探索AI漏洞修复的潜在空间

Anthropic 的 Mythos AI 找微软漏洞的速度，比微软工程师打补丁还快。

Anthropics' Mythos AI is finding Microsoft bugs faster than Microsoft's engineers can patch them.

Google 的 SynthID 水印撑过了三百次修改，但碎片化可能让它彻底没用。

Google's SynthID watermark survived 300 edits, but fragmentation might actually make this completely useless.

这个我们待会儿细说。

We'll get into that.

Encore AI 刚刚融资三千万美元，要把销售通话录音转成语音助手。

Encore AI just raised $30,000,000 to turn sales call transcripts into voice agents.

从数据角度看，这个流程挺有意思的。

This is an interesting pipeline, as far as the data goes.

现在写东西的人都开始故意保留错别字了，我也一样，就是想对抗无处不在的 AI 写作——那玩意儿简直成了互联网上的瘟疫。

Writers right now are embracing typos, myself included, and all of the idiosyncrasies basically to fight the AI writing that we're seeing everywhere that has just become, you know, basically a plague on the entire Internet.

Sierra 要花十亿美元收购 Oasis Security，专门管 AI 代理的身份认证。

Sierra is going to acquire Oasis Security for a billion dollars to police AI agents' identities.

如果你经常用 AI 工具重复做某些任务，不妨试试 AI Box Builder 平台。

If you have tasks that you do over and over again using AI tools, I'd love for to check out the AI Box Builder Platform.

这是我自己的创业项目，网站是 aibox.ai，你可以把多个 AI 模型串联起来，输入提示词，自动完成那些重复性任务，省得你手动一遍遍复制粘贴，来回切换不同文档或 AI 工具。

This is my own startup at aibox.ai, and it allows you to link together multiple AI models and put prompts in and automate entire processes of repetitive tasks so you don't have to do them, and, you know, put in your prompts and copy and paste between different, documents or different AI models over and over again.

厉害的是，平台上我们集成了八十多个不同的 AI 模型，它们都能协同工作。

And the cool thing is that all of all the different AI models, we have over 80 on the platform, work together.

比如你可以让 11 Labs 生成音频。

So you can have 11 labs creating audio.

用 Google 的 V3 做视频，用 ChatGPT 生成图片，再用 Claude 来写文案。

You can have Google v o three creating video, you can use ChatGPT to generate images, and you can use Claude for your writing.

你可以在 AI Box Builder 平台上轻松搭建各种自动化流程和实用工具。

You can build out workflows and really cool tools all on the AI Box Builder platform.

链接在简介里，每月只要 8.99 美元。

It's a link in the description, and it's only $8.99 a month.

而且，你还能用同样的价格，在一个地方直接跟所有 AI 模型聊天。

In addition, you can chat with all the different AI models in one place for that same price.

咱们聊聊 Anthropics 的 Mythos 最近在搞什么。

Let's talk about what's going on with Anthropix Mythos.

他们有一个专门找漏洞的AI，仅在四月就发现了微软SharePoint上90个严重漏洞和141个重要漏洞。

They have a bug hunting AI, and it found 90 critical and a 141 different important bugs in Microsoft SharePoint just in April.

问题在于，这些漏洞被发现的速度，比微软工程师修补的速度还要快。

And this is basically, the problem here is that it's finding these bugs faster than Microsoft's engineers can patch them.

所以微软在7月14日一次性发布了600个补丁，内部会议记录显示，公司当时赶在敌对政府开发出类似工具之前，必须赶在5月31日前完成修复。

So Microsoft shipped 600 fixes in on July 14, and their internal meetings show that the company was racing a May 31 deadline before hostile governments build equivalent tools.

而且这还不是全部，微软SharePoint只是他们关注的其中一个目标。

And I'll also say that this isn't the only thing, you know, like, Microsoft Microsoft SharePoint isn't the only thing it's been working on.

今年年初以来，这个AI已经在Microsoft 365、Teams、Copilot等多个产品里发现了数百个严重漏洞。

It has found hundreds of critical bugs in Microsoft three sixty five, Microsoft Teams, Microsoft Copilot, all of that since the beginning of the year.

目前还有大约300个中等严重程度的SharePoint漏洞在排队等待修复。

So I think they have about 300 moderate severity SharePoint flaws that are still queued for patching.

这事儿真不是小问题。

So, I mean, this is kind of a big deal.

就在本月14号，微软发布了超过600个漏洞的补丁。

And on July of this month, or the fourteenth of the month, Microsoft released patches for over 600 bugs.

其中只有七个属于低或中等严重级别，还有一个已经被黑客在现实中利用了。

Only seven were low or moderate severity, and one was already being exploited by hackers in the wild.

也就是说，它真真切切地发现了一个正在被攻击的漏洞。

So, I mean, it literally found something that was actively being hacked.

去年六月底，五眼联盟警告称，防御方赶超AI攻击者的时间窗口将在几个月内关闭。

The Five Eyes Intelligence Alliance warned in late June last month that the window for defenders to outpace AI equipped attackers would close within months.

微软内部文件显示，这个窗口可能已经关闭了——毕竟，现在已经有漏洞被实际利用了。

Internal Microsoft documents show that it might have already closed, right, because we already have one exploit being, used.

