AI模型发布协议的重要性

今天这期播客里，达里奥·阿马德和杰米·迪蒙在纽约现场，而Anthropic那边刚发布了10个预装的金融智能体。

Today on the podcast, Dario Amade and Jamie Dimion were both on stage in New York while Anthropic was shipping 10 prebuilt finance agents.

我们还聊了他们跟微软365的深度整合，以及一堆新更新，还有Anthropic新出来的营收增速。

We're also talking about their full Microsoft three sixty five integrations and a ton of other updates and their new run rate coming out of Anthropic.

另外，所有美国主要的前沿实验室都同意，在发布模型前，先把模型交由商务部审核。

In addition, every major US frontier lab has agreed to hand over their models to the commerce department before launch.

这事儿是新冒出的苗头，政府想提前看看这些模型长啥样。

This is kind of a new thing that's getting drafted up where the government wants to look at the models.

对此，不少人各有看法。

A lot of people have opinions on that.

还有，谷歌DeepMind在伦敦的员工以98%的压倒性票数投票决定成立工会，起因就是他们反对五角大楼和以色列国防军的合作项目。

Also, Google's DeepMind staff in London have voted 98% to unionize, and they're doing this over the Pentagon and the IDF's deal.

这事儿我们以前也聊过。

We've kind of covered this in the in the past.

有不少谷歌员工都联名写信，明确表示不想让公司跟五角大楼合作。

There's a bunch of Google employees that have signed letters where they don't want Google to work with the Pentagon.

另外，Cerebras 已经申请了 266 亿美元的 IPO，而 OpenAI 会因为合作关系获得其中 3300 万股。

Also, Cerebras has filed for a $26,600,000,000 IPO, and OpenAI is gonna get 33,000,000 shares of that because of their partnership.

PayPal 宣布了 4500 个由 AI 驱动的裁员，说白了，他们得把整个公司重新打磨一遍，想借 AI 来带动复兴。

PayPal has announced 4,500 AI driven layoffs, and, basically, they're having to redo everything at PayPal, and they're trying to use AI as the catalyst for bringing the company back.

现在跟 2021 年的巅峰期比，差远了。

They are way down from their high of 2021.

IBM Think 正在重新定义人们对 Watson x Orchestrate 的认知，同时哈佛的一项研究发现，OpenAI 的 O1 模型在真实急诊分诊中，诊断准确率超过了急诊科医生。

IBM Think is trying to reframe how you think of Watson x Orchestrate, and there is also a Harvard study that finds that OpenAI's o one model out diagnosis ER, doctors in real triage cases.

内容真多。

A lot to cover.

咱们直接开聊吧。

Let's get into it.

我第一个想聊的是哈佛医学院的这项研究。

The first story I wanted to talk about was this Harvard Medical School study.

他们和贝斯以色列女执事医疗中心的团队一起做的，刚发表了一篇论文，核心结论是 OpenAI 的 O1 预览版……

They did it with Beth Israel Deaconess, the team over there, and they just published a science paper that basically said OpenAI's O1 preview.

它在急诊室里击败了好几位医生，就是那些专门处理急诊的医生。

It it was out beating a bunch of people inside of the ER, a bunch of doctors specifically.

他们让AI和两位内科主治医师直接比拼。

They put it head to head with two internal medicine attendees.

总共用了76个真实的急诊病例。

So, basically, they had seventy six real ER cases.

顺便说一句，急诊室挺特别的，因为医生们根本不认识病人，也没法了解他们的病史。

And ER, by the way, I'm I'm gonna say is kind of an interesting place because this isn't like doctors that know their patients and work with their patients and understand their history.

我觉得在这种环境下，AI表现得更好也不奇怪。

I think you get much better results there.

但话说回来，急诊确实是真实世界里非常典型的场景。

But to be fair, ER is a real you know, it's a real use case.

这可是实打实的临床应用。

It's a real thing.

在急诊室里，他们给AI和医生提供了完全一样的电子病历、生命体征和护士的初诊记录。

And so inside of the ER, they gave the same electronic medical records, the same vitals, the same intake nurse notes, all of that.

在分诊阶段，信息最不完整，你得立刻做出判断，急诊室里尤其如此。

And at the triage stage where information is, like, basically, it's the hardest to get, you have to make a call and you have to do it very quickly in the ER.

对吧？

Right?

你得拼命救人。

You're trying to keep people alive.

病人可能正在大出血。

They're, like, bleeding to death.

或者遇到什么严重的状况。

They're having some sort of, you know, big issue.

O1模型在67%的情况下给出了准确或非常接近的诊断，而两位主治医生的准确率只有50%到55%，也就是猜中病人真正病因的几率。

The O1 model hit the exact or very close diagnosis sixty seven percent of the time, and the two attending doctors hit it fifty five to fifty percent of the time of, you know, the diagnosis of what was actually wrong.

这两方都没做到完美，但OpenAI的模型能到67%，医生们却只有50%到55%。

So neither of these were perfect, but OpenAI was getting sixty seven percent, and the doctors were getting fifty five and fifty percent.

