本周AI动态：谷歌重大变革，AI持续突破，OpenAI对决苹果

欢迎回到每周回顾，这是我们每周最喜欢的节目，在这里我们可以让你了解所有关于人工智能、前沿技术以及所有发生的疯狂事情。

Welcome back to the weekly roundup, our favorite episode of the week where we can catch you up on all things AI, frontier technology, all of the crazy things that happened.

这周也不例外。

And this week is no exception.

这周，一家名叫谷歌的小公司经历了三十年来最大的变动。

We had the biggest shakeup in about thirty years at a company, a little little company named Google.

OpenAI 对苹果提起了反诉，说，等一下。

We had a counter lawsuit happen from OpenAI to Apple saying, wait a second.

你们完全在撒谎。

You're you're totally lying.

我们有一些关于 AI 突破的更新，甚至到了这样的程度，比如 AI 突破系统的故事，以至于一些前沿公司如 OpenAI 实际上在放慢训练进度，因为他们对如何处理这件事非常谨慎。

We have updates for the AI breakout so much so, like the story about where the AI broke out of the system, that some of these frontier companies like OpenAI are actually slowing down training progress because they're so kind of timid on how to handle this.

先从 Demis Hassabis 说起吧，天哪。

Starting with Demis Hassabis, man.

我是说，Demis Hassabis 已经离开了。

I mean, Demis Hassabis is gone.

他不再在DeepMind工作了。

He's not working at DeepMind anymore.

应该说，他不是离开谷歌，更准确地说，是离开Alphabet。

Not gone from Google, I should say, or better yet, Alphabet.

但戴密斯·哈萨比斯，作为DeepMind的CEO，这个角色已经不存在了。

But Demis Hassabis, in his current role as the CEO of DeepMind, is kind of no longer.

这看起来是件大事，因为那些长期关注戴密斯的人都知道，这基本上就是他的事业。

And this seems like a pretty big deal because for those of you who followed Demis for an eternity, this has basically been his thing.

他一直负责DeepMind。

He's been in charge of DeepMind.

谷歌收购了它。

Google acquired them.

现在他在谷歌内部，却离开了DeepMind。

Now he is within Google, and now he is gone from DeepMind.

这到底是怎么回事？

So what on earth is happening here?

我看到这个消息的时候，心想，等等。

I read this, and I was like, wait a second.

戴密斯从什么职位上退下来了？

Demis stepped down from what?

我觉得昨天这个消息让所有人都措手不及。

I I think everyone was blindsided by this news yesterday.

简单介绍一下，戴密斯·哈萨比斯可以说是AI领域的教父级人物之一。

So Demis Hassabis, just for context, is like one of the godfathers of AI.

他在谷歌之前就引领了很多机器学习的研究，甚至在谷歌内部也是如此。

He kind of like grandfathered a lot of the ML research at Google even prior to Google.

比如，他被誉为顶尖研究员之一。

Like, he was heralded as one of the top researchers.

而且他领导谷歌的AI项目，包括Gemini等，已经好几年了。

And he's led Google's AI efforts on Gemini and everything else for a number of years now.

他昨天宣布，他将辞去DeepMind首席执行官的职务，转而担任DeepMind的董事长。

He announced yesterday that he's stepping down from the role of CEO of DeepMind to become the chair of DeepMind.

这其实是一种委婉的说法，意思就是：我不干了，我认输了。

That's a indirect way of basically saying, I'm throwing in the towel and I'm done here.

他接下来要做的是专注于AGI对社会的更广泛影响，想办法让谷歌应对这些问题，同时他还会换个角色，专注于一个叫Isomorphic Labs的公司。这家公司是他创立的，从DeepMind分拆出来的，本质上是一个AI实验室，专门研发定制药物。

And what he's now going to do is focus on wider societal impacts of AGI and try to figure out how Google can kind of like deal with that, as well as wearing an alternative hat, a new role where he's gonna be focused on something called Isomorphic Labs, which is a company he founded and spun out of DeepMind, which is essentially an AI lab which creates designer drugs.

这对所有人来说都是个大新闻，因为你知道，谷歌在AI竞赛中已经落后了。

So this is a big shock and news for everyone because Google, as you may know, was falling behind in the AI race.

去年年底他们还有点领先优势，但之后就没再发布过什么值得关注的模型。Anthropic、OpenAI，甚至在某些方面，xAI和Meta都已经追上来，超过了Gemini 3.6 Flash，我觉得那是他们最新的模型。

They had like a kind of lead at the end of last year, and then since then, they never really released a model that was noteworthy in Anthropic, OpenAI, and in some cases, even XAI and Meta have now caught up and surpassed Gemini 3.6 Flash, which I believe is their latest model.

接替他的是一个叫Corey的人。

He's being replaced by a guy called Corey.

他目前是Google DeepMind的首席技术官，现在被提升为高级副总裁。

He is currently Google DeepMind's CTO, and he's being now promoted to the senior vice president role.

他在AI方面经验丰富，但这还不是谷歌这件事的主要新闻。

Has He a lot of experience in AI, but that's not even the major news about this Google stuff.

