Claude Mythos: Anthropic's Leak That's Too Dangerous to Release

三周前，有传言称一家主要的AI实验室研发出了一款比我们此前见过的任何AI模型都更强大、更危险、更昂贵的模型。

Three weeks ago, rumors broke that a major AI lab had built a model more powerful, more dangerous, and more expensive than any AI model that we had seen before.

我们不知道会是哪家实验室的模型。

We didn't know which model lab it would be.

我们也不知道这个模型叫什么名字。

We didn't know what the model was called.

就在几天前，Anthropic泄露了一个名为Claude Mythos的模型，据称其性能超越了他们以往构建的任何模型，比我们今天看到的Opus 4.6还要高一个层级。

And then just a few days ago, Anthropic leaked a model called Claude Mythos, which is supposedly more powerful than any model that they've ever built before, a tier above Opus 4.6, which is what we see today.

这个模型实际上如此强大，以至于被视为网络安全威胁，目前还不能向公众发布。

This model is actually so good that it is considered a cybersecurity threat and can't be rolled out to the public just yet.

但不仅仅是Anthropic在研发接近AGI的此类模型。

But it's not just Anthropic that's building a model that is close to AGI like this.

OpenAI有一个代号为Spud的模型。

OpenAI has a model codenamed Spud.

谷歌有一个代号为Agent Smith的模型，今年还将有更多类似模型问世。

Google has a model codenamed Agent Smith, and there's many more to come this year.

但Anthropic的泄露并非有意为之。

But the Anthropic leak wasn't intentional.

这是在上星期四，即3月26日，由《财富》杂志的一名记者偶然发现的，他发现Anthropic的内容管理系统存在配置错误。

This was discovered by accident last Thursday, March 26, by a Fortune reporter who discovered that Anthropic's content management system had a configuration error.

对于不太熟悉的人，内容管理系统是网页服务器提供文件的方式。

And for those who aren't familiar, the content management system, it's how the web server serves files.

而在该系统中，有一个配置错误，导致近3000份未发布的资产被泄露到这个公开可搜索的数据库中。

And within that, was a config error that leaked nearly 3,000 unpublished assets sitting in this publicly searchable database.

任何人都能找到这些文件。

Anyone could find them.

于是，两位独立的安全研究人员进行了核查，确认这些文件中包含两篇关于名为Claude Mythos和一个新层级Capybara的模型的博客文章。

So two independent security researchers, they went through, they confirmed, and among these files were two blog posts of two models named Claude Mythos and a new tier named Capybara.

Anthropic在问题曝光后立即撤回了所有访问权限，但随后一名Anthropic发言人确认，这代表了AI性能的重大突破，是我们迄今为止构建的最强大的模型。

Anthropic immediately removed access to all of this as soon as it came out, but then later on, an Anthropic spokesperson confirmed that it represents a step change in AI performance and is the most capable model we've ever built.

因此，他们证实了我们所看到的内容是真实的。

So they confirmed what we're seeing here is real.

现在的问题是，就像屏幕上的这张图片所显示的，我们缺少大量信息。

Now the problem is, like this image suggests on screen, we're missing a lot of information.

这次泄露只是让我们知道这样的东西存在，但我们并不确定具体是什么。

This is a leak that something like this exists, but we don't we're not sure exactly what.

我们所知道的是，有一个新的模型层级，正如你提到的，名为Capybara。

What we do know is that there's a new model tier, Ejaaz, like you mentioned, named Capybara.

它是位于Opus之上的新层级。

It is the new tier that sits above Opus.

所以现在的系列将大致变为Haiku、Sonnet、Opus，而Capybara位于最顶端。

So now the lineup will kind of look like haiku, sonnet, Opus, and then Capybara at the top.

这听起来不太对劲。

It doesn't really sound quite right.

也许那只是一个实验性的名称。

Maybe that's an experimental name.

他们可能会找到更好的名字。

They might find something better.

而Mythos是该层级中的具体模型名称。

And then Mythos is the specific model name within that tier.

你可以把Capybara看作重量级，而Mythos则是具体的选手。

So you could think of Capybara as the weight class and Mythos as, like, the fighter.