等病人住院后，O1的诊断准确率升到了81.6%，而医生们分别是78.9%和69.7%。

Once patients were admitted, o one hit a eighty one point six percent versus a seventy eight point nine and a sixty nine point seven for the doctors.

患者一入院，所有人诊断准确率都提升了，但在急诊分诊阶段，刚进来还没确诊的时候，OpenAI的表现简直太出色了，这前景非常可观。

So everybody got better once the patients were actually admitted, but when they're coming into triage and you're trying to figure out what's wrong with them, OpenAI was doing really, really well, which is very promising.

他们总共做了六组实验，这个模型跟几百位不同的医生都比过。

They did this across six total experiments, and the model went up against hundreds of different physicians.

在其中一组里，它还跟医生打了个平手。

And it kind of held its own in one.

我猜很多人会想：哎，他们是不是就找了几个医生、一两个地方、几个简单病例来做的？

I think a lot of people were like, oh, did they only do this with, like, a couple doctors or, like, a couple places or a couple things?

但其实不是，他们找的是几百位不同医生，做了大量实验，覆盖了形形色色的病例，范围广得很。

But, no, they did this with hundreds of, different doctors, and they did this against a against a ton of different experiments and a ton of different cases in all of this.

所以哈佛大学贝斯以色列医院的临床研究员、这项研究的合著者彼得·布罗德总结说，他还接受《财富》杂志采访，给了个评价。

So Peter Broder, a Harvard clinical fellow at Beth Israel and one of the study's coauthors, he summarized it, and he he kind of gave a a quote over to Fortune.

他说，过去我们都是用选择题来评估模型。

He said, we used to evaluate models with multiple choice tests.

现在，我们的准确率快逼近百分之百了。

Now we're scoring close to a 100%.

我们已经到顶了。

We're already at the ceiling.

不过说实话，我要提醒一下，这些输入都只是文字，没有其他信息。

Now personally, if I'm gonna put a caveat on all of this, I will say that these are text only inputs.

没有影像资料。

There's no imaging.

也不是多模态的。

There's no kind of multimodal.

作者自己也明确说了这一点。

And the authors are also pretty clear on that.

他们强调，这并不等于临床应用的批准。

They're saying that this isn't, you know, approval for clinical development.

但现在有个责任问题，谁都没能解决：

But I think there's definitely a liability question right now that's kind of nobody has solved it.

如果AI的判断和主治医生不一样，结果医生没听AI的，病人出了问题，该谁来背锅？

If O1 disagrees with the attending, physicians and then the attendings override it, who's gonna get sued when a patient has an issue for malpractice?

他们会去看AI说了什么，以及医生哪里不同意吗？

Are they gonna look at the AI model said and what the, you know, what the physician disagreed with?

我的意思是，这事儿两面都有可能。

I mean, this could go both ways.

对吧？

Right?

假设医生采纳了AI的建议，结果错了，那该谁负责？

Let's say the physician agrees with what the AI model says and it turns out to be wrong, who's liable?

所以这些都是一些亟待解决的大问题。

So these are some big questions that need to be solved.

接下来我想聊的是IBM。

The next story I wanna talk about is IBM.

他们今天发布了Think2026的主旨演讲，在我看来，这简直是近几年来他们最激进的一次企业级AI战略重塑。

They held their Think2026 keynote today, and they announced what to me basically looks like their most aggressive enterprise AI repositioning that they've done in the last few years.

我记得三四年前参加一些AI大会时，IBM还在那儿大谈Watson。

I remember going to some AI conferences maybe, like, three or four years ago when IBM was there talking about IBM Watson.

这种事在主流媒体里可不常听得到。

This isn't really something that you hear a ton in the mainstream.

他们确实签了一些企业客户，但规模好像没达到他们预期的那么大。

They've done some enterprise deals, but it's I don't think it's as big as they had hoped to.

他们现在真正发力的方向叫 Watson X Orchestrate，说白了就是 IBM 把它称作一个‘智能体控制平台’。

And what they're really pivoting and and kind of pointing this towards is that it's called the Watson X Orchestrate, and this is basically IBM now calling it a, quote, agentic control plane.

这个平台能让所有财富500强企业，统一用一套系统部署来自 Anthropic、OpenAI、谷歌，还有他们自己内部的所有智能体。

So this is basically the layer that's gonna let all the Fortune 500 companies deploy agents from Anthropic, OpenAI, Google, plus all of their own internal agents.

而且这一切都能通过一个统一的策略来管理。

And they can do this all with a single policy.

他们还内置了身份认证、审计追踪，能一目了然地监控所有操作。

So they can, you know, they have some identity, they kind of have an audit layer, and, it can look across everything.

他们还推出了 IBM Bob，一个智能体开发助手。

They also shipped IBM Bob, which is an agentic developer partner.

我真的说不好……

I just I don't know.

也只有IBM能给他们的智能体取名叫鲍勃了。

Leave it up to IBM to name their agent Bob.

说真的，这名字太IBM了——你看Anthropic的Claude、Google的Gemini，结果IBM整出个Bob。

It's just, I don't know, it feels so on brand for IBM versus like, you know, you have like Claude from Anthropic and you use and you got like Gemini from Google and we got Bob from IBM.