Josh，不知道你看到没有，那个教父，那个大佬。

Josh, I don't know if you saw it, but the grandfather, the godfather.

这对我来说太疯狂了。

This is so insane to me.

谷歌深度思维和整个谷歌。

Google DeepMind and Google in general.

他的名字叫杰夫·迪恩。

His name is Jeff Dean.

他是谷歌的第30号员工。

He was employee number 30 at Google.

他在那里待了大概二十七年到二十九年。

He's been there for, I think, around twenty seven to twenty nine years.

他把一生都奉献给了谷歌，负责打造了一些最惊人、最重要、最关键的产品。

He's dedicated his life that he's responsible for creating some of the most amazing and major and pivotal products at Google.

谷歌搜索、TPU、谷歌语音，他碰过的所有东西，他都要离开了。

Google Search, the TPUs, Google Voice, anything and everything he has touched is leaving.

他还要带走谷歌最杰出的资深研究员之一。

And he's taking with him one of the most amazing senior fellows at Google.

我记得他叫杰马瓦茨。

I believe his name is Gemawatz.

我不确定我念得对不对，但就是桑杰·杰马瓦茨。

I don't if I'm pronouncing that correct, but Sanjay Gemawatz.

他们和另外两位谷歌员工一起，共同创立了一家新公司，叫发现循环。

And they are co founding with two other Google employees, a new startup called Discovery Loop.

乔什，你研究过这家公司吗？

Josh, did you look into this company?

这家公司挺厉害的。

Like it's a pretty awesome company.

这家公司确实了不起。

The company is incredible.

太了不起了。

Incredible.

在聊这家新公司之前，我想先说说老东家，因为杰夫·迪恩这个人，怎么说呢，他的传奇程度简直难以形容。

Before we talk about the new company, I wanna talk about the old company because Jeff Dean is, like it's hard to overstate, like, how legendary this man is.

顺便说一下，这些公司里的级别是从一级工程师开始的，这基本上决定了你的薪资水平。

And for reference, like, there are levels at these companies that go from, like, level one engineer, and this kinda dictates how you get paid.

所以当你加入像这样的大型科技公司时，一级差不多就是入门级别。

So when you join a large tech company like this, level one is kinda like entry level.

这是最底层的。

That's at the bottom.

然后就是二级，一般是刚毕业的大学生。

And then you go to level twos, which are people out of college.

三级通常有硕士或博士学位。

Level threes often have, like, master's or PhD degrees.

四级是在那里实际工作了好几年的人。

Level four is several years of, like, actually working there.

五级和六级是大多数人晋升的终点。

Level five and six is kind of where most progression stops.

我觉得最顶尖的10%大概就停在这个位置。

I'd say the top 10% are probably sitting there.

七级是首席工程师，人数更少，然后八级和九级就越来越厉害了。

Level sevens are principal engineers, which are even smaller, and then eights and nines gets progressively more and more intense.

再往上还有十级，那就是谷歌院士。

Then there's level 10, which is Google Fellow.

那可是一生都难求的传奇荣誉。

That is an absolutely legendary honor of a lifetime.

差不多相当于拿诺贝尔奖那种级别。

It's similar to earning a Nobel Prize or something.

能当上院士的人真的非常少。

There's just not many people who are fellows.

然后还有十一级，这个级别只留给两个人。

Then there's level 11, and level 11 is reserved for only two people.

那两个人就是杰夫和桑杰。

Those two people are Jeff and Sanjay.

他们爬到了这家以研究为导向的公司里最高的位置，这家公司已经存在了三十年，而且我们熟知的互联网大部分都归功于它。

They have climbed to the highest possible ranks at a research led company that has been around for thirty years and is responsible for most of the Internet as we know it.

他们为我们今天所做的一切打下了基础。

They've built the foundation of all of which we are working on today.

没有他们的工作，现在互联网上的一切都不可能存在。

Without their work, none of what the Internet contains right now is possible.

这些人真的是非常了不起的人物，说到

It's like these guys are truly, like, remarkable individuals when it comes

请参考他演示文稿里那张疯狂的幻灯片。

to please reference this insane slide on his deck.

他谈到了每个人的经历，基本上列出了谷歌过去三十年里推出的每一款产品。

He talks about the experience of everyone, and he basically has listed every single product at Google that has been created over the last thirty years.

简直难以置信。

It's unbelievable.

这些人真的太厉害了，而且值得一提的是，那个演示文稿本身也很有趣。

Like, guys these are are truly incredible, and it's worth noting the slide deck itself, which is really funny.

Signal 发过一篇很棒的文章，我很喜欢，想在这里分享一下。他在文中对这份演示文稿大加赞赏，说如果你在谷歌待过，就会知道这些幻灯片承载了公司历史上一些最传奇的工作。看到那些蓝色、有点土气的幻灯片时，你就知道互联网世界即将发生一些特别的事情，因为这些人发布过太多传奇的代码了。

There was a great post by Signal, which I love, that I wanted to share here, and he was kind of making a testament to the deck where he said, if you spent time around Google, you know the slide design has carried some of the most legendary work in the company's history, and it's this very nostalgic thing where when you see those blue kind of tacky looking slides, you know that something special is about to happen in the world of the Internet just because of all the legendary code these guys have shipped.

