Anthropic瞄准9000亿美元：接下来会怎样？

五角大楼刚公布了七家可以接入其机密网络的AI公司，结果出人意料，Anthropic被拒之门外。

The Pentagon just named seven AI companies that it will let into its classified network, and Anthropic, surprise surprise, is the one company that it shut out.

就在Anthropic给投资人48小时确认是否参与其9000亿美元估值融资的同一时刻，这事就发生了。

And this is all happening at the exact same moment that Anthropic has given its investors forty eight hours to confirm if they want in on its $900,000,000,000 valuation round of funding.

在此之前，马斯克在最近的庭审中承认，X AI是用OpenAI的技术来训练Grok的。

Before that, Elon Musk admitted in his most recent trial hearings that x AI used OpenAI to train Grok.

Anthropic联合CrowdStrike和Palo Alto推出了Claude Security公测版，而谷歌的Gemini已经悄悄进驻了四百万辆通用汽车的中控系统。

Anthropic dropped Claude Security in the public beta with CrowdStrike and Palo Alto on the board, and Google's Gemini is taking over the dashboard inside of 4,000,000 GM cars.

好多品牌都在悄悄接入这项技术。

A ton of different brands getting this.

这些事儿我们待会儿在播客里细聊。

We're gonna get into all of that on the podcast.

我第一个想聊的是Gemini进驻四百万辆通用汽车的事，顺便说，沃尔沃也在这批名单里了。

The first story I wanted to cover was this Gemini rollout into 4,000,000 GM cars, and also, we've just got news that Volvo is included in this.

他们昨天刚宣布，Gemini将取代这四百万辆车里的谷歌助手。

So they announced yesterday that Gemini is replacing the Google Assistant in about 4,000,000 cars.

这可不是什么新合作。

This isn't like a new deal.

这用的就是Android Auto系统。

This is Android Auto that's in there.

所以2022年之后生产的车，都内置了谷歌系统。

So anything that is from buff or from after 2022, will have the Google built inside of it.

凯迪拉克、雪佛兰、别克、GMC，还有沃尔沃，全都在里面了。

So everything with Cadillac, Chevrolet, Buick, GMC, Volvo's in there.

你不用再跑去换车里的硬件了。

So you don't have to go and, you know, swap out your hardware.

我还记得以前好多次痛苦的经历，为了升级Android Auto或者换新屏幕，把整个车载娱乐系统都拆了重装。

I I I'm still remembering many painful experiences where I swapped out the entertainment system inside of the car to a new Android Auto or to a new screen.

那真叫一个折腾，不过嘛，也算车生活的一部分了。

It's a painful experience, but, you know, it is part of the fun.

现在不用再这么麻烦了。

You don't have to do that anymore.

这一切都是通过空中升级完成的。

This is all over the air.

沃尔沃也宣布了同样的转变，从2020年起的16款车型都会支持。

Volvo announced the same shift, is gonna happen for 16 of its models going back to 2020.

这对很多车型来说都是一次大升级。

So this is gonna be a big upgrade for a lot of different vehicles.

再说，咱们谁还受得了跟Siri说话它却听不懂啊？

Beyond, you know, I think we're all kind of sick of trying to talk to Siri and it not working.

亚马逊和Alexa也做了不少大改进。

Amazon and Alexa have made big upgrades.

谷歌这次更是下了大功夫，我最近在《AI应用》播客里刚提过，谷歌地图来了个超强升级，以后Gmail和Docs也能直接集成进车里。

Google has made some massive upgrades here, specifically, like I recently mentioned on my AI applied podcast because Google has an incredible AI upgrade with Google Maps, and they're gonna be bringing in Google, Gmail and Google Docs all into that.

你只要动动嘴，就能跟车说话——要是真开了自动驾驶，说不定还能顺便把工作给干了。

So you're gonna be able to just talk to your car, And, I mean, technically, probably get work done if it was a self driving car.