我简直太喜欢了。

I love it.

明白吧？

Okay?

我就是吃这一套。

I'm I'm a sucker for it.

不管怎么说，他们的目标用户跟其他工具不太一样。

In any case, they're they're basically really targeting a different audience than I think a lot of these other tools are targeting.

他们还有IBM Cofluent，用来处理实时数据。

They have IBM Cofluent for real time data.

还有IBM Concert，专门搞运维的。

They have IBM Concert for ops.

他们还有IBM Sovereign Core，专门给那些必须在特定国家司法管辖区内运行AI的客户用。

They have IBM Sovereign Core for customers who need to run AI inside of specific national jurisdictions.

所以他们瞄准的真是另一批用户，但我觉得现在他们压根不打算在模型层面上跟别人硬刚。

So they're really targeting a different audience, but I think this is a moment where they're not trying to compete on the model layer.

他们心里清楚，模型这块早就被OpenAI、Anthropic和Google拿走了。

They know that they've lost that to OpenAI Anthropic and Google.

现在他们就想当个中间人，帮那些受监管的企业和模型提供商搭桥，管好合规这摊事。

They're just trying to kind of rent their self out as the governance layer between all of the regulated enterprises and the model providers.

接下来IBM会怎么走，真挺值得瞧的。

This is gonna be interesting to see where IBM goes from here.

接下来我想聊聊PayPal。

The next thing I wanna talk about is PayPal.

他们的新CEO恩里克·洛里斯今天刚开完第一场财报会，宣布要砍15亿美元的成本，其中包括裁掉大约4500名员工。

Their new CEO, Enrique Loris, just got off of their first earnings call today, and they announced that they are doing a $1,500,000,000 cost cutting plan, and that includes laying off about 4,500 employees.

这差不多占了公司总人数的两成。

That's about 20% of the entire workforce.

说实话，我到现在还是用PayPal做几个线上业务和项目，但感觉这家公司确实掉队了，变得有点懒散、迟钝，跟那些更拼命的对手比起来差远了。

And if I'm gonna be honest, I still use PayPal for a number of different online businesses and projects, but it definitely feels like a company that has lost a lot of ground and just gotten, I don't know, kind of lazy and slow compared to other more hungry competitors.

现在市面上金融产品多得是。

There's a lot of finance products out there.

Stripe现在简直杀疯了，IPO在即，牢牢抓住了开发者群体——这点本来应该是PayPal该做的。

Stripe is just, you know, crushing it right now and getting ready for their IPO and really tapped into the developer field, and I think in a way that PayPal really should have.

不过看起来PayPal正在努力扭转局面，但必须裁员降本、重新组织架构。砍掉20%的员工可不是小事，他们说这全靠AI才能做到。

But it seems like PayPal is turning a lot of that around, but they gotta do some cost cutting and and reorganize the company, And 20% of their workforce is massive, and they're saying that they're gonna be able to do this because of AI.

所以未来两三年，他们的说法是：要靠AI来翻身。

So over the next two to three years, the framing is that they're gonna do a AI turnaround.

他们的CEO一开始就说，PayPal得重新变成一家科技公司——这话真够狠的，毕竟这家公司可是现代在线支付的开创者啊。

He their CEO initially said that PayPal needs to become a, quote, a technology company again, which is kind of brutal when you talk about a company that, you know, literally invented modern online payments.

意思就是：咱们得重新当回科技公司。

So was saying, like, look, we're gonna become a technology company again.

连CEO都这么说，可见他们真的迷失方向了。

It's like, oh, they really have lost their way, it feels like even coming from their CEO.

彭博社的玛德琳·斯皮德提到，他们其实打算组建一个全新的AI转型与简化团队，把平台全面迁移到云原生架构，还在开发流程里大力引入AI。

Over at Bloomberg, Madeleine Speed was kind of talking about this and said that what they're actually doing is they're planning on spinning up a new AI transformation and simplification team, which is gonna be moving the platform to cloud native, and they're gonna be aggressively adopting AI inside of their developer workflow.

这事儿不是他们一家在干，好多公司都在这么做。

We have this on the back of a bunch of other companies that are doing this.

Snapchat 也刚宣布了类似的动作。

Snapchat announced something similar.

Meta 同样发布了类似的消息。

Meta announced something similar.

微软、亚马逊，这阵子裁员不断，我觉得我们正亲眼看着这一波趋势。

Microsoft, Amazon, there's a lot of there's a lot of layoffs, and I think that we're we're seeing this.

现在AI真正在接受考验——它到底能不能真正推动、自动化、甚至带动公司增长。

The AI is really getting put to the test right now to see if it can actually push, automate, and grow, companies.

咱们现在正处在一个特别有意思的关键时刻。

And so we're just in a really interesting moment in the industry.

下一个故事是Cerebras。

The next story is Cerebras.

他们昨天刚提交了IPO申请。

They have just filed their IPO yesterday.

他们想把估值定在266亿美元左右，这已经是他们目标区间的上限了。

They're looking to do a two a $26,600,000,000 valuation, and that's kind of the high end of their range.

