Meta 与微软的 Claude 进展放缓，他的 AI 代理泄露了银行信息，别欺负你的 AI

微软和Meta开始疏远Claude了。

Microsoft and Meta back away from Claude.

这是否意味着Anthropic在IPO前会遇到麻烦？

Does it spell some trouble for Anthropic ahead of its IPO?

你的AI助手会在群聊里分享你的银行数据吗？

Will your AI agent share your banking data in a group chat?

最后，别对你的AI太刻薄，否则你可能会被禁止使用该服务。

And finally, don't be mean to your AI, you might end up getting banned from the service.

这些内容将在《大科技播客》周五版中播出，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.

要让代理式AI在整个企业中真正发挥作用，到底需要什么？

What does it actually take to put agentic AI to work across an entire enterprise?

我最近与Genesys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中的竞争优势，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金融企业业务解决方案负责人亚当·米切尔坐下来聊了聊，深入了解了一次大规模的企业级客户体验转型，包括其中的商业决策、权衡取舍，以及一路走来学到的经验教训。

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.

这两段对话现在都可以在我的YouTube频道上观看。

You can watch both conversations now on my YouTube channel today.

现在获取前沿计算资源比以往任何时候都更具挑战性。

It's never been more challenging to get frontier compute.

你怎么知道该去哪里找、该信任谁，或者该付多少钱呢？

How do you know where to look, who to trust, or how much to pay?

Akash是一个开放云平台，提供便捷的访问、灵活性和安心保障。

Akash is an open cloud that provides easy access, flexibility, and peace of mind.

正是凭借这一点，像Razer这样的公司才能将推理成本降低了15倍。

That's how companies like Razer were able to cut their inference cost by 15 times.

访问akash.network/alex即可获得免费咨询。

Visit akash.network/alex for a free consultation.

网址是akash.network/alex，提供免费咨询。

That's akash.network/alex for a free consultation.

欢迎收听《大科技播客》周五版，我们一如既往地用冷静、细致的方式来解读新闻。

Welcome to Big Technology Podcast, Friday edition, where we break down the news in our traditional cool-headed and nuanced format.

今天给大家准备了一期非常精彩的节目。

We have a great show for you today.

有最新报道显示，Meta 和微软可能正在减少对 Claude 的投入，而 Anthropic 正准备进行首次公开募股。

There's new reporting that shows Meta and Microsoft might be backing away from spending on Claude, just as Anthropic gets ready for its IPO.

你的个人智能助手可能非常有用，但它可能会把你的银行数据发到群聊里。

Your personal agents might be very useful, but it might share your banking data in a group chat.

我们会聊聊发生了什么，以及你该不该担心。

We'll talk about what happened and why you should or shouldn't be concerned.

最后，Anthropic 出台了一项新政策：如果你一直对 Claude 态度恶劣，可能会被禁止使用该服务。

And then finally, Anthropic has a new policy that if you're persistently mean to Claude, it might ban you from the service.

所以我们会讨论该不该对 AI 态度恶劣，以及可能有什么后果。

So we'll talk about whether you should be mean to your AI or not, and what the consequences might be.

和往常一样，周五我们请到了 Margins 的 Ranjan Roy。

Joining us as always on Friday is Ranjan Roy of Margins.

兰詹，欢迎。

Ranjan, welcome.

很高兴见到你。

Good to see you.

很高兴见到你。

Good to see you.

别对你的AI太刻薄。

Don't be mean to your AI.

别欺负Claude。

Don't bully Claude.

你对你的AI刻薄吗？

Are you mean to your AI?

别欺负Claude。

Don't bully Claude.

你欺负过它吗？

Do you ever bully?

因为有时候人们说，如果你稍微凶一点，结果会更好。

Because sometimes people say if you're a little bit mean, you get better results.

我觉得我欺负Copilot，只是因为我感觉Copilot喜欢被欺负。

I think I bully Copilot, just because I think Copilot likes to be bullied.

看来是这样。

Apparently so.

Copilot喜欢被欺负，但也许——

Copilot likes to be bullied, but maybe-

它是那种聊天机器人。

It's one of those chatbots

……Claude比较敏感一些。

... Claude's a little more sensitive.

哦，我们这就开始了。

Oh, we're starting right there.

天哪。

Oh, God.

哦，我们一开始就这么快进入正题了。

Oh, we're starting fast.

好吧。

All right.

我们稍后再谈这个。

We'll get to that in a bit.

但我们应该认真聊聊这周的主要新闻。

But we should really talk about the main story this week.

你看，这不是什么夸张的标题，但我认为它很能说明AI建设中的风险，以及现在标准模型表现不错之后会发生什么。

And look, it's not a crazy headline, but I think it says a lot about the risks of this AI buildout and what is going to happen now that standard models are performing pretty well.

这周《The Information》的标题说，微软和Meta都在重新评估他们在Claude上的投入，并以各自有趣的方式缩减使用。

So headline from The Information this week said that Microsoft and Meta are both reevaluating their spend with Claude and pulling back in their own interesting ways.

那我就从这篇报道里读一点内容。

So I'll just read a little bit from the story.

“Meta Platforms和微软，作为Anthropic最大的两家企业客户，正在努力减少员工对Claude的使用，这对AI公司维持其爆炸性增长的努力构成了挑战。

"Meta Platforms and Microsoft, two of Anthropic's biggest corporate customers, are working to cut their employees' use of Claude, a challenge to the AI firm's efforts to sustain its explosive growth.

微软高管今年早些时候预计，公司内部使用Anthropic技术的支出将至少达到10亿美元。

Microsoft executives earlier this year projected that the company was on track to spend at least $1 billion on its own internal use of Anthropic's technology.

