Anthropic研究人员警告的AI风险

欢迎收听本期播客。

Welcome to the podcast.

今天科技和AI领域发生了好多大事。

There is so much happening in tech and AI news today.

我会尽量挑最有趣的内容讲给你们听。

I'll try to get you the most interesting stuff.

我一直在刷X上的评论，看到不少好笑的梗，还有对各种事情的看法。

I've been scrolling through the comments on X, finding some funny memes, and takes on a lot of what's going on.

我想先聊的这件事是刚发生的，苹果刚刚发布了全新的iPhone Duo，就是一款折叠屏iPhone。

The thing I wanna kick it off with because it just happened is that Apple has just released their brand new iPhone Duo, which is like a folding iPhone.

里面会有一些很酷的AI功能，我觉得很多人都会感兴趣。

And there's some cool AI features that are gonna be in there that I think, a lot of people are are gonna be interested in.

其中一项功能，他们会更广泛地推出，但会随iPhone 18一起发布，就是苹果所谓的参考图像功能，这主要是为了帮助打击AI伪造内容。

One of the things, and they'll be rolling this out more broadly, but this is gonna be with the iPhone 18, is that Apple has something called reference image features, and this is to help basically fight AI fakes.

这些我都会在今天播客里详细聊聊。

I'll get into all of that on the podcast today.

OpenAI 给新的 ChatGPT 图像 2.5 加了一个叫 Sketch 2.5 的功能。

OpenAI has added something called Sketch 2.5 to the new ChatGPT image 2.5.

他们推出了一个新的图像模型，里面有个很酷的功能叫 Sketch。

So they have a new image model that came out, they have a cool feature called Sketch.

我今天试了一下。

I was testing it out today.

OpenAI 还声称他们解决了纳维-斯托克斯方程。

OpenAI also claims that they have solved Navier Stokes.

你知道，这个数学问题有一百万美元的奖金。

This is, you know, there's, like, a million dollar prize for this math problem.

这里面有很多戏剧性的事情。

There's a lot of drama in here.

我都不想全讲，但我觉得挺搞笑的。

I don't even wanna get into all of it, but it's kind of hilarious to me.

Anthropic 有个叫 Jacob Coxton 的研究员辞职了，今天早上在 X 上火了。

Anthropic had a researcher named Jacob Coxton who quit, went viral on x this morning.

我看到他那篇帖子有超过40万个赞，他说他离开Anthropic是因为他觉得那个AI模型会杀死世界上所有人，或者有可能毁灭全人类。

I saw he had over 400,000 likes on his post saying he quit Anthropic because, he believed that the AI model was going to kill everyone in the world or had the potential to kill all of humanity.

所以这件事，至少可以说，里面的戏剧性更足了。

So this, to say the least, had even more drama in it.

另外，Suno也刚刚发布了他们的v0.6版本模型。

We also have Suno that just released v point six, their model.

他们的模型全部是用有授权的音乐训练的，因为他们面临太多版权诉讼了。

It was all trained on licensed music because they have so many copyright suits.

这里也有不少戏。

More drama there.

各位，我不是在开玩笑。

Guys, I'm not joking.

今天对科技圈来说真是疯狂的一天。

This is just a wild day for tech.

还有，有一大群黑客正在耗尽Claude订阅用户的token额度，因为他们基本上拿到了你的密码，然后把你账户里的所有token都耗光了。

We also have, a bunch of hackers that are draining Claude subscribers token allowances because they're basically getting your password and they're draining all of your tokens from your account.

另外，Meta 也推出了 Muse，这是一个以安全为首要卖点的个人 AI 助手。

We also have Meta that has launched Muse, which is a personal AI agent built on a security first pitch.

我今天就注册了。

I signed up for it today.

而且，说实话。

And, I'm not gonna lie.

我对 Meta 还挺期待的。

I'm kind of excited for Meta.

我觉得他们正在推出一些有趣的新技术，我对这些都挺兴奋的。

I think that they're they're pulling out some interesting new technology, and I'm excited for all of that.

我还想提一下，如果你还没用过，想访问全球 90 多个顶级 AI 模型的话，可以去看看 aibox.ai。

I also wanted to mention, if you haven't already, if you'd like to get access to over 90 of the top AI models in the world, you need to go check out aibox.ai.

那是我自己的创业公司。

That's my own startup.

我们有一个 MCP 连接器。

We have an MCP connector.

所以基本上，这个连接器能让世界上每一个AI模型在任何其他AI模型里运行。

So, basically, it's a connector that lets every single AI model in the world work inside of any other AI model.

这些AI公司彼此之间是出了名的互相看不顺眼。

All of these AI companies notoriously hate each other.

比如山姆·奥特曼和达里奥，在印度那个大型科技峰会上，大家都手拉手鞠躬，因为那是大型科技峰会，而且是在别人的国家，但他们就是不肯握手。

You have Sam Altman and Dario, you know, at the big summit in India, and everyone's holding hands and bowing because it's a big tech summit and they're you know, another person's country, and they refuse to hold hands.

这些人互相恨得要死。

These people hate each other.

埃隆和山姆·奥特曼，至少可以说，也是互相看不顺眼。

Elon and Sam Altman, to say the least, also hate each other.

所以他们永远不会合作。

So if you they're never gonna work with each other.

他们永远不会让彼此的模型互相兼容。

They're never gonna let their models work with each other.