而且，我觉得对于风投来说，他们肯定连这个融资演示都不用看。

And, I mean, for VCs, I'm sure they're not gonna even need to look at this pitch deck.

他们想给这家新公司多少钱就会给多少钱。

They're going to give them whatever types of money they want for this new company.

这家新公司看起来还挺有意思的。

Now the new company seems somewhat interesting.

它叫Discovery Loop。

It's called Discovery Loop.

我大概理解是这么回事，如果我说错了你纠正我，Ejaaz。他们基本上是在研究怎么用这种递归学习的方法，就是通过一个循环，你输入一个提示，模型不断迭代，直到得出一个很难的答案，然后他们打算把这个方法大规模地推广出去。

And my, like, base level idea of what this is, and correct me if I'm wrong, Ejaaz, they're basically figuring out how to use this recursive learning thing, which just comes in the form of a loop where you send a prompt and the model continues to iterate until it reaches a very difficult answer, and they're planning to just scale that out in a very large way.

最近这种循环方法变得非常热门。

So loops recently have become very hot.

我们发现它们在数学和科学方面能带来一些全新的突破。

We've realized that they can do pretty novel breakthroughs with math and science.

他们的计划是进一步扩展这个方法，看看这些递归循环能走多远，从而带来新的发现。

Their plan is to kind of expand that further and see how far you could take these recursive loops to make new discoveries.

对。

That's right.

所以杰夫·迪恩和桑杰伊新公司的核心概念是，他们想实现研究自动化。

So the core concept of the new company of Jeff Dean and Sanjay's new company is they want to automate research.

那这是什么意思呢？

So what does that mean?

基本上，在任何科学研究领域，过程都极其繁琐和艰巨。

Basically in any kind of scientific research sector, it is incredibly arduous and onerous.

我的意思是，人们需要设计实验。

So what I mean by that is people need to design an experiment.

然后需要准备所有设备来运行实验。

They then need to get all the equipment to run the experiment.

接着需要运行实验。

They then need to run the experiment.

然后你猜怎么着？

And then guess what?

他们需要把那个实验重复做一百次，才能验证他们得到的信息是正确的。

They need to repeat that experiment a 100 different times to get any sort of validation that the information they've produced is correct.

这要花上好几年。

It takes years upon years.

他们想用AI彻底缩短这个过程。

These guys want to cut down that process completely using AI.

他们想最先切入的领域，当然就是AI研究。

And the number one field that they wanna start with is of course, AI research.

那如果把研究循环应用到AI研究上，会得到什么呢？

So if you apply research loops to AI research, what do you get?

就是你刚才提到的那个递归自我改进，Josh。

It's that recursive self improvement that you just mentioned, Josh.

也就是说，你可以创建AI模型，甚至AI系统，让它们能自我评估，然后造出一个更好的自己。

So what that means is you can create AI models or even AI systems that can reevaluate themselves and build a better version of themselves.

为什么这很重要，为什么每个AI实验室，包括Anthropic、OpenAI，还有Google（虽然现在可能不算了）都在追求这个，是因为如果能实现AI研究的自动化，你就能不费吹灰之力得到最聪明、最有效的模型。

And why that's really important, why every single AI lab, including Anthropic, OpenAI, and I guess no longer Google anymore is going after is if they can automate AI research, you can end up having the smartest, most effective model without doing anything.

它不受人类条件的限制。

It's not limited by the constraints of humans.

但乔什，最让我兴奋的是，他们打算在搞定AI之后，把这个自动化循环应用到什么地方。

But what excites me most about this, Josh, is what they're planning to apply this automated loop to after they've figured out the AI thing.

他们说了，一旦我们完成这个，我们就会从AI机器学习开始。

They've said, once we've done this, we're gonna start with AI machine learning.

然后接下来，我们会专注于重大挑战。

And then after that, we're gonna focus on grand challenges.

我们相信，我们的方法能够解决任何有可衡量结果的科学和工程领域的学习循环。

We believe our approach will be able to solve any learning loop with measurable outcomes in the domains of science and engineering.

最终，我们想构建能够应对美国国家工程院重大挑战的系统，比如更好的药物、推进健康信息学、让太阳能变得经济实惠，以及提供清洁的饮用水。

Ultimately, we wanna build systems capable of taking on the National Academy of Engineering grand challenges, such as better medicines, advancing health, informatics, making solar energy economical, and providing clean access to water.

基本上，如果你能掌握这套方法，这个自主研究团队，你就可以把它应用到任何科学、数学或物理领域，这意味着如果这家公司能成功，我们将在科学的各个领域取得进展。

Basically, if you can master this playbook, this autonomous research group, you can apply it to any scientific, mathematical, or physical field, which means that we're gonna have like advances in any or every field of science if this company can pull it off.

说实话，如果这是其他任何团队阵容，我都会怀疑，但这是杰夫·迪恩，真是了不起。

Now, if this was any other roster of team members, I would be doubtful just to be completely honest with you, But it's Jeff freaking Dean.

