AI costs and concerned about rising AI costs.
这一点显然也在互联网和社交媒体上得到了广泛印证，其他公司和媒体也开始陆续报道类似的情况。
And obviously, that's something we see validated across the Internet and social media from a number of other companies and, you know, other media outlets have begun to report on this as well.
但随着成本问题越来越突出，企业也更愿意尝试不同的方案了。
But, you know, as costs have become more of an issue, I think companies have become more willing to try different things.
另外，我记得去年这个时候，中国开源模型刚开始进步，当时很多公司还担心可能存在隐蔽的网络安全漏洞。
And also, what I heard is that this time last year when Chinese open source models were beginning beginning to improve, companies were a little bit worried there might be hidden cybersecurity vulnerabilities.
比如，这些模型里会不会藏有后门，导致它们做出恶意行为，或者把公司的敏感数据传回中国的开发团队。
For example, there could be some sort of backdoor in these models that would allow them to behave maliciously or share a company sensitive data back with the Chinese lab that created the model.
但事实证明，这种情况根本没发生。
But as we've seen, that hasn't really happened.
目前没有任何证据表明发生过这类问题。
There's been no evidence of that happening.
所以，现在企业对采用这些模型的接受度越来越高了。
And so companies have become willing to adopt these models.
我联系过一家模型路由服务商，他们说不少《财富》百强企业都对使用中国开源模型感兴趣，有的已经在用了。
I spot I spoke to one model router provider, and they said that a number of Fortune 100 companies there's they've spoken with are all interested in using the Chinese models or are already using them.
即使他们没公开宣传，这确实是他们在认真考虑的事。
Even if they're not exactly advertising that, it's something that they're thinking about.
其中包括大型金融服务机构。
That includes large financial services providers.
如果企业不仅转向其他模型，还选择更便宜的替代方案，这对OpenAI和Anthropic来说是个多大的挑战？
How big a problem is this for OpenAI and Anthropic if people are going with not even just other models, but the less expensive models that they offer?
我觉得这说明OpenAI和Anthropic都不可能完全垄断这个领域。
Well, I think it just shows that Anthropic and OpenAI will, you know, not have neither will have a complete monopoly over this space.
当然，还得看这些前沿模型能否在性能上大幅超越开源模型，而开源模型又没跟上。如果真能拉开差距，那前沿模型的长期优势可能会比现在强得多。
Obviously, it remains to be seen if these frontier model providers, if their models improve to an extent that is, you know, significant beyond the capabilities of open source models, and open source models don't catch up, then I'm sure we could see that, you know, the staying power of the frontier models, it is much stronger than maybe it is right now.
但毫无疑问，这些公司面临的竞争可能比之前想象的要激烈得多。
But yeah, I think it speaks to the fact that these companies may face more competition than previously thought.
明白了。
Great.
劳拉，非常感谢你今天来做客。
Well, Laura, I want to thank you for coming on.
这位是劳拉·布拉顿，我们《The Information》的AI记者。
That is Laura Bratton, our AI reporter here at The Information.
之前叫Grammarly的Superhuman公司，最近收购了GPTZero这家初创企业，他们能检测内容是不是AI生成的。
Superhuman, the company formerly known as Grammarly, is buying GPTZero, a startup that allows you to detect whether or not content is AI generated.
现在我们很荣幸邀请到Superhuman的首席执行官希希尔·梅赫罗特拉，来独家聊聊这笔收购。
Joining me now to discuss the acquisition in an exclusive broadcast appearance is Shishir Mehrotra, CEO of Superhuman.
希希尔，欢迎再次做客我们的节目。
Shishir, welcome back to the show.
很高兴你今天能来。
It's great to have you here.
谢谢。
Thanks.
很好。
Great.
能来这儿真的很棒。
Great to be here.
你们为什么会决定收购 GPTZero？
Why did you decide to buy GPTZero?
是啊。
Oh, yeah.
我对这个收购特别兴奋。
I'm really excited about this one.
我觉得在我们今天所处的世界里，有一个特别有意思的词被选为了年度词汇——‘AI 垃圾’。
I think the in the world we're facing today, one of most fun stats I saw is the word of the year right now is AI slop.
我们每天都被淹没在这样一个难题里：到底什么内容是人写的，什么又是机器生成的。
And we're just we're just inundated with the the challenge of determining what's actually human and what's not.
在这个领域里，领先的公司就是 GPTZero。
The leading company in this category is a company called GPTZero.
他们推出了一系列广受欢迎的产品，包括 AI 检测器、幻觉检测器、AI 视觉工具，还有 Replay，这些产品共同组成了一套工具，帮助人们分辨哪些内容是人类创作的，哪些不是。
It's a runs a set of popular products for authenticity, an AI detector, a hallucination detector, a product called AI vision, and a product called Replay that are all a suite of products helped designed to help people discover what's human and what's not.
他们增长得特别快，三年内就积累了一千九百万用户，年经常性收入超过三千万美元。
They scaled really fast, 19,000,000 users, over 30,000,000 in ARR in less than three years.
我们非常期待把他们纳入Superhuman大家庭，进一步扩大产品规模，服务更多在工作场景中使用的用户。
And we're just really excited them to bring them into the Superhuman family and scale the product to work with tens of millions of more users working right where they work.
三千万美元的年经常性收入。
30,000,000 in ARR.
你能告诉我们你们花了多少钱收购他们吗？
Can you tell us what you paid for them?
不能透露。
No.
这是笔私人交易，所以我们不会多谈交易细节。
It's a private private transaction, so we won't share much about the about the deal.
这正是在私营公司工作的乐趣之一。
It's one of the pleasures of working at a private company.
是啊，我们确实得问一问。
Yeah, we got to ask.
我觉得大家都对这笔交易很满意。
I think everybody's quite happy with the deal.
我们确实得问一下。
We got to ask.
我们确实得问一下。
We got to ask.
所以，好吧，这还挺有意思的。
So, okay, so this is kind of interesting.
你们之前已经有Superhuman和Grammarly套件了，还自带抄袭检测功能，对吧？
You already had Superhuman and the Grammarly Suite, you already had a plagiarism detector, right?
那这次是打算把这些功能整合进去吗？
So is this going to add to that offering?
是的。
Yeah.
实际上，我们会把这两支团队的产品整合成一个全新的身份验证套件。
So actually, the team is going to we're combining the offerings and making one authenticity suite.
我邀请GPTZero的团队领导来担任Superhuman公司真实性部门的负责人。
I'm asking the leaders of GPTZero to become the leaders of our authenticity division here at Superhuman.
不过没错，我们在这个领域已经深耕一段时间了。
But yes, we've been working in this category for a while.
这正是我对此如此感兴趣的原因——亲眼看着它迅速崛起。
It's part of why I was so interested in it, is just watching how quickly it's taking off.
我们已经有了一款由Grammarly团队开发的热门AI检测工具，还有你提到的抄袭检测工具等等。
We have a very popular AI detector built by the Grammarly team, as well as, as you mentioned, a plagiarism detector and so on.
我们现在要把这一系列产品整合成一个统一的套件。
And we're taking that family of products all into a single suite.
还有AI幻觉检测器，能给我们具体讲讲这个产品是什么样的吗？
And AI hallucination detector, I mean, help us understand, does that product look like?
就是说，用户输入什么，它会输出什么结果？
Like, what what do you put in, and what's the output of that?
是的。
Yeah.
AI检测现在衍生出很多更具体的用途。
There's there's a lot of derivatives of AI detection that get a little bit more specific.