但如果你想所有不同的AI模型一起工作，就去看看AI Box的MCP，它让你知道，我们有ChatGPT刚推出的全新图像模型。

But if you wanna have all of the different AI models working together, go check out the AI box MCP that lets you know, we have the brand new image model coming out of ChatGPT.

你就可以用那个在Claude里面生成图片了。

You're going to be able to use that to generate images inside of Claude.

或者如果你用Eleven Labs来生成音频，你也可以在Claude里面搞定。

Or if you have something like Eleven Labs that you use for generating audio, you can generate all of that inside of Claude.

特别是像Claude这种模型，虽然不错，但它不擅长图像、音频、视频生成，连OpenAI都在砍掉他们的Sora视频模型。

Especially for something like Claude, is a pretty good model, but it doesn't do any of the image audio video generations or even OpenAI that's killing their Sora video model.

如果你想在OpenAI里生成视频，你可以去用Google的Veo 3来生成。

If you wanna generate videos inside of OpenAI work, you can go and get Google v o three to generate them.

所以，反正所有模型都能一起工作。

So, anyways, all the models work together.

而且每个月不到9美元就能开始用，我们还有好几个套餐，能让你接触到很多很酷的东西，链接在描述里，aibox.ai。

It's only it's less than $9 a month to get started, and we have a bunch of tiers where you can get access to a ton of really cool stuff, aibox.ai linked in the description.

好。

K.

我们先聊聊苹果，因为我觉得他们推出的这个工具挺有意思的，是用来对抗假货的。

Let's talk about Apple first because I think this is, an interesting tool that they've rolled out to kind of fight against fakes.

苹果刚刚发布了一个叫参考图像的东西。

Apple just announced something called reference images.

这个跟他们的iPhone 18 Pro发布有关。

They have kind of this is tied to their iPhone 18 pro launch.

基本上，它会捕获带签名的传感器数据。

Basically, what it's gonna do is it is going to capture signed sensor data.

也就是说，你拍照的那一刻，它就会捕捉这些数据。

So the moment that you take a photo, it's gonna capture that.

然后通过私有云计算，也就是苹果的计算系统，生成一张不可篡改的图像。

It's gonna route it through private cloud compute, so Apple's compute, and it's gonna produce an unalterable image.

苹果采用了Synth ID标准，帮助用户识别AI生成或修改过的图像。

Apple is adopting the Synth ID standard to help users detect AI generated or AI altered images.

我觉得OpenAI和谷歌也在跟这个标准合作。

This is something that I I that I think OpenAI and Google have also been working with.

iPhone 18 Pro将配备8400万像素的融合主摄像头，采用可变光圈，由六片激光切割的叶片组成，每片比头发丝还薄。

The iPhone 18 Pro is gonna have an 84 megapixel fusion main camera featuring a variable aperture built from six laser cut blades, each thinner than a human hair.

所以苹果，今天我们得好好聊聊这些八卦。

So Apple, you'll get we're gonna get into all the drama today.

比如，Anthropic的研究员说它会毁灭全人类。

There's, like, Anthropic's researcher says it's gonna kill all of humanity.

OpenAI好像在偷所有人的数学研究。

OpenAI apparently is stealing everyone's math research.

然后苹果刚刚说，嘿。

And we have Apple that just said, hey.

看。

Look.

我们做了个更好的相机，而且更难被深度伪造。

We made a better camera, and it's gonna be harder to deepfake.

所以苹果现在做得挺不错的。

So Apple, doing pretty good right now.

参考图像也会从照片应用内部开始使用。

Reference image is also gonna start inside of the inside of the photos app.

还会有面向第三方开发者的API推出。

There's gonna be an API coming for third party developers as well.

所以其他人也能用上这个功能，这会很有意思。

So other people are gonna be able to use this as well, which is gonna be interesting.

我们会持续关注这个进展。

So we'll keep up with that.

然后我还有一个表情包要分享，如果你是看视频版的话，因为我们确实会分享这些。

And then I just have one meme to share if you're watching the video version of this because we do share this.

有人发了一个关于新款苹果折叠屏手机——苹果Duo iPhone Duo——的表情包。

Someone, posted a meme of the new Apple folding the Apple Duo iPhone Duo with the folding screen.

基本上，它有一个主屏幕，你可以看到，然后背面还有一个小屏幕。

And, basically, there's a feature on it where it has the main screen that you can see, and then there's a little screen on the backside of it as well.

所以这玩意儿就像是有两个屏幕。

So it's kinda like there's two screens on this thing.

但自拍的方式是这样的：如果你给别人拍照，他们能在手机背面看到自己，然后可以告诉你调整拍照角度。

But the way selfies work with this is if you're taking a picture for someone, they can actually see themselves on the back of the phone, and they could tell you to adjust how you're taking the picture.

有人说，这玩意儿会毁掉情侣关系。

And someone said, this is gonna destroy relationships.

完了。

It's over.

男朋友彻底没救了，被炸飞了，烤焦了，炸透了。

Boyfriends are cooked, nuked from orbit, toasted, fried.

基本上就是因为，你知道，男朋友给女朋友拍照，她能看到前面的自拍屏幕，然后就在那儿指挥他怎么调相机。

Basically because they're you know, the boyfriend's taking a picture of the girlfriend, and she's, like, telling him how to adjust the camera because she can see into, like, the selfie screen on the front.