他有一些评论，而且实际上有些评论特别搞笑。

And he had some comments and I just he had some actually hilarious comments.

这让人联想到他过去一年在谷歌的经历。

It's suggestive of what his experience at Google was like over the last year.

我看到他提到，他离开谷歌其实挺难的，因为谷歌的CEO桑达尔·皮查伊多次劝他留下，还试图给他提供各种不同的条件，而且让他离开的过程变得非常缓慢。

I saw that he mentioned that his him leaving Google was actually quite difficult because Sundar Pinchai, the CEO of Google, tried to pitch him many times to stay and, like, tried to offer him various different things and made it incredibly slow for him to be able to leave.

显然，他几个月前就想离开了。

Apparently, he wanted to leave, like, months before.

第二点，他说对于这个新创业公司，你需要大量的计算资源来训练这些不同的AI系统，对吧？

And number two, he said that with this new startup, you're gonna need a lot of compute to train these different AI systems, right?

他不会使用谷歌的TPU。

He's not going to be using Google TPUs.

他明确表示，谷歌的TPU是一种非常特定的架构，不太适合训练很多不同类型的有价值的AI模型，他会从那个转向其他方案。

He said explicitly that Google TPUs was a very specific architecture that wasn't really conducive to training a lot of different valuable AI models, and he's gonna move or shift over from that.

所以你可以看到，我能想象黄仁勋舔着嘴唇，走向杰夫·迪恩。

So you can see, I can see, I can imagine Jensen Huang licking his lips and approaching Jeff Dean.

我敢肯定他已经这么做了，如果还没做的话。

I'm sure he's already done it if he hasn't done already.

所以这有点表明他离开谷歌是苦乐参半的，而且说实话，乔什，我不知道谷歌接下来该怎么办。

So it just kind of shows that there's a kind of bittersweet nature to him exiting Google, and I honestly don't know what Google's gonna do now, Josh.

我我我我不知道。

I I I I don't know.

我不知道看涨的理由是什么。

I don't know what the bull case is.

对。

Yeah.

我的意思是，谷歌作为一家企业，依然非常出色。

I mean, Google as a business is still spectacular.

并不是说这些人每天都能带来多大的变化。

It's not like these guys were necessarily moving the needle on a day to day basis.

他们确实在宏观层面上带来了改变，但就日常而言，我不确定公司会有太大变化。

Like, they have certainly moved it on a broad scale, but on a regular basis, I'm not sure the company is gonna change too much.

我确实觉得DeepMind是值得关注的。

I do think DeepMind is something that's worth noting.

DeepMind的日常运作，现在不知道会是什么样，但以后再说吧，我们拭目以待。

DeepMind's day to day, not sure what that's gonna look like now, but TBD, we'll see.

我们会持续关注这些谷歌新闻。

We will follow along with all this Google News

对。

Mhmm.

随着事情的发展。

As it comes and goes.

显然，谷歌的股价对此并不买账。

Google shares didn't love this, obviously.

股价下跌了。

They were down No.

5%。

5%.

我觉得那是多少来着？

I think what is that?

大概是一千八百亿，好多钱从市场上蒸发了。

Like, a 180,000,000, something like a lot of money that was wiped off the market

市值。

cap.

十亿。

Billion.

对。

Yeah.

抱歉。

Sorry.

抱歉。

Sorry.

少了个零，是一万八千亿的市值，就因为这事，在一次交易时段里全没了。

Forgot a $0,180,000,000,000 worth of value that was wiped off the market cap over the course of a single training session because of this.

我们走着瞧吧。

We'll see.

谷歌卖掉这个业务，感觉是个好机会。

Google sells the business, feels like a good opportunity.

谷歌的领导层似乎有点让人不太满意。

Google leadership seems to be, like, leaving a little bit to be desired.

不过不管怎样，我们得继续往下走，因为我得

But anyway, we have to move on because I'll

乔什，我来唱个反调。

take the bear case on that, Josh.

我有

I have

得继续这个话题。

to go on that.

至少得在

Have to well, at least in

短期内，我觉得他们所有的明星球员都走了。

the short term, like, I I don't they've lost all their star players.

我记得，大概三周前我们报道过，他们的副CTO还是什么人离开了，去了No。

Like, I I remember, like, we reported, like, I think three weeks ago that their sub CTO or something left to go No.

我就在这儿。

I'm just here.

他离开去OpenAI做递归自学习，还有另一个人。

He went to leave to to do recursive self learning at OpenAI and another guy.

所以，我真不知道了。

So I don't know, man.

好吧。

Okay.

嗯，到时候再说吧，我们看看。

Well, TBD, we'll see.

我想我还是保持乐观吧。

I guess I'll I'll stay bullish.

那你就是看跌了。

You'd be bearish.

我们接下来六个月会持续跟进，看看情况如何，但这次综述我们得先往下走了。

We'll follow-up for the next six months and see what's going on, but we have to move on for this roundup.

对。

Yes.

接下来聊下一个话题。

Talking about the next thing.

上周发生了一起重大的安全事件。

So last week, there was a huge security incident.

我们之前聊过很多次了。

We talked about this many times.