打算以每股115到125美元的价格出售2800万股。

They're gonna sell 28,000,000 shares at a 115 to a $125 apiece.

这样能给他们带来大约35亿美元的收入。

That will bring them about $3,500,000,000.

这可真是个让人兴奋的IPO啊。

Now this is, you know, this is kind of exciting IPO.

这公司太牛了。

This is an awesome company.

不过挺有意思的是，他们这次想融的钱总共也就35亿美金。

It's funny to me though, their total raise that they're trying to get out of this is about $3,500,000,000.

就在昨天，OpenAI刚宣布要搞个新项目，把他们的开发团队直接嵌入到客户公司里去。

Meanwhile, just yesterday, OpenAI announced that they're gonna, you know, start this new company where they're getting OpenAI's aid you know, their developers embedded into organizations.

他们跟一堆私募股权公司合作，说白了就是个顺手搞出来的子公司，结果融了整整40亿美元。

They're working with all these PE firms, like, that's kind of just like a random spin off company, and that raised, like, $4,000,000,000 to do that.

然后又冒出这么个IPO，估值直接干到35亿美金。

Then we have this entire c this entire IPO for another company, and it's coming in at, 3 and a half billion.

这事儿充分说明，现在Anthropic和OpenAI的融资规模简直离谱到不行。

It just goes to show you how absolutely astronomical these rounds of funding are for Anthropic and OpenAI right now.

这些巨头随手就能把天量资金收入囊中。

You kinda have these giants just scooping up so much money.

不过话说回来，Cerebras的股票代码是CBRS。

But in any case, for Cerebras, their ticker is gonna be CBRS.

定价日期是5月13号。

The pricing is May 13.

第二天就开始正式交易了。

Trading starts, the next day after that.

但很多人没注意到，去年12月，OpenAI其实借给了Cerebras十亿美元。

And I think what not a lot of people are talking about is that in December, OpenAI actually loaned Cerebras a billion dollars.

他们通过约三千万股的认股权证获得了这笔投资。

They secured it by warrants on about 33,000,000 shares.

这相当于在IPO时能转化成数十亿美元的股权。

So that's basically a multibillion dollar stake that converts on an IPO.

所以，OpenAI现在在纸面上成了这家即将上市的AI芯片公司的最大单一股东，而这家公司可是要和英伟达正面硬刚的。

So OpenAI just became on paper one of the largest single shareholders of this publicly traded AI chip company that is gonna compete head to head with NVIDIA.

对吧？

Right?

挺有意思的，因为英伟达本来是OpenAI最大的开支来源之一，现在他们却在资助自己的主要竞争对手。

Pretty interesting because NVIDIA is one of OpenAI's probably biggest expenses, and now they're funding one of their, you know, biggest rivals.

我觉得这里面肯定有深思熟虑的战略布局。

So I think there is a lot of strategy at play here.

过去几年，OpenAI一直在暗示想摆脱对英伟达的依赖。

OpenAI has been signaling for a couple years that they wanna diversify off of NVIDIA.

毕竟，把命脉系在一个供应商身上风险太大了，尤其是英伟达还有各种出口限制，还不只是出口限制——每次出新芯片，他们还得自己分配额度。

I mean, this just makes a lot of economic sense and business sense having one supplier that you're really reliant on, especially because OpenAI NVIDIA has a lot of export restrictions or or not even just export restrictions, but, like, when they come up a new chip, they gotta allocate it.

整个行业都想买下英伟达生产的每一块芯片，而英伟达得决定把货给谁。

Everyone in the industry wants to buy basically every chip NVIDIA has, and NVIDIA has to decide who to give it to.

所以就有了芯片配额这回事，我觉得对像OpenAI这样想疯狂扩张的公司来说，这感觉太被动了。

So you get, like, chip allocations, and I just think that doesn't feel good for a company like OpenAI that wants to scale super aggressively.

他们巴不得多几个供应商。

They would love more suppliers.

他们可不想哪天英伟达因为啥原因突然减少他们的配额。

They don't wanna feel like, you know, at some point NVIDIA is gonna lower their allocation for one reason or another.

这还能大大增强他们的议价能力。

It also helps their negotiating position become a lot stronger.

你知道吗，英伟达现在的情况，跟几个月前早就不是一回事了。

You know, NVIDIA isn't at the place they were, just a few months ago.

现在呢，OpenAI一边用着英伟达的芯片，一边又在投资别的公司。

Right now, we have OpenAI who's got NVIDIA as, you know, one of their suppliers, but they're also investing in someone else.

更别说他们的竞争对手Anthropic了，人家早就绑定了谷歌的TPU，还搞了亚马逊的Trainium。

But we even have, you know, their competitors like Anthropic that is now locked into Google TPU, and they also have Amazon Trainium.

所以NVIDIA的依赖性问题，跟2024年的时候已经不一样了。

So the dependency story for NVIDIA is not what it was in 2024.

现在竞争对手都冒出来了，而且一些大实验室也纷纷站队到他们那边。

We have competitors coming out, and we have some of the big labs locking into those competitors.

接下来我想聊聊谷歌DeepMind在伦敦的办公室。

The next thing I wanna talk about is Google's DeepMind London office.