然后他就一直在调。

And so he's adjusting it.

但就算她告诉他怎么调，他最后还是会把照片拍歪。

And inevitably, he is still gonna take the picture wrong even with her telling him how to adjust it.

所以，反正我觉得这挺搞笑的。

So, anyways, that's pretty funny to me.

现在OpenAI已经推出或者发布了Image 2.5。

So OpenAI has now added or has launched image 2.5.

这个模型真的很厉害。

This is a really impressive model.

我今天一直在玩它。

I've been playing with it today.

除了那个，这里最酷的功能之一就是——

One of the coolest features on here besides okay.

我简单介绍一下，我觉得他们打算直接和Canva竞争。

I'll just like the quick rundown on it is that I think they're gonna start competing directly with Canva.

他们内置了叫做模板的东西。

They have these things called templates built into it.

比如你生成一个logo，它就会弹出提示，问你想不想看看它在T恤、杯子、帽子上是什么样子。

So it's like if you generate a logo, it's like you know, it has these pop ups where it's like, do you wanna do you want to see what this looks like on a t shirt, on a mug, on a hat?

有各种各样的模板、风格、海报，还有你能生成的各种东西，它会给你灵感。

There's, like, all of these different templates and all these different styles and all these different posters and things you can generate, it gives you ideas.

我觉得他们这是要直接跟Canva对着干。

I think they're going straight after Canva with all of this.

好的。

Okay.

不过，他们推出的一个很酷的功能叫 Sketch，我今天试了一下。

One of the cool features they rolled out, though, was called Sketch, and I was playing with it today.

基本上，你只需要输入 @ 符号，然后说 @Sketch。

Essentially, what you do is you you say just the at symbol, say at Sketch.

一旦这样做了，就会出现一个小画板。

And once that happens, there is a little draw pad that comes up.

对我来说，我就去随便涂鸦了一个小人，拿着麦克风，坐在红色椅子上，背景是紫色的，然后我说，你知道的，根据这个生成一张图片。

For me, I went and, like, kind of doodled a little guy holding a microphone in a red chair with a purple background, and I said, you know, generate image of, you know, this.

我大致描述了一下我图片里的内容，然后它就生成了一张图片，和我要求的完全一样。

I kind of explained what was in my image, And it went ahead and generated an image which was exactly like I had, you know, requested.

太完美了。

It's perfect.

我喜欢这个功能的一点是，有时候你在写提示词的时候，我承认，我确实喜欢用参考图片。

The thing that I like about this is sometimes when you're doing a prompt, I I do I will admit, I do like taking a reference image.

所以我就会截一张我喜欢的图片，然后说，像这样生成一张图，但要有那个和这个。

So I'll screenshot an image I like and say, like, make an image like this, but it has that and this.

然后我会告诉它，哪些地方需要改。

And I tell it, like, some of the some of the things that should be changed on it.

这真的很酷，因为你直接画个草图就行。

This is really cool because you can just doodle it.

如果你没有截图，但脑子里有个想法，你也可以这么做。

If you don't have a screenshot, but it's just a idea in your mind, you can do that.

而且OpenAI确实给出了一些很好的建议和例子，说明这个功能什么时候特别有用。

And OpenAI actually had some really good recommendations and examples of, like, when this become really useful.

他们展示的一个例子是，比如你画个后花园的草图，或者你可能在想，我想重新弄一下我的花园。

One thing that they showed is, like, if you just doodled, like, a picture of the garden in your backyard and and or, you know, maybe you're like, I'm trying to, like, redo my garden.

然后你就画个草图，他们有个例子是，你画了个泳池，然后把你想要放在花园里的所有东西都画出来。

And so you do, like, a doodle where you they had one where it was, you doodled the pool and you kind of doodled everything that you wanted to to be inside of your garden.

然后它就能生成，首先，它生成了那个花园的俯视图。

And then it was able to generate first of all, it generated, like, a top down view of that specific garden.

所以就像，你看。

So it's like, look.

这就是你的花园可能的样子，有游泳池、你的遮阳伞，还有你所有的东西。

This is what your garden could look like with, like, a pool and your umbrellas and all your stuff.

然后它实际上还生成了，比如，一张超逼真的照片，就像，你知道，从站在那里的某个人的视角。

And then it actually went and generated, like, a ultra realistic photo from, like, you know, just the perspective of someone standing there.

所以，总之，如果你有个想法，你把它画出来，它就能帮你生成各种示意图，各种角度的照片，各种方式。

So, anyways, if you had an idea and you doodled it out, it could help you with all sorts of schematics and all sorts of angles of this photo generated all sorts of ways.

我喜欢这个功能。

I love this feature.

我觉得这非常酷。

I think this is very cool.

我觉得OpenAI的图像生成器就是最让人上瘾的功能之一。

I think it's like, OpenAI's image generator is one of just the most sticky features.

这就是我为什么一直没离开OpenAI。

It's why I never left OpenAI.

我用了Claude大概六个月，重度使用Claude Code和Claude Cowork，但我一直没取消OpenAI的订阅，因为我总是需要生成图片。

I went to Claude for, like, six months heavy with Claude Code and with Claude Cowork, And I never was able to get rid of my subscription to OpenAI because I always needed image generation done.

所以我觉得他们在这方面投入，添加新功能和新模型，做得非常出色。

And so I think by investing in it, adding new features and new models, they're doing an incredible job.