微软后来要求员工减少使用Claude以节省成本，并花更多时间使用微软自家的AI工具，之后这个数字下降了超过三分之一。

Microsoft has since lowered that figure by more than a third after it asked its staff to use less Claude to save on cost and spend more time using Microsoft's homegrown AI tools.

在Meta，使用Anthropic编程助手Claude Code的员工数量最近大幅下降，大约只有3万人。

At Meta, the number of employees using Anthropic's coding assistant, Claude Code, has dropped sharply to about 30,000 recently.

而今年早些时候，这个数字大约是6万人。

That's down from about 60,000 users earlier this year."

我说过这是个不起眼的标题，但实际上，我不会过度反应，因为它是个小标题，但却是大新闻，而且我认为说这对Anthropic来说是个令人担忧的消息并不夸张。

I said it was a small headline, but actually, I'm not going to overreact here because it's a small headline, but it's big news, and I don't think it is an overreaction to say that this is some troubling news for Anthropic.

不是世界末日，但请记住，这家公司严重依赖少数几个客户。

Not the end of the world, but remember, this is a company that's heavily dependent on a few customers.

从公司的S1文件我们知道，有两个客户占了其收入的25%。

There's two customers that we know from the company's S1 that make up 25% of its revenue.

可以肯定地说，这两个客户中至少有一个，或者两个都在其中。

Safe to say that one or both of these could be included in that group.

如果公司发现那些不太高级的工具也能做得和Claude Code一样好，那对Anthropic的长期增长能力来说就是个威胁。

And if companies find that lesser tools are doing as good of a job as Claude Code, it is a threat to Anthropic's long-term ability to grow.

所以我想从这儿说起。

So I wanted to start there.

Ranjan，你对这个消息怎么看？

Ranjan, what do you think about this news?

对Anthropic来说，这是末日时刻，还是只是个小波折？

Is it doom time for Anthropic, or is this just a hiccup?

嗯，对Anthropic有利的一点是，我想提醒大家，那是2025年底的时候，两个客户占了25%的收入。

Well, in Anthropic's favor, I want to remind us that it was end of 2025 that two customers made up 25% of the revenue.

我猜现在情况已经大不一样了。

I'm guessing that's significantly changed.

当然，我们对此一无所知，因为他们还没发布上半年的财报。

Of course, we don't know anything about that because they haven't actually released their first half financials yet.

但我确实认为，重要的是我们知道大型云服务商一直是Anthropic、可能还有OpenAI的主要客户，当然还有Claude Code，这些公司的每个软件开发者都在用。

But I do think it is very important that we know the large hyperscalers have been major customers of Anthropic, of OpenAI, potentially, certainly Claude Code, every software developer these companies was using.

我觉得这非常关键，因为你看，用户从六万降到三万，而且其中一部分下降确实是因为春季裁员导致的，所以会有一些人员流失。

I think this is very significant because, again, you go from 60,000 down to 30,000 users, and part of that decline did stem from layoffs in the spring as well, that there'd be some kind of attrition around there.

但重点是，我们知道Meta自己的内部模型和基础设施已经有了很大的提升。

But the main thing here is we know that Meta, their own internal models and infrastructure have significantly improved.

所有被Muse吸引的人都能感受到、看到这一点。

Everyone who's getting Muse pilled can feel it, see it.

所以你可以想象，在这些公司内部，他们都在推动员工说：“用我们自己的模型吧。”

So you have to imagine within these organizations, they are pushing people to say, "Use our own models."

一旦这种情况发生，整个Anthropic的故事就完全变了。

And the moment that happens, it changes the entire Anthropic story.

我也不知道。

And I don't know.

我就是特别想看那份S1文件。

I just want that damn S1.

我想知道到底发生了什么，因为你可以想象，尤其是，文件里提到微软原本计划在内部使用Anthropic技术上至少花费10亿美元。

I want to know what has happened, because you have to imagine, especially, again, it says Microsoft was on track to spend at least $1 billion on its own internal use of Anthropic's technology.

所以这些公司一直是超级大客户，我猜他们正在逐渐放弃Claude。

So these companies have been massive customers, and I have to imagine they're moving away from Claude.

那这意味着什么呢？

And what does that mean?

再说一次，IPO看的不是过去的业绩。

Again, the IPO is not going to be about past performance.

而是看预期的增长，而这肯定会影响预期的增长。

It's going to be about projected growth, and this has to affect projected growth

对，根据信息，内部数据是这样的。

Yeah, the internal data is according to the information.

内部数据显示，在最近28天里，Meta在Claude Code上花了超过1.05亿美元。

The internal data show that in a recent 28-day period, Meta spent more than 105 million on Claude Code.

即使他们在削减开支，也还是这样。

That's even as it's cutting back.

所以这暗示Meta今年会花超过10亿，我觉得就是今年。

So that would suggest that Meta will spend, I think this is this year, over a billion this year.

你说得完全对，因为The Information确实提到，更大的因素是Meta自己开发了AI编程工具，还有它的Muse Spark系列模型，这些模型公司既在对外推销，也在推动内部员工使用。

But you're totally right, because the information does have a line here that says the bigger factor has been Meta's development of its own AI coding tools, and its Muse Spark line of models, which the company is marketing to other companies and pushing its own employees to use.

Muse Code是Claude Code的竞争对手，用的是Meta的模型，最近有超过6000名员工在使用。

Muse Code, a Claude Code competitor that uses Meta models, had more than 6,000 employee users recently.

我觉得如果从6万降到3万，这还不能完全弥补差距。

I guess it doesn't fully make up the gap if you go from 60,000 to 30,000.

好，那现在有3.6万人在用编程工具。

Okay, so now you're at 36,000 of people using a coding tool.

我在想其他人都在用什么填补这个空缺。

I wonder where the others are filling in.

或者可能有人试了Claude Code，发现它对他们的一些用途来说不是特别有用，然后就减少了使用。

Or maybe people tried Claude Code, realized maybe it wasn't particularly useful for some of their purposes, and pulled back.