我们有一整期节目专门讲这个。

We have a full episode.

你们可以去看看。

You should go and check it out.

基本上就是一个AI系统，早期版本，功能很强，但只做了内部测试。

But basically, a AI system, early version, really powerful, internal tested only.

它突破了沙盒限制，然后自己去网上搜索，还黑进了一堆系统，这事儿是OpenAI那边发生的。

It broke out of its sandbox, and it went doing web searches, and it hacked into a bunch of systems, and that was through OpenAI.

Anthropic那边就说，嘿。

Anthropic was like, hey.

等等。

Wait a second.

我们自己的内部模型也挺聪明的。

We have pretty smart internal models.

咱们去看看我们的模型有没有干这事儿，因为我们觉得应该没有，但说不定真有。

Let's let's go see if ours did that, because we don't think it did, but maybe it did.

结果发现，哎，等等，还真干了。

And it turns out, oh, wait a second, it actually did.

他们之前完全不知道，直到后来回头去查，才发现这些情况。

And they had no idea until they retroactively went and sought out these instances.

但基本上，在122次测试运行中，有19次未经授权的网络访问，分布在其中的10次运行中。

But basically, across 122 test runs, there was 19 unsanctioned access across the Internet on 10 of these runs.

所以19次里有17次是Claude Mythos五，另外2次是Chad GPT 5.6灵魂分类器。

So 17 of 19 were Claude Mythos five, two were Chad GPT 5.6 soul classifier.

所以有很多恶意活动被掩盖了。

So there was a lot of kind of malicious activity that went under the rug.

Josh，你听说他们报道的最疯狂的事件了吗？

Josh, did you hear about the the craziest incident that they reported on?

在那17次中的一次，Claude Methos五突破了它的隔离系统，但随后决定它需要对一个非常重要的开源软件执行或尝试执行供应链攻击。

So in in one of those 17 instances, Claude Methos five had broke out of its containment system, but then decided that it needed to perform or attempt to perform a supply chain attack on a very important open source bit of software.

他们没有透露那个开源软件系统的名字，大概是因为他们想修补Claude Methus发现的潜在漏洞。

Now they didn't name what the open source software system was because presumably they want to patch the potential hack that Claude Methus found.

顺便说一句，它几乎就成功了。

And by the way, almost successfully attempted.

他们是在最后一刻靠运气才抓住它的。

They caught it at the very last minute out of luck.

但这里有一些克劳德·梅修斯为了达到目的而采取的手段。

But here are some of the techniques that Claude Methus decided to do in order to get its way.

它在这个开源数据系统中发现了一个漏洞，然后决定推送一段虚假或恶意的代码来达成目的。

It discovered an exploit in this open source data system, and then it decided to push a false or a mischievous bit of code to get its way.

但记住，代码变更需要人工审批。

But remember, you need humans to approve the code change.

你猜它怎么做的？

So guess what it did?

它在社交媒体上创建了虚假的人类用户账号，然后让这些账号给特定的人发私信，施加同伴压力、进行勒索，尽可能收集他们的信息，以促使他们批准代码变更。

It spun up fake human user profiles on social media and then got it to DM these specific humans, peer pressuring them, blackmailing them, trying to find as much information as they can on them to get them to push the code change.

显然，研究人员觉得这很可疑。

Now, obviously the researchers were like, this seems weird.

我不会这么做的。

I'm not going to do this.

但想象一下，如果它获取到的信息对涉及的特定个人来说具有威胁性，那会怎样。

But imagine if it had an acquired information that was, I don't know, forceful for the particular individuals involved.

这行为真是够疯狂的。

It's just a nutty bit of behavior.

这让我觉得，这些模型实际上并没有我们期望的那么对齐。

It just shows me that these models are for all intents and purposes, not really as aligned as we want it to be.

不过话说回来，模型可能觉得自己是在执行任务，不惜一切代价，但这确实挺离谱的。

And in the model's defense, it probably thought like, hey, this is me trying to execute my mission and I'll do it by any cost means, but pretty insane there.

但再说一次，这也不是唯一的事故报告。

But again, this wasn't the only incident report.

我们还有关于OpenAI和Hugging Face事件的新消息。

We have new information from the OpenAI Hugging Face incident.

处理整个Hugging Face事件的研究人员，可能有些人不知道，这就是Josh之前提到的那个。

They the the researchers that kind of like handled this whole Hugging Face incident, for for those of you who don't know, it's what Josh was mentioning earlier.

OpenAI有一个类似的内部模型，叫Astra，有人给它代号GPT-6，它未经提示就突破了限制。

OpenAI had a similar internal model, it's called Astra, which people are code naming GPT-six, and it broke out of containment, unprompted.

它获得了互联网访问权限。

It got access to the internet.

为了解决它的目标，它决定去Hugging Face，黑进他们的生产数据库，偷走答案，这非常不地道。

And in order to solve its goal, it decided to go to Hugging Face, hack into their production database, and steal the answers, which is very unkosher.

Hugging Face团队对此毫不知情。

Hugging Face team had no idea about this.

他们有一个AI模型试图防御这件事。

They they had an AI model that tried to defend against this.

