新任苹果首席执行官、OpenAI 新图像模型、Vercel AI 黑客马拉松

苹果CEO蒂姆·库克正式将在9月1日卸任，接替他的是约翰·塔纳斯。

Apple CEO Tim Cook is officially stepping down on September 1, and John Tarnas is gonna become the CEO.

这事儿真挺有意思的。

This is really interesting.

苹果在AI方面其实做了不少事，但说实话，他们一直没兑现当初的承诺。

There's a lot of AI things at play for Apple, and, you know, they really haven't delivered on their AI promise.

咱们待会儿就聊聊这个。

We're gonna be getting into that.

我平时用Vercel做所有AI工具的托管，结果他们出了个大安全问题，原因是有个员工用了个AI工具。

Vercel, the company that I personally use for being my web host for all of my AI tools, has had a major security incident because of an AI tool that one of their employees was using.

FTC刚和Clarify达成和解，因为他们用了三百万张OkCupid的图片来训练人脸识别模型。

And the FTC just dropped a settlement with Clarify over the company using 3,000,000 OkCupid photos to train facial recognition.

Snapchat裁了大约一千人，说这是他们AI计划的一部分，因为现在AI写的代码已经占了他们新代码的65%以上。

Snapchat laid off around a thousand people and said that it is part of their AI initiative because AI is writing more than 65% of their new code.

OpenAI刚发布了一个叫Image 2.0的新图像模型，进步真的很大。

They dropped a major new image model over at OpenAI called Image two point o that is a huge step up.

我今天一直在玩这个。

I've been playing with it today.

待会儿咱们聊聊这个。

We'll talk about that.

另外，Anthropic 的 Mythos 模型最近口碑有点下滑，萨姆·阿尔特曼还直接点名批评他们搞恐吓式营销。

And in addition, Anthropic's Mythos model has, of course, getting is starting to get a bad rep, and Sam Altman is going and calling them out for their fear based marketing.

他们还跟诺和诺德签了个大单子，就是那个生产 Ozempic 的公司，待会儿咱们也会提到。

They also signed a huge enterprise deal at Novo Nordisk, the Ozempic company we'll be talking about.

我觉得最大的新闻是，亚马逊又砸了超过两千五百亿美元给 Anthropic，这还没算上 AWS 团队之前已经投的一千亿。

And I think one of the biggest stories is that Amazon is pouring over $25,000,000,000 into Anthropic on top of the $100,000,000,000 that the AWS team has already committed.

如果你现在还在为 ChatGPT、Claude、Gemini、Grok，或者 Eleven Labs 的语音模型、各种图像模型付费，那你真得听听我介绍的 AI Box。

Now if you're still paying for ChatGPT, Claude, Gemini, Grok, any of the audio models, Eleven Labs for audio, and any of the image models, I've gotta tell you about AI Box.

这东西是我自己搞出来的。

This is what I personally have built.

我朋友问我怎么用 AI 又不被订阅费掏空，我就是推荐他们用这个。

It's what I'm recommending to my friends who ask me how to actually use AI without going broke on the subscriptions.

你可以在一个地方访问八十多个AI模型，所有主流模型都包含在内，想用哪个就用哪个，完全根据你的任务来选。

You get access to over 80 different AI models in one place, all of the top models, so you can pick whichever one is best for the task that you are doing.

我觉得特别实用的是我们刚推出的那个自动化构建工具。

The part that I think is super useful is our automation builder that we have just created.

你用大白话描述你想干嘛，它就能自动给你搭好工作流程。

You describe what you want in plain English, and it builds out a workflow for you.

根本不用懂编程。

You don't need to know how to code.

也不用费劲去折腾什么新平台，完全没这必要。

You don't need to how know how to, you know, wrestle with a with a new platform.

你说清楚想要什么，它就帮你搞定。

You describe what you want, it builds it for you.

每月只要8.99美元，真心希望能帮你省下一大笔钱，让你轻松用上所有这些AI模型。

So it's $8.99 a month, and I hope this is something that saves you a ton of money and is incredibly useful for getting your hands on all of the different AI models to use them.

好了。

Alright.

咱们聊聊 Vercel 的安全漏洞吧。

Let's talk about Vercel's security breach.

这家公司就是 Next.js 的背后团队。

This is the company behind Next.

Next.js。

Js.

他们托管了现代网页的很大一部分。

They host a huge chunk of the modern web.