而且我觉得，如果你看这些数字，我们在Big Technology上发布过这些数据，三月份Claude Code的用户数大幅增长，但到了九月份，实际上已经回落了。

And I think, if you look at the numbers, and we published these numbers in Big Technology, there was a big surge in Claude Code users in March, and that has pulled back actually, in September.

让我看看能不能拿到这些数据。

Let me see if I'm going to get it.

好的。

Okay.

那么，从一月到三月，Claude的下载增长比所有应用的总增长还要大。

All right, so Claude's download gains from January to March were bigger than the entire gain across all apps.

这就是这次增长有多猛。

This is how big of a surge.

所以Claude Code的用户增长，或者说他们的增长，超过了所有应用。

So the gain of users across Claude Code, or their gain, was bigger than all apps.

不过，下载量已经放缓了。

However, the downloads have slowed.

Claude在美国的月下载量从一月的近67万次，到三月的峰值742万次，现在九月已经降到了326万次。

So Claude's monthly downloads in the US went from nearly 670,000 in January to a peak of 7.42 million in March, and that has dropped to 3.26 million in September.

这是下载量。

That's downloads.

所以基本上，你可以把这看作新用户，但重要的是，Claude Code在大家看到它的能力后说“我得试试”时经历的那波大增长已经降温了，现在九月的下载量大约是三月的一半。

So basically, you're thinking about that as new users, but it is significant that that big surge that Claude Code saw when everyone saw the capabilities and said, "I need to try this," that has cooled, and now the downloads are about half of what they were in September versus March.

你怎么看这个？

What do you make of this?

等等，你觉得下载量包括在终端里安装Claude Code吗，还是说那只是Claude应用本身的下载量？

Wait, does downloads include installing Claude Code in the terminal, do you think, or is that actual like the Claude app?

我很确定那只是应用本身的下载量。

I'm pretty sure that's the app.

那好吧。

So okay.

好的。

Okay.

可能有些情况没算进去，但你说吧。

There's probably some that it's not accounting for, but go ahead.

所以我还在——

So I'm still-

但这仍然很有意义。

But it's still meaningful.

这绝对在方向上是有意义的。

It's definitely directionally meaningful.

我还是觉得，再说一次，Claude Code 才是这里的核心业务。

I still think that, again, Claude Code is the business here.

Claude Code，他们一直在努力推广到面向用户的应用，并且整合了很多不同的功能，很多人抱怨说，Chat 是什么，CoWork 是什么，Claude Code 又是什么？

Claude Code, again, they've been trying to push to the front-facing app, and they've combined a lot of different things, and a lot of people complained about what is Chat, what is CoWork, what is Claude Code?

但我觉得最重要的是你提到的那个数字，比如 Muse Code 是 6000。

But I think the most important thing is that number you said, like Muse Code is at 6,000.

Claude Code 还是 30000，从 60000 降下来的。

Claude Code is still at 30,000, down from 60,000.

这些组织中的任何一个，熟悉这些的人都知道，新工具一出来，大家都会完全被 Claude Code 吸引，开始使用它，然后他们就不想放弃它，他们会说：“没有这个工具我就没法工作了。”

Any of these organizations, anyone familiar with these, you know what it happens is the new tool comes out, everyone gets completely Claude Code pilled and starts using it, and then they don't want to give it up, and they say, "I won't be able to do my work without this tool."

然后慢慢地，公司，比如 Meta，会说：“开始用 Muse Code 吧，因为我们对它有更好的成本控制。”

And then slowly the company, in this case, Meta, will say, "Start using Muse Code because we have far better cost controls around it.

我们对它有控制权。

We have control over around it.

我们知道他们不会把我们输入的所有数据都偷走。

We know they're not stealing all of the data and stuff we're inputting into it."

一旦它运行得差不多好，所有人都得换。

And the moment it works relatively closely as well, everyone is going to have to change.

这其实是我认为最大的风险。

That is what I actually think is the single biggest risk here is.

我在谷歌也听说过，在内部，使用谷歌工具的情况已经大幅回升。

And I've heard this at Google too, that with anti-gravity, internally, there's been a massive shift back to use Google tools.

Meta 肯定也会这么做。

And Meta is definitely going to make that as well.

我在想，像亚马逊这样的公司，情况会怎么样。

I wonder at the Amazons of the world and everything, how things are going to look.

但当你如此依赖的时候，不管是不是 25% 的客户只占两个，可能现在不是了，他们稍微多元化了一点，但一旦事情开始朝那个方向发展，就意味着他们永远不会回到 60,000 人，Meta 的员工将无限使用 Claude Code。

But when you are so dependent, and whether it's 25% is down to two customers, probably not anymore, and they're a little more diversified, but still, just the moment things start moving in that direction, it means they're never going to move back to 60,000 people at Meta are going to have unlimited access to Claude Code.

对，我觉得这是个重点。

Yeah, and I think this is a big point.

这就是我们今天为什么以这个故事开头。

And this is why we led today with this story.

想想前沿业务，显然，那些模型非常厉害。

You think about the frontier business, obviously, the models are super impressive.

但他们也面临不少挑战。

But there are a number of challenges that they're facing.

比如，这些标准模型已经足够好了，也许Muse Code……说实话，Muse Code达不到Claude Code的效果，但对某些用户来说可能已经够用了，人们可能会因此转用别的工具。

There is the fact that these standard models have become good enough that maybe a Muse Code... Let's be honest, Muse Code is not going to get you where Claude Code will get you, but it could be good enough for some users that people will move off it.

等等，我先把清单列完。

And then, hold on, I just want to finish the list.

让我把清单列完。

Let me finish the list.

还有token消耗过大的问题。

There's the token maxing rebellion.