总之，接下来有意思的事儿可多了。

Anyways, interesting things are gonna be happening.

好的。

Okay.

咱们聊聊Anthropic最近在搞什么。

Let's talk about what's happening with Anthropic.

他们昨天刚公开上线了Claude Security的测试版。

They just launched Claude Security in public beta yesterday.

目前只有Claude企业客户能用，就在claude.ai的侧边栏里。

It's available right now to Claude Enterprise customers with the claude.ai sidebar.

接下来，团队版和最高档套餐也会陆续上线。

Teams, max plans are all gonna start getting it next.

我觉得有意思的是，Security Week和SiliconANGLE的Ryan Nareen最近提到过这个功能。

And essentially, the thing that I think is interesting about this, Ryan Nareen, who's over at Security Week, and also SiliconANGLE, was talking about this.

它的工作方式就像安全专家一样，扫描你的代码库，追踪所有数据流向和读取源头。

And essentially what it does is it scans your code base the same way that a security researcher was would, and then it traces all of the data flows, the read sources.

它能分析不同文件里各组件之间的交互，然后像出一份风险评估报告一样，把漏洞给你列出来。

It looks at how components interact across files, and it hands back vulnerabilities, kind of like an impact report.

然后它还会给你提供具体的修复步骤和建议。

And then it's gonna give you production steps, and it's gonna recommend fixes.

这其实是在Anthropic宣布他们新模型之后推出的。

Now this is kind of coming on the backs of Anthropic saying, look.

他们推出了一个叫Mythos的新模型。

We have this new Mythos model.

这个模型在发现和利用安全漏洞方面特别厉害，现在主要给顶级公司用。

It's really good at security vulnerabilities, finding them, exploiting them, and so it's giving it to the kind of the top firms.

所以这次推出来的，感觉就像是给普通用户简化版的功能。

And so this seems almost like a dumbed down version that they're giving to, everybody.

不过说‘简化’可能不太准确，因为Mythos其实能跟你聊天，让你指挥它做各种事。

And and maybe dumbed down is a bad, you know, version for because Mythos, essentially, you can chat with and get it to do what you want.

而这个新功能更像是：我们只是在扫描你的代码。

This feels more like it is like, look, we're just scanning.

然后直接把发现的安全漏洞列给你看。

We're just gonna give you a list of security vulnerabilities.

我的理解是，你不可能光靠一个软件，盯着代码看看就能直接黑进去。

My understanding on this is you're not gonna just go to a software and it's gonna, you know, be able to go and just hack into it from from viewing it.

它必须能接触到源代码才行。

It's gonna have to get access to the code.

所以你得连上你的 GitHub 仓库，让它去扫描里面的内容。

So you would add, you know, you'd you'd connect your GitHub repo and be able to go scan that.

然后把扫描报告直接丢给 Claude Code，它就能自动帮你修复漏洞。

You then go drop its report straight into Claude code, and it's actually gonna be able to go and patch it.

这跟 Claude Design 的集成方式差不多——它能给你设计出超棒的网站或 App，然后生成一份简单的交接文档，你直接拿给 Claude Code 去实现就行。

This is kind of the same integration that we see with Claude design where it's gonna it it designs something amazing for your website or your app or whatever, and then it has this little, like, handoff document that it makes that you can just give to Claude Code to go and actually implement.

我真的超喜欢这个功能。

So I really love this.

我觉得这是 Anthropic 做得特别聪明的一招。

I think this is this is a really smart move by, Anthropic.

他们根本不需要搞什么 API。

They don't need an API.

他们根本不需要搞什么真正的系统集成。

They don't need to, like, have any sort of, like, literal integration.

它就生成一个文档，一个 Markdown 文件。

It just creates a document, an MD file.

对吧？

Right?

这个文件里写清楚要做什么，你把文件丢给 Claude Code，它就能直接执行。

That just explains what needs to happen, and you just give that file to Claude Cove that can go and execute on that.