它是具体的模型。

It's the specific model.

根据泄露的文档，它在几乎所有方面都大幅超越了Claude Mythos 4.6，尤其是在编程、学术推理和网络安全基准测试上。

Now according to the leaked documents, this dramatically outperforms Claude Mythos 4.6 on basically everything, but particularly coding, academic reasoning, and the cybersecurity benchmarks.

我认为网络安全这一点尤其有趣，因为它在网络安全方面如此强大，以至于他们无法发布它的主要原因是为了防止人们恶意使用。

And I think the cybersecurity one is one of the more interesting points here because it's so powerful in cybersecurity that one of the main reasons why they can't release it is to actually prevent people from using it maliciously.

是这样吗？

Is that right?

是的。

Yeah.

所以，如果我们回溯到一个半月前，Anthropic的AI安全负责人——这位在业内堪称传奇的人物——曾就刚刚发布的Claude Opus 4.6发表过演讲。

So actually, if we rewind to about a month and a half ago, Anthropic's head of AI security, who's actually a legend in the industry, gave a talk about Claude Opus 4.6 when it had just released.

他的演讲描述了这个模型被指向五个到十个非常流行的开源代码库，且没有给予任何指令。

And his talk described how the model was pointed at five to 10 very popular open source code bases with no instructions given.

而这个模型所做的事情非常、非常有趣。

And what the model did was very, very interesting.

它扫描了所有这些代码库，发现了500个重大安全漏洞，而这些漏洞是专家级的人工智能安全研究人员在数十年持续研究和使用这些确切代码库的过程中都没能发现的。

It scanned all those code bases and discovered 500 major security flaws that expert human AI security researchers couldn't discover in decades that they'd been staring and using these exact code bases.

所以Claude在几个小时内完成的事，许多安全研究人员几十年都做不到。

So Claude did in a couple hours what many security researchers couldn't do.

我们说的是数百万小时的计算时间和长时间盯着这些代码库进行测试，而Claude Opus 4.6却成功找到了这些问题。

We're talking about like millions of compute hours and time spent staring at these code bases, testing it, Claude, Opus 4.6 managed to figure this out.

这引发了极大的兴奋，但也带来了许多担忧。

Now, this created a lot of excitement, but also a lot of concern.

因为这些人工智能安全研究人员心地善良，并没有恶意使用它。

Now, because these AI security researchers had a good heart, they weren't using this maliciously.

但你可以想象，如果这个模型落入恶意行为者手中，他们可能会出于多种原因利用这些漏洞。

But if you could imagine that if this model had been placed to say a malicious actor, they could have exploited these for many different reasons.

因此，这些漏洞被曝光并得到了修复。

And so these exploits were surfaced and they were fixed.

但现在的问题是，如果一个更强大的模型被更广泛地提供给任何人或攻击者，比如外国对手，他们可以发现并利用未来的任何漏洞呢？

But the question now becomes what if a more powerful model was made more readily available to anyone or an attacker, for example, a foreign adversary that could discover and exploit any future bugs?

这正是我个人对Claude Mythos或Capybara模型的担忧所在。

That's the concern that's around that I have personally around Claude Mythos or Capybara.

这个模型据称旨在超越我们之前见过的任何模型。

This model is supposedly meant to be a tier above anything that we've ever seen before.

显然，它在发现和利用漏洞方面表现极为出色。

Apparently, it is amazing at discovering and exploiting exploits.

所以，如果它真的比Opus 4.6强两个数量级——我们保守一点说。

So if it is, let's say, two orders of magnitude, let's be conservative.

如果真的强两个数量级，那我们可能真的面临大问题了。

Two orders of magnitude better than Opus 4.6, we could have a real problem on our hands.

因此，Anthropic现在开始逐步向网络安全专家发布这个秘密模型——Mythos和Capybara。

And so what Anthropic has done now is they've started to slow release this secret model, Mythos and Capybara, to cybersecurity experts first.

为什么？

Why?