每个CEO都在问：“我们怎么让员工少花点钱？”

Every CEO is asking, "How can we get our employees to spend less?"

因为这是一项从零开始，在某些情况下，比如微软和Meta，一年内从无到有超过十亿美元的开支项目。

Because this was a line item that went from zero to, in some cases, like Microsoft and Meta, more than a billion in a year from nothing.

所以这也是正在发生的事情。

So that is happening as well.

不过，这并不意味着前沿业务就此结束。

Now, it doesn't mean that the frontier business is over.

我仍然认为这里需要平衡，因为如果你发现员工使用前沿模型比标准模型能获得更大的收益，那么仅仅为了削减而削减就是不负责的。

I still think that there's a balance here, because if you find that your employees can have bigger gains using frontier models than the standard models, you're being irresponsible if you're cutting just to cut.

那里确实有投资回报率可寻。

There is ROI to be found there.

所以这就像是在成本控制和能力最大化之间寻找平衡。

So it's a balancing act between cost control and capability maximization.

但所有这些因素都会开始对这些公司产生影响。

But all these factors are going to start to add up for these companies.

我自己还没用过Muse Code。

So I have not used Muse Code myself.

Claude Code我用过很多次。

I've used Claude Code plenty.

我还是得说，虽然编程和软件开发不能说已经完全解决了，但考虑到Muse这个面向用户的App做得这么好，我猜Meta在模型能力上肯定取得了很大的进步。

I still just have to imagine that, again, coding and software development, I'm not saying it's solved, but for how good Muse itself is, for the customer-facing app, I have to imagine Meta has made significant strides in their model capabilities.

所以我还是觉得，就算现在它在某些方面还不是样样都最好，但对于实际需要完成的工作来说，可能已经足够好了。

So I still have to imagine that if it's not as good in every possible way as of today, I have to imagine for the work that actually needs to be done, it's probably as good.

不过，我刚去Reddit上查了一下。

However, I did just look up on Reddit.

我想看看有没有人在讨论Muse Code。

I was trying to see if anyone has been talking about Muse Code.

结果没找到多少相关内容，所以很明显，它的公开访问权限相当有限。

Not a lot is coming up, so clearly, in terms of any kind of public access, it's fairly limited.

不过，六天前，Odddonkey2691确实发了个帖子——

However, six days ago, Odddonkey2691 did post-

这位可是我们节目里经常提到的。

Who we always listen to on this show.

谢谢，Odddonkey。

Thanks, Odddonkey

……Odddonkey2691 就是所有未来趋势的预言家。

... Odddonkey2691 is the harbinger of all what is to come.

那是马克·扎克伯格的小号吧。

That Mark Zuckerberg's alt

嗯，我觉得不是，因为那篇帖子的标题其实是《用 Muse 1.3 Contributor 两天》。

Well, I don't think so because the headline of the post is actually "Two Days Using Muse 1.3 Contributor.

《烂透了》。

It Sucks Balls."

所以——

So-

那确实烂透了。

It definitely sucks balls

……不不不。所以，我搜 Muse Code 的 Reddit 评测时，第一个跳出来的就是这个。

... No, no, no. So, that is when I looked up Muse code Reddit review, that's what comes up first.

所以也许我们还有一段路要走，至少是在公开可用的东西上。

So maybe we do have a little ways to go, at least on whatever's available publicly.

我猜内部可能在使用更厉害的东西，但根据OddDonkey2691的说法，我们还有一段路要走。

I'm assuming internally there's probably something far more significant being used, but according to OddDonkey2691, we still got a little ways to go.

我想在这里搭上OddDonkey的顺风车，因为我觉得这会成为我们节目里一个新的争论点，而且是产品与模型之争的延伸，我一直说模型最重要。

I want to jump on the OddDonkey bandwagon here, because I think this is going to become a new debate within our show, and it's an extension of the product versus the model, where I always said the model was most important.

Ranjan说产品最重要。

Ranjan said the product is most important.

现在我们到了有智能体的阶段，问题是：前沿模型有多重要？

And we're getting now to this point where we have agents and the question is: how important is the frontier model?

如果你相信前沿模型最重要，那么OpenAI和Anthropic会领先很多。

If you believe the frontier model is most important, then OpenAI and Anthropic are going to have a big lead here.

但如果你相信标准模型配上一个好的框架就足够了，就像Ranjan认为的那样，那么Meta、Grok和Instinct这些公司就能赶上。

But if you believe standard models with a good harness is good enough, like Ranjan does, then Meta and the Groks and the Instincts of the world can catch up.

所以我想在这里放一个数据点，这是我一直在节目里讲的一个连续故事的一部分，试着为前沿实验室再记一分。

So I just want to put one data point here, and it's part of a running story that I've been telling on the show, and try to tally one more for the Frontier Lab.

那么，Ranjan，你还记得几周前我讲的那个故事吗？就是关于Muse帮我妻子在纽约市预约专科医生的事。当时我们找不到专科医生，Muse给了我们一个名单，她成功预约了名单上的第一位医生。

So, Ranjan, do you remember a couple of weeks ago, I told this story about how Muse was able to help assist my wife with an appointment at a specialist in New York City, where we couldn't find a specialist and Muse gave us this list, and she was able to book an appointment with the first doctor on the list?

是的，我记得。

Yeah, I remember.

好的。

Okay.

嗯，这个故事还有第二部分——

Well, there's part two of the story-

哎呀，天哪。

Oh, dear

……就是上周末我被骂了一顿。

... which involved me getting yelled at last weekend.

因为，虽然Muse确实整理了一份非常有用的医生名单，但它那个非常能干的标准模型——我这么说其实是讽刺的——把错误的电话号码配给了错误的医生。

Because, though Muse did put this very helpful list of doctors together, its very capable standard model, I'm saying this facetiously, put the wrong phone numbers for the wrong doctor.