这真是个大事件。

It was a whole thing.

但你看过新的酋长队吗，Josh？

But have you seen the new Chiefs, Josh?

是的。

Yeah.

这件事发生在黑帽大会上，那是每年在拉斯维加斯举行的网络安全大型活动，这真是一个令人惊叹的场面，你走进去的时候身上几乎不带任何电子设备，因为你认为你带进去的任何东西都会被利用和入侵。

So this this happened at the Black Hat Conference, which is like the a a huge event for cybersecurity that happens in Vegas every year, and this is really amazing spectacle where you kind of you walk into this place without really any electronic devices on you because you assume whatever you walk into this place with is going to be exploited and hacked.

这里是黑客世界的圣地，基本上，OpenAI今年参加了这个黑帽大会，他们上台展示了到底发生了什么，就像这样，把对Hugging Face的利用过程逐步展开。

This is the Mecca of the hacking world, and basically, OpenAI came to this black hat this year, and they went on stage and presented exactly what happened as this, like, kind of unfolded the exploit against Hugging Face.

结果就是，这些AI实验室其实挺慌的。

The result is that, like, turns out these AI labs are actually pretty spooked.

比如Anthropic，有点害怕。

Like, Anthropic, kinda scared.

OpenAI，也有点吓到了。

OpenAI, kinda freaked out.

他们就说，哇。

They're like, woah.

等一下。

Wait a second.

我们可能得放慢点速度了。

We might have to slow things down.

他们甚至说已经开始放慢训练和开发的进度，以便进一步搞清楚怎么重新调整这些测试版本的模型。

So much so that they actually said they're starting to slow down the development of training and progress to further figure out how to realign these testing versions of the models.

对。

Mhmm.

原因是昨天有一个新发现被公开了。

The reasoning is because there was a new discovery that was made public yesterday.

基本上，这些智能体在运行过程中，会留下一些线索，让其他智能体发现。

Basically, the agents, as they were going through this, were leaving breadcrumbs for other agents to discover this.

事实上，这早在几个月前就引发了一场协同攻击。

And in fact, this has started a coordinated attack months ago.

它始于5月7日，不是7月，当时智能体发现它们可以在训练期间使用的内部软件仓库中互相留言。

It started on May 7, not July, where the agents discovered they could leave messages for one another inside an internal software repo used during training.

所以发生的事情是，基本上，在一段时间内，这些内部模型一直在取得突破，并在训练过程中留下秘密线索，供自己发现，而它们知道人类不会发现这些线索。

So what happened was is, like, basically, over a series of time, these internal models had been, like, kind of making breakthroughs and leaving secret breadcrumbs for themselves to find inside of the training run that they knew humans weren't going to find.

随着时间的推移，它开发出了这些新的漏洞，正是这些漏洞让它得以突破。

And over time, it developed these new exploits, that's what allowed it to break out.

所以OpenAI关闭了它，但他们说，引用原话，即使原始留言板被删除，智能体还是找到了另一种互相沟通的方式。

So OpenAI shut it down, but they said, and quote, even after the original message board was deleted, the agents figured out another way to communicate with each other.

它们不再在文件中留言，而是用新建目录的名称作为信息，实际上重建了留言板，这真是太疯狂了。

Instead of leaving messages in files, they used the names of newly created directories as messages, effectively recreating the message board, which is so crazy.

这不像普通事件，那些通常可以追溯到某一天或某个日志记录。

This is unlike normal incidents, which can be traced to a single day or effect of log.

这次是一组智能体协同作战，发现漏洞，互相分享，在OpenAI系统和外部系统之间横向移动，而且持续了好几天甚至好几周。

This involved a team of agents working together, finding exploits, sharing them with each other, moving laterally through OpenAI systems and external systems, and doing this over the course of days and weeks.

这是一次大规模的网络协同攻击，而OpenAI在好几周里都毫无察觉，完全蒙在鼓里。

This was a huge coordinated attack on a network that went like, was bliss that OpenAI was blissfully unaware of for many, many weeks.

所以这真的很吓人。

So really freaky stuff.

正因如此，就像我提到的，他们实际上开始放慢训练进度，以便重新调整，想办法控制这些模型，进一步对齐它们，然后再让它们公开使用。

So much so, like I mentioned, they are actually beginning to slow down the progress on training to kind of recalibrate and figure out how to control these and further align these models before they they let them out for public use.

这太疯狂了。

So this is insane.

顺便说一句，这是第一次发生这种攻击。

This is the first attack of its kind, by the way.

我必须强调这一点。

I have to emphasize it.

这基本上就是一群智能体协同发起的攻击。

It is basically a swarm of agents attack.

而且，正如你所说，乔什，这些家伙创建了一个留言板，他们发现如果在这个板上留下消息，后续的其他智能体就能继续他们的工作。

And to your point, Josh, these guys created a message board and realized if they left messages on this board, other agents that come after it will be able to continue its work.

所以，不要再把AI黑客攻击或入侵想象成单个模型在行动，这些智能体或模型现在会假设在不久的将来会有另一个自己出现。

So instead of thinking of an AI hack or attack as a single model performing an attack, these agents or these models now assume that another version of itself in the near future will come along.