因为他们希望这些网络安全专家在模型公开发布前，先找出如何加固自己的防御系统，以免被恶意攻击者利用以获取非法利益。

Because they want them to figure out how they can harden their own defense systems before they publicly release this model, and someone, maybe a nefarious attacker, might use it for unachievable gain.

我认为这很讽刺，这家自称构建了具有前所未有网络安全能力的AI公司，却因为有人错误配置了博客而泄露了它。

I think it's ironic that the company building what it describes as an AI with unprecedented cybersecurity capabilities leaked it because someone misconfigured their blog.

这种讽刺太过强烈，你不得不怀疑。

And, the irony there is too strong, and you have to wonder.

你真的得问自己一个问题：如果这个模型如此聪明，以至于它在自我泄露呢？

You have to really ask yourself the question, well, what if this model is so smart that it's leaking itself?

如果它在故意制造漏洞，好让别人秘密发现它呢？

If it's, like, poking holes to, like, let people secretly find it?

我不知道。

I don't know.

但有一点是肯定的：首先，至少目前来说，运行这个模型的成本会极其高昂。

But the the one thing for sure is that, one, this model is going to be incredibly expensive to run currently, at least.

这正是我们现在看不到它的部分原因。

That's part of the reason why we're not seeing it now.

但第二点是，它将变得难以置信地强大，而我们在过去一年取得的进展，与未来三个季度将要取得的进步相比，可能根本算不了什么。

But the second is it's going to be unbelievably powerful, and the progress that we've had in the last year is going to probably look like nothing compared to what we're going to get for the next three quarters.

市场也明显感受到了这一影响，天啊，这些股价图表看起来糟糕透顶。

The market also very much felt the effects of this because, oh my god, these stock charts look absolutely horrendous.

是的。

Yeah.

CrowdStrike，这家主要的网络安全公司，在消息公布后股价下跌了数十亿美元，另一家同样在此领域竞争的公司Palo Alto Networks也受到了影响。

CrowdStrike, which is like the major cybersecurity firm, was down a couple billion on the news and Palo Alto Networks, which is another similar company that competes in this firm also suffered from this.

现在，我正在查看的这两家公司的股价图表，Josh，让我有点创伤后应激反应或似曾相识的感觉，因为我想四周前我们刚讨论过Anthropic发布他们的安全审查功能Claude的时候，那和Mythos无关，但主要是帮助审查你用Claude生成的代码。

Now, these two charts that I'm looking at right now for these specific companies, Josh, gives me a little PTSD or deja vu because we were talking about this, I think four weeks ago when Anthropic released their security review Claude feature, which wasn't anything to do about Mythos, but basically helped review the vibe code that you produced using Claude.

于是，网络安全股再次暴跌。

And so cybersecurity stocks dumped again.

到目前为止，这种情况似乎每个月都在发生。

This is happening seemingly on a monthly basis at this point.

尽管这些图表大幅下跌，但我并不确定市场是否需要立即感到担忧，因为看起来这个即将推出的新模型——这个新的网络安全专家——需要大量的计算资源。

Even though these charts are down quite a bit, I'm not sure how concerned the market needs to be immediately because it appears as if this new model that's coming, this new cybersecurity specialist is really compute intensive.

是的。

Yeah.

以至于如果不进行严重的压缩、迭代，并找到更优化的运行方式，几乎不可能在所有现有账户上运行它。

So much so that it's almost going to be impossible for them to run across all of the accounts currently without some serious compression and iteration and figuring out how to run this more optimally.

而且我们似乎已经开始看到这些成长的阵痛。

And it seems like we're starting to see those growing pains.

对吧？

Right?

就像他们在训练这类模型、在自己的服务器上运行时，已经开始影响普通用户。

It's like, as they're training models like this, as they're running them on their own servers, it's starting to affect the average user.

我知道有时候我早上醒来会感觉我的Opus比前一天笨了一些，而我们实际上有数据支持这一点。

I know sometimes I'll wake up and I'll feel like my Opus is running a little bit dumber than it was the day before, and we actually have data that backs this up.

对。

Yeah.

所以基本上，上周末Claude的服务器彻底瘫痪或严重受损。

So basically over the weekend, Claude servers basically went down or were majorly impaired.