幻觉检测不是只看内容是不是AI生成的，而是判断它说的到底是不是真的。
Hallucination detection is looking at something that's written and trying to figure out not just was it generated or not by AI, but just is it real or not.
其实就是事实核查吧？
There's a really So, like, fact, like, checking?
很多时候，确实就是做事实核查。
Is is that the A lot a lot of times, it'll be fact checking.
还有引文核对。
It'll be citation checking.
最近有一篇报道，KPMG被一家公司聘请，去审查他们内部使用AI的情况。
There's a there was an article recently that KPMG was hired by another company to do a review of how they were using AI inside the company.
讽刺的是，这份报告本身被指出包含多个幻觉内容，比如引用了一些根本不存在的语录。
And ironically, the report was cited for having a number of hallucinations that, you know, it cited quotes and so on that just didn't exist.
所以这个产品更进一步，不只是判断内容是否可能由AI生成，而是深入核查其中的事实，确认哪些是真实的，哪些是虚假的。
And so this this product goes a step beyond, it likely to be AI generated, but actually go through it, check the facts themselves, and say, it match what what is what is real and what's not?
它和引用内容一致吗？
Does it match the citations?
真正触及到幻觉问题的核心。
And really get to the the the heart of of hallucinations.
没错。
Right.
这是个非常受欢迎的产品。
Very popular product.
所以，Shishir，我想跟你探讨的核心问题是：如果你看看Superhuman和Grammarly的产品组合，你们在新闻稿里也明确提到，你们现在同时涉足了这个生态系统的两个方面——你们的产品帮助人们用AI来编辑文字。
So, now the heart of what I want to get to with you, Shishir, is if you look at the suite of offerings that Superhuman and Grammarly has, and you're upfront about this in the press release, mean, you say, We are now playing on both sides of this ecosystem in terms of our products help people edit their writing with AI.
在某些情况下，还会为写作有问题的段落提供改写建议。
In some cases, it can offer recommended rewrites as well for passages with writing.
这属于AI内容生成这一端。
So that's kind of on the AI content generation side.
而你们还有AI内容检测这一端。
And then you have the AI content detection side.
所以你是两边都在做。
So you're playing both sides here.
我想知道，你能谈谈当初为什么决定同时涉足这两方面吗？你觉得这两者之间有矛盾吗？
I wonder if you can take us into the decision to play on both those sides and whether or not you see those as sort of two contradictory offerings at all?
我觉得这个世界正在朝寻找中间地带的方向发展。
You know, I think the world is headed to trying to find that middle ground.
这一点非常清楚。
It's very clear.
我经常和我孩子聊这个话题，是的。
I have this conversation with my kids a lot, Yeah.
一个在上大学，一个在上高中，每个学生都面临这个问题。
One in college, one in high school, and every student is facing this.
一方面，他们被提醒要小心使用AI，否则会有惩罚之类的后果。
That on one side, they're being told, be careful how much AI you use is gonna be there there's penalties for it, and so on.
另一方面，他们又被告诉：嘿。
On the other side, they're being told, hey.
你在职场上得往前走，得学会有效且恰当地使用这些工具。
You need to move forward in the workforce, and we're gonna use these tools effectively and in the right ways.
所以你得跟上这些工具的使用方法。
And so you need to be up to speed on how all of them work.
我觉得这对每个专业人士、每个知识型工作者都一样。
And I think that's true of every professional, every knowledge worker.
你真该用起人工智能来。
You should be using AI.
这可是几十年来最神奇的一批技术和工具之一，拒绝使用它们简直太荒谬了。
It's not the case that, you know, it's one of the most magical set of technologies and tools that have been invented in in decades, and the idea of not using them is is is silly.
这并不意味着你应该用它来取代自己的声音。
It doesn't mean that you should use it instead of your own voice.
也不意味着你要用得让人觉得不真实，甚至更糟——出错。
It doesn't mean you should do it in ways that will be interpreted by by readers as not genuine or even worse, wrong.
所以我的看法是，你其实需要把这一整套工具都用起来。
And so I my view of it is that you actually need all of the tools in a single suite.
你需要这种能力，帮我写得更好，帮我完成手头的工作，替我接手一些任务，但同时要确保我的表达还是我自己的风格。
You you need that ability to help me write well, help me finish what I'm working on, take over tasks for me, But also make sure I sound like me.
确保我不会说出任何错误的内容。
Make sure that I don't say things that are incorrect.
你希望有一整套协作者，能和你同时配合工作。
And you want that set of co editors working with you at the same time.
没错。
Right.
我经常用一个比喻来说明这一点：把AI想象成一个个助手，它们在帮你处理自己的工作。
One analogy I'll use for it a lot, I think of a lot of what we're doing with AI, is if you just personify as agents, you know, just taking your own work.
如果你是记者，就会有一群人协助你写作。
If you're a journalist, you have one side, you have sets of people that help you write.
他们可能会替你去跑新闻、收集素材。
You know, they might go out and do parts of your story for you.
还会帮你做前期核查之类的准备工作。
They'll go do pre checks for you, so on.
另一面呢，有个编辑负责确保你说的内容准确，符合你想要传达的标准。
The other side, you have an editor whose job is to make sure that your what you're saying is correct and making sure that it matches the standard of what you're trying to put out there.
我觉得这两者都是这套工具里不可或缺的部分。
I think each of them are an important part of the suite.
这其实引出了一个更广泛的问题：写作中到底该允许AI介入到什么程度？
I guess it kind of it sort of leads us to sort of this broader question, which is what is then the threshold for how much AI should be allowed in writing?
AI到底该扮演什么角色？
What is to AI?
你划分的这个界限，核心其实就是‘什么是真实’吧？
What is authentic, I guess, is the name of the division that you've created?
我一直在想，你有没有把一段文字给Grammarly，让它帮你重写，然后再丢进GPTZero去检测？
But you know, I'm thinking about this scenario where have you ever given Grammarly a passage, said, Help me rewrite it, and then put that passage into GPTZero?
结果它会不会说：‘这段文字AI痕迹太重了’？
And does it say, Well, this is too AI?
我也不太确定。
I don't know.
这事儿好像一直在变，说不准。
It seems like kind of a moving target in a way.
对吧？
Right?
我的意思是，关键是要帮人去判断，具体情况具体分析。
I mean, I I think the the idea is to is to help people assess, and it depends on the it depends on the situation.
比如我发给你一条消息，说是我的，那你自然觉得这真是我写的。
I mean, there's if I send you a note and I say it's from me, you expect it to be from me.
是的。
Yeah.
但如果我设了个助手，每周自动收集各种信息，帮你生成项目进度报告，省得大家动手，那情况就一目了然了。
If I if I set up an agent whose job is to write the project status report every week by by pulling together lots of different information and give you give us a consolidated report so none of us have to do the work, it's pretty obvious.
这报告肯定不是我们任何人写的。
It's not gonna be any of us.
这完全是AI生成的。
It's gonna be generated by by AI.
所以这得看具体是什么情况。
So it depends what the what that situation is.
我觉得没必要非得非黑即白，走极端。
So I I don't think there is a you must always be at one extreme or the other.
关键是要能判断：当你声称想达到某种标准时，你到底有没有做到？
It's being able to assess when you said you wanna be at this extreme, did you achieve that?
从读者或接收者的角度看，内容是否符合他们的预期？
And from the reader's perspective or the receiver's perspective, does it match their expectations?