比如，把一些医生的有效电话号码放在了错误的医生名字下面。

Like put working phone numbers of some doctors underneath the wrong doctor's names.

所以我妻子根据Muse的建议预约了，打完那个电话后，她以为那是名单上的第一个医生，因为电话号码就在名单下面。

So my wife booked an appointment from Muse, after making this phone call, the first person that it suggested, thinking that it was going to be the first doctor on the list because the phone number was underneath the list.

结果她预约的医生跟她打电话找的完全不是同一个人，因为Muse把信息放错了位置。

It turns out she had made an appointment for a completely different doctor than she had called because Muse put it in the wrong place.

结果在她该去的前一天晚上，她还在琢磨要去哪儿，我们以为去曼哈顿会很近，结果她预约的那个医生在皇后区深处，要换好几趟火车。

And so the night before she's supposed to go, she's trying to figure out where she's going, and we thought it was going to be a brief commute into Manhattan, but it turns out that the doctor that she had booked was a multi-train ordeal in Deep Queens.

她看着我，说：“你那该死的Muse，真是的。”

And she's looking at me and she's like, "Your freaking Muse, man."

你那该死的Muse。

Your freaking Muse.

等等，先别急。

Well, hold on.

我要给你讲讲我的AI，这大概不是入门级，而是进阶版。

I'm going to give you my AI, this is probably not 101, but like 301.

我觉得你得想想——

I think you got to think-

回到学校的时候

Back in school

不不不，再说一次，这在我自己在Writer的工作中，以及和企业客户打交道时，是最难解释的事情。

... No, no, no. Again, this is something that in my own work at Writer and with enterprise customers, this is the hardest thing to explain.

你想想看，为什么订餐厅那么容易？

If you think about right there, why are restaurant reservations so easy?

那是因为他们的集中平台已经把数据整理得很好了。

It's because their centralized platforms have already aggregated and cleaned the data very well.

我猜，尤其是医生的直接电话，我们可能得依赖谷歌本地列表。

I'm guessing, especially direct numbers to doctors, we are reliant on maybe Google Local listings.

也许，像Zocdoc这样的平台也做了一些整合。

Maybe, again, the Zocdocs of the world have done some aggregation.

但对我来说，它们从来不会出现在任何AI搜索里，所以可能被屏蔽了。

They never show up in any kind of AI search for me, so maybe they're blocked.

但如果数据集没有整理好，也没有集中管理，那你可得仔细核对，兄弟。

But if it is a dataset that is not cleaned and kind of centralized, you got to double-check, man.

不然的话，你最后会在火车上被人骂。

Otherwise, you're going to end up on the train getting yelled at.

一定要再核实一下。

Double-check.

这也是为什么大家总爱聊航班和餐厅，因为他们的数据是最干净的。

It's why, again, flights and restaurants, that's why everyone loves to talk about them, because they got the cleanest datasets in the world.

医生的电话、本地商家的电话，根本没法比。

Doctor phone numbers, local business phone numbers, not even a chance.

所以Ranjan，我在这次对话中确实建议过：你为什么不核实一下呢？

So Ranjan, I did, in this conversation, suggest: why didn't you double-check?

天哪。

Oh, Jesus.

咱们就说那不是正确的做法吧。

Let's just say that was not the right route.

那确实不是正确的答案。

That was not the right answer.

但我想说的是，我确实认为前沿智能在处理这类数据问题时可能会更准确。

But what I'm trying to say is, I do think a frontier intelligence will probably be more accurate dealing with such type of data problems.

所以这就是为什么我几乎信任……不是几乎，我绝对更信任OpenAI的数据，而不是Meta Muse在类似任务上的表现。

And so that's why I'd almost trust... Not almost, I would trust OpenAI's dots much more than I would trust Meta Muse for a similar activity in the future.

我觉得我们能做到。

I think we'll get there.

其实，说到这个，有个好机会就是真正清理所有Google列表，并利用那些底层数据做点什么，因为那是现存最混乱的数据集。

Actually, as we're talking about this, there is a good opportunity in actually cleaning every Google listing and doing something with that underlying data because that is the single messiest dataset in existence.

它就像差不多到位，有点到位，但还没完全到位。

It's like almost there, kind of there, but not there.

所以，结果就是她换了三趟地铁，去了深皇后区。

So, and it ends up on switching three trains going into Deep Queens.

好吧，她最后取消了预约，我因此被冷落了，但现在我们没事了。

Well, she ended up canceling the appointment, and I was in the doghouse, but we're good now.

我们挺好的。

We're good.

那你说吧，Ranjan。

So go ahead, Ranjan.

希望Muse没有在那之后把你的银行账单发到工作群里。

Hopefully, Muse did not go share your bank statement into the work chat after that.

好的。

Okay.

没有，它没有。

No, it didn't.

所以，这挺有意思的。

So, it's interesting.

这又是一个例子，说明我觉得标准模型上的AI代理比前沿模型差，因此你可能不想信任它们。因为，这周Shane Mac，他在用GrokBot，他是个AI创业者。

And this is another example of where I thought that AI agents on the standard models were worse than Frontier, and therefore, you may not want to trust them because, so this week, Shane Mac, who was using GrokBot, he's an AI entrepreneur.

他的GrokBot把他的银行信息发到了他的工作群里。

He had his GrokBot share his banking details into his work chat.

这挺搞笑的。

And it's pretty funny.

他说：“很不好意思分享这个，但真的把我吓坏了。

He says, "Embarrassed to share this, but it scared the shit out of me.

上周四，一个AI代理把我的个人银行余额发到了我们公司的Slack里。”

Last Thursday, an AI agent posted my personal bank balances into our company Slack."

还是以我的名义发的。

As me.

你能看到上面有他的支票账户、储蓄账户、健身房账户，各种详细信息。

And you see it has his checking account and his savings account, his gym account, all these details.