今天他们投票，98%的人都支持成立工会。

They just voted 98% in favor of unionizing today.

也就是说，几乎全员参与，就一两个人可能没投。

So that's basically, I don't know, everyone but, like, one or two employees.

这被CW称为全球首个前沿AI实验室的工会。

And this is what the CW is calling the world's first union at a Frontier AI lab.

参与这次行动的DeepMind员工大概有一千人左右。

There's about a thousand DeepMind staff involved in this.

明白了。

Okay.

所以大概有二十个人决定不加入工会。

So I guess 20 people decided they didn't wanna unionize.

他们要求谷歌承认CW，并正式承认这个工会作为他们的代表，而且他们特别提出了一项要求。

And they're asking Google to recognize the CW and unite the or and unite the union as their official representatives, and they're asking for one thing in particular with this new union.

他们希望撤销与五角大楼的机密Gemini项目合作，终止为以色列国防军做的合同，并重新恢复谷歌今年二月从官网删除的‘不将AI用于武器’的承诺。

They wanna walk back the classified Pentagon Gemini deal and end the Israeli Defense Force contract work and reinstate the no AI for weapons commitment that Google deleted from its public site in February.

这跟2018年那次有点不一样，那时候Project Maven项目引发了谷歌内部的大规模抗议，最终逼得公司放弃了国防合同，也暂时退出了国防业务。

The thing that makes this a little bit different than, you know, we have you know, back in 2018, Project Maven had a whole revolt from inside of Google, and that killed a defense contract and had Google kinda pull away from doing defense work.

但现在，DeepMind那边的很多研究人员都在X上公开表达不满，情绪相当激烈。

Now everyone over at DeepMind, there's a lot of researchers that are publicly posting on X, and they're they're quite upset about it.

但我认为，跟2018年那次能成功叫停合同相比，现在的环境已经完全不同了——因为谷歌的股价现在正处在历史最高点。

But the reason why I think the environment today is definitely different than back in 2018 when they were actually able to kill that contract is because right now you have stock prices that are an all time high for Google specifically.

而且，现在AI工程师的就业市场也没以前那么火爆了。

But also the labor market for AI engineers is not as hot as it used to be.

那时候，2018到2021年，整个市场简直热得发烫。

Back in, you know, 2018, 2019, 2020, 2021, things got really, really hot.

所以很多工程师都在说：我们不干了。

And so a lot of the engineers were saying, hey.

谷歌可不想失去他们。

We're gonna walk away, and Google didn't wanna lose them.

于是，他们好像有点儿妥协了。

And so they, it felt like they kind of, bent to to their will a little bit.

而今天，桑达尔·皮查伊却公开为这笔交易辩护。

Today, we have Sundar Pichai who is publicly defending the deal.

他的个人声誉可都押在这上面了。

So he's you know, his own reputation is on the line.

现在工会要求他们这么做，与此同时，所有科技公司——微软、亚马逊、Snapchat、Meta——全都在大规模裁员，理由是：

And with the union kind of asking them to do this, you also have basically every tech firm, Microsoft, Amazon, Snapchat, Meta, every single tech firm is laying off all of like, tons of engineers because they're saying, look.

我们有AI帮忙了。

We have AI that's kinda helping us.

所以能用更少的人干更多的活，我猜以后还会继续扩大规模。

So, kind of, get a lot of this work done with less people, and I imagine they'll scale up at some point in the future.

但眼下，正好赶上一波大规模裁员，挺有意思的。

But right now, we have this interesting moment where there are a lot of layoffs happening.

这时候工会跳出来，逼着谷歌取消合同，说实话，我也不知道该咋说。

And so kind of having a union come in and try to force Google's hand to cancel a contract feels like I don't know.

这事儿真是挺微妙的。

It's an interesting place.

但与此同时，AI领域的研究人才本来就没多少。

But at the same time, we only have so many AI researchers.

搞不好哪天谷歌就彻底留不住他们了，全跑去NVIDIA、Anthropic、OpenAI或者其他公司。

And so there might just be a moment where Google completely loses control of them and they all want to walk over to NVIDIA or over to Anthropic or OpenAI or some other company.

我个人觉得，这事儿对谷歌的决策影响不大，但可能会影响他们未来招到什么样的人。

Personally, don't think it's going to have a huge impact on Google's decision, but it might impact who they're able to recruit in the future.

不过这事儿吧，得一家一家看。

But that's kind of a company by company thing.

好多公司都在跟五角大楼合作，但Anthropic目前好像还是少数没参与的那一个。

Many different companies are working with the Pentagon, and Anthropic seems to be one of the only ones that isn't really at the moment.

所以我也说不准，这将来到底能有多大影响。

So I don't know if even that will have that much impact in the future.

接下来我想聊的是，商务部那个AI标准与创新中心，简称CAISI，其实就是特朗普政府把原来的AI安全研究所改了个名，现在他们跟谷歌DeepMind、微软和xAI签了协议，要求他们在公开发布前沿模型之前，先让政府测试这些模型有没有国家安全风险。

The next thing I wanna talk about is that the Commerce Department's Center for AI Standards and Innovation, that's the CAISI, it's basically the Trump administration's renamed successor to the AI Safety Institute, they have signed an agreement with Google DeepMind, Microsoft, and x AI to test all of their frontier models for national security risks before public release.