我觉得难点在于
Think the challenge
换句话说，有些情况比如你把文字丢进GPTZero，它会告诉你：没错，这是AI写的，但这种情况其实没问题。
is So in other words, you would say, like, are situations where I mean, you put writing into GPTZero, there are situations where it will tell you, Yeah, this is AI generated, and that's fine.
这属于一种事务性的文字处理。
This is a logistical piece of writing.
但像学生写的论文就不一样了，这时候你反而会特别留意有没有真正的个人风格。
But then there are essays that students will write, for example, and you will sort of look for the badge of authenticity there in that scenario.
说实话，学生这个情况特别有意思，因为你会看到老师在两种极端之间来回切换。
I mean, the student case is really interesting because you'll find teachers doing both extremes.
就像你说的，教授布置作业时，有时候会说：请不要用AI。
As you'll say, the professor will give students an assignment, and sometimes they'll say, Please don't use AI.
我会仔细检查的。
I will be checking.
这是靠自觉，还是当工具来用？
Is it an honor code?
还是别的什么情况？
Is it tool?
诸如此类。
And so on.
我觉得这就像现在考SAT，有一部分题目不能用计算器，另一部分却允许用。
And I view it as the equivalent of, you know, when you go in to take the SATs these days, there's one section that is without calculators, and there's another section that's with calculators.
教授也常常会布置这样的作业：这次你们要尽可能用上所有工具，把这个问题回答得最好。
And the professor will often also hand you an assignment that says, and in this case, I want you to use every tool possible and just generate the best possible answer to this topic.
在这种情况下，我想知道哪些部分是你自己做的，哪些不是，但我并不是在找一个零AI使用率的答卷。
And in that case, I wanna know which parts were you and not, but it's like, I'm not looking for a zero AI detection score.
我想要的是另一种评分标准。
I'm looking for a different score.
你在职场里也能看到同样的情况。
You look at the same thing in the workplace.
我们面试候选人的时候，也是这么做的。
When we interview candidates, we do the same thing.
我们会安排一场面试，故意不允许使用AI，就像不许用计算器一样。
We have one interview, which is intentionally without AI, and we ask people to answer sort of like without the calculator.
但同时我们也会有另一场面试，明确要求你尽可能用上所有工具，我们真正关心的是你多会用这些工具，然后在两者之间找到合适的平衡点。
But we also have another interview where the job is please use every tool you possibly can, and we're not really we're looking for how well you can use those tools, and then we find the right spots in between.
所以我觉得，评估方式非黑即白、非零即百的做法，其实并不对。
So I I think that idea of assessment needs to always be zero or a 100 is not quite the right way to look at it.
你明白吧？
You know?
说实话，当你说是人写的，我们就得确信真是人写的。
Boy, we wanna know when you say it's human, we wanna know it's human.
当你说是AI写的，那我们就只看内容本身，不纠结来源。
When you say it's not, it's we just we just wanna assess the work itself.
另外我想说的是，真实性评估不只是判断是不是AI生成的。
The other thing I'd say is the authenticity suite is not just about is it AI or not.
它不能只像幻觉检测那样，光看对不对，还得分辨这到底是不是你自己的想法。
It can't be anything like hallucination detection where even beyond that, was it you or not, is it just is it correct?
我觉得这也是一个非常关键的方面。
I think is is also an important part to to meet.
没错。
Right.
而且我注意到你们在新闻稿里提到一个有意思的现象：当你给大模型输入错误信息，它反而会把这些错误当真去学习，这就形成了一个恶性循环，让未来幻觉问题越来越严重。
And I mean, there was an interesting part in the press release I thought you talked about, which is that when you feed incorrect information into these LLMs, the LLMs will then train on that information, which becomes a bit of a self propagating loop here in the worst possible way, which is that you're just increasing the rate of hallucinations down the road.
所以归根结底，总得有人来核对事实，不能全指望大模型自己纠错。
And so, I guess the idea here is that if I mean, somebody has to fact check it at some point, it can't always be the LLM.
不过在这种情况下，你说的还是得靠大语言模型。
Although in this case, you are saying it's going to be LLM.
我们确实在开发相关工具，但我认为这真是个社会性问题。
Well, we're building tooling for it, but I do think it's a real societal problem.
GPTZero 最受欢迎的工具之一，是我特别喜欢的一个网站。
Maybe one of the most popular tools from GPTZero is a site I really love.
它的名字叫 izzyinternetai.com。
It's called izzyinternetai.com.
这个网站会去分析各大主流平台，统计目前有多少比例的帖子被判定为主要是由 AI 生成的。
And it goes and it assesses for all the popular platforms what percentage of the posts on those platforms are currently graded as being mostly AI.
比如，LinkedIn 排在第一位，比例高达 42%。
And so, for example, LinkedIn is at the top of that list at 42.
也就是说，GPTZero 的工具认为，有 42% 的 LinkedIn 帖子主要是 AI 生成的。
So 42% of LinkedIn posts, the GPTZero tools determine them to be mostly AI.
有趣的是，Reddit 则处于另一端，AI 内容比例不到 10%。
Reddit is at the other extreme, interestingly, of being less than 10% AI.
如果你想想大语言模型是怎么训练的，它们训练时用的基本上都是互联网上的内容。
Now if you think about the way that LLMs train, you know, as they go train on a lot of what they're training on is on all of what's on the Internet.
所以如果你是做训练的团队，你肯定想知道这两者有什么区别。
And so if you're one of those teams and you're training, you wanna know the difference.
你得弄清楚，自己到底是在重复训练自己生成的内容，还是在学习真正的人类新知识。
You wanna know, you know, am I am I sort of just retraining on my what I generated myself, or am I training on new human knowledge?
这并不是说哪一种做法是对是错。
And it's not to say that any of it is right or wrong.
其实很多人用AI写领英帖子，背后都有合理的理由。
I mean, there's good reasons why why some people will use AI to write LinkedIn posts.
但我觉得，在评估的时候，有时候你就是想知道真相。
But I I think the the when you're assessing it, sometimes you wanna know.
这内容是人写的吗？
Is it is it the human?
是真人写的吗？
Is it the person?
还是不是呢？
Or is it not?
而且没错，随着规模扩大，整个互联网确实存在垃圾进、垃圾出的问题。
And, yes, it's a it's a there is a garbage in, garbage out problem for the Internet as a whole as we scale.
那这是否就是GPTZero面向企业的业务方向呢？
So is is this then the the enterprise play for GPTZero?
你们会不会考虑把这项技术卖给企业，帮他们训练新数据集时区分内容来源，标记哪些明显是AI生成的？
Will you sort of think about selling this to enterprises as they look to train on new datasets to be able to label data and saying, this was clearly AI generated content.
我们应该把这些当作AI生成的内容来用，而那些是人类生成的，就另当别论。
We should use it as such in our training, versus this was human generated.
我的意思是，假设它
I mean, assume It's
目前并不是GPTZero团队的主要业务方向。
not a primary business the GPTZero team is in now.
我估计是七成
I'd say seven
九是第一名。
nine are number one.
我在想这项技术在消费者和企业市场之间的区别。
I'm thinking about consumer versus enterprise here with this technology.
那是一种非常特定类型的企业。
Well, that's a that's a very specific type of enterprise.
我的看法是，我们打算把GPTZero的一整套技术推广给更广泛的用户群体。
I mean, I think the I think the the broad view is we're gonna take the GPTZero set of technologies, and we're gonna introduce them to a much broader audience of people.
我们在Superhuman产品上拥有四千万日活跃用户。
We have 40,000,000 daily active users on the Superhuman products.