真的，不是吹的。

Like, no hype.

在我看来，OpenAI的图像生成器是最好的。

OpenAI's image generator, in my opinion, is the best.

我从来没——抱歉，谷歌。

I've never I'm sorry, Google.

我对Nano Banana一直不太感冒。

I've never been very impressed with Nano Banana.

我觉得它被过度炒作了。

I think there's a lot of hype around it.

在我看来，它永远比不上OpenAI。

It'd never be OpenAI in my opinion.

所以现在他们有了这个新模型，他们就在这方面加倍投入，你知道的，这对他们来说是好事。

So now that they have this new model, they're just kind of doubling down there, and, you know, good on them.

干得漂亮。

Good work.

好的。

Okay.

Anthropic的研究员雅各布·科克斯顿辞职了，他警告说实验室们，引用他的话，是在“拿我们的生命赌博”。

Anthropic researcher named Jacob Coxton has quit, and he warns that labs are, quote, gambling with our lives.

科克斯顿在OpenAI和Anthropic花了三年时间做预训练，然后他说，私下里开发这项技术的人相信，到2030年它可能会杀死所有人。

So Coxton spent three years on pre training at OpenAI and Anthropic, and then he said that people building the tech privately believe it could kill everyone by 2030.

所以，你知道的，倒计时开始了。

So, you know, countdown.

在AI接管之前，我们只剩下不到四年的寿命了。

We have less than four years left to live before AI takes over.

关于这篇特别的帖子，我确实看到了一些非常有趣的阴谋论。

Now I've actually seen some really interesting conspiracy theories about this particular post.

他们所有的核心概念基本上都是一样的，就是这样。

The main concept for all of them is basically the same, and that's it.

这是一个相当复杂的监管捕获手段，基本上就是，或者说是Anthropic或者其他人在花钱请人做这件事。

This is a pretty complex regulatory capture move where essentially you can or essentially Anthropic or or someone is paying to have.

也许他辞职的时候是真诚的，但他的帖子很快就被放大了。

Maybe he was sincere in his, like, leaving, but it was his post was very quickly amplified.

所以这基本上是一个在X上今天之前零粉丝、零发帖历史的人。

So this is basically someone that has zero followers, zero, you know, like, history posting on x before today.

他一发帖，马上就有很多领导AI监管公司的人，在二十分钟内就转发和转贴他的内容。

And as soon as he posted, he immediately had a whole bunch of people that lead, like, AI regulatory companies retweeting him and reposting within, like, twenty minutes.

所以这帮助帖子爆火，获得了大量关注。

And so this kind of helped the post blow up, get a lot of notoriety.

所以很多人说，这不过是人们为了把AI监管到死而故意放大的。

And so a lot of people are saying that it is, you know, just people amplifying it because they want to, you know, they want to regulate AI into oblivion.

但在相关的阴谋论中，有人提到，在他发帖之前，他去找了《华尔街日报》，让他们写了一篇关于这件事的文章或评论文章。

But in the conspiracy theory around, one of the things that someone said about this was before he ever made the post, he went to The Wall Street Journal and had them write, I believe, like, an article or an op ed about this.

发布之后，他才发了推文。

And after it launched, then he tweeted.

所以这是有协调的，他显然在和媒体合作。

So it's it's, like, coordinated, and he's obviously working with the press.

所以有些人认为这更复杂，因为有点像是所谓的随机。

So some people believe it's more sophisticated because it's kind of a, you know, quote, unquote random.

他说这是他的推文，其实是个长帖。

He said this is his tweet, which was kind of a thread.

但第一句话说的是，认真搞AI的人真的相信，到本十年末，AI可能会杀死我们所有人。

But the first thing said, the people building AI earnestly believe that it could kill all of us by the end of the decade.

这不是营销噱头。

This is not a marketing stunt.

实际上，很多高管和资深研究员在媒体上说话都会措辞谨慎，显得理智，但我私下听到同样的人表达恐惧。

If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible, but I hear the same people express fear privately.

没有其他人类活动构成这种级别的威胁。

No other human activity poses this level of threat.

所以，哦，对。

So and oh, yeah.

我猜他主要的推文是说，我今天从Anthropic辞职了。

And I guess the main tweet he said was, I I resigned from Anthropic today.

我过去三年在OpenAI和Anthropic做预训练研究。

I spent the last three years building or the the last three years doing pretraining research at both OpenAI and Anthropic.

两家公司都没有负责任地行事。

Neither company is acting responsibly.

他们正直接冲向自我改进的超级智能，拿我们的生命赌博。

They are racing straight to self improving superintelligence and gambling with our lives.

更多想法见下文。

More thoughts below.

总之，然后他列了一堆东西，他说，看。

Anyways, then he kind of goes on the list, and he's like, look.

比如，也许Anthropic在做点什么，但他们没有什么，你知道，他们没有真正的递归超级智能自我改进的计划。

Like, maybe Anthropic's sorta doing something, but there's nothing that they're you know, they don't really have a plan for recursive superintelligence self, you know, self improving.

然后大家都在说，Anthropic的公关策略是有史以来最糟糕的，但也许这其实是天才之举。

And then everyone's just saying that Anthropic has the absolute worst PR strategy of all time, but maybe it's maybe it's a genius.

我也不确定。

I don't know.