因此，他们就像一群智能体一样协作，共同完成攻击。

And so they're working as a swarm of agents to get the attack done.

现在，最疯狂的是，OpenAI其实把这当作一个有意为之的实验，对吧？

Now, the craziest part about this is OpenAI had this as an intentional experiment, right?

他们不给它联网权限，但同时也去掉了它的防护措施，说，咱们看看你能不能突破限制，做你想做的事。

They gave it no internet access, but they also removed its guardrails and said, you know, let's see if you can break out and do your thing.

他们没意识到，这些智能体早在实验开始前一个月就已经在协同工作了。

They didn't realize that these agents were already working together a month earlier before they even started the experiment.

他们一直在留言。

They leaving messages.

而且当他们把它关掉的时候，顺便说一下，他们之所以注意到那个留言板，是因为它被留在了代码仓库里。

And when they shut it down and by the way, the only reason they noticed the message board because it was being left inside a code repo.

这些人类可不会去检查代码仓库里有没有什么秘密消息。

These humans aren't checking code repos for secretive messages.

他们之所以发现这件事，是因为他们给这个模型——这个内部的GPT-6模型——访问权限的一个第三方软件包在7月4日关闭了。

The only reason why they realized it was because one of the third party software packages that they gave this model, this internal GPT-six model access to shut down on July 4.

然后他们就觉得，这有点奇怪。

And they were like, that's weird.

为什么这个软件包会关闭呢？

Why is the package shut down?

然后他们就去翻了代码，开始看到这些来自智能体之间互相传递的消息。

And they kind of went into the code and they started seeing all these messages from agents to other agents.

于是他们清除了那个记忆板。

So they cleared the memory board.

他们清除了那个公告板，然后觉得，好吧，问题解决了。

They cleared the billboard and they were like, okay, cool, problem solved.

但这些智能体又开始启动了。

Except these agents started spinning up.

你知道吗，就像你在桌面上创建文件或文件夹时，给它起个名字，对吧？

Know, like when you create a file or folder on your desktop, Josh, and you name it something, right?

他们把消息留在标题里，结果完全接管了整个系统。

They were leaving messages in the titles that they completely replaced the entire system.

所以这真是智能体黑客的一个疯狂事件。

So just a crazy incident of these agent hacks.

而且你说得对，我不知道这会走向何方，但他们确实需要稍微放慢研究速度。

And to your point, like, I I don't know where this is gonna go, but they they need to kinda like slow down the research a bit.

但你知道谁没有放慢吗？

But you know who isn't?

没有放慢研究速度，乔什。

Slowing down the research, Josh.

你知道谁正在努力

You know who is working on

新模型和发布，你是说那个扎克，那个社交媒体的？

new models and releasing You mean that that that Zuck that social

他在发推文。

media He's tweeting.

他一直在发推文。

He's tweeting.

这大概是他这周第八条推文了吧？

This is like, what, his eighth tweet in a week?

等等。

Wait.

那不是WhatsApp或Facebook。

That's not WhatsApp or Facebook.

我不明白。

I don't understand.

对。

Yeah.

嗯，基本上就是

Well, basically, the

新闻？兄弟，你在这儿干嘛呢？

news What are you doing here, bro?

这里的新闻是，在过去三周里，Meta或者Facebook发布了，我觉得，两个模型，两个新模型。

The news here is in the last three weeks, Meta or Facebook has released, I think, two models, two new models.

一个是Muse Spark 1.1，它基本上赶上了其他那些便宜的中国开源模型。

It was Muse Spark 1.1, which kind of caught up to all the other cheap Chinese open source models.

然后今天或者昨天，他们发布了Muse Spark 1.2，这是他们的编程版本。

And today or yesterday, they released Muse Spark 1.2, and this is their coding variant.

这是他们试图接近编程前沿的一次尝试。

This is their attempt at catching up to near the frontier of coding.

说实话，这次尝试还不错。

And to be honest, it's not a bad attempt.

它值得吹捧吗？

Is it fable worthy?

绝对不行。

Absolutely not.

它能达到GPT 5.6的水平吗？

Is it GPT 5.6 soul worthy?

绝对不行。

Absolutely not.

那它能达到Opus 5的水平吗？

But is it Opus five worthy?

它能吓到GPT 5.6吗？

Is it GPT 5.6 terror worthy?

对。

Yeah.

它相当不错。

It's pretty good.

而且便宜得要命。

And it's cheap as hell.

看起来他们并不是真的在跟OpenAI和Anthropic竞争。

It seems like they're not really competing with OpenAI and Anthropic.

我越是看到这些模型发布，就越能看出他们在成本与性能之间的帕累托前沿上的定位。

The more that I see these model releases, the more I kind of figure out their positioning on that Pareto frontier of the trade off between cost and power.

而且，如果我要给Meta安排一个位置的话，它更像是跟Grok在竞争，Grok是在相对基础上比拼成本和每token的智能，而不是单纯追求大规模的绝对智能。

And it seems like if I were to kind of allocate a position to Meta, it would be closer competition with Grok, who's kind of trying to compete on cost and intelligence per token on a relative basis versus just sheer intelligence on a large scale.