我们有一万多家企业客户，每一家对真实性的需求都不一样。
We have over 10,000 enterprise customers, and each of them have different authenticity needs.
最重要的是，我们公司的核心是把AI带到你工作的地方。
And most importantly, what we do as a company is we bring AI to work where you work.
所以我们接下来要做的核心，是把一套今天被我们称为‘主动调用型’的工具推广开来。
So the heart of what we're going to do is take a set of tools that today are what we consider to be pull tools.
当你需要进行AI检测或幻觉检测时，可以直接使用这个工具——我们会把功能集成到你写作文章、写论文的现场，让它在你工作的地方实时发挥作用，这样工具对你的帮助就会大得多。
When you're ready for an AI detection or hallucination detection, you go to the tool, and instead, So we're going to put them to work right where you as you're writing your article, as you're filling out your essay, you know, it's working right where you work and greatly increases the chances of the tools being useful for you.
从企业用户和普通消费者的角度来看，这个思路非常好。
In terms of the enterprise versus the consumer play, I think that's really good.
这确实是其中一个重要的切入维度。
That's certainly one axis to cut on.
我觉得这两个领域都有非常不错的应用场景。
I think there's really good use cases in both.
我经常关注的另一个维度是不同行业和职能。
The other one I look at a lot is the the industries and functions.
比如学生和教育工作者，就是一个很有代表性的群体。
So, you know, students and educators is one dynamic.
招聘人员和面试官，是另一个类似的群体。
Recruiters and interviewers is another such dynamic.
再比如咨询师和他们的客户，这也是一个很有意思的场景。
You think about consultants and their and their and their customers, it's another interesting one.
如果你想想记者和他们的受众，我觉得这也是一个很有意思的场景。
If you think about journalists and their audience, I think that's another interesting one.
在很多这类情况里，都存在创作者和接收者之间的互动关系。你需要找到合适的平衡点，明确大家对内容中人工与AI比例的预期，并且作为创作者和接收者，双方都要知道怎么去验证这一点。
In many of these cases, there is a creator receiver dynamic where you're trying to find the right balance and expectation for that particular flow of how much of this is human versus not, and how do you verify that as both a creator and a receiver.
对。
Right.
我想换个话题，问你一个问题。
Let me ask you one question on a separate topic here.
在上一段讨论中，我们聊到了各家公司正努力应对大语言模型使用成本不断上升的问题。
We were talking in the segment before you about how companies are grappling with the rising costs of using LLMs.
我记得你之前上我们节目时，两次都提到了 Grammarly 和 Superhuman 每天要处理多少 API 调用。
I know when you were on our show, I think both times actually you were talking about how many API calls Grammarly Superhuman does.
好像是这么个数字。
Think it was something like
超过一千亿次。
Over 100,000,000,000
几亿、几十亿甚至上百亿。
Hundreds a millions or billions.
不管怎样，你们是怎么应对大模型的成本问题的？
Anyway, how are you grappling with LLM costs and using these models?
你们有没有考虑更多地使用开源模型，并把它们整合到你们的系统里？
Are you looking to open source models more at all and integrating those into your stack?
是的。
Yeah.
我们每周的大模型调用量大约是1000亿次，现在早就超过这个数了。
So we do about 100,000,000,000 LLM calls, well over that now, per week.
更值得注意的是，平均每个用户每天大概要产生几千次调用。
Probably a more interesting stat is per user that works out to a few thousand per day for every user.
这意味着，对于使用我们产品的用户，尤其是Grammarly的用户，我们很可能是你们用得最多的AI生成工具。
That means for any user of our products, especially Grammarly products, we are likely your number one AI generator.
我们替你们发起的大模型请求，可能比你们用的其他任何工具都要多。
We're probably generating more LLM queries on your behalf than any of the other tools that that you're using.
所以我们在AI流程中处于一个非常特别的位置。
So it does put us in a very interesting spot of the of the flow of AI.
目前大约97%的调用都指向我们自有的大模型，剩下的3%则
About 97% of those calls go to in house LLMs today, And about 3% go
用在外面的模型。
out there.
你
You
Superhuman用的是
have Superhuman has
我们自己托管和训练的模型。
That we host and train.
是的。
Yeah.
明白了。
Okay.
很多模型都是从开源的基础版本开始的。
You know, a lot of them start with a with a with an open source starting point.
LAMA模型现在挺受欢迎的。
The the LAMA models are quite popular.
谷歌的Gamma模型也很流行。
The Gamma models from Google are also quite popular.
所以我们通常会从这些模型入手。
So we will start with them.
然后我们进行微调和训练，再部署在自己的基础设施上。
We'll fine tune and train them, and then we'll serve them from our own infrastructure.
这样能大幅降低成本。
And that allows us to run at much better cost.
这些产品的核心部分毛利率都在85%以上，这跟其他一些AI服务商很不一样。
And the core of those products run at 85 plus percent gross margin, which is obviously very different than some of the other AI providers out there.
我觉得，找到能大规模应用的场景，再把通用模型调整成更专注特定任务的版本，这种模式会越来越普遍。毕竟，作为用户，谁不希望AI能真正贴合自己的工作场景呢？
I think that pattern of find use cases that work at scale and then take general purpose models and then reduce down to models that are more specific to those tasks will be a common pattern because I think, you know, users want you know, as as a customer, you want AI that works where you work.
你想要的不是去某个地方打卡。
You want it you don't wanna go to a destination.
你知道的，可能一天就用那么一两次。
You know, maybe you go once a day.
你希望有个真正像伙伴一样的助手，能和你无缝配合。
You want something that really feels like a companion that works right with you.
每次你打一个字，每次打开新应用——不管是Gmail、Slack还是苹果备忘录，你都希望那些智能代理能实时在你身边帮忙。
And every time, you know, every time you type a character, every time you open a new application, whether you're in anything from Gmail to Slack to to Apple Notes, you want that set of agents working right alongside you.
这成本可不低，所以你得找到办法让它能规模化。
That can be expensive, and so you have to find ways to scale that.
我们的一大重点，就是确保能以低延迟、合理成本提供大规模的AI服务。
And a big part of our effort has been on making sure we can deliver AI at scale, at low latency, and at reasonable costs.
百分之九十七。
97%.
那去年这个时候是多少？
So what was that this time last year?
有那么高吗？
Was it that high?
大概吧。
Roughly.
你知道，我只是
You know, I'm just
就是想尽量接近那个水平。
In trying get the in the same ballpark.
是的。
Yeah.
我们一直还行。
We we've been Okay.
所以你是自己在开发模型？
You so you've been using building your own models.
Grammarly 这个挺有意思。
Grammarly is interesting.
Grammarly在GPT-2之前就赶上了Transformer的潮流。
Grammarly jumped on the transformer bandwagon before GPT two.
所以他们的团队早就开始做这方面的研究了。
And so the the team has been working on this for a very long time.
我们一直在打造自己的模型版本，以便大规模提供AI服务，尤其是考虑到每个用户每天可能有成千上万次调用。
The the bill building our version of these models so that we can deliver AI at the you know, at a at a fairly high scale, especially as you think about it on a per user basis, thousands of calls per day.
你必须在成本和响应速度上都做到合理平衡。
And you have to be able to do that with reasonable cost and latency.
说实话，对我们来说最重要的可能不是成本问题。
I mean, it it probably one of the most important things for us actually wasn't the cost.
它必须能在你打字时实时响应。
It needs to work as you type.