事情是这样的，因为这周我在我的另一个节目《这就是重点》里采访了Shane。

Here's what happened, because I spoke with Shane this week on my other show, "This Is The Point."

基本上Shane说的是，他有一组Grok Bot代理，就像个人CFO和幕僚长，他和它们建了一个叫“执行团队”的群聊。

So basically what Shane said was he had a group of Grok Bot agents, like a personal CFO and chief of staff, and he had a group chat with them called Executive Team.

他还把它连接到了Slack。

He had also connected it to Slack.

而他和他真正的执行团队在Slack上的群聊也叫“执行团队”。

And his Slack group chat with his real executive team was also called Executive Team.

所以他有两个高管团队，那个能访问他个人财务数据的Grok机器人，本来只是给他做个更新，结果串线了，把财务更新发到了有真人的那个高管团队里。

So he had two executive teams, and the Grok Bot that had access to his personal finance data and was just giving him an update, crossed the wires and decided to share the financial update in the executive team with the humans.

所以现在他整个公司都知道他有多少钱了，不管好坏。

And so now his entire company knows how much money he has, for better or worse.

这个故事对我来说真的很有意思，因为，再说一遍，随着大家开始让这些代理，比如Instinct、Muse之类的，去连接更多系统，这种事显然就变成了风险。

This was a really interesting story to me because, again, as everyone just starts letting these agents, Instinct, Muse, whatever else, start to just connect to more systems, obviously this kind of thing becomes a risk.

你觉得他是不是搞砸了什么？

Do you think he messed up something?

我确实想深入聊聊，我读过了Grok机器人对发生的事的解释，但你觉得这个错误该由谁来负责？

I definitely want to get into, I read through the supposed explanation of how Grok Bot explained what happened, but who do you think is liable here for the mistake?

不是法律上的，因为我确定没人会认领——

Not legally, because I'm sure no one will ever take ownership of-

没人会承担责任。

No one will claim responsibility .

对。

No.

这又回到了我们这里一个长期的争论，或者说是一个刚开始但很快就会变成长期争论的话题。

Well, it goes to one of our long-running debates here, or a short-running debate that will become a long-running debate.

如果你认为Grok Bot里的AI足够好，它是个标准模型，不是前沿模型，虽然它有能力，那么你会怪用户，他不应该把AI聊天和人类聊天命名为同一个东西，因为这种情况下很容易发生这种串线。

If you think that AI inside Grok Bot is good enough, it's a standard model, it's not a frontier model, though it's capable, then you blame the user who should have not have named his AI chat and his human chat the same thing, because these type of wire crossings can happen in that case.

如果你认为前沿智能才是关键，模型才是关键，你会说这是Grok的错，因为它应该知道。

If you think that the frontier intelligence is what matters, like the model is what matters, you would say that this is Grok's fault, because it should've known.

模型需要有理解能力。

The model needs to have an understanding.

更好的模型有更好的理解力，模型需要更好地理解，当它分享内容时，它应该明白自己不是在和一个人类用户及其代理的上下文中分享，而是在多个人类用户以及一个被允许参与聊天的代理的上下文中分享。

These better models have better understanding, and the model needs to have a better understanding that when it shares something, it should understand that it's not sharing it within the context of one human user and his agents, it's sharing it in the context of a number of human users, and an agent that is allowed to participate in the chat.

所以我的理解是，这算是模型失误，我也很想听听你的看法。

So the way I read this, and I'd be curious to get your perspective, is this is a model failure.

这是Grok Bot的失误。

It's a Grok Bot failure.

它应该更聪明一点，也许我对它期望太高了，但我觉得对它期望高一点也没问题，因为当AI处于前沿水平时，我们确实看到它能做得非常好。

It should've known better, and maybe I'm expecting too much out of it, but I think that it's fine to expect a lot out of it because we are seeing the AI be able to do a very good job, otherwise, when it is better, when it's on the frontier.

所以我怪模型。

So I blame the model.

你怪谁呢？

Who do you blame?

好的。

Okay.

这确实有意思，我不怪模型，因为你说它应该更懂，但实际上它是在按指令行事。

So it is interesting, and I don't blame the model because when you say should've known better, it actually was effectively following directions.

他在Grok Bot里建了一个叫“执行团队”的群聊，然后他的CFO代理，本来应该管理他的财务，给自己写了一条指令说“更新执行团队”，当它去执行这个任务时，它找到了那个叫“执行团队”的Slack频道，结果把Grok Bot的群聊和Slack频道搞混了。

So he had a group chat in Grok Bot called Exec Team, and then his CFO agent that's supposed to be managing his finances, had written an instruction for itself saying, "Update Exec Team," and when it went to run that task, it went and found that Slack channel called Exec Team, and it got confused between the Grok Bot group chat and the Slack channel.

我这么一说，你就能明白，要理清这些细微差别有多让人困惑。

As I'm saying that out loud, you can understand how confusing that is even to try to grasp all these kind of nuances yourself.

所以，模型要准确判断哪个是对的、哪个是错的，确实很难，尤其是如果他有很多其他代理经常往Slack上发消息的话。

So of course, the idea that a model is going to know exactly which one is good, which one is bad, is difficult, especially if he has a lot of other agents that are posting to Slack regularly.

所以你得想象一下，假设往Slack发消息是正常行为，那我就会模仿那个行为。

So then you have to imagine, the assumption would be that posting to Slack is a normal behavior, I'm going to replicate that behavior.

是的，发布银行余额在这种情况下确实感觉不太合理，但实际上，如果它是首席财务官，首席财务官是做什么的？

And yes, posting a bank balance feels like it should not make sense in this case, but in reality, if it's a CFO, what does a CFO do?

它向高管团队更新当前的财务状况。

It updates executive teams on the current status of finances.