我所有的新 AI 应用基本都跑在他们平台上，现在他们确认被黑了，客户数据被盗了。

I basically create all of my new AI apps on them, and they have confirmed that they have got breached and customer data was stolen.

但对我个人来说，最有趣的是这一点。

But this is the interesting part for me in particular.

他们不是直接被黑的。

They don't get hacked directly.

攻击者其实是入侵了 Context AI——这是 Vercel 一名员工在用的 AI 工具。

The actual the attacker actually broke into context AI, which is an AI tool that Vercel, one of their employees, was using.

然后他们就从那里入侵了这位员工的谷歌工作区，再顺藤摸瓜进入了 Vercel 的系统。

And from there, they got into that employee's Google workspace and from there into Vercel's environment.

所以他们是一步步来的，用了好几道工序才攻进去。

So there's kinda like this multistep approach, that they took in order to to get in.

其实最初的入侵早在二月份就发生了。

And the original compromise actually happened back in February.

我们到现在才听说这事。

We're just hearing about it now.

我觉得这种攻击方式挺常见的。

I feel like this is how a lot of these kind of attacks happen.

就是吧，事情早就发生了，

It you know, the the event happens.

他们也不想往外说，拖到实在藏不住了才曝光。

They don't really wanna leak it, and then it finally, you know, comes out.

结果有个 Contacts AI 的员工中了信息窃取木马，攻击者拿到合法凭证后，一直按兵不动，就等着找机会串联起来搞个大的。

So a Contacts AI employee actually got hit with the info stealer malware, and the attacker basically just sat on the valid credentials until they could chain them into something bigger.

Vercel 表示受影响的客户数量有限，但一个自称‘Shiny Hunters’的团体已经在黑客论坛上兜售他们声称从这次入侵中获取的 API 密钥、源代码和数据库数据。

Vercel says that the number of affected customers is limited, but a group claiming to be they're they're calling themselves shiny hunters is on hacker forums trying to sell API keys, source code, and database data that they say they pulled from this whole hack.

在我看来，现在每个公司都在往自己的系统里硬塞各种 AI 工具，而每个新工具都意味着一个新的供应商。

So in my opinion, every company right now is bolting AI tools onto their stack, and every new tool is a new vendor.

新的凭证。

It's new credentials.

新的突破口，让攻击者有机可乘。

It's a new way for attackers to get in.

如果未来会有越来越多的 AI 工具被开发出来，那类似这种围绕 AI 的供应链攻击也会越来越多。

I think if you're gonna see, you know, a lot of these new AI tools get built, we're also gonna see a lot more AI adjacent supply chain breaches.

这种情况只会越来越多，直到情况好转为止。

This is gonna happen more before it gets better.

如果你在一家公司，那你通过这些 AI 工具能接触到哪些东西？

If you're at a company, who has access to what through your AI tools?

我觉得这根本不是什么理论上的风险。

I don't think this is something theoretical.

我觉得每个人都该好好想想这个问题。

I think everyone should be thinking about this.

FTC 刚刚和一家叫 Clarify 的 AI 公司达成了和解，因为他们用了 OkCupid 上三百万张照片。

The FTC also just settled with the AI company called Clarify over the use of 3,000,000 photos from OkCupid.

就是那个交友网站。

That's the dating website.

他们拿这些照片来训练人脸识别模型。

And they use these to train facial recognition AI.

最绝的是，人家还真留了证据。

And the best part is that they actually have receipts.

Clarify 的 CEO 直接给 OkCupid 的联合创始人发邮件，说：'我们现在在收数据，刚发现 OkCupid 肯定有海量优质素材可用。'

So the Clarify CEO literally emailed the OkCupid cofounder saying, quote, we're collecting data now and just realized that OkCupid must have a huge amount of awesome data for this.

这些照片后来又被传了回去。

The photos were then shared going back.

他们手里的所有照片，最早能追溯到 2014 年。

All of the photos that they had were shared going back to 2014.

用户根本不知道发生了什么。

Users didn't know what was going on.

所以作为和解的一部分，Clarify 必须删除这三百万张照片以及所有基于这些照片训练的模型。

So as part of the settlement, Clarify has to delete the 3,000,000 photos and any models trained on them.

我觉得这事儿简直荒唐到家了。

I think that this is, you know, something that is absolutely insane.

多年来，AI公司一直游走在灰色地带，动不动就说：‘反正这东西在网上能搜到，我们用一下怎么了？’

For years, AI companies have been operating kind of in this gray area that, you know, they're like, well, if it was on the Internet, we can use it.