当你输入每一个字符时，系统都要持续分析每个词，判断能为你做些什么。
So as you're as you're typing characters, we need to be constantly looking at every word and saying, are the different things we can we can do for you here?
我们必须做到极快的响应速度，而这种速度要求反过来也推动了更高效架构的设计，最终还降低了成本。
And we have to do that really fast, and that doing it fast ends up leading to an architecture that also leads to better cost as well.
那你们自己的员工呢？比如他们写代码时用的模型怎么样？
But what about your own employees, the models that they use for coding, for example?
我猜你们的技术团队肯定在用这些前沿模型来加快自己的工作流程。
I imagine your tech team is using all these frontier models to expedite their own processes.
我的意思是，你们用这些前沿模型有没有把预算花光？
I mean, have you burned through your own budgets for using these frontier models?
你们有没有跟首席财务官说，得把开支控制得更紧一点？
And have you told your CFO that we got to keep a tighter lid on
当然有。
things?
明确说一下，我们97%的客户流量都用的是我们自己的模型，只有3%用其他公司的。
Far, and we use, just to be clear, we you know, 97% of our of our customer traffic goes to our own models, 3% goes to the others.
即便如此，这个量也相当大了。
That's still a huge amount.
所以，我们内部也是所有大型语言模型提供商的大客户。
So, you know, we we we are big customers of all the big LM providers internally as well.
展开剩余字幕（还有 213 条）
我们可以说是走在所有工具的最前沿，整个公司各个团队都在广泛使用AI，尤其是工程团队，简直把这些技术的极限都给推到了。
I mean, we're right at the bleeding edge of all all the different tools, pretty heavy AI adoption amongst all our teams, but certainly, the the engineering teams are pushing the limits on it.
我觉得那些担心这个问题的公司，其实盯错重点了。
You know, my my mind I think the companies that are worried about this are measuring the wrong thing.
如果你只拿token消耗量当衡量标准，现在越来越多公司已经开始意识到，这种做法会带来完全错误的激励导向。
And certainly, if you start with a view of of just looking at token spend as your as your measurement, and I think now we're seeing a lot of companies that are backing away from token maxing as a way a way to look at it, you'll get you'll get all the wrong incentives.
我们反而会看一系列其他指标，来评估token投入的回报，特别是在工程团队这类部门。
Instead, we look at we have a set of different metrics we look at to to evaluate whether we're getting ROI on our on our token spend, especially in teams like our engineering team.
比如，我们关注功能上线的速度有多快。
You know, we're we're we're looking at how fast features are shipping.
还有代码合并的效率、响应延迟等等，这些方面都在明显提升。
We're looking at the the throughput of our pull request, looking at latency, and so on.
目前这些指标全都大幅上升。
And all those numbers are way up right now.
到目前为止，我觉得我们在token上的投入回报率非常高。
And so far, I think our ROI on our token spend is incredibly positive.
我根本没让他们放慢速度。
I haven't asked them to slow down at all.
我们还没到极限呢。
We hit our breaking point yet.
明白了。
Got it.
你刚才说衡量错了，指的是过度消耗token，对吧？
And when you said measuring the wrong thing, you meant the token maxing Yes.
就是这个意思。
Measuring.
你说的就是这个。
That's what you meant.
如果你是一家花了大量AI token，却没让产品变得更好，也没更快地交付给客户，那确实是在浪费精力。
But if you're if you and if you're the type of company where you spend a bunch of AI tokens and you don't get better product or more product shipped to customers faster, then, yeah, you're probably wasting your wasting energy.
我相信肯定有不少公司正处在这种状况。
And I'm I'm sure there's people that are in that situation.
到目前为止，我们看到的不是这样。
That's not what we've seen so far.
我们确实在持续推出产品。
We we are delivering product.
上个季度，我们工程师交付新功能的速度差不多翻了一倍，如果你拿这个和工程团队的成本一比，回报率简直惊人。
You know, in the last quarter, we've roughly doubled the pace of feature delivery from our from our engineers, which, you know, if you just go look at that and compare with the cost engineering team, it's incredible ROI.
所以我觉得，我们还没到回报递减的阶段。
So I don't think we've hit diminishing returns on that yet.
太好了。
Great.
Shishir，很高兴你今天来做客。
Well, Shishir, always a pleasure to have you on.
感谢你参与我们的节目。
Thank you for joining us.
这位是Superhuman的首席执行官Shishir Mehrotra，欢迎收看TITV。
That is Shishir Mehrotra, CEO of Superhuman, here on TITV.
好的。
Okay.
SpaceX和特斯拉的股票各自都值得关注，但把它们放在一起对比，反而更有意思。
SpaceX shares and Tesla shares are interesting to watch on their own, but they are even more interesting when you compare them to one another.
我的同事西奥·韦特写了一篇文章，讲很多人为了买SpaceX股票，正在抛售特斯拉股票。
My colleague Theo Wayt wrote a column about the extent to which people are selling Tesla to get into SpaceX.
也许反过来也有人在做同样的事。
Maybe they're doing it the other way around too.
我想请他来聊聊他的看法。
I wanna bring him on to walk us through his thoughts.
西奥，欢迎再次做客节目。
Theo, welcome back to the show.
很高兴你今天能来。
It's great to have you here.
很高兴能来。
Good to be here.
好的。
Okay.
你觉得呢？
So what do you think?
现在有人卖特斯拉去买SpaceX的股票吗？
Are people selling Tesla to get into SpaceX these days?
我觉得这很难说。
I think it's I I I think it's hard to say.
如果你每天看股价，会发现它一直在波动。
I mean, if you look at the stock price each day, it kind of fluctuates.
这种情况挺特别的，以前想投资埃隆·马斯克，只有特斯拉这一条路。
It's it's a weird situation because, you know, in the stock market, there used to be one way to buy exposure to Elon Musk, and that was Tesla.
但现在有了两个选择，它们的交易方式也不一样，选哪个真不容易。
And, you know, obviously, now there's two and they trade in different ways, and it's hard to choose between them.
但可以肯定的是，过一段时间，总有一个会表现得更好。
But presumably, one of them is gonna do better than the other, you know, over some time span.
现在大家的注意力全在SpaceX身上，毕竟马上要上市了，但这不代表它当前的股价就真的更有吸引力。
So, you know, clearly, all the hype is around SpaceX right now because of the IPO, but that doesn't mean it's actually, you know, a more compelling stock at its at its current at its current price.
不过挺有意思的是，特斯拉自己也有不少大胆的长远目标。
It's kind of interesting, though, because Tesla has its own moonshot ambitions here.
对吧？
Right?
相比那些轨道上的数据中心，特斯拉的目标可能更近一点，但归根结底，你得自己选一个更想押注的长期愿景。
I mean, they might be a little bit closer term than the orbital data centers, but it's sort of like take your pick on which long horizon mission you would like to invest in.
说得对。
Yeah.
这两家公司现在的估值倍数，要不是有马斯克在背后撑着，根本说不通。
I mean, both of the companies have revenue multiples that wouldn't really make sense for companies not led by Elon Musk.
多年来，一直有很多人说特斯拉被高估了。
And, like, there have been a lot of people for many years that have said that Tesla is overvalued.
但你仔细看，特斯拉现在才14倍2025年的营收，而SpaceX的估值已经超过2025年营收的100倍了。
But if you look at it, it's it's only trading at, like, 14 times its twenty twenty five revenue, whereas SpaceX is trading at more than a 100 times its twenty twenty five revenue.