测试版新增了很多定时扫描功能。

The beta is adding a bunch of scheduled scans.

对吧？

Right?

不只是能扫一次，还能设成定时扫描，自动发现各种漏洞。

So beyond just getting to do it once, you can actually have it run a a scheduled scan and and find any sort of vulnerabilities.

而且你还能标记那些不重要的问题，直接忽略掉。

And then you also get to dismiss anything that's not that important.

还有一个CSV和Markdown导出功能，你可以直接对接到你的工单系统。

There's a CVS and also a markdown export that you can also route to your ticket system.

所以现在推出了一大堆超棒的功能。

So a lot of really cool things rolling out.

他们目前的合作伙伴名单包括CrowdStrike、Palo Alto Networks、SentinelOne、Trend Micro和Wiz。

And the partner list that they have with this right now, they have CrowdStrike, Palo Alto Networks, SentinelOne, Trend Macro, Wiz.

这些公司都在把Cloud Opus 4.7直接整合进自己的平台里。

All of them are integrating Cloud Opus 4.7 directly into their platforms.

埃森哲、BCG、德勤、印孚瑟斯、普华永道，都在忙着搭建这套系统的实施团队。

Accenture, BCG, Deloitte, Infosys, PwC, they're all spinning up deployment, practices around this.

我觉得现在企业安全领域的大牌，几乎全都在这儿了，连那些顶级咨询公司都上了。

And I think that's basically every name we have in enterprise security right now, all the big consulting firms.

所以我觉得，这个事儿里被忽略的关键点其实是时机。

So I think, you know, the under discussed angle on this is just the timing.

在网络安全这个领域，Anthropic在联邦市场是畅通无阻的。

Cybersecurity is one category where Anthropic isn't blocked from the federal market.

就连五角大楼现在也正闹得不可开交。

Even with the Pentagon right now, you know, they have this big fight.

他们被挡在其他几笔大单之外，但安全领域政府还是允许他们参与的。

They're they're blocked out of some other big deals coming out of the Pentagon, but security is something that they're still allowed to to do for the government.

所以感觉他们在这块儿是铆足了劲在推。

And so it feels like this is a an area they're pushing on very heavily.

明白了。

Okay.

咱们聊聊马斯克和阿尔特曼那场官司现在咋样了。

Let's talk about what's going on with the Elon Musk, Sam Altman trials.

这事儿接下来几周肯定没完没了地发酵。

I mean, this is gonna be rolling around for the next couple weeks.

我估计每天都能冒出点新料来。

I'm just gonna have I base I feel like, basically, daily tidbits coming out of this.

昨天马斯克在加州联邦法院出庭，有人问他，XAI是不是用OpenAI的模型做蒸馏来训练Grok的。

But Elon yesterday, was in federal court in California in the lawsuit, and he was asked whether XAI used distillation on OpenAI models to train Grok.

对吧？

Right?

其实就是问OpenAI模型一个问题，拿到它的回答，再把问题和回答一起喂给Grok，用来训练它。

So basically asking OpenAI model a question, getting their answer, and then feeding the answer, the question and the response back to Grok to help train it.

他说他们确实部分这么做了。

And he said that they partly did.

他还想轻描淡写地说，这在业内挺常见的——说实话，确实如此。

He tried to kind of wave it off saying it's kind of general industry practice, which to be fair, yeah, it is.

DeepSeek就是这么干的。

DeepSeek, that's how DeepSeek did it.

实际上，所有AI公司都这么干，虽然违反了OpenAI的服务条款，但并不违法。

Like, actually, every AI model company does it, and it's against the terms of service of OpenAI, but it's not illegal.

对吧？

Right?

OpenAI根本不想让人这么做。

Like, OpenAI doesn't want you to do it.