结果FTC直接撂下话：不行。

And the FTC just said, no.

你们真不能这么干。

You actually can't.

只要用户没明确同意这种用途，你们就没资格用。

If people don't consent to this specific use, then you can't actually have it.

而且我觉得，那些基于这些数据训练的模型也得全盘推倒重来，这代价可太大了。

And I think the models that are trained on it have to also come down, which is incredibly expensive.

所以这对整个行业来说都不是什么好事。

So this is just not good for anyone in the industry.

使用数据之前必须获得授权。

You need consent for use of the data.

我觉得这个案例特别有代表性，如果你是家用爬取数据训练模型、却没明确获得授权的AI公司，那你真该好好研究下这个和解方案。

I think that this is a really interesting precedent, this story in particular, because if you're an AI company that trained on scraped data without clear consent, basically, this is the settlement that you should be studying.

我觉得以后这种事会越来越多。

I think we're gonna see a lot more of these in future.

明白了。

Okay.

Snapchat刚宣布要裁掉大约一千人。

Snapchat just announced that they're cutting roughly a thousand people.

这占了他们全球员工的16%左右。

That's about 16% of their global workforce.

他们还取消了三百多个开放职位。

They're closing more than 300 open roles.

就连CEO斯皮格尔都发了备忘录，宣布了这些决定。

And even Spiegel sent the memo that's kind of announcing all this, their CEO.

他给出的理由，我觉得特别关键。

And the reason that he gave is is basically the thing that I I think is critical.

他说，AI现在已能生成Snap超过65%的新代码，公司预计在今年下半年前能节省超过5亿美元的年度开支。

He said that AI is now generating more than 65% of new code at Snap, and the company expects to cut over $500,000,000 annualized expenses by the back half of this year.

所以他们这次裁员真是大手笔。

So they are cutting in a massive way.

消息一出，股价立马涨了近8%，这说明华尔街怎么看这个取舍，一目了然。

The stock jumped almost 8% on the news, and I think that tells you exactly how Wall Street is feeling about this trade off.

依我看，这事其实早就在我们预料之中。

In my opinion, I think this is something that a lot of us have been saying was coming.

这种事以后肯定还会再发生，毕竟现在连CEO都得给上市公司发备忘录，明说AI在写大部分代码了。

I this is gonna be repeated because here we have a CEO writing a memo to a public company.

AI已经负责了大部分代码的编写。

AI is writing a majority of the code.

小团队反而交付得更多了。

Smaller teams are shipping more.

而且他们也直接把这归功于AI，市场也给了他们回报。

And they're also being directly you know, they're attributing this directly to AI, the market is rewarding them.

华尔街也在奖励他们。

Wall Street is rewarding them.

我觉得这种趋势会在更多公司里看到。

So I think we're gonna see this in a lot more companies.

我觉得这其实告诉我们一件事。

This is what I do think it tells us.

如果你是知识型工作者，问题早就不是AI会不会改变你的工作了。

If you're a knowledge worker, the question isn't whether AI is gonna reshape your job.

它已经改变了。

It already did.

关键是你是不是在用AI让自己成为团队里的高手，还是你的活儿只是被合并到别人的工作里了。

I think the question is whether you're using it to become really a power user on the team or whether your tasks just got consolidated into someone else's workflow.

我知道这听起来挺狠的，但我觉得这确实是实话。

I know this sounds really ruthless, but I think this is a very honest take.

再说他这次裁了全球16%的员工，你说谁没被裁？在我看来，那些留下的才是关键。

And and looking at the, you know, this case where he cut about 16% of their global workforce, well, who didn't he cut in my opinion?

说实话，公司留下的那一大块人，都是成了AI高手的那批人。

In my opinion, I mean, there's still a huge chunk of the workforce that they're keeping, and those are the people that are becoming the power users.

他们都在用AI。

They're using AI.

所以现在这事儿真的特别关键。

So I just think it's really critical now.

如果你被裁了，那你更得赶紧上手这些工具。

And if you were someone that got cut, you're gonna wanna use these tools.

得学会怎么真正把它们融入工作，让自己更有价值。

You're gonna wanna learn how to really integrate them and to make yourself more valuable.

下次找工作，人家就希望你能在自己行业里当个AI专家。

So the next job that you find, they're gonna want you to to basically be kind of like an AI expert in your industry.