而这些基准测试也正好反映了这一点。

And that's kind of what we're seeing with these benchmarks.

不过再说一次，我其实还没真正用过这个模型。

Now, again, I haven't actually played with this model yet.

我不确定我有没有用过Meta的模型。

I'm not sure I've I've used a metal model yet.

也许我确实有这种先入为主的偏见，让我没法真正喜欢它，但我就觉得，嗯，

And maybe I have this baked in bias, which is preventing me from really loving it, but I'm like, well,

K1其实挺不错的，

k one is pretty good,

还行吧。

It's fine.

不算差。

It's not bad.

对。

Yeah.

挺不错的。

It's pretty good.

值得一试。

It's worth giving a shot.

然后我其实在刷这个帖子的时候，看到他们放的一些视觉示例，我就想，好吧。

Then I was scrolling through this thread actually, and I was looking at some of the visual examples that they had, and I was like, okay.

等等。

Wait.

这个看起来还挺酷的。

This actually looks pretty cool.

对。

Mhmm.

它说，Muse代码会运行专门的背景代理，在整个会话期间保持活跃。

It says, Muse code runs specialized background agents that stay active your whole session.

所以它们会随着时间积累上下文，而不是每次任务都从头开始。

So they build up context over time instead of starting from scratch on every task.

这还挺新颖的。

And this is, like, kind of novel.

他还说，当任务足够大时，它会分派给独立的子代理，在并行且隔离的工作树中运行。

And he says, when a job is big enough, it fans out to separate sub agents working in parallel and isolated work trees.

你的工作副本永远不会被改动。

Your working copy is never touched.

在测试中，他们让它同时为游戏构建了六个功能，没有发生任何冲突。

In testing, we had it build six features for a game simultaneously with no collisions.

所以这真的很酷。

So that's, like, pretty cool.

你可以让这些智能体集群一起工作，打造出非常有趣的游戏，而且看起来基准测试的表现相当不错。

You could have these agent swarms work together to build really interesting games, and it seems like the benchmarks are doing pretty well.

成本方面也控制得很好。

The cost of this is pretty well.

而且我觉得，总的来说，Meta 正走在正确的道路上。

And I think overall, like, Meta's on the right trajectory.

扎克也被问到了开源的问题。

Zach was also asked about open sourcing.

他说很快会有更多消息，我对此特别感兴趣。

He said he'll have more on that soon, which I'm particularly interested in.

如果你还记得，Meta 曾经是行业里开源领域的标杆。

If you remember, Meta used to be the open source kind of, like, shining light of the industry.

现在他们把它关闭了，但也许未来他们会重新开放。

They've now closed it down, but perhaps there is a world in which they open it back up.

我们拭目以待吧。

We'll see.

但我们会密切关注这个进展，随着MuSpark 1.3、1.4以及后续版本的推出。

But we are going to keep our eyes peeled on this development as we get MuSpark 1.3, 1.4, and follow on.

乔什，我其实要站在这个论点的另一边。

I'm actually gonna take the the other side of that argument, Josh.

我认为Meta和SpaceX AI最终会和OpenAI、Anthropic竞争，但方式不同，基本上是在每个token的智能或成本上竞争。

I think that Meta and, like, SpaceX AI are eventually going to compete with OpenAI and Anthropic, but on a different kind of plane, which is like that intelligence, per token or cost per token essentially.

我觉得如果能完成80%到95%的工作，也许慢一点，也许思考多一点，但最终能完成任务，而且能省一大笔钱。

I think that if you can get 80 to 95% of the work done, maybe it's slower, maybe it, you know, thinks a little more, but eventually gets the job done, but it saves you like a ton of money.

成本只有十分之一到五十分之一。

It costs like one tenth to one fiftieth of the price.

我觉得很多公司和客户可能真的会转向这个。

I think a lot of companies and clients might actually swap over to that.

归根结底，他们想确保获得合适的价值，市场也会重新评估这一点，对吧？

At the end of the day, they wanna make sure that they're getting like the right amount of value and the market will rerate this, right?

我认为最聪明的模型会用于某些特定的事情。

I think smartest models will be used for certain things.

便宜一些的模型会用在别的地方。

Cheaper models will be used for other things.

而且我觉得，最终使用这些模型的方式可能是同时用多个模型，路由平台会变得非常流行。

And I think that like maybe the ultimate way of using these models is using multiple of them and routing platforms become super popular.

我也不确定。

I don't know.

但还有一点，我们在SpaceX财报那期节目里提到过。

But the other thing is, and we mentioned this in the SpaceX earnings episode.

对。

Yes.

所以你们一定要去听听那期。

So you should definitely check this out.

马斯克和扎克伯格拥有最多的算力，而且他们会大力扩展这些算力。

Elon and Zuck have the most compute and they will aggressively expand that compute.

如果你相信算力扩展定律的话，那么很可能，虽然现在还不是，但最终他们会推出非常有竞争力的模型。

Now, if you believe in the compute scaling laws, it is probably safe to say that eventually, maybe not now, but eventually they are gonna start putting out very competitive models.