然后Evan Hubinger，他现在还在做对齐科学，他是Anthropic的对齐科学负责人，他回应了Jacob的帖子，说Jacob说得对，还加了个破折号。

Then Evan Hubinger, who still currently works on alignment science he's the alignment science lead at Anthropic, he responded to Jacob's post, and he said, Jacob is correct here with a m dash.

我不知道他是不是用AI来回复的，但我觉得大概率是。

I don't know if he's using AI to to respond, but, I mean, I would guess yes.

但也许有人会因为我这么说而生气，觉得破折号很正常。

But maybe some people are mad at me saying that m dashes are normal.

总之，他说Jacob说得对。

Anyways, he said, Jacob is correct here.

我们真的认真相信AI可能会杀死全人类，感叹号。

We really do earnestly believe AI could kill all humans, exclamation point.

我个人认为，未来十年内这个概率不到百分之十。

I personally think it is less than ten percent within the next decade.

我相信Anthropic在尽力，但我们还没有解决超级智能对齐问题的方案，而且显然也没有在朝着这个方向明确前进。

I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligences and are not clearly on track to.

好吧，那真是太好了。

Like, okay, awesome.

所以负责这事的人说，是的，我觉得，你知道，未来十年内AI有可能杀死全人类，有一定的概率。

So the guy in charge of this says, yeah, I think it's, you know, less there's a there's some percent chance that within the next decade, AI could kill all humans.

而Anthropic在尽力，但实际上我们并没有解决问题的方案。

And Anthropic is doing its best, but we don't actually have a problem a plan to solve it.

所以这个回应相当惊人。

So that was a pretty wild response.

下面有人说，这是不是那部人工智能电影的赞助推文？

Someone below that said, is this a sponsored tweet for that artificial film?

总之，是的，关于这一切，至少可以说有很多戏剧性的事情发生。

And, anyways, yeah, there's a lot of a lot of drama going on to say the least about all of this.

雅各布原始帖子下的大多数评论都是某种版本：说实话，我不太把他当回事。

Most of the comments on Jacob's original post are some version of, I'm not really taking him seriously if I'm being honest.

然后还有很多人就是那种AI末日论者，在那儿使劲炒作这事儿。

And then there's a lot of people that just are sort of AI doomers that are hyping it up.

总之，很多人都在截图。

Anyways, a lot of people are screenshotting.

不知道你们还记不记得，2022年的时候，有个叫布莱克·莱曼的人说，那个系统有感知能力，能表达思想和情感，跟人类小孩差不多。

I don't know if you remember, but there was a moment where back in 2022, a guy named Blake Lehman said that sis said that system has perception of the ability to express thoughts and feelings equivalent to a human child.

那是个谷歌工程师，因为说AI聊天机器人有了自我意识，被停职了。

It's a Google engineer who was put on leave after saying AI chatbot had become sentient.

所以，如果你们都记得那事儿，那真是2022年挺疯狂的一个时刻。

So, anyways, if you guys all remember that, that was kind of a wild moment from 2022.

感觉好像过了很久了，但有些人说这让他们想起了那个人。

Feels like it was longer ago, but some people are saying that kinda reminds them of that guy.

还有些人说《终结者》不是电影。

And some people are saying Terminator wasn't a movie.

那是预言性的纪录片。

It was a prophetic documentary.

说实话，这有点被过度炒作了，但他现在有62万个赞。

Honestly, this is pretty overhyped, but he's at 620,000 likes right now.

所以也许我们真的都活不过这几年了。

So perhaps we are all on track to die in the next couple years.

这个说法真是够夸张的。

Pretty pretty wild claim.

好吧。

Okay.

他们正在发布他们的v6模型。

Is launching their v six model.

我很喜欢Suno。

I love Suno.

我用过很多次。

I've used it a ton.

真的很有意思。

It's a lot of fun.

我还用它发了几张专辑，其中一些每个月能给我带来几百美元的收入，都是AI生成的音乐。

And I've released a few albums with it, some of which make me, you know, hundreds of dollars a month from AI generated music.

我最常跟人提起的那张专辑是《帮助你的猫入睡的放松钢琴曲》。

My most famous one that I always tell people about is I have relaxing piano music to help your cat fall asleep.

那张专辑每个月能赚600美元，而且完全是AI生成的。

That album makes me $600 a month, and it's just AI generated.

真的，就是AI生成的。

It's actually, like, AI generated.

不是那种粗制滥造的东西。

It's not slop.

就是纯粹的放松钢琴曲。

It's it's just relaxing piano music.

听起来很棒。

It sounds great.

有意思的是我当初怎么宣传它的。

It's funny how I marketed it.

这张专辑表现非常好，很多人把它加到播放列表里放给他们的猫听。

It did really well, and a lot of people just add it to their playlist to play to their cat.

所以猫显然很喜欢AI生成的音乐，我也不觉得抱歉，因为你知道，如果它们喜欢，那就喜欢呗。

So cats apparently love AI generated music, and I'm not even apologizing because, you know, if they love it, they love it.

现在我要说的是，这些音乐以及其他所有东西的问题在于，Suno承担了全部责任，所以这不是我的问题，但Suno之前的所有模型都是用有版权的音乐训练的。

Now one thing that I'll say is the problem with that music and and everything else, which Suno takes the full liability for, so it's not my problem, but Suno trained all of their previous models on licensed music.