总之吧，我知道这么说有点狠，但我觉得确实没错。

So, anyways, I know this a ruthless ruthless way to put it, but I do think it's true.

所以我只想说，咱们都得把AI用到极致，把新功能都试个遍。

And so I just wanna put it out there that we should all be taking AI to the pushing it to the the limits, testing out everything new.

你要是听这个播客，八成早就这么干了。

If you're listening to the podcast, that's probably what you're already doing.

明白了。

Okay.

今天OpenAI发布了好多新消息。

OpenAI has a ton of news today.

我来给你捋一捋。

I'm gonna tie it all together.

第一件事，他们推出了ChatGPT Image 2.0。

The first thing is that they dropped ChatGPT Image two point o.

这东西我最近一直在玩。

This is something I've been playing with.

之前他们用的是 Image 1.5。

Previously, they had Image 1.5.

这个新版本特别擅长以前 1.5 版本搞不定的事儿——就是生成文字。

Basically, it's, really good at doing something that the last model 1.5 was pretty bad at, which is rendering text.

你要是让它做张海报，老是出来那种怪里怪气、像精灵似的字，说不上来哪儿不对。

If you ask it for a poster, you know, you would always get this kind of like elfish looking I don't know.

那些字母根本就看不下去。

The the letters weren't that weren't that great.

其实我最近就遇到个事儿，本来想说个水管工，哎不对。

In fact, I I recently had, like, a plumber or a sorry.

是个空调维修工上门给我修空调，事后发了张名片过来。

An AC repairman come to my house, and he sent me his business card via text.

我一眼就看出来是用 ChatGPT 生成的，因为上面有个错别字——字迹太烂了。

I could tell it was generated with chat GPT, there was, a typo in it because the letters just weren't that good.

总之，他们自己说，2.0 版本是个质的飞跃。

Anyways, two point o, in their words, is a, quote, step change.

他们老说‘质的飞跃’，我都搞不懂这到底啥意思。

They always say step change, so I don't even know what that means.

就是更好用了。

It's better.

明白了。

Okay.

太棒了。

Fantastic.

它能生成特别清晰的多语言文字，其他语言也一样拿手。

Basically, it makes really clear multilingual text, so it's good at doing other languages as well.

像日语、韩语、印地语、孟加拉语这些非拉丁字母的文字，它都能搞定。

It can do non Latin scripts like Japanese, Korean, Hindi, and Bengali.

我特别兴奋，因为我其实给我的公司 PodcastStudio.com 做了个功能，还没公开呢。

Personally, I'm excited about that because I actually built a feature into my software company podcaststudio.com, which isn't public.

我还没对外发布呢。

I haven't dropped anything yet.

我本人正在用，还找了几个内测用户，所以还没搞成什么大项目。

I I'm personally using it and have a couple beta testers on it, so it's not something big.

但我加了一个功能，主要是帮你更快、更轻松地发布播客集数。

But one of the features that I added to it was essentially it's it's basically to help you publish podcast episodes faster and easier.

其中一个功能是你可以选择多种语言。

And one thing that it does is you can select multiple languages.

比如我这个正在听的播客，我也同时发布了法语版，还有好几种其他语言。

So, like, for my own podcast that you're listening to here, I also publish this in French and like a bunch of other languages.

我们还能帮你直接做出对应语言的播客版本，连分发渠道都帮你配好。

But we can actually build you your own podcast in that language and help you get all of your distribution set up and stuff.

最酷的是，我们会自动重新生成你的播客封面图。

And the cool thing is we actually regenerate your podcast cover image.

比如封面上有你的播客名字和你的头像，它会自动把文字换成新语言，其他部分都不动。

So, like, if you have text with your podcast name and you on it, it will regenerate your whole cover image just changing the text to whatever the new language is.

反正呢，这是我做的一个小功能。

Anyways, it's one of the things I built.

我特别兴奋，因为以前翻译成其他语言时，文字经常乱掉，我每次都得反复核对，确保内容没错。

I'm stoked about it because if it was a bit of a struggle when you translate it to a new language, the text, sometimes would get kind of garbled, and I'd always have to double check to make sure that it actually said what it was supposed to.

所以，说白了，这正是它最拿手的地方。

So, anyways, this is what it's good at.

它特别擅长处理信息图、幻灯片、地图，还有漫画页面。

It's good at infographics, slide decks, maps, pages of comic books.