出现了多次不同的服务中断。

There were a bunch of different outages.

人们报告称，与Claude的交互质量大幅下降。

People were reporting very, very reduced quality in their interactions with Claude.

在过去几周里，这种情况一直反复出现。

And this has been kind of like a repeating trend over the last couple of weeks.

现在我们可能找到了原因。

And now we might have the answer why.

通常，大型AI实验室最近一次公开的信息来自OpenAI在2025年进行的一次大型模型训练。

Typically major AI labs, the last public bit of information that we had was from OpenAI's 2025 run of a major model.

他们将30%的可用算力投入到了这次训练中。

They dedicated 30% of their available compute to a training run.

现在有传言称，对于Claude Mythos，他们投入的算力更多，这正是他们取得的重大架构突破。

Now the rumors state that for Claude Mythos, they've dedicated even more, and that's like the major architectural breakthrough that they've made.

如果他们真的这么做了，那可能就是为什么我们消费者无法使用Claude最佳版本的原因——因为他们把算力都用在训练下一代模型上了。

If they've done that, that might be the reason why we aren't being able to use the best version of Claude as consumers because they're too busy using the compute to train the next step or tier in model.

我不知道这是好事还是坏事，但有一件事让我非常清楚：我们需要更多的算力。

I don't know if this is a good or bad thing, but one thing it definitely like screams at me is like, we need a ton more compute.

确实如此。

Big time.

想想看，就在过去三个月里，我们已经取得了多么巨大的进步。

And it's amazing to think about how far we've come just in the last three months leading up to this moment here.

我的意思是，人们真正开始认真对待氛围编程，是从寒假期间开始的。

I mean, when you think about over the winter break is when people really start to take vibe coding seriously.

从那以后，公司生成的代码比例从极低的百分比，几乎上升到了接近100%。

And since then, companies have gone from a very small percentage of code to almost a 100% of code.

这意味着，超过80%的部署代码都是由Claude Co生成的。

I mean, this is saying 80% plus of all code deployed is written by Claude Co.

仅就Anthropic而言，这简直难以置信。

Just for Anthropic, it's unbelievable.

我们从Opus 4.5开始，它于十一月发布，随后在二月推出了Opus 4.6，将我们的上下文窗口从20万token提升到了一百万。

We started with Opus 4.5, which was released in November, and then Opus 4.6 came in February, which took us from a 200,000 token context window to a million.

而现在这个新版本将会极大提升编码能力。

And now whatever this new thing is is going to really drive up the coding capabilities in a really big way.

我认为值得关注一下，在六月之前哪个模型会成为最强模型，哪家公司拥有最好的模型，而得益于Polymarket，我们有一些有趣的统计数据。

And I think it's probably worth checking in on which model is going to be the strongest model, which company has the best model through the June, and thanks to Polymarket, we have some interesting stats on this.

人们押注Anthropic在六月拥有最佳AI模型的概率为66%，这非常惊人。

So the people are betting that Anthropic has a 66% chance of having the best AI model in June, which is huge.

这个数字最近大幅上升了。

And that number has increased very significantly recently.

如果你回看二月，当时谷歌是大热门，拥有近80%或70%的概率成为最佳模型。

If you look just back in February, it was Google who was the heavy favorite with a almost 80% chance or 70% chance of having the best model.

最近这一情况发生了巨大变化，或许是因为这次泄露事件。

That has changed recently in a big way, perhaps because of this leak.

但我也不确定这些数据是否完全最新，可能遗漏了一些信息，因为我们还听说OpenAI和谷歌也计划发布一些非常重要的内容。

But I'm not sure if this is fully up to date, and it may be missing some information because we have some news on OpenAI and Google who are planning to release something really important too.

感谢 Polymarket 赞助节目的这一部分。

And thank you for Polymarket for sponsoring that part of the show.

但让我们聊聊 OpenAI。

But let's talk about OpenAI.

有一个新的代号为 Spud 的模型即将推出，这很可能将是 Mythos 的竞争对手。

There's a new code name Spud model that's coming, and this is probably gonna be the Mythos competitor.