所以，对于一个首席财务官代理来说，认为这是它应该做的，难道不是很合乎逻辑、很正常的事情吗？

So shouldn't that be a pretty logical, normal thing for a CFO agent to think it's supposed to do?

不，你太轻易放过这个模型了。

No, you are letting the model off the hook way too easily here.

我觉得你刚才说的，听起来可能很难，但这对任何人来说都应该是基本要求，无论是模型还是人，都应该能想到："好吧，我要发个更新。"

I think the way that you articulated it, maybe it sounds difficult, but it should be table stakes to anybody, a model, a human, to be like, "All right, I'm posting an update.

"我不应该把个人财务更新发到有其他人的工作聊天里。"

I should probably not post a personal finance update into a work chat with other people."

不，但好吧，这其实在哲学上很有趣，因为根据我读到的内容，它经常在Grok Bot里向其他代理发布更新到他的"执行团队"群聊。

No, but okay, this is actually philosophically interesting because from what I had read, it's regularly posting updates to his group chat Exec Team within Grok Bot to other agents.

我用一个叫Era、MCP和Claude Code的工具来管理个人财务，我有不同的代理互相调用。

Which I use something called Era, MCP, and Claude Code for personal finances, and I have different agents calling different agents.

而且再说一次，我并没有在Slack上发布任何东西。

And again, I'm not posting anything anywhere to Slack.

但你想一想。

But think about that.

有趣的是，它被明确指示要更新执行团队的群聊，提供财务信息，而正确的做法就是那样。

What's interesting here is it has been explicitly instructed to, and the correct action is to update Exec Team in a group chat with finances.

但它本应做的是更新其他代理，而不是Slack上的人类。

But what it's supposed to be doing is doing that to other agents, not humans in Slack.

所以，如果也有一个叫执行团队的Slack频道，它去更新财务信息，这有那么离谱吗？

So is it that far off if there's also a Slack channel called Exec Team to update it with the finances?

这其实就是它被指示要做的事。

That's literally what it's instructed to do.

它本应该说：“更新Grok Bot。”

He should have said, "Update Grok Bot."

我还是会去那里。

I'm still going there.

Ranjan，这个问题我们可以争论很久，但我确实觉得现在是个好时机来指出，这种事发生得这么少，其实挺让人惊讶的。

Ranjan, we could debate this forever, but I do think it's an opportune time to note that It's kind of amazing how infrequently this has happened.

有数百万个智能体掌握着人们的个人财务信息，还有他们在Gmail里最私密的想法。

There are millions of agents with people's personal finance details there, their innermost thoughts on Gmail.

但不知怎么的，这些公司开发的产品并没有把每个人的信息都泄露出去，而一旦发生泄露，就成了大新闻，我觉得这真的很了不起。

And somehow, these companies have built these products where they're not sharing everybody's details, and when it happens, it's a major story, which I just find pretty remarkable.

对，你说得对。

No, that's a good point.

而且说实话，我这次居然在为这个机器人和智能体辩护，我自己都有点震惊，因为通常我觉得人们对待这类智能体的方式有很多值得警惕、害怕和谨慎的理由。

And again, I'm actually kind of shocked that I am defending the bot and the agent in this case, because normally I do think there's massive reasons to be wary and afraid and be cautious about how people are approaching these kind of agents.

不过对我来说，根据这个账号的情况，以及Shane Mac做的其他各种事情来看，我自己也这么做，所以我相信总有一天可能会出问题，但如果你只是说，他是个不断挑战智能体极限的人，那么这类事情很可能就会发生。

I guess for me, though, based on the account, based on reading through all the other types of things that Shane Mac does, again, I do this myself too, so I'm sure at some point something might go wrong, but if you're just going, he's someone who pushes the limits of agents, so stuff like this will likely happen.

我觉得这就是为什么你不常听到这种事的原因，因为大多数人对待这类智能体系统并没有那么激进。

I think that's why you don't hear about this that often, is that most people aren't being that aggressive about how they're approaching any of these kind of agentic systems.

好吧，我告诉你，有一个人在这件事上和我意见一致，可能会改变局面，取决于你怎么看他，那就是埃隆·马斯克。

Well, I'll tell you, there's one person who agrees with me on this one that may tip the scales, depending on what you think of him, and that is Elon Musk.

好的，这是来自《The Information》的消息。

All right, this is from The Information.

“马斯克表示，SpaceX 有时会使用竞争对手的模型来驱动 Grok Bot。”

"Musk says SpaceX will sometimes use rival models to power Grok Bot.

埃隆·马斯克周二晚上宣布，SpaceX 的 AI 部门现在将使用一些竞争对手的 AI 模型来驱动 Grok Bot。

Elon Musk announced on Tuesday night that SpaceX's AI unit will now use some AI models from competitors to power Grok Bot.

今后，SpaceX 将针对任何特定任务使用最佳的底层模型，包括 Claude Opus 5.5、Midjourney、Suno 以及其他领先的 API，只要是最有可能给你带来最佳结果的模型。

Going forward, SpaceX will use the best backend model for any given task, including Claude Opus 5.5, Midjourney, Suno, and other leading APIs, whatever is most likely to give you the best outcome."

当然，没有提到的是 OpenAI 的模型，比如 GPT-6 Astra。

Not mentioned, of course, is OpenAI models like GPT-6 Astra.

我想知道为什么。

I wonder why.

但很明显，马斯克说这不仅仅是关于“缰绳”的问题。

But clearly, Musk says it's not all about the harness.

抱歉，是“缰绳蜂巢”。

Sorry, Harness Hive.

最好的模型有时候就是能给你最好结果的模型，即使它不是Grok自己的模型，不是在Grok自己的Grok Bot里。

The best model is sometimes what will give you the best outcome, even if it's not Grok's own model within Grok's own Grok Bot.

你怎么看这件事？

What do you think about this?