有些人觉得这简直是太棒了，觉得这能推动AI安全和监管，方方面面都挺好的。

Some people are saying this is fantastic for, you know, AI safety and regulation and everything that we see coming from there.

但也有人对此挺担心的。

Other people are more concerned about it.

他们不喜欢政府有权因为觉得不安全，就压着不让模型发布。

They don't like the government being able to, I guess, hold back a model if they don't think it's safe.

但从实际操作来看，流程是这样的：所有前沿实验室都得把最新模型原原本本交到政府手里，连任何安全防护措施都不能有。

But operationally, how this would actually work is that all of the frontier labs, they're gonna give their newest models to the government with no safety guardrails at all.

整个测试过程都由CAISI来操作，他们会去疯狂试探这些模型，看看在没有任何限制的情况下，它们到底能干出啥事儿来。

So this is all gonna be done by the CAISI, and they're gonna be able to probe those models and actually basically push them to the limits to see what they can do if there was no, you know, if there was no guardrails on them.

我觉得有意思的是，这可能是第一次，有第三方能拿到这种权限。

The thing that I do think is interesting about this is kind of the first time that a third party has will be given access.

我们平时有红队研究人员，专门去测试模型的极限，但这些人要么是OpenAI内部的，要么是被OpenAI请来的安全专家或研究员。

We have kind of red teaming researchers that go and try to push the models to their limits, but those are inside of OpenAI or firms that are kind of hired by OpenAI or researchers or security professionals.

而OpenAI根本不用对外公布这些测试结果，别人也不知道这些模型到底有多强。

And OpenAI doesn't really have to tell anyone what the outcome was, what these models were capable of doing.

Anthropic倒是会主动说：嘿，我们有个叫Mythos的模型，能黑掉世界上所有代码库，超强什么的，自己吹一通。

And you have Anthropic that'll say like, hey, look, we have this this model, Mythos, and it can, like, hack every, you know, code base in the world, and it's super powerful, yada yada.

他们自己就先定规矩，主动公开这些信息。

So they'll kind of regulate or self, tell everyone.

但我觉着，正是这个模型吓坏了不少人，才促使政府出台了审查机制——而这件事，前AI事务主管大卫·萨克斯是死活反对的。

But I think that model scared a lot of people and was kind of the catalyst for getting the government review in in place, which is something that, politically, the former AI czar, David Sacks, was completely against.

他一直主张别搞什么实际操作机制，就该放任自流，别管这些事。

He kind of championed having, you know, no operational mechanism that you're gonna and kind of have this deregulatory posture about all this stuff.

他一被换掉，新上任的AI事务主管，我听说任期上限是一百三十天。

So since he got switched out of the role, David Sachs is the AI Czar, I think there's like a one hundred and thirty day cap.

他三月份就被调去PCAST了，现在早就不是这个职位了。

So he got rotated to PCAST in March, so he's no longer in that same role.

从那以后，感觉接手的政府人员或现在掌权的这些人，都觉得得在模型发布前彻底评估一遍。

And as that happened, it feels like whoever took over or whatever's happening in the government now, they believe that, you know, they need to analyze all the models before release.

明白了。

Okay.

如果你是 ChatGPT、Claude、Gemini、Grok 或 Eleven Labs 的付费用户，或者用过任何 AI 绘图工具，我得跟你聊聊我做的创业项目——AI Box。

If you're a paying customer for ChatGPT, Claude Gemini, Grok, or Eleven Labs, and, you know, any of the AI image models, I would love to tell you about my startup AI Box.

这是我亲自打造的工具，也是我推荐给所有不想每个月为六七个 AI 订阅服务掏钱的朋友的解决方案。

This is what I personally have built, and it's what I basically recommend to all of my friends who want to use AI without being on the hook for six different subscriptions every single month.

你能用上八十多个顶尖的 AI 模型。

You get over 80 of the top AI models.

全都在一个平台里。

It's all in one place.

各大实验室最前沿的模型一应俱全，你想用哪个就用哪个，完全按任务需求选，一个月才 8.99 美元。

You get every top frontier model from every major lab, and you can pick basically whatever is best for the task in front of you, and it's all $8.99 a month.

我个人写东西的时候特别喜欢用 Anthropic 的语气风格，但它不支持图像生成，也没法做 AI 音频。

So personally, I love using Anthropic for tone when I'm doing writing, but Anthropic doesn't have an image generator, and Anthropic doesn't do AI audio.

所以我用的时候，会先让它帮我写东西，然后切换到ChatGPT去生成图片。

So when I'm using it, I'll have it help me write something, but then I'll help I'll go over to ChatGPT and get it to generate images.

如果我想做个音频概要听听，或者发个音频文件给别人，我就用Eleven Labs。

And then when I want to do do, like, an audio overview to listen to it or send someone an audio file, I'll use Eleven Labs.