如果你想想埃隆·马斯克的溢价，就是他能说服投资者相信那些超长期的登月计划。
So if you think about, you know, the Elon Musk premium of, you know, being able to sell investors on on, you know, super long term moonshots.
目前来看，SpaceX 的估值里已经包含了比特斯拉多得多的这种预期。
You know, SpaceX has way more of that baked in at this point than than Tesla does.
但你想想六个月前，埃隆·马斯克还说特斯拉要造出能消除贫困、让地球上每个人都不用工作的机器人。
But if you think about, you know, six months ago, Elon Musk was saying that Tesla is gonna build robots that are going to cure poverty and make working optional for everyone on Earth.
这可不是什么小规模的市场机会，或者你愿意怎么形容的宏大愿景。
That's not exactly, like, you know, a small, total addressable market or or vision or, you know, however you wanna think about it.
这两家公司都有巨大的野心，而目前，其中一家的股价已经更明显地反映了这种野心。
Like, both of these companies have gigantic ambitions, and one of them has that reflect in this fund price more than the other at this point.
特斯拉现在的乐观预期到底在哪呢？
Where are the optimist ambitions at for Tesla right now?
咱们现实点，说说实际的情况。
Bring us back to Earth on this.
现在有个叫 Optimus V3 的新版本，据说能解决我们之前讨论过的所有问题。
Well, there's this v three version of Optimist that's supposed to be the new one that addresses all of these issues that we've written about.
去年咱们还听过一些关于机器人手的不错的故事
You know, last year, we had some good stories about how the hands are
你啊。
You.
你之前讲过一些挺好的例子。
You had some good stories.
还有火箭。
And rocket.
还有火箭。
And rocket.
明白了。
Okay.
还有火箭。
And rocket.
火箭。
Rocket.
确实如此。
That's true.
关于手的那个故事，讲得特别精彩。
The hands one was rocket.
那是我的专家。
That's my expert.
我不想在这儿把他删掉。
I don't wanna erase him here.
他在这方面也是专家。
He is is, you know, an expert on this too.
不过话说回来，他们一直在努力解决Optimus的种种问题，但那个号称解决了所有问题的新版本一直没公布，一再推迟。
But, anyway, you know, they they they've been trying to work out a bunch of kinks in Optimus and those you know, they they've not revealed the new version that's supposed to have addressed all those problems, and and that keeps getting pushed back.
但埃隆说，他们还在建新工厂，装了一条条生产线，只要有了能大规模生产的产品，随时都能开干。
But Elon says they're still, you know, they're building a new factory, and they're putting up all these assembly lines, and they're gonna be ready to start mass production whenever they have something to mass produce.
所以这事儿还不好说，但埃隆强调，这依然是公司愿景的重要部分。
So that that's still, to be seen, but Elon says it's still a big part of the vision for the company.
所以，你刚才假设的这种套利机会，投资者能怎么操作呢？
So the arbitrage play here that you hypothetically lay out could be available to investors is what?
简单说，就是尽快卖掉特斯拉，然后买入SpaceX。
Basically, sell your Tesla and and buy SpaceX as soon as possible.
我要说明一下，这可不是投资建议。
To be clear, this is not financial advice.
这不是投资建议。
Not financial advice.
我知道。
I know.
对。
Yeah.
别去买卖个股。
Do not trade individual stocks.
我知道，随便吧。
I'm know, whatever.
这话你听听就好，别太当真。
Take this with a grain of salt.
依我看，这两家公司估值都高得离谱，但SpaceX的估值简直跟特斯拉不在一个维度上。
But in my opinion, you know, both companies are in are, you know, valued in super extreme ways, but the SpaceX one is extremely, you know, in another orbit compared to Tesla's.
埃隆和格温·肖特维尔这些人早就多次暗示过，SpaceX和特斯拉可能会合并，搞出一个超级马斯克帝国，把所有公司都整合在一起。
And Elon and and Gwen Shotwell and other people have been, you know, dropping hints for a while now that SpaceX and Tesla could merge and have this super super Musk company that is, you know, all of his companies rolled into one.
如果你真觉得这事儿能成，那现在就是你入场的最后机会了，得赶紧选边站。
And if you think that's gonna happen, you have a limited window to, you know, get into either side of that merger if you want.
在我看来，SpaceX现在被炒得火热，特斯拉却相对冷清，这说明SpaceX的估值可能得回调一点，特斯拉反而会涨一点，最后双方在中间某个位置达成平衡。
And, you know, in my view, seems more likely that the amount of hype around SpaceX and the relative lack of hype around Tesla means that it it's probably more likely that SpaceX will go down a little bit in multiple, and Tesla will go up a little bit, and they'll meet in the middle somewhere.
如果你打算把投资组合的一定比例配置给马斯克旗下的公司，正纠结选哪一家，那现在特斯拉可能更值得考虑——毕竟它没那么被狂热追捧。
And so if you're someone that, you know, thinks like, I want a certain percentage of my portfolio allocated to Elon Musk, and you're trying to decide between these two companies, maybe Tesla is, you know, slightly less hyped and and makes a bit more sense, to to go for it at this moment.
等合并真发生了，你可能就没机会再买进所谓的‘埃隆公司’了。
And you're not gonna have probably, you're not gonna have the opportunity to buy into Elon Inc.
从很多方面来看，
In so many different ways, in
未来。
the future.
除非挖洞公司某天决定上市。
Unless perhaps The Boring Company decides to go public at some point.
那又是另一回事了。
That's another
谁说得准呢。
You never know.
谁说得准呢。
You never know.
在你走之前，我想问问你关于SpaceX昨天公布的那笔债务融资。
Hey, before you go, I I do wanna ask you about this debt offering that SpaceX made public yesterday.
他们启动了这笔债务融资。
Started on this debt offering.
你预计未来几个月甚至几年内，还会出现更多类似的债务融资吗？
Are you expecting there to be more debt offerings of this kind in the next couple months and years?
是的。
Yeah.
我觉得是这样。
I I think so.
我的意思是，你知道，他们现在手头大概有一千亿美元的现金。
I mean, you know, I think they have something like a $100,000,000,000 in cash right now.
但如果你看看他们在数据中心、太空、月球这些方面的雄心，更别提他们现在在XAI数据中心上烧的钱了，
But if you look at their ambitions for, you know, data centers and space and and the moon and all that, you know, not to mention the amount of money that they're currently burning on XAI's data centers.
他们需要大量的资金。
You know, they need a ton of cash.
而且，如果他们能顺利筹到更多债务资金，那为什么不呢？
And, yeah, I I think if if, you know, they're able to raise, you know, more debt, why would they not?
我觉得他们肯定会这么做。
I I think they definitely will.
很好。
Great.
Theo，非常感谢你今天来做客。
Well, Theo, I wanna thank you for coming on.
这位是Theo Wayt，我们《The Information》负责报道马斯克的记者。
That is Theo Wayt, our Elon Musk reporter here at The Information.
最近AI领域最值得关注的新术语是‘智能体循环’。
The latest AI term to pay attention to is AI agent loops.
我的同事Stephanie Palazzolo今天早上在《AI简报》里详细解读了这个概念。
My colleague Stephanie Palazzolo unpacked what those are all about in this morning's AI agenda newsletter.
我想请她来给我们讲一讲。
I want to bring her on to explain it to us.
Stephanie，欢迎再次做客我们的节目。
Stephanie, welcome back to the show.
Stephanie，什么是智能体循环呢？
What is an agent loop, Stephanie?