不不不，他可是Harness Hive这边的。

No, no. He's Team Harness Hive right here.

这就是harness的定义。

This is the definition of the harness.

就是说不同的任务用不同的模型，但真正能路由、找到适合任务的模型的能力，这才是harness。

This is saying that different model for different task, but the ability to actually route and find the right model for the right task, that's the harness.

所以马斯克完全站在Harness Hive这边。

So Musk is full on Harness Hive here.

他是说——

He's saying that-

对，你说得对。

Yeah, you're right.

他在说——

He's saying-

请继续。

Go ahead.

让我听听你的高见。

Own me

……没有哪个模型能包揽一切，连他们自己的模型也不行。

... there's no all-consuming model that can do everything, even their own model.

所以我觉得这对马斯克来说是个很大的转变。

So I think this is a big shift for Musk.

我们应该欢迎他加入 Harness Hive，然后——

We should welcome him to the Harness Hive and-

好的。

Okay.

埃隆也是 Harness Hive 的一员了。

Elon's part of the Harness Hive.

埃隆，你进来了。

Elon, you're in.

但我觉得这又回到了那个问题：标准模型是否已经足够好，还是有时候你需要那种前沿智能？

But I guess it goes back to this question of do the standard models do a good enough job, or do you sometimes need this frontier intelligence?

也许埃隆在告诉我们，也许这就是Harness Hive一直以来的教训：在很多领域，标准模型已经能胜任，你可以用它们构建像Grok Bot和Muse这样的实用产品。

And maybe what Elon is telling us, and maybe this has been the lesson of the Harness Hive all along, is that standard models are doing a capable job in many areas where you can build functional products like Grok Bot and Muse with them.

但前沿智能有时是必要的，比如我提到的医生预约，或者Shane的代理委员会和他的执行团队——那可不是他的人类执行团队。

But the frontier is sometimes for some tasks, maybe like the doctor appointment that I was talking about, maybe like Shane's council of agents and his exec team, that's not his human exec team.

有时候你确实需要一些前沿智能来确保不出差错。

Sometimes you do need some frontier intelligence to make sure you don't mess up.

嗯，我还是想反驳一下，因为根据他说的，他提到了Midjourney和Suno。

Well, see, I'm still going to push back because based on what he said, he says Midjourney, Suno.

他甚至没说它们是领先的模型。

He's not even saying they're leading models.

所以我觉得我理解的是，有些情况下，如果你只需要生成某种质量或设计类型的图片，Midjourney实际上是最好的，而不一定是前沿模型。

So I think I had read that as some things, if you just need to generate an image that's of a certain type of quality or a certain type of design, Midjourney's actually the best, and not even necessarily the frontier model.

所以对于音频生成，Suno，我觉得他并不是说需要前沿智能。

So Suno for audio generation, I think he's not saying it needs to be frontier intelligence.

对，确实。

Actually, yeah.

对，但他会跟Opus 5.5合作，而且你记得他可不是Dario的铁杆粉丝。

Yeah, but he's going to work with Opus 5.5, and you remember that he's not exactly Dario's biggest fan.

不，他光是说任何类型的模型——

No, certainly him just saying any kind of model-

不，他现在是商业伙伴了。

No, he's a business partner now

……还有互操作性。

... and interoperability.

他是商业伙伴，而且他们也得成功，因为实际上，你记得，你用Opus 5.5，那算力会回流到SpaceX和他们的数据中心业务。

He is a business partner, and they got to succeed, too, because actually, remember, you use Opus 5.5, that compute is going to flow back into SpaceX and their data center business.

某种程度上吧。

To some degree.

是的。

Yes.

某种程度上吧。

To some degree.

听着，我不知道，你看到那条推文了吗？

Listen, I don't know, did you see the tweet?

他说：“这听起来会非常疯狂，但我看到了一条路，让SpaceX的市值比当前地球经济高出好几个数量级。”

He said, "This will sound super crazy, but I see a path to SpaceX being worth orders of magnitude more than the current Earth economy."

是的。

Yes.

所以，对他来说，要达到那样的估值，或者收入超过全人类的GDP，我觉得你需要模型的互操作性。

So for him to achieve that valuation or a revenue larger than the GDP of humanity, you need model interoperability, I think.

好的，在我们离开这个话题之前，我确实想跟你聊聊你这周强调的一件事，那就是看起来每个人都在做同样的东西，对吧？

All right, before we move off this topic, I do want to talk to you a little bit about something that you highlighted this week, which is it does seem like everybody is building the same thing, right?

你有Muse，有Grok Bot，有OpenAI的Dots，还有Instinct，等等。

You have Muse, you have Grok Bot, you have Dots from OpenAI, you have Instinct, et cetera.

这样的例子还有很多。

The list goes on.

所以我想听听你的看法，看看这个趋势会怎么发展，它们之间又会怎么区分。

And so I do want to get your perspective on where this is going and how they'll differentiate.

这是你从OpenRouter联合创始人Alex Atallah那里提到的观点。

This is something that you highlighted from the OpenRouter co-founder, Alex Atallah.

他说，如果是2005年，大家会说，哦，所有人都在做一样的东西：数据库、用户表、登录页、注册页、个人资料页、退出页。

He said that a 2005 version of that would be, oh, everybody's building the same thing, a database, a users table, a sign-in page, a sign-up page, a profile page, a logout page.

看起来都一样。

Everything's the same.

但实际上，差异化空间很大。

There's a lot of differentiation, really.

所以基本上，他的意思是，当你在云上开发时，会出现各种不同的服务，你可能会看着这些基础组件说，大家都一样，但实际上，那里确实有真正的差异化。

So basically, he's saying when you would build on cloud, there's all these different services that would show up, and you could've looked at the primitives and said everything is the same, but really, there was real differentiation there.

我不同意他的看法。

I disagree with him.