他们就是把世界上所有的音乐都拿来训练，然后所有唱片公司都起诉了他们。

They just trained it on all the music that existed in the world, and all of the record labels sued them.

有华纳、BMG、Believe、索尼和环球。

You had Warner, BMG, Believe, Sony, and Universal.

所以他们决定训练他们的全新模型。

And so they decided to train their brand new model.

顺便说一下，这是我的一个非常反传统的观点，我觉得很多人也会对此不满，因为这适用于Suno以及基本上所有其他公司。

And and by the way, here's, like, my very contrarian take that I think also a lot of people would be mad about because this goes for Sunu and basically everyone else.

他们把所有东西都拿来训练了。

They trained it on everything.

他们惹上麻烦了。

They got in trouble.

他们和发行商达成了授权协议。

They made licensing deals with the distributors.

但与此同时，他们的公司和品牌壮大了很多。

But in the meantime, their their company and brand grew a ton.

也许他们之前没有足够的钱，或者无法获得许可，但他们的品牌壮大了很多，现在他们有能力支付了，对吧，因为他们筹集了超过8.19亿美元。他们能付清所有这些罚款。

Maybe they wouldn't have enough money or wouldn't have been able to get permission before, but their brand grew a ton, and now they were able to pay, right, because they've raised over 819,000,000 doll They're able to pay off all of these fines.

他们能支付授权费，现在他们能用真正获得授权的音乐来训练模型了。

They're able to pay for licensing, and now they're able to train a model with actual licensed music.

关于这件事，有很多不同的看法。

Now there's a lot of different takes on this.

我一直在用AI音乐生成器，到现在大概有十多年了吧，我觉得。

I've been do I've been using AI music generators for, like, ten plus years, I think, at this point.

对。

Yeah.

大概十年了。

About ten years.

所以在Chekipati出来之前，就已经有很多音乐生成模型了。

And so way before Chekipati ever came out, there's been a bunch of music generator models.

我不知道那些早期的模型用的是什么东西。

I don't know what those early ones were using.

它们不是很好，但有些还行。

They weren't that great, but some of them were okay.

它们不算糟糕。

They weren't horrible.

那时候玩起来还挺有意思的。

It was it was pretty fun back in the day.

总之，这个新模型用的全是授权数据。

Anyways, with this new model, it's all licensed data.

我觉得这太棒了。

I think that's fantastic.

这就是我们看到其他所有人都在做的事情。

This is what we've seen everyone else do.

你先做出一个模型，把它做大，赚到一些收入，证明这个概念可行，然后再去给所有人付钱。

You get a model, you get it big, you get some revenue, you prove the concept, and then you go pay everybody out.

是不是应该提前给所有人付钱呢？

Should you have paid everyone out ahead of time?

也许吧。

Maybe.

但如果模型还没被证明可行，那确实挺难的。

But if the model wasn't proven, it's kinda hard.

他们当时也不会有钱去做这件事。

They wouldn't have had the money to do it.

所以，反正那是另一个话题了。

So, anyways, that's another argument.

这次发布将会分三个阶段进行。

This release is gonna happen in three tiers.

所以他们有v6，这个是付费的。

So they have v six, which is paid.

他们有v6 wild，这个是付费且实验性的，然后还有v6 mini，这个是免费的。

They have v six wild, which is paid and experimental, and then they have v six mini, which is free.

所有旧的Suno模型都会被淘汰，原因是数据许可问题。

All of the older Suno models are gonna be retired, the reason is because of the data licensing issue.

Suno筹集了那么多钱，而且你知道，显然还有那些诉讼在进行中。

Sunu has raised all of that money, and they, you know, obviously have all these lawsuits going on.

这次发布是在Suno承认用YouTube视频训练早期模型的同一天，或者仅仅一天之后。

This launch is on the same day or just one day after Sunu admitted to training earlier models on YouTube videos.

所以除了所有的音乐，他们实际上还用了YouTube视频，这可是大忌。

So beyond just all the music, they actually went to YouTube videos, which is a huge no no.

但公平地说，OpenAI从未承认用YouTube训练他们的视频模型Sora，不过我有点怀疑这可能是他们关闭它的部分原因。

But to be fair, OpenAI never admitted to train Sora, their video model, on YouTube, although I kinda wonder if that might maybe partially why they shut it down.

但是Miriam Maradi在接受《纽约时报》质询时，被直接问到，你们有没有用YouTube视频训练OpenAI的Sora视频模型？

But Miriam Maradi, when she was questioned by the New York Times and asked point blank, did you train the OpenAI Sora video model on YouTube?

她基本上就是回避了那个话题，那真是那种特别尴尬的时刻，她根本不想回答那个问题。

She kind of just, like, shied away from the topic, and it was, like, one of those famously awkward moments where she didn't want to, you know, didn't wanna answer the question, basically.

所以我们大概能猜到OpenAI可能也干了同样的事。

So we we get that OpenAI probably also did that.

好的。

Okay.

黑客正在利用窃取的会话，耗尽Claude订阅用户的令牌配额。

Hackers are draining Cloud subscriber token allowances with stolen sessions.

Anthropic确认，Infosealer恶意软件正在劫持Claude Code的OAuth令牌。

Anthropic confirmed that Infosealer malware is hijacking Cloud Code OAuth tokens.

有一位Max用户发现，自己什么都没干，使用量却从45涨到了55。

One max user watched usage climb from 45 to 55 while doing no work.