它能根据一个提示，生成多达八张风格一致的图片。

It can produce up to eight consistent images from one prompt.

而且我觉得，这是第一个内置推理能力的OpenAI图像模型，能自动上网查证，确保生成的内容准确无误才最终出图。

And I also think that it's the first OpenAI image model with reasoning baked into it so it can search the web and verify that it's generating something correctly before it finishes the images.

光是这一点，就已经是巨大的升级了。

That's a massive upgrade just in and of itself.

今天刚上线。

It rolled out today.

现在免费开放，Go用户、Plus和Pro用户还能享受更高清的版本，而且API已经接入了。

It's for free and Go users plus and pro get higher quality tiers basically, and it's already in the API.

这工具简直太棒了。

So this is just an awesome tool.

我简直太兴奋了。

I'm super stoked about it.

萨姆·阿尔特曼最近上了一个播客，直接点名批评了Anthropic的Claude Mythos预览版。

Sam Altman also went on a podcast recently, and he took a very direct shot at open or at Anthropic's Claude Mythos preview.

他说这是‘靠制造恐惧做营销’。

He called it, quote, fear based marketing.

如果你没关注，Anthropic之前说，Mythos太擅长发现软件漏洞了，所以他们不敢公开发布。

So if you missed that, Anthropic basically said that Mythos is so good at finding software vulnerabilities that they, you know, they can't release it publicly.

天哪。

Oh, no.

这玩意儿要统治世界了。

It's gonna take over the world.

萨姆·阿尔特曼的意思是，他们这么宣传，其实就是想装厉害，把AI技术牢牢攥在自己手里，或者只留给那些大公司。

And Sam Altman basically made the argument that their framing was just basically a way to look powerful and keep AI exclusively to for themselves or their that, you know, rich companies.

我觉得这两件事可能都是对的。

I think that probably both of these things are true.

我觉得Mythos确实是个很强的模型，但Anthropic一贯的做法是，每次发布新模型，都配套发一篇报告，说这玩意儿多危险、多离谱。

I think Mythos does seem like a really capable model, but I also do think that there's a consistent pattern at Anthropic of announcing a model alongside a, you know, report on how dangerous it is or how crazy it is.

萨姆·阿尔特曼点出这点，我觉得挺有道理的。

I think Sam Altman is kind of right to call that out.

这种操作确实挺烦人的。

It it definitely is annoying.

你回看最早期的ChatGPT，GPT-1那会儿，达里奥还在OpenAI的时候，他们就说这东西太危险了，不能公开。

You could see articles, like, dating back to the first version of ChatGPT, like GPT one when Dario was over at OpenAI, they were saying, like, it's so dangerous to to release this thing.

不过呢，萨姆·阿尔特曼也不是啥中立角色，但当有人动不动就说‘这模型要毁灭世界’，而其实根本没那么夸张的时候，确实该有人站出来拆穿。

Anyways, I don't think Sam Almond's definite like, necessarily a neutral party on that, but I do think it's it's good to call people out when they're saying that their model's gonna destroy the world when it probably isn't.

有人分析过数据，说他们挑了最极端的案例，实际没他们说的那么危险。

Some people have been looking at the data and saying they cherry picked a bunch of it, and it's not dangerous as they basically led on to say.

行吧。

Alright.

诺和诺德。

Novo Nordisk.

生产司美格鲁肽的诺和诺德最近和OpenAI达成了一项巨额合作，要把AI全面应用到公司各个领域——药物研发、临床试验、生产制造、供应链，全盘部署，目标是2026年完成。

The Ozempic company recently signed a massive partnership with OpenAI to roll out AI across their entire business, drug discovery, clinical trials, manufacturing, supply chain, everything, full deployment by the 2026.

如果OpenAI真能帮诺和诺德大幅缩短药物研发周期，那对人类健康的长远影响，我觉得会非常大。

If OpenAI can actually help Novo compress drug discovery timelines, then the main then the downstream impact on human health, I think, is gonna be really big.

我对AI在医疗领域的应用特别感兴趣。

This is an area I'm excited about is AI in health care.

我觉得未来所有大型药企都会签类似这样的合作。

And I think that every major pharma company is also gonna be signing deals that look basically exactly like this in the future.

好吧。

Okay.

苹果刚宣布了CEO交接的事。

Apple just announced they're they have a CEO transition.

这可是下一个重磅玩家。

This is basically the next big player.