而且Anthropic最近真的试图游说政府，逼着公司们别这么干，因为这简直就是偷懒的训练方式——能快速拿到大量输出结果。

And and Anthropic actually tried to, I think, lobby recently to to get the government to, like, force companies to stop doing it because it's a it's kind of like a cheap way to train a good model because you can just get a lot of these outputs really fast.

这还能大大加快训练进程，因为你直接告诉模型：当有人问这种问题时，我们就希望你这么回答。

And it kind of speeds up the whole training process because you're telling it, look, when someone asks you this type of question, this is the kind of response that we would like you to give.

有意思的是，如果你想得到最顶尖、最深思熟虑的回答，用这种方法可能还真得不到，不过这事儿吧，我觉得另说。

Now it's kind of interesting because if you want the absolute best cutting edge, well thought out answers, you're not probably gonna get it per se in this in this way, but anyways, another conversation in my opinion.

说到怎么让AI模型发挥出最佳效果，如果你已经在用ChatGPT、Claude、Gemini或者Grok，我强烈推荐你试试我的平台——aibox.ai。

And speaking of getting the best results out of these AI models, if you're already paying for ChatGPT or Claude or Gemini or Grok, I would love for you to check out my software platform, which is called aibox.ai.

你能一次性访问80多个顶级AI模型，最新最全的工具全都有，比如Eleven Labs的语音、音乐生成、视频生成、图像生成，全都在aibox.ai上，现在每月只要8.99美元。

You get access to over 80 of the top AI models, all of the latest stuff, including all of the amazing audio tools like eleven Labs, music generation tools, video generation tools, image generation tools, all of that is over on aibox.ai, and it's currently $8.99 a month.

你要是现在锁定这个价格，以后就再也不会涨了。

If you get that price, you'll get locked in, to that.

我们保证不加价，而且所有80多个模型都集中在一个地方，随你用。

We won't raise it on you, and you're able to get access to over 80 different models all in one place.

我们还提供年费方案，能再打八折。

We also have an annual plan which makes that 20% off as well.

所以，赶紧去试试吧。

So anyways, go check it out.

你还可以用我们的氛围构建器，它能把这些模型连起来，帮你自动完成各种任务。

You can also, talk to our vibe builder which links together all of these different models and automates workflows for you.

希望这能帮你省一大笔钱，而且特别实用。

I hope this is something that saves you a ton of money and is incredibly useful.

链接在简介里。

Link is in the description.

别忘了去试试，告诉我你的感受。

I hope you check it out and let me know what you think.

接下来是新闻：五角大楼刚刚发布声明，宣布有七家AI公司被允许接入他们的机密网络。

Next up on the news, we have the Pentagon, which has just released a statement naming seven different AI companies that they're gonna be letting into their classified network.

分别是OpenAI、谷歌、英伟达、SpaceX、Reflection、微软和AWS。

So they have OpenAI, Google, NVIDIA, SpaceX, Reflection, Microsoft, AWS.

这些公司都能接入他们的机密网络了。

All of these are getting access into kind of their classified networks.

路由器和《美国新闻》都提前泄露了这个正式公告。

Routers and US News both kind of broke the formal announcement.

这些公司现在已经被接入六级和七级安全等级，这可是最高机密级别。

And the companies right now are integrated into impact levels six and seven, which is basically the highest classification tier.

你能接触到机密和绝密数据，甚至任何涉及实际军事行动的信息。

You get, you know, secret and top secret data and anything that touches an actual active operation.

这些公司 basically 拿到了最顶级的资源，帮美国政府处理一些最重要、最关键的项目。

So these companies are, you know, basically getting the the top of the line, information and helping with some of the biggest projects the US government is working on, some of the most critical things.

不过在我看来，名单里没提的公司才是真正的重点——那就是Anthropic。

Now in that list, is probab the like, what's not included in that list is the bigger story in my opinion, which is Anthropic.