然后他就想，这是怎么回事？

And so he's like, what's going on?

为什么我的令牌一直在消失？

Why are my tokens disappearing?

有人黑进来，就用这些令牌去做一些项目。

People are hacking in and just using tokens to do certain projects.

Anthropic 确认了这件事，还有另一个用户报告说，在短短十二分钟内，他的使用量从零跳到了49%，而期间他只发了几个提示词和做了一次网络搜索。

Anthropic confirmed this, and there is a separate user that reported their usage jumping from zero to 49% in twelve minutes after only a couple of prompts and a web search.

Anthropic 暂停了受影响的账户。

Anthropic suspended the affected accounts.

他们让所有会话失效，并且给了部分退款。

They invalidated all of the sessions, and they issued a partial refund.

有一个用户每月付200美元订阅，结果只收到了44英镑的退款。

One user received £44 for their $200 a month subscription.

账户支持只追踪总使用量，不追踪明细使用量，这意味着令牌被盗可能几个月都发现不了。

Account support tracks account support tracks total usage, but not itemized usage, meaning token theft could run undetected for months.

对吧？

Right?

这让我想起那些搞音乐农场的人，他们偷了一大堆人的 Spotify 密码。

It honestly reminds me of all of the people that were doing, like, music body and farms where they would steal a whole bunch of people's Spotify passwords.

他们通常能从暗网上搞到这些，你知道的，随便什么渠道。

They would just get them on the dark web because they're usually, you could get a a you know, whatever.

所以他们会从暗网偷来这些账号，然后随机登录一堆陌生人的Spotify账户，反复播放自己的歌。

So they'd steal them from the dark web, and they would log in to a whole bunch of random people's Spotify accounts randomly and just play their own song over and over again.

然后他们就退出登录。

And then they would log out.

这很难追踪，Spotify花了很多时间和金钱，最后总算打击了整个骗局。

And it was really hard to track, and Spotify spent a lot of time and money and eventually kinda cracked down on the whole scheme.

我觉得很多非法操作的人可能进了监狱，或者根据欺诈的严重程度，面临各种不同的后果。

And I think a lot of people doing it illegally might have went to prison or some, I don't know, different things depending on how nefarious all of the the body and and and kind of fraud went.

但这让我想起了很多类似的情况。

But this reminds me a lot of that.

这是同样的道理。

It's the same idea.

你可能会从暗网拿到组合列表。

You you get maybe combo lists from the dark web.

你偷一大堆人的账户令牌和密码。

You steal a whole bunch of people's account tokens and and passwords.

你登录进去，随便干点事，或者如果他们有API，你登录进去，用几个令牌，然后退出，这样他们可能几乎不会注意到。

You log in, just get a couple things done, or maybe if they have an API, you log in, use a few tokens, and then you log out and to then you know, maybe they maybe they'll hardly notice.

今天关于这一切，我最后一个抱怨，跟那个完全没关系，因为这是完全合法的，没人黑我，就是今天我让Claw，或者说我让ChatGPT Astra，就是他们的GPT六模型。

Now my last complaint of the day in regards to all of this, which has absolutely nothing to do with that because this was completely legal, no one hacked me, is that today, I asked claw or I asked ChatGPT Astra, so their, you know, GPT six model.

我把它调到最高使用级别，对吧，最高的思考能力，然后让它去审查我的网站，给我一些好主意，来改进我的网站的品牌和视觉效果。

I put on the absolute highest usage tier, right, the highest thinking capability, and I asked it to go and review my website and give me some good ideas for some good ideas to, you know, review and and make the branding and and the visuals on my website better.

在它工作的两分钟里，它烧掉了90%，或者我觉得实际上是80%的我整个星期的所有令牌。

And over the course of two minutes that it worked for, it burnt through 90 per or I think actually 80% of all my tokens for the entire week.

所以千万别碰那个最高级别的Asterisk 6，用六英尺长的杆子都别碰。

So do not touch asterisk six on the absolute ultra max tier with a nine with a six foot pole.

如果你担心你的令牌用量，这就是我给聪明人的建议。

That's my advice to the wise if you're worried about your token usage.

除非你到了周末，必须赶紧用完，不然就重置了，或者你有一个特别大的问题要解决。

Unless you're at the end of the week and you gotta use it quick before it resets or you have one really big problem.

我不知道。

I don't know.

我就是特别震惊，两分钟就能用掉我80%的额度。

Like, I just am so blown away that in two minutes, it could use 80% of my tokens.

这件事让我超级不爽。

I'm super annoyed by that.

我登录我的ChatGPT账号，他们确实有个很酷的功能，这点真的要给ChatGPT点个大大的赞。

I'm in my ChatGPT account, and they do have a very cool feature, which huge shout out to ChatGPT for, for adding.

顺便说一句，Claude也有同样的问题。

And by the way, Claude's got the same problem.

不过Claude的话，基本上就是直接把你账号锁了。

With Claude, though, you basically get locked out of your account.

ChatGPT这边呢，它会提示你现在用的是Luma，一个更快但适合简单任务的模型。

With ChatGPT, they say you are now using Luma, a faster model for simpler tasks.

要么用重置，要么花80买一次重置，才能继续用最先进的模型，要么就等它到重置日期自动恢复，然后它会显示我的重置日期。

Use your reset or buy a reset for 80 to continue using the most advanced models or wait for it to reset after, and then it has my reset date.