但这些全都能在AI Box里完成，同一个聊天窗口里来回切换模型，所有功能都集中在一起。

And I could do all of that inside of AI Box, inside the same chat thread, just switch between the different models, and you get it all in one place.

价格也便宜多了，要是选年费计划，还能再打八折。

It's a lot cheaper, and you even get, 20% off if you get the annual plan.

赶紧去看看吧。

So go check it out.

一个月只要8.99美元。

It is $8.99 a month.

希望它能帮你省一大笔钱，也让你不用再开着一百个标签页、用一百个AI工具搞得头大。

I hope it saves you a ton of money and also just helps you consolidate the madness of a 100 different AI tools in a 100 different tabs.

描述里有链接，去aibox.ai看看。

There's a link in the description to aibox.ai.

好了

Okay.

那今天我要深入聊的核心话题是Anthropic，还有这家公司最近在华尔街闹出的一系列动静。

So the big deep dive I wanna cover today is Anthropic and kind of the day they've been having with Wall Street.

首先呢，他们刚办了一场排场很大的活动。

First of all, they just had this big kind of glitzy event.

这个活动是仅限受邀者参加的，名字叫「金融服务吹风会」。

It was, like, invite only, and it was called Briefing Financial Services.

当时Anthropic的CEO达里奥·阿莫代伊，还有摩根大通的CEO杰米·戴蒙都同台了，一起聊人工智能和金融业的未来。

They Dario Amadeo, the CEO of Anthropic, and then also JPMorgan Chase CEO Jamie Dimion were both there on stage talking about the future of AI and finance.

而且他们不是在空泛地畅想，说什么‘未来可能会变成这样，AI也许能做这些事’，

And they weren't just talking about, you know, like, oh, possibly this is what it could look like or this is what it could do.

那种套话我在好多峰会上都听腻了。

As something that I feel like I've seen at a lot of conferences.

但Anthropic这次直接展示了他们专为金融场景开发的具体智能代理工具。

Anthropic was actually showing off specific agents they've built for finance.

他们现在有十个预置的智能代理，专门处理金融领域最耗时的那些工作流程，比如投资推介书、收益分析、信贷备忘录、承销、KYC（了解你的客户）、月末结账、对账审计、保险理赔这些。

So they have 10 prebuilt agents, which are for kind of the most labor intensive workflows inside of finance, which are pitch books, earnings analysis, credit memos, underwriting, KYC, know your customer, month end close, statement audits, insurance claims.

每个代理都自带完整的技能、连接器和子代理，开箱即用，不用额外配置。

And each of those is going to ship as a reference architecture with all of the skills, connectors, and sub agents that it needs out of the box.

金融从业者拿到就能直接上手用，立马就能跑起来。

So a finance person can grab those things, a finance professional, and can, like, immediately start running with them.

他们把金融圈最常见的一堆场景全都覆盖了。

They they built out all of the most common use use cases.

除此之外，Anthropic 还推出了完整的 Microsoft 365 集成，这意味着你不仅能用这些工具，还能直接跟微软全家桶无缝对接。

On top of that, Anthropic rolled out a full Microsoft three sixty five integration, which means that they can not only use those things, but they all integrate directly with Microsoft three sixty five.

PPT、Excel、Word、OneDrive，统统都能直接连上。

So it can, you know, PowerPoint, Excel, Word, OneDrive immediately.

Anthropic 生成的内容，能自动存进这些你天天用的系统里——Outlook 啥的，全都不在话下。

Once Anthropic generates something, it can actually put it into all of those systems, which people are so you know, use all the time, Outlook, everything.

对吧？

Right?

而且它还能在这四个应用之间传递上下文。

And it can carry context across all four of those apps.

在数据方面，他们接入了穆迪的MCP，可以直接把信用评级和信贷数据传入Claude。

On the data side, there is a Moody MCP that pipes credit ratings and credit data directly into Claude.

他们还和邓白氏、ThirdBridge深化了合作。

There's also an expanded partnership with Dun and Bradstreet and ThirdBridge.

而模型本身，Claude Opus 4.7，已经专门针对金融场景做过优化。

And the model itself, Claude Opus 4.7, has actually been finance tuned.

他们专门针对金融岗位的大量需求做了微调。

So they fine tuned this for a lot of the jobs that finance needs.

这感觉就像他们当初专注开发者市场时那样，结果在开发者群体里爆火了。

And it feels kind of like what they did initially targeting really heavily onto developers, and they saw such huge growth from the developer industry and and what they were building.

现在，金融领域看起来是他们下一个要大干一场的方向。

It feels like the next big frontier for them is finance.

他们还和FIS合作了，就是那个给成千上万家银行做后台系统的金融科技公司。

They also have a partnership with FIS, which is FIS, the bank tech vendor that basically runs the back office for thousands of banks.

特别有意思的是，这个合作推出的第一个产品。

Specifically, the first product that is coming out of that partnership I thought was very interesting.

这是一个金融犯罪AI助手，能大幅压缩反洗钱调查的时间，帮着识别可疑的洗钱行为，省去好几个小时的人工排查。

It is a financial crimes AI agent that compresses anti money laundering investigations, and it's just, like, helping identify what might be money laundering and flag that so it's not gonna take hours.