那么这个模型会是什么样子呢？

So what is this looking like?

是的。

Yeah.

这正是问题所在。

That that's the issue.

我们其实并不清楚。

We we don't really know.

这些所有模型，我们都没有具体的规格信息。

All of these models, we don't have the the specs.

我们需要具体的规格参数才能讨论它们。

We need the specs to talk about them.

但目前在最顶尖的那两三个AI实验室中，出现了一些趋势或模式。

But there's a few trends or patterns that are happening amongst the hottest, or should I say top two or three AI labs.

我们看到Anthropic发布了Mythos，这是他们的AGI或准AGI模型，实现了巨大的飞跃。

We've got Anthropic releasing Mythos, which is their AGI or pre AGI model, a massive, massive leap ahead.

OpenAI也在做同样的事情。

OpenAI is working on the same thing.

他们一直在秘密地开发一个更大的模型。

They've been secretively working on a larger model.

这个模型已经换过几个代号。

This has gone through a few different names.

你还记得吧，Josh，到年底的时候，它曾被称为代号Sprout，现在则被称为Spud。

If you remember, Josh, by the end of the year, I think it was referred to as code name Sprout, and now it's referred to as Spud.

所以我不确定这是否意味着它自那以后已经大幅升级了。

So I don't know if that implies that it's grown massively since then.

它在增长。

It's growing.

但这些模型据称参数规模将在1万亿到20万亿之间。

But these models are supposedly meant to be anywhere between 10 to 20,000,000,000,000 parameter models.

作为参考，我们目前所知的最大模型参数规模在1万亿到2万亿之间。

Now, for context, the largest models that we currently look at right now is between one to 2,000,000,000,000.

所以这比现有模型大了几个数量级。

So this is like major order of magnitude larger model.

它们将需要大量计算资源。

They're gonna be compute intensive.

运行它们的成本将极其高昂。

They're gonna be very expensive to serve.

因此我们需要想办法扩展AI基础设施以及其他诸多方面，但OpenAI的模型代号为Spud，旨在与Mythos竞争。

So we need to figure out kind of like how to scale AI infrastructure and a bunch of other things, but OpenAI's model is code name Spud, it's and meant to be the competitor to Mythos.

人们预计它可能会是GPT 5.5，或者更可能是GPT-6。

People are anticipating that it might be something like GPT 5.5 or rather GPT six.

所以，这比我们今天看到的还要高一个层级。

So again, a tier above what we see today.

它在编程、推理以及Anthropic擅长的许多方面都将更加先进。

It's gonna be advanced in coding, reasoning, and a lot of the things Anthropic is as well.

当我看到这一点时，乔什，我个人觉得，这首先是双方试图互相超越的一次巨大努力。

When I look at this, Josh, personally to me, this seems to be one, a massive bid to try and leapfrog each other.

其次，可能是为了在潜在的IPO前提升自己的数据表现。

And number two, maybe try and juice their numbers ahead of a potential IPO.

我不知道你的反应是否和我一样，但当我看到这样的新闻时，这就是我的直觉反应。

I know whether your reaction to this is the same, but that's like my gut reaction when I read news like this.

是的。

Yeah.

很可能两者都有。

It's probably both.

他们想在IPO前提升各项指标，但同时也只是想赢。

They want to juice up things before the IPO, but they also just wanna win.

基于一些直觉，我对这将会是什么样子有一些相当强烈的猜测。

And I have some pretty strong speculations just based on vibes of what this is gonna look like.

我认为我们最近一直在见证OpenAI在专注力上的趋同，特别是他们正在精心聚焦于核心方向。

I think we've been seeing this recent convergence around OpenAI of like, particularly on focus and on really dialing in what they're focused on.

上周他们发布时移除了Sora，这是一次巨大的动作。

And we saw a big move last week when they released when they removed Sora.

他们彻底放弃了Sora。

They totally destroyed Sora.

嗯。

Mhmm.

他们把很多团队整合到了一起。

They moved a lot of the teams together.