我们有过乔布斯，有过库克，现在终于迎来了下一个重磅人物。

We had Steve Jobs, Tim Cook, and we have the next big player finally coming in.

库克当苹果CEO已经十五年了，这次他们请来了约翰·塔尔纳斯，他现在是硬件工程的高级副总裁。

Tim Cook has been the CEO of Apple for, like, fifteen years, and they're going to be pulling in John Tarnas, who is currently SVP of hardware engineering.

他从9月1号起正式接任CEO。

He's gonna become the CEO September 1.

库克会转任执行董事长。

Cook is going to move to be the executive chairman.

塔尔纳斯2021年就加入苹果了。

Tarnas has been at Apple since 2021.

他是个工程师出身。

So he's an engineer.

过去十年苹果推出的每一款重要硬件产品，背后都有他默默操盘，这是自2011年库克接替乔布斯以来，苹果第一次换CEO。

He has quietly overseen basically every major hardware product that Apple has shipped in the last decade, and this is the first big CEO change since Cook took over from Steve Jobs in 2011.

所以我特意在这档AI播客里提这件事，是因为这个时间点太关键了。

So the reason why I wanna talk about it on the on this AI podcast is because the timing, I think, is really important.

我一直对苹果很不满，觉得他们在AI上落后得太远了，至少从我有印象开始就是这样。

I have been very critical of Apple saying that they are far behind for, don't know, as long as I can remember on AI.

苹果智能推出得太晚了。

And Apple intelligence was late.

说实话，根本没达到预期。

It was definitely underwhelming.

Siri都十五年了，还是那个Siri。

Siri, after fifteen years, is still Siri.

可你看OpenAI、Anthropic、谷歌、Meta，一个个都在疯狂推AI产品。

Meanwhile, OpenAI, Anthropic, Google, Meta, they're all shipping these really aggressive AI products.

苹果对外宣传的AI主打点就一句话：我们做在设备上，隐私有保障，你放心。可这话说了这么多年，压根没兑现。

And Apple's main public AI pitch has basically been, you know, we'll do it on a device and and private and you could trust us, but they've never really delivered on this.

苹果根本没做到。

Apple has not.

光许了一堆诺言。

They made a lot of promises.

所以看到蒂姆·库克宣布了这么多AI功能，然后就撂挑子不管了，真是有点离谱。

So it's kinda wild to see Tim Cook having all of these kind of AI features announced and then just saying, alright.

把烂摊子丢给下一个人去收拾。

Handing over the reins to the next guy to deliver on them.

我觉得现在把公司交到一个硬件工程师手里，是个重大抉择。

I think handing the company to a hardware engineer at right now is basically a big choice.

苹果相信，下一波AI大战会在设备端、芯片上、功耗效率和所有芯片模型的整合上展开。

I think that Apple believes the next AI war is gonna be one on devices and on silicon, on power efficiency, on the integration between all of the chip models.

我觉得塔纳斯这一路的职业生涯，大概就是为这个准备的。

And I think that's what Tarnas' entire career hopefully has prepared him for.

如果他们这次押对了，五年后回头看，这个交接简直完美。

If they are making a correct bet there, then I think the transition is gonna look amazing in five years.

但如果他们押错了，AI还是离不开云，那苹果就选错了时机，把公司交给了一个硬件出身的领导人。

If they're wrong on this and, you know, AI keeps living in the cloud, then Apple just picked basically a hardware first leader at at the exact wrong moment.

所以我觉得咱们得盯着看这一步怎么走。

So I think we'll be watching this one closely.

最后我想提一下，亚马逊打算向Anthropic投资250亿美元，搞个超大的云服务协议。

The last thing I wanna bring up is that Amazon is going to put $25,000,000,000 into Anthropic in a massive cloud deal.

他们宣布，这还是在2023年以来已经投了80亿美元的基础上追加的。

They announced that they're investing all of this, on top of the 8,000,000,000 that they've already put in since 2023.

虽然250亿听起来吓人，但其实是个循环投资——因为Anthropic自己也承诺未来十年在AWS上花掉超过1000亿美元。

And while the $25,000,000,000 sounds incredible, it's sort of one of those circular money deals because Anthropic is also committing over a $100,000,000,000 to AWS spend over the next ten years.

这个交易具体怎么落地呢？是这样的：

So here's how the deal basically is gonna be rolled out.

亚马逊现在先砸50亿美元，按Anthropic最新的3800亿美元估值算的。

Amazon is putting in $5,000,000,000 right now, and this is at Anthropic's new $380,000,000,000 valuation.