所以基本上，他们会把你切换到他们最基础的模型之一。

So, basically, they switch you on to, like, one of their most, I don't know, basic models.

你还能完成一些基本的任务。

You can get basic tasks done.

但就我个人而言，我其实挺高兴的，因为至少他们没有完全封掉我的账户。

But personally, I'm actually very happy about that because at least they're not completely shutting off my account.

对，就把我切换到那些不太好的模型，紧急情况下还能帮我处理点事情。

Just, yeah, switch me to one of the terrible models, and it can sort of do some stuff for me if I have if I'm in a pinch.

我其实还有一个Claude的备用账户，这段时间可能会用它来处理很多事情。

I actually have a second Cloud account that I'll probably be using for a lot of stuff in the meantime.

但总的来说，有好有坏吧。

But overall, yeah, good good pros and cons to that.

好的。

Okay.

最后快速总结一下，我们得聊聊Meta今天发布的东西。

Really quickly wrapping up, we gotta talk about what Meta has launched today.

这就是Muse。

This is Muse.

它是一个以安全为首要卖点的个人AI助手。

It's a personal AI agent built on a security first pitch.

这个超级智能实验室的产品在iOS、Android和WhatsApp上都有。

The super intelligence lab product ships on iOS and Android and WhatsApp.

它支持Stripe支付，而且漏洞赏金上限高达30万美元。

It has Stripe checkout and a $300,000 bug bounty ceiling.

所以如果你去发现漏洞，他们真的会付钱给你，我觉得这招挺聪明的。

So if you go find bugs, they're actually paying you for it, which I think is genius on their part.

付费版是每月20美元，还有每月100美元的档位，以及免费版。

The paid tier is for people that are paying $20 a month, and they have a $100 a month tier and a free tier.

当然，所有这些都是按使用量计费的。

And, of course, all of those is usage based.

他们说，单个用户的提示注入漏洞能拿到大约13万美元的赏金。

They said that you get about a $130,000 for a single user prompt injections.

所以，他们就是花钱让你去测试这个系统的安全性。

So, like, they're just paying you to red team this thing.

如果你能利用它，让它做出坏事，他们就会付你13万美元。

If you can use it, get it to do something bad, they'll pay you a $130,000.

我觉得这真是个超棒的交易。

That's a pretty awesome deal in my opinion.

他们这么做是因为他们要在里面加入支付功能。

Checkouts because the and the reason why they're doing this is because they're adding payments into it.

他们和Stripe合作了。

They partnered with Stripe.

他们的结账流程是通过Stripe的链接来运行的。

They have checkouts that run through Stripe's link.

一次性卡号、一个密码和Shop Pay的集成都在进行中。

Single use card numbers, one password, and shop pay integrations are all underway.

他们正在处理所有这些。

They're working on all that.

所以这一切意味着，你可以去Muse说，嘿。

So what all this means is that you can go to Muse and say, hey.

我需要去，你知道，我午餐想吃塔可钟。

I need to go you know, I wanna eat Taco Bell for lunch.

帮我下单。

Go and order it for me.

而且它能访问你的Stripe账户。

And it has access to your Stripe account.

它会去把所有东西准备好。

It will go get everything ready.

它会弹出一个窗口，问你是否确认花14美元买这个卷饼和塔可钟的饮料。

It'll have a pop up that's like, do you confirm spending $14 on this burrito and drink from Taco Bell?

你点击确认，然后它就会通过DoorDash直接送到你家门口。

You click yes, and it gets shipped straight to your door with DoorDash.

我个人很喜欢这个。

I personally love this.

我是说，这确实是个大赌注。

I mean, it's a big bet.

我知道这是在冒险。

I know it's a gamble.

我知道这有风险。

I know it's risky.

我的意思是，OpenAI和Anthropic都没有做这种直接的Stripe集成，能帮你买东西。

I mean, OpenAI and Anthropic are not doing this direct Stripe integration where it can buy stuff for you.

但我认为，第一，这对Stripe来说很明智，因为他们会在这里获得不错的交易量。

But I think, number one, it's smart for Stripe because they're gonna get some good volume here.

他们会推广这个概念。

They're gonna this concept.

这几乎就像有保险一样，因为OpenAI提供了13万美元的提示注入漏洞赏金。

It's almost like there's insurance because OpenAI is, like, offering a $130,000 for prompt injection.

所以，你知道，我是说，那是一个巨大的赏金计划。

So, they're you know, I mean, that's a massive bounty program.

如果你能黑进这个东西，他们会给你很多钱。

They'll pay you a lot of money if you can hack this thing.

所以会有很多人尝试去黑它。

And so a lot of people will try hack it.

他们会解决很多问题。

They'll solve a lot of problems.

而且，是的，他们会有权限让你的AI模型去帮你买东西。

And, yeah, they're gonna get access to be able to tell your AI model to go buy things for you.

你需要批准并给它权限。

And you gotta approve it and give it permission.

Muse是由Meta内部的MuseSpark模型驱动的，由Meta超级智能实验室发布，这个实验室大约一年前成立。

Muse is powered by Meta's in house MuseSpark model, and it's shipped by Meta Superintelligence Lab, which actually came about about a year ago.

我知道有些人可能对这个感到害怕。

So I know some people are probably freaked out about this one.

我觉得这挺酷的。

I think it's kinda cool.