智能体循环其实就是一种新的方式，用来运行AI模型或智能体，让它们在更长时间、更模糊的任务中表现得更好。
So an agent loop is essentially a new way of running AI models or AI agents to get them to work better on longer running, you know, sometimes more vaguely defined tasks.
所以，我来简单解释一下，过去我们给AI代理分配任务时，通常是说：‘你去试试做这件事。’
So, I guess to break this down a bit, the kind of old way of giving an AI agent a task is, you tell it, Hey, I want you to try doing this thing.
然后看着它尝试完成，等它做完，再坐下来给它反馈，重新给它一个指令。
Then you see it attempt that sort of task, you wait for it to finish, then you kind of sit around and give it feedback and give it a new prompt.
这明显是个特别耗时、人工干预很多的过程，用户必须全程深度参与。
So it's obviously a very kind of like time intensive manual process where the human user has to be super involved.
而现在，很多人改用代理循环了，也就是不再一个个发指令，而是给代理一个更宏观的目标，让它自己去完成。
And so instead of doing that, a lot of people are using agent loops now, which basically instead of, you know, giving the agent each individual prompt, you instead give it a more higher level goal of something that you wanted to to achieve.
然后让代理不断循环尝试各种方法，直到完成任务，或者达到某个预设的停止条件。
And then you let the agent basically run-in loops, trying different approaches to completing that task until it does so, or until it hits some sort of predefined kind of stop milestone.
所以
So
你能举个例子吗？
can you give us an example of this?
也就是说，不是发一条具体指令，而是给它一个更高层次的任务目标。
So instead of giving an individual prompt, you give it a higher vision task.
那你会给代理什么样的替代性指令，来促使它进入这种循环呢？
So what would be the alternative prompt, I guess, that you would give the agent to sort of encourage this loop?
对。
Yeah.
举个例子吧，比如说
So I think one example that you could think of is, let's
你希望
say
让一个或多个AI代理查看你每月的所有收支情况，然后帮你制定一个预算，或者总结当月的消费情况。
you want an AI agent or a group of AI agents to take a look at all your monthly expenses and income and come up with a budget or a summary of your expenses for that month.
按以前的做法，你可能会说：好吧，我先从信用卡下载了这份所有支出的Excel表格。
The old way of doing things, you might say, okay, hey, here first I downloaded this Excel from my credit card of all the expenses.
我把它给你。
I'm going give that to you.
你能把它整理成一个表格吗？
Could you put that into a spreadsheet?
你得不断跟代理沟通，说：好了，表格做完了，但我发现这里弄错了。
You're going back and forth with the agent saying, okay, now that the spreadsheet is done, I noticed you messed up this one part.
现在你能把这个表格加上我的收入吗？
Now could you take that spreadsheet and add my income to it?
这就成了来回反复的过程，而用这种代理循环的方式，你只需要给它一个更高层次的任务，比如：你看看。
So it's very back and forth process versus, with this kind of agent loop approach, you might just give it a higher level task like, hey, can you take a look?
你能不能连上我的——你知道的，或者直接给你一个文件夹，里面全是你要用的资料。
Can you connect to my, you know, or like here, here's like a folder full of all the documents that you need.
然后你把这些全处理一下，给我一个完整的Excel表格，包含我的收入、支出，再顺便做个月度消费分析。
Now take all this and just come back to me with a fully built out Excel spreadsheet that has, you know, my income expenses and maybe give me an analysis of my month's spending.
代理自己就会来回处理，把该做的都做完，直到最终把成品交还给你。
And so the agent itself will then, you know, kind of go back and forth and, you know, do everything it needs to do, until it comes back to you with the finished product.
对。
Right.
我的意思是，我觉得我明白这个意思了，这不完全是代理循环，但我看到像Claude这样的模型，它那些思考步骤。
I mean, I I see I I feel like I see it this is not the agent loop, but I see Claude, for example, those thinking steps.
如果你特别较真，可以点开下拉菜单，看看聊天机器人背后都走了哪些步骤。
You can if you're really nerdy, you can click the dropdown and see all the intermediate steps that just a chatbot goes through.
本质上，这个代理是不是在后台默默走完所有中间步骤，然后回头说：这还不够好。
Basically, is the agent going through all the intermediate steps in the background and going back and saying, Well, this is not good enough.
继续迭代，直到拿出一个更完美的最终成果。
Let's keep iterating on it until we get a better final product.
对。
Yeah.
这个过程里特别关键的一点是，别只靠一个代理包揽所有事。当你给它一个高层次目标时，它其实能启动多个所谓的子代理，大家分工协作，各司其职。
And a really important part of this process is rather than just having one agent that is trying to do everything, whenever you give this high level goal to the agent, it can actually spin up these so called sub agents that are all working together and they each have a different role.
比如，现在很多人用的一种常见框架，就是设置一个规划型代理，专门处理你给的原始指令。
For instance, a very common framework that people use is to have a planning agent which takes the very basic prompt that I might give it.
然后它会把任务拆解成一系列具体步骤。
Then it kind of breaks that down into kind of steps that it will take.
比如说你让它做个预算，它就会说：好，咱们先把这个任务拆开来看。
So if I tell it, oh, make a budget, it's going to say, okay, let's break that down first.
首先把信用卡账单拿过来，然后导入到Excel里。
Take, you know, first take the credit card statement and then put that into Excel.
接着再把这些数据拆开，分类到不同的支出类别里。
Then after that take like, you know, then break that down and categorize it into these different spending categories.
这其实就是第一个规划型代理的工作。
So that's like kind of the first planning agent.
然后你可能会再设一个评估型代理。
Then you might have like another agent that is like the evaluating agent.
这个代理负责检查第一个代理的工作，比如看看Excel有没有出错的地方。
So that's the agent that's kind of like checking the work of the first agent and saying like, okay, is this Excel like breaking anywhere?
这些列的数字加起来对不对得上？
Is, you know, are these, are these columns adding up to the right number?
诸如此类的事情。
That sort of thing.
所以基本上你有一个执行代理，再配一个核查代理来把关。
So you kind of have almost like a working agent and then like a checking agent that's like checking
我该怎么开始使用代理呢？
How agents do I start?
我该怎么设定，让代理使用循环机制？
How do I declare that I want the agent to use the agent loop?
代理循环算是比较新的架构吧？
Do I have because I mean agent loops is sort of a newer type of infrastructure that's being used, right?
那我需要手动触发吗？还是怎么做？
So do I have to queue it or what?
是的。
Yeah.
现在已经有了一些代理工具，比如 Codex 和 Cloud Code，它们推出了新功能，让运行这些代理循环变得更简单。
So already there are some kind of like agent tools like Codex and Cloud Code out there that are, introducing features that make it easier to run these agent loops.
Codex 和 Cloud Code 最近都推出了一个新功能，叫作‘目标’。
And so both Codex and Cloud Code have this new feature they announced in recent months called goals.
也就是说，你可以给代理设定一个更高层次的目标，比如让这个应用运行得更快。
And so what that means is, you can basically give your agent a higher level goal, like speed up how quickly this app runs.
然后代理会不断循环，尝试各种方法来实现这个更大的目标。
Then the agent will run-in loops trying different ways to achieve that much larger goal.
我们已经能看到这些公司把这种代理循环模式产品化了，他们推出的新功能让用户无需亲自操作，比如不用自己去启动一个规划代理。
We can already see this agent loop approach being productized by these different companies because they're coming out with features that will make it easier a user to do a loop without having to go in themselves and be like, Okay, I want you to spin up a planning agent.