他们让首席产品官兼任AGI发布负责人，根据传闻来看，他们似乎正在打造一个超级应用。

They made their chief of product, the chief of, like, AGI release, and it appears as if they're building a mega app based on the rumors.

所以，我很难使用OpenAI的产品，部分原因是因为它们分散在各个地方。

So part of the reason why I have a difficult time using OpenAI's products is they're kinda spread out everywhere.

比如，Sora 应用就是其中一个。

There's like, the Sora app was one.

还有 Codex，然后是他们的浏览器，还有 ChatGPT，以及很多不同的软件。

There's Codex, then there's their browser, then there's ChatGPT, and there's a lot of different software.

他们的模型也是如此，至少以前是这样，有 GPT 5.3 Codex，还有 5.3 高、中、低不同版本。

And the same is true with their models, or it was at least, where there was GPT 5.3 Codex, and there was 5.3 high, mid, low.

这些不同的模型让事情变得非常复杂和混乱。

There's this all these different models that really complicate things and confuse things.

在 5.4 版本中，他们推出了一款单一模型。

With 5.4, they made a singular model.

现在 5.4 可以同时处理编码和推理，全部整合在一起。

Now 5.4 does your coding, and it does the reasoning all in one.

我猜测这个代号为 Spud 的新模型，将是这种专注的巅峰之作，我希望他们能将这款模型与他们的新应用一起发布，采用单一模型。

And what I suspect with this new model, code named Spud, is going to be the kind of pinnacle of this focus, where I'm hoping they release this with their new application with a singular model.

所以，现在有一个无所不知的模型。

So there's one model that is all knowing.

有一个应用程序类似于Anthropic的Claude桌面应用，将所有功能整合在一个平台上，我认为他们可能会以此为契机，全力聚焦于此，而不是将资源分散到众多不同领域。

There's one application similar to what Anthropic does with the Claude desktop app that has all of the functionality under one roof, and I think they're gonna probably use this as a point to really lean into that focus instead of distributing this across a lot of different areas.

我期望这能真正改变OpenAI，甚至比改变Anthropic更显著，因为它真正改变了用户与产品的交互方式，使产品变得更好。

And I'm hopeful that that will meaningfully change OpenAI more so than it'll change Anthropic because it actually changes the way that users interface with the product, and it becomes a much better product.

是的，我认为去年大部分时间里，我对SAM和OpenAI同时关注太多事情感到非常不满。

Yeah, I think for the majority of last year, I was pretty upset with the way that SAM and OpenAI were focusing on so many different things.

我只是觉得，你们就专心打造一个真正优秀的模型吧。

I was just like, just focus on creating a really good model.

你们在编程领域已经被落后了，Anthropic都快把你们的午餐吃光了，赶紧解决这个问题吧。

You're being left behind in coding, Anthropic Seating Your Lunch, like figure this out.

从他们去年十一月左右的红色警报开始，他们就开始重新分配计算资源、资金、数据和所有资源，专注于打造最优秀的通用模型和编程模型。

And then since their code red of like, what was it, November, they've been like reallocating compute, money, data, and all their resources to focus on building the best general model and the best coding model.

所以我们现在开始看到这些努力的成果了。

So we're starting to see the fruits of that labor.

我现在对OpenAI充满信心，相信他们会推出一款优秀的产品，与一直占据优势的Anthropic一较高下。

I have a lot of faith now in OpenAI that they're gonna produce a really good product that will compete with the likes of Anthropic, which have been eating their lunch.

当我回顾过去一周时，似乎对OpenAI来说相当不利。

When I look at like the last week, it seems like it's pretty negative for OpenAI.

你提到他们取消了Sora项目。

You mentioned that they killed Sora.

他们还终止了与迪士尼签署的10亿美元交易，同时关闭了ChatGPT的成人模式以及一系列消费者购物应用和他们的应用市场。

They also killed the $1,000,000,000 deal that they had signed with Disney and they also shut down ChatGPT adult mode and a bunch of like consumer shopping apps and their like app marketplace as well.

他们现在只专注于这几件事。

They're just focused on these few things right now.

但另一件事是，SAM也在某种程度上违约了几项重要的GPU和数据中心交易，对吧？

But then the other thing is SAM is also kind of defaulting on a few of the major GPU and data center deals, right?