这事儿其实源于国防部和Anthropic之间的矛盾，后来他们把Anthropic定为供应链风险，这标签一般只给外国对手，因为Anthropic不肯让国防部随便访问他们的系统，双方闹得挺僵的。

Now this is all, of course, coming from the dispute that the DOD had with Anthropic over and, you know, they eventually labeled them as a supply chain risk, which is, a terrible designation for a company that's, a foreign adversary usually because they wouldn't give the, you know, the DOD unrestricted access to Anthropic to do whatever they want with it, and there was kind of this whole legal spat.

艾米丽·米歇尔，也就是国防部的副部长，在X上直接骂达里奥·阿马德奥是个自大狂，撒谎精。

Emily Michel went on x, and she's the the Pentagon undersecretary, and she called Dario Amadeo a liar with a god complex.

总之，这背后一堆狗血剧情。

Anyways, there's lots of drama in that.

但有意思的是，就在这一切发生的同时，加利福尼亚的一位联邦法官早在三月就已叫停了这个供应链风险的认定。

What's interesting though is with all of that, a federal judge in California already blocked the supply chain risk designation back in March.

更有趣的是，如果你看下实际数据，Claude 其实是国防部内部用得最多的工具之一，而这东西的威力可不小。

What's interesting is if you look at all of the data, Claude is actually one of the most used internal tools at the Department of War, and, of course, this is something that is, you know, very powerful.

所以它这么抢手也不奇怪，尤其是对开发者来说，这会儿简直就是行业里的标配工具。

And so it's no shocker that it would be, you know, highly sought after, especially for developers who this is kind of the the tool of choice right now in the industry.

Axios 的迈克·艾伦本周早些时候报道，白宫办公厅主任苏西·威尔斯和财政部长斯科特·贝塞特已经跟阿马德奥见过面，想找个台阶下。

Axios Mike Allen actually reported earlier this week that the White House chief of staff, Sussie Willis, and the treasure secretary, Scott Bessette, already met with Amadeo to try to find an exit.

对吧？

Right?

他们现在琢磨的是，怎么才能把 Anthropic 拉回来。

So they're trying to find, like, how could they how could they bring them back in.

原话是说，怎么既保全面子，又把他们请回来。

The quote is how to save face and bring them back in.

这事还没定下来，咱们就等着瞧吧。

This hasn't shipped yet, so we're gonna see what happens there.

但考虑到Anthropic目前的时机，这简直太惊人了。

But this is also really incredible considering the timing for Anthropic right now.

他们现在估值高达9000亿美元，据说正在筹备一轮约500亿美元的融资。

They have a $900,000,000,000 valuation, round of funding that people are looking at, that they're allegedly looking at doing with about a $50,000,000,000 in, funding that they're going to raise.

据TechCrunch报道，他们刚给投资人下了48小时的最后通牒，要求尽快确认投资额度。

And apparently, according to TechCrunch, they just gave investors forty eight hours to submit allocations and agree to the amount that they're gonna be putting in.

看来这笔交易很快就要敲定了，这操作也太猛了，直接跟投资人说：

So this seems like something that might be wrapped up here very quickly and they're you know, I mean, this that's pretty aggressive to just tell your investors, hey.

听好了。

Look.

你们只有48小时。

You have forty eight hours.

到底投不投，给个准信。

Are you in or are you out?

但我认为这特别关键，因为之前很多人都说，Anthropic根本没主动找过融资。

But the reason why I think that that is important is because a lot of people have been saying that Anthropic hadn't been out actively seeking funding.

他们其实刚融过一轮钱。

They'd they'd just raised some.

本来大家以为他们不缺钱，结果一堆人主动找上门，几乎是在求着他们收投资。

They'd been allegedly fine, but they had a bunch of people just, like, coming to them and basically begging them to take their money.

这事是不是真的，或者只是宣传话术，现在还不好说，但Anthropic这很可能是最后一次私募融资，之后就要准备IPO了。

So whether that's true or part of the narrative for all of this is, left to be said, but Anthropic is probably the last private, you know, this is probably the last private round for Anthropic before they have their IPO.