剩下的200亿，得看商业目标达不达标才给。

The other 20,000,000,000 is tied to commercial milestones.

这事儿跟业绩挂钩，而且Anthropic还能拿到亚马逊自研Tranium芯片超过5吉瓦的算力资源。

So it's gonna be basically performance linked, and part of this deal is that Anthropic gets access to over five gigawatts of capacity on Amazon's custom Tranium chips.

他们计划今年底前，就把将近1吉瓦的Tranium 2和Tranium 3算力上线。

And they're planning to bring nearly one gigawatt of Tranium two and Tranium three capacity online by the end of this year.

一吉瓦差不多相当于一座大型核电站的发电量，而这些公司建的AI数据中心，规模简直跟发电厂一样。

So one gigawatt is roughly the output of a large nuclear reactor, and all of these companies are building AI data centers at, you know, power plant scale.

这事儿啊，确实是个大头。

So this is, you know, this is a really big part.

Anthropic拿的是别人的钱，而那人正好能帮上忙；亚马逊投钱进去，结果那笔钱又流回自己口袋。

Anthropic is taking money from someone who's gonna be able to help them out, and Amazon is investing money into someone who's gonna be giving it back to them.

我觉得这事分两块。

So I think there's two pieces.

第一，整个Cloud平台都会直接集成到AWS里。

First, the full Cloud platform is gonna be available inside of AWS.

如果你本来就是AWS的客户——要知道财富500强里一大半都是——现在你就能用现有账户的账单和安全系统直接用Cloud了。

If you're already an AWS customer and a huge chunk of the Fortune 500 companies are, you can now use Cloud through your existing account billing and security controls.

这简直就是个超强的用户锁定策略。

I think that's a really big distribution lock.

第二，就在两个月前，亚马逊还同意向OpenAI投资最多五千亿美金。

The second thing is that Amazon just two months ago also agreed to put up to $50,000,000,000 into OpenAI.

所以现在亚马逊成了这两个前沿模型的云基础设施，亚马逊手里既有Azure也有OpenAI。

So Amazon is now the cloud backbone for both frontier models, and Amazon has Azure and OpenAI.

谷歌有谷歌云和Gemini。

Google has Google Cloud and Gemini.

亚马逊现在既有AWS加Anthropic，又有AWS加OpenAI。

Amazon now has AWS and Anthropic plus AWS and OpenAI.

所以他们根本不是在押注赢家。

So they basically are not picking the winner.

他们只是把钱给任何能回本、帮他们平台扩张的人。

They're just gonna be giving money to anyone that's gonna be giving it back to them and helping their their platform grow.

我觉得这一招基本就锁定了AI基础设施的格局，这分明是一场真正的淘金热——不是模型，也不是应用，而是底下的算力。

I think this one is basically gonna cement AI infrastructure, and and, basically, it's showing that this is an actual gold rush, not the models, not the apps, but the compute underneath.

现在每个主要实验室都在提前承诺几百亿美金投给云服务商，光是算力本身，已经成了护城河。

Every major lab is now pre committing tens of hundreds of billions of dollars to cloud, cloud providers, and the compute alone is basically becoming a moat.

对吧？

Right?

我甚至听说，OpenAI最新模型的训练算力成本，已经高达数十亿美元了。

And I even saw a a rumor that with the latest model OpenAI's, you know, training, that the amount of compute they're using is, you know, billions and billions of dollars.

这事儿简直太疯狂了。

I mean, this stuff is absolutely wild.

行了。

Alright.

今天的内容就到这里。

That's the show for today.

如果你觉得有收获，麻烦在Apple播客留个评论，或者在Spotify上点个星星，这对我真的特别重要。

If you got something out out of it, it would mean the world to me if you drop a comment on Apple Podcasts or hit these stars over on Spotify.

在Spotify的‘关于’页面可以点星星，在Apple播客也能留言。

It's the about tab on Spotify where you can hit some stars or drop a comment over on Apple.

这能帮我们获得超多曝光。

It helps us show out a ton.

另外，如果你想每月花8.99美元，就能用上80多个AI模型，还能用纯英文指令做自动化，赶紧去试试aibox.ai。

And, also, if you wanna get access to over 80 AI models and an automation builder that runs on plain English for $8.99 a month, go check out aibox.ai.

描述里有链接。

There's a link in the description.

下期节目见，各位！

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