我们之前提到的OpenAI与甲骨文的Abilene合作项目就告吹了，原因是多种多样的融资问题。

So we had the OpenAI and Oracle Abilene deal fall through where they couldn't finance it for a variety of different reasons.

另外，他们还未能履行购买全球高达40%内存供应的承诺，因为他们目前还没理清财务状况。

And then the other thing is they're defaulting on purchasing up to 40% of the world's memory supply because they haven't still figured out their finances right now.

所以我认为OpenAI正处于一个类似青春期的阶段，正在摸索如何调整资源分配，但我相信他们最终能挺过来。

So I think that OpenAI is going through kind of like a puberty period where they're figuring their stuff out and where to reallocate resources, but I think they're gonna pull through.

而且这确实看起来是一个重大的突破。

And it also seems like this is indeed a serious breakthrough.

我的意思是，萨姆在一份泄露给员工的内部备忘录中提到，事情进展比我们许多人预期的要快，他还称这是一个非常强大的模型，能够真正推动经济发展。

I mean, Sam, in an internal memo that got leaked out to employees, he said things are moving faster than many of us expected, and he called it a very strong model that can really accelerate the economy.

对于内部那些同样了解情况的员工来说，这样的说法似乎过于夸大了。

That seems like pretty large claims for to make internally with employees who are also kind of in the know and aware of what's going on.

我只是觉得，我们这些身处实验室之外的人，还没有完全理解即将降临我们的进展究竟有多大，嗯。

And I just think that a lot of us who are sitting outside these labs are not entirely wrapping our head around how much progress is actually about to hit us Mhmm.

在未来几个月，随着这些新模型的发布。

Over the next couple of months with these new model releases.

这似乎是阶梯式的飞跃，而OpenAI的一名员工甚至暗示，Spud具备一种与我们此前见过的完全不同的能力。

It seems like they're step function improvements, and one of the employees from OpenAI actually hinted that Spud contains a capability that is very different from what we've seen before.

因此，尽管没有具体细节，但显然有许多重大的全新突破即将到来，值得关注。

So while there aren't specifics, there are clearly a lot of these huge novel breakthroughs incoming, which is worth looking out for.

还有一个最终的模型发布和泄露来自谷歌，他们一直稳步前进，这个模型被称为Agent Smith。

There's one final model release, model leak that we have from Google, who has been doing well, kind of chugging along slowly in the background, and this is called Agent Smith.

这是一个秘密的AI工具。

It's a secret AI tool.

关于这个，你有什么信息吗，Ejaaz？

Do you have any information on this one, Ejaaz?

是的。

Yeah.

据一位谷歌内部人士泄露的报告称。

So there was like a leaked report from an insider at Google.

据三位知情人士透露，谷歌员工正在使用一个名为Agent Smith的新内部工具，可以自动化编程等任务。

Apparently Google employees are using a new internal tool called Agent Smith that can automate tasks such as coding, according to three people that were familiar with it.

这个产品的运作方式是在他们现有的Vibe编程平台——名为Anti Gravity的平台上进行的，但该平台已经大约两个月没有重大升级了，而在AI领域，这简直就像过了一个世纪。

The way that this product is supposed to work is within their Vibe coding platform called Anti Gravity, which exists today, but hasn't really had a major upgrade for, let's say, a couple months now, which is like an eternity in the AI world.

因此，他们正在发布一个名为Agent Smith的新AI模型，该模型将采用多智能体方法，并使用升级版的Gemini 3.1。

So they're releasing a new AI model called Agent Smith that is supposed to take a multi agent approach and use an upgraded version of Gemini 3.1.

所以它可能并不是3.1版本。

So it's probably not gonna be 3.1.

可能是3.5，甚至可能是4。

It might be 3.5 or maybe even four.

再次实现一个数量级的飞跃。

Again, another order of magnitude leap up.

所以我们看到的是，谷歌正在开发一个AI编码模型，试图追赶Anthropic和OpenAI的Codex等竞争对手。

So what we're seeing here is Google working on an AI coding model competitor to try and catch up to Anthropic and the likes of OpenAI's Codex.