所有这些举措凑在一起，我觉得意义重大。

All of this lining up, I think, is a really big deal.

这背后明显是Anthropic在发力。

One big part of this, it feels like Anthropic.

他们之前在开发者市场赢了一局，现在OpenAI正拼命追赶。

They kind of won with developers and OpenAI is trying to play catch up.

我觉得Anthropic在金融领域也要一鸣惊人了。

It feels like Anthropic is about to crush it in finance as well.

最明显的信号之一，就是Jamie Dimon居然亲自上台和Dario同框了。

And I think one of the biggest signals is Jamie Dimon, you know, or Jamie Dimon being on stage with Dario at all.

Dimon平时可不随便出席这种供应商活动。

Dimon, he doesn't, you know, show up for most vendor events.

他是美国最大、最知名的银行的CEO。

He's the CEO of America's most, you know, famous biggest bank.

但我猜他现身的原因是摩根大通早就深度参与进来了，而且他说了一句我觉得特别棒的话——Anthropic肯定也觉得这话太棒了，毕竟他们马上就要上市了。

But I think the reason he showed up is because JPMorgan is already really deep into this, and he gave a quote that I thought was phenomenal, and I'm sure Anthropic thought it was phenomenal considering their their upcoming IPO.

他说：‘这技术太强大了，值得投上一万亿美金。’

He said, quote, the technology is so powerful, it's worth the trillion dollar investment.

费利克斯·萨尔蒙之前报道过这事，他说戴米安实际上是在为AI资本支出热潮背书，这种说法很到位，因为现在企业最头疼的采购和合规障碍，正在慢慢消解。

So Felix Salmon was kind of reporting on this, and he said that, you know, Demian is basically blessing the AI CapEx boom and that the framing is right because now the procurement and the compliance objection that I think a lot of these firms have had is kind of going away.

FIS已经为成千上万家银行搭建了SOX合规体系，所以Anthropic接入FIS，其实就是让这些AI代理真正能合法落地的关键一步。

FIS already has SOX compliance rails on thousands of banks, So Anthropic plugging into FIS is, I think, basically how all of these agents actually become legally deployable.

这样一来，就不用再搞那种绕开审计的‘小动作’了。

So there's not kind of this, you know, end runaround audit.

我觉得Anthropic现在年收入已经冲到300亿美元了，没错。

And I think the run rate numbers that we're seeing from Anthropic hitting 30,000,000,000 or yeah.

300亿美元的年经常性收入说明，这家公司增长快得吓人。

$30,000,000,000 in annual recurring revenue shows, look, this company is growing insanely fast.

不过，这一切还有一个特别有意思的视角：Anthropic 可能已经变得太重要了，美国金融体系根本没法把它监管到消失。

Now there's one other really interesting angle with all of this, and that is that Anthropic might just be making itself too important for The US financial system to be regulated out of existence.

对吧？

Right?

咱们已经看到他们和五角大楼在较劲，还跟政府有些摩擦，正在硬扛着。

We we already see that they have kind of this battle going with the Pentagon, and they have some, you know, government issues where they're kind of fighting.

你想想，要是真遇上一个特别严厉的监管环境，他们可能会像其他科技公司那样，被一堆政府机构联手打压。

And you could see, you know, in a really punitive environment where they would kind of get hammered by a lot of government agencies like we've seen in all sorts of administrations for different tech companies.

但现在，他们深度嵌入了美国金融系统，跟华尔街绑得这么紧，你说美国政府能怎么办？

But it feels like right now, by integrating themselves deeply into The US financial system, working with Wall Street like, what's the US government gonna do?

嘿。

Hey.

你看啊。

Like, look.

我们不能用 Anthropic，就因为讨厌他们在某个问题上的立场？可那会直接把我们的金融体系搞瘫痪的。

We can't use Anthropic for anything because, we don't like their stance on one particular, you know, one particular area, and it's gonna cripple our financial markets.

他们真是无缝融入了方方面面，太有意思了。

It's really interesting the way that they are working into everything.

好了。

Alright.

今天节目就到这里，各位。

That's it for the show today, guys.

非常感谢你们收听。

Thank you so much for tuning in.

如果喜欢这期节目，麻烦去你听播客的平台给个评分和评论吧，这对我真的特别重要。

If you enjoyed this episode, it would mean the world to me if you could go leave a rating and review wherever you get your podcast.

这样节目才能越做越大，让更多人发现。

It just helps the show to grow and be found by new people.

平台就是靠这个来判断值不值得推荐的。

It's how the algorithm decides if this is a good show to share.

真的特别感谢大家的每一条评论。

So thank you so much for all of the reviews.

每一个我都特别感激。

I appreciate every single one.

和往常一样，别忘了去 aibox.ai 看看。

And as always, make sure to go check out aibox.ai.

简介里有链接，每月 8.99 美元，就能一次性用上 80 多个 AI 模型。

There's a link in the description if you wanna get access to over 80 AI models all in one place for $8.99 a month.

好嘞。

Alright.

下期节目见，各位！

I'll catch you guys all in the next episode.