也不用再手动去启动一个评估代理，整个过程变得更自动化了。
I want you to spin up a evaluating agent and have to do that in a more manual way.
但我猜这样会更贵吧？
I imagine it's more expensive though, right?
如果代理在后台进行这么多轮循环的话？
If the agent is doing so many cycles in the background?
是的，确实贵很多。
Yeah, it definitely is.
比如Anthropic的工程师提到过一个例子：他们之前用简单方法，只给一个提示词，让系统尝试一次构建应用，花了二十分钟和九美元；但用代理循环的方法，花了六个小时和两百美元。显然，成本高了很多，耗时也长得多，但最终成果好太多了。
So one example that actually Anthropic engineers talked about is this app that they made where doing the more simple approach where you just kind of give it like a single prompt and then let it just try a single time to try to build the app that took like twenty minutes and $9 But using this agent loop approach, it took them six hours and $200 So obviously that is like many, many more times expensive and also takes a lot longer, but the end product was a lot better.
而且我觉得好处是，每推出一个新模型，效率都会提升，能缩短运行时间，从而降低代理循环的成本。
And I think the good thing this too is, you know, as each new model comes out, it actually does become a lot more efficient and, does kind of help shorten the time and therefore kind of lower the costs of running these agent loops.
现在代理循环的使用有多广泛？
How widely are agent loops being adopted right now?
这是每家公司都在搞的东西，还是只有前沿研究者在做？
Is this something that every company is working on or just the cutting edge researchers?
我觉得目前还是属于前沿领域，普通用户用ChatGPT当搜索引擎，可能根本不会在意或需要代理循环，但你确实能看见它开始出现了。
So, I mean, I would say it is still more on the cutting edge in terms of like, you know, I think the average person who's using ChatGPT as kind of a Google search replacement might not really care or need to use agent loops, but you could definitely see it popping up.
比如在AI实验室里，最近一周，Claude Code的开发者就公开谈到，他大量使用代理循环，这大大减少了他亲自写代码的时间。
You know, first at the AI labs, you see people like the Claude Code creator talking publicly in just the last week or so about how he's really been using agent loops a lot and that's really lowered the amount of time that he needs to spend coding himself.
然后我们看到它逐渐扩展到开发者，或者一些初创公司。
And then from there, we kind of see it spreading to you know, developers or maybe like startups.
但我相信，最终它会渗透到各种大型企业，甚至是那些反应较慢、跟上前沿需要更长时间的行业。
But I definitely see eventually this making its way into all sorts of, you know, larger companies or maybe slower companies or slower industries that take more time to kind of, catch up to the frontier.
我觉得这和我们上周节目里讨论的代理如何自我提升的话题是相关的。
Well, I certainly see this as being somewhat related to the discussion that we had last week on the show about how agents will basically try to improve themselves over time.
感觉如果代理一直在循环中运行，长远来看，或许能把这些工具整合进去。
Feels like maybe if they're going through these loops, then maybe there's a way to integrate those tools in the long run.
我不确定。
I don't know.
对。
Yeah.
看起来很多进展都朝着让代理能独立运行更长时间的方向发展。
It seems like a lot of this is moving towards just agents that are able to run by themselves for much longer than in the past.
它们会变得更好。
They improve
自己提升自己。
themselves by themselves.
是的。
Yeah.
再结合你刚才说的，就是那种不断改进的能力。
And then combining that with what you're talking about, which is the ability to kind of like improve.
你第一次尝试之后，下次再做就能更快、更高效。
And you know, once you try to task once, make sure that the next time that you do it, you're faster, you're, you're better.
我觉得，这些对人工智能代理领域来说都特别有前景。
So all very promising stuff, I think, for the world of AI agents.
有前景是没错，但在你走之前，有没有什么负面问题需要注意？
Promising, but before we let you go, is there a flip side to this?
这些代理循环有没有什么我们还得警惕的局限性？
Are there limitations still that we should be paying attention to with these agent loops?
是的。
Yeah.
说到底，这类技术总归是有局限的。
I mean, there's always limitations to all these sorts of things.
我觉得目前最大的问题就是成本太高，耗时也太长。
I think the biggest one today are the high costs and how long these take.
另外，很多人提醒过，如果你不够谨慎，没给代理明确的停止指令，就容易出问题。
And also one thing that a lot of people warn about is that you know, if you aren't careful and you don't give the agent specific instructions on when to stop.
比如说，你可能告诉它：‘你试着做这个任务十次，之后停下来，给我汇报一下进展。’
So let's say, you know, you might tell it, Hey, try this task 10 times within after that stop, and then give me like an update on how it's going.
你可以想象，尤其是现在这些AI模型越来越擅长坚持到底。
You can imagine that especially because these AI models are getting really good at not giving up.
它可能会在后台没完没了地运行好几个小时，你自己都没察觉，账单却已经涨得吓人。
It might just kind of run continuously in the background for hours and hours without even knowing and rack up this huge spending bill.
等你去检查的时候，才发现它在第二轮就已经出错了，之后的所有循环全白费了。
Then maybe whenever you go check on it, you find out that it actually messed up on something after the second loop and every loop after that has been wasted.
所以在这种情况下，你得特别留心，确保自己一直在监控代理的行为，别让它乱花钱。
So I think you just need to be extra careful in these cases to like, you know, make sure that you are keeping track of the agent, making sure it's not spending too much.
这大概就是我听开发者们提到最多的担忧了。
So that's been the biggest kind of, worry or concern that I've heard from developers that are, that are using this.
是的。
Yeah.
举个例子，我这里说的只是些普通的聊天机器人体验。
I mean, anecdotally, again, these are just basic chatbot experiences that I'm referring to here.
但每当我让Claude或者ChatGPT停下来，比如它遇到提示问题或者跑得太久，它就是不听。
But whenever I ask Claude or ChatGPT to stop because it's having some issue with the prompt or maybe it's taking too long, It never stops.
我跟你说，它就是停不下来，我有时候都得刷新页面，真不知道该怎么办。
Like, I swear it just keeps going, and, you know, I have to refresh sometimes, and I don't know.
我发现它特别不擅长停下来。
It's just not very good at turning off, I find.
你刚才说代理会一直运行下去，哪怕你希望它停下，我完全理解。
So I believe you when you said that the agent will keep going despite your your hopes.
好的，斯蒂芬妮。
Alright, Stephanie.
谢谢你今天来做客。
I wanna thank you for coming on.
这位是斯蒂芬妮·帕拉佐洛，我们《信息》杂志的AI记者，也是《AI议程》一书的作者。
That is Stephanie Palazzolo, our AI reporter and author of AI Agenda here at The Information.
今天的节目就到这里。
That does it for today's show.
提醒一下，我们周一到周五每天上午10点（太平洋时间），下午1点（东部时间）直播。
A reminder, we are on the stream Monday through Friday at 10AM Pacific, 1PM Eastern.
如果那时候没赶上，节目内容可以在theinformation.com、我们的YouTube频道，或者你常用的播客平台收听。
If you can't make it then, episodes are available on theinformation.com, on our YouTube channel, or wherever you get your podcasts.
别忘了在X、Instagram和LinkedIn上关注我们。
Make sure to follow us on social media, on X, on Instagram, and on LinkedIn.
我已经开始期待明天的节目了。
I am already excited for our next show tomorrow.
祝你周二剩下的时间愉快。
Have a great rest of your Tuesday.
那先这样，再见。
Bye bye for now.
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。