所以现在想入局的人，趁估值还没涨，赶紧下手，这大概就是Anthropic在跟投资人说的潜台词。

So if anyone wants to get in now before the valuation probably goes up, now would be the moment, or at least that's the pitch that Anthropic is giving.

现在整个AI行业真是太疯狂了。

So wild times for the entire AI industry right now.

不过还有件事，大家讨论得太少了——就是国防部那笔大单，还有他们点名的那些公司里，有一家叫Reflection AI的也在名单上。

One other thing that I think is not talked about enough though with the whole, with the whole deal that the DOD did and all the different companies that they announced that they're going to be that are gonna be working on them is because there is a company called Reflection AI that is currently on the Pentagon's list.

Reflection的两位创始人是米沙·拉斯金和罗尼斯·安图奥格鲁。

Reflection's two founders are Misha Laskin and Lonis Antuoglou.

这俩人以前都是DeepMind的。

Both of them are from DeepMind.

以前，拉斯金在 Gemini 上做奖励模型。

Formerly, Laskin ran their reward modeling on Gemini.

安图奥格鲁是 AlphaGo 的核心架构师。

Antelogo was a core architect at AlphaGo.

所以，这些显然是非常传奇的研究员。

So, you know, these are obviously very legendary researchers.

他们去年十月以八十亿美元的估值融了二十亿美元，但这些公司可不是世界上最大的。

They raised $2,000,000,000 at an $8,000,000,000 valuation last October, but these aren't the biggest companies in the world.

而且说实话，大多数人可能根本没听过 Reflection 这家公司。

And I would also say, you know, most people probably haven't even heard of Reflection.

我觉得五角大楼把他们和 OpenAI、微软放在一起，却把 Anthropic 删了，这本身就是个信号。

I think the Pentagon basically putting them on the same list with, like, OpenAI and Microsoft and leaving Anthropic off of that is kind of their own way of saying, look.

瞧，我们还有你压根没听说过的前沿实验室选项。

Like, we have some Frontier lab options that you've never heard of.

如果非得如此，我们就自己掏钱支持这些 Anthropic 的替代者。

We're gonna fund these alternatives to Anthropic ourselves if we have to.

我觉得这其实是个挺有意思的策略。

And so I think that's kind of a interesting strategic move.

他们把一堆大公司列上去，又加了几个全新的，一看就是重点扶持的对象，估计能拿到一些定制化的东西，别的公司可能没这个能力或者不愿意干。

They put a lot of these big companies on there, then they put some brand new ones on that feels like they're kind of funding, and they'll probably get some custom stuff done that they need, that others might not be willing or able to do.

我觉得Anthropic被排除在外，本身就是一个信号。

I think the Anthropic exclusion is kind of a symbol.

对吧？

Right?

说不定Reflection就是被当成替代品放上来的。

Reflection is basically getting waved in as a substitute perhaps.

我也不确定。

I don't know.

但不管怎样，接下来会发生什么，真挺让人关注的。

But anyways, it's very fascinating to see what's gonna happen here.

好吧。

Okay.

今天节目就到这里了。

That's it for the show today.

如果你觉得有收获，能不能去苹果播客留个评论？那对我真的特别重要。

If you got something out of it, it would mean the world to me if you could drop a comment over on Apple Podcasts.

如果你在Spotify上听，就点个五星好评吧。

Or if you're on Spotify, hit the stars.

在‘关于’页面就能找到，这能帮节目大忙了。

It's on the about tab, and it would help the show out a ton.

想用80多个AI模型，再配上一个超简单的自动化工具？每月8.99美元，去aibox.ai看看吧。

If you wanna get access to over 80 plus AI models and an automation builder that runs, super simply, it is eight ninety nine a month to go check out aibox.ai.

链接在简介里。

There's a link in the description.

好了。

Alright.

咱们下期再见！

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