你看到OpenAI正试图重新分配资源，专注于打造最好的通用模型，并追赶在编码领域领先的Anthropic。

You've got OpenAI trying to reallocate resources and focus on building the best general model and catch up with Anthropic, which they have at coding.

而Anthropic则试图抵御这两家的攻势，全力投入计算资源，追求下一个数量级的突破，但这却以牺牲现有用户为代价——他们每天新增百万用户，却报告Claude服务器宕机、使用质量下降。

Then you have Anthropic trying to keep these two at bay and make the next order of magnitude up, spending all their compute, but coming at the expense of serving their existing users, which they're adding like a million a day reporting, you know, Claude servers being down and reduced quality of usage.

所以现在这三家公司之间的紧张气氛，我都能真切地感受到。

So this is a very, I can like feel the tension in the air between these three companies right now.

我不知道Mess在做什么。

I don't know what Mess is doing.

我不知道Grok在哪里。

I don't know where Grok is.

我支持他们。

I'm rooting for them.

我希望他们能赶上来，但目前看来，这三大竞争对手正在争夺这场竞赛的胜利。

I hope they catch up, but it seems to be these three major competitors right now that are in the running for winning this race.

他们正在全力推进。

They're firing.

我的意思是，从去年年初到现在这90天里，我们的上下文窗口从20万增加到了100万。

I mean, in the last ninety days since we started this year to now, we went from 200,000 context windows to a million.

我们从简单的代码辅助工具，发展到能够完全独立编写代码的代理，现在这些代理已经能编写谷歌超过四分之一的生产软件，以及Anthropic超过80%的软件。

We went from these coding assistants to compiler writing agents who are completely capable of writing a very small amount to now over a quarter of Google's production software and 80% plus of Anthropic software.

我们这周学到的一切表明，前沿技术的发展速度会越来越快。

Everything we learned this week, the frontier is going to keep moving faster and faster.

所以，我们即将迎来一个疯狂的第二季度、第三季度和第四季度，2026年简直会疯狂到极点。

So we're in for a crazy q two, q three, q four, just a crazy 2026.

随着这一切发生，随着这些公司陆续上市并获得更多融资来大规模部署AI数据中心，事情会迅速变得异常起来。

And as all these things happen, as these IPOs start to happen and they get even more fundraising to deploy these AI data centers at scale, things are really gonna get weird in a hurry.

但我们会一如既往地持续报道。

But we will be here to cover it as always.

如果你喜欢这一集，请别忘了分享给你的朋友。

If you enjoyed this episode, please don't forget to share it with your friends.

在YouTube上点赞。

Like it on YouTube.

别忘了订阅。

Don't forget to subscribe.

如果你在Spotify或RSS等播客平台收听，可以在那里给我们打五星评价。

If you listen on a podcast player like Spotify or RSS, you could rate us five stars there.

我们非常感激。

It's always really appreciated.

Ejaaz，我们在结束今天的内容前还有什么最后的补充吗？

Ejaaz, any final notes before we sign off for the day?

我们在自己的这边简直大获全胜。

We've been absolutely killing it over on our side.

我们这边新增了大量订阅者和听众。

Loads of new subscribers, loads of new listeners.

非常感谢大家加入我们。

Thank you guys so much for for joining us.

而且，是的，我有个请求，因为我们总喜欢在每集结束时布置一些作业。

And, yeah, I have a a request because we always like to give out homework at the end of the episode.

如果你正在收听这期节目，并且你是Anthropic、OpenAI或Google的内部人员，愿意向我们的账号提供匿名爆料，请在xTwitter上注册一个匿名账号并给我们发私信。

If you're listening to this and you are a insider at Anthropic, OpenAI, or Google, and you are willing to give an anonymous tip to our accounts, please spin up an Anon account on xTwitter and DM us.

我非常期待收到你的消息。

I would love to hear from you.

那太好了。

That'd be great.

嗯，感谢大家的观看。

Well, yeah, thank you guys for watching.

我们下期再见。

We'll see you in the next one.