真正的危险在于，AI能把多个小漏洞串联起来，组成人类根本发现不了的完整攻击链。

The real danger, I think, is that mythos can chain multiple small bugs into working exploits that humans miss.

所以，高级别Anthropic顾问、前NSA人工智能主管文·纽根说，当AI能将低优先级漏洞变成武器时，微软那种‘先修高危、后修低危’的漏洞分级处理方式就失效了。

So Vin Nugen, who's a senior anthropic advisor and a former NSA AI chief, said that Microsoft's standard triage system, which is patch critical flaws first, low severity last, no longer works when AI can weaponize the bottom of the queue.

所以这场竞赛确实已经打响了，光靠速度可能还不够。

So the race is definitely on and speed alone might not be enough.

目前，谷歌的Synth ID水印技术也已经上线了。

Right now, also have Google's Synth ID watermark.

所以本质上，这是一种嵌入在AI生成图像中的隐形水印，主要来自谷歌，但其他不少公司也联手采用了这项技术。

So basically this is an invisible watermark embedded on AI generated images primarily from Google, but there's some other a bunch of other companies have kind of all worked together to adopt this.

不过并不是所有人都用了。

Not everybody has.

我记得Grok好像就没做，但很多其他平台都已经用了。

I think Grok, if I if I'm remembering correctly, does not do this, but many others do.

当他们推出SynthID水印时，实测显示，这种水印在经历了三百次压缩和缩放后依然完好无损。

So, when they were when they did this, these AI generated images, when it had the SynthID watermark, apparently, it actually survived 300 rounds of compression and resizing and real testing.

而且系统仍然能准确识别出这张图是AI生成的。

And so and it still was able to detect that an image was AI generated.

所以有人可能会说，你看——

So people might say, like, oh, look.

我拿一张AI生成的图片，

I'm gonna, you know, take this AI generated image image.

然后把它压缩一下。

I'm gonna compress it.

我要把它缩放一下。

I'm gonna resize it.

我还可以放大，或者用各种方式编辑它，但这个水印依然能被检测出来，真的很厉害。

I'm going to zoom in and, like, do you you can try to edit it in all sorts of ways, but that watermark is still detectable, which is really cool.

谷歌表示，这证明他们的技术比预期的要耐用得多。

And that, Google says, proves that their tech is way more durable than they were expecting.

我觉得还有一个大问题，就是谷歌的检测器无法识别OpenAI、Runway或NVIDIA的水印。

I think there's still one big problem, and that's Google's detector can't read watermarks for OpenAI, Runway, or NVIDIA.

如果我没记错的话，他们正在和Meta合作，可能还有其他几家公司在参与。

So I I believe if I'm correct, they're working with Meta on this, and maybe a couple other players.

但即使用的是完全相同的技术，OpenAI、Runway、NVIDIA这些平台的水印，检测器就是跨不了平台，这有点奇怪。

But even though they're using the exact same technology, OpenAI Runway NVIDIA, like, the the detector doesn't go cross platform, which is kinda weird.

所以SynthID确实非常出色，在对完全AI生成和AI编辑的图片进行300次压缩后，只有在额外裁掉20%边缘的情况下才被破坏。

So SynthID is obviously really impressive, and with those 300 compressed cycles on both fully AI generated and AI edited images, It was only breaking after a 20% border crop was added on top.

那是唯一能把它搞垮的办法。

That was the only thing that was killing it.

谷歌限制每个用户每天只能验证大约十张图片，并且会对相似图片的快速查询进行限速，防止攻击者反向破解他们的系统。

Google limits SynthID verification to about 10 image checks per day per user, and they throttle faster lookups on similar images to block attackers from reverse engineering their system.

这就有点像Anthropic抱怨大家到处搞蒸馏攻击的时候。

It's kind of like when you when Anthropic is complaining about everyone going and, doing distillation attacks.

对吧？

Right?

他们觉得，人人都在问Anthropic问题，把答案拿去训练自己的AI模型。

They're like, everyone's, asking Anthropic questions, getting the answers, and training their AI models on it.

谷歌担心有人会利用验证功能，比如问‘这张图是不是用SynthID生成的’，来反向推算他们的技术原理。

Google's worried that people are going to use the verification, like, hey, is this was this generated with SynthID to reverse engineer how they're doing it.

我不确定这能不能长期奏效，毕竟迟早你会积累足够多的数据，最终摸清门道，但至少现在，这还是个严密保守的秘密。

I don't know if that's really a long term play because eventually, right, you're gonna get enough data and and figure it out, but for now, there it's kind of a closely guarded secret.

这感觉就像CAPTCHA一样，得不断升级才行。

It feels kind of like it's gonna be like these CAPTCHAs, how CAPTCHA has had to evolve a lot, right?

以前就是简单的字母组合，后来变得复杂了，再后来变成让你在照片里找出所有桥梁，或者旋转一个3D模型，让它和图片里的视角对齐。

It used to just be like really simple letters, then it got more complicated, and then it got turned into like identify all the bridges in this photo and, you know, zoom this, you know, spin the three d model till it lines up in the image.

现在市面上有各种稀奇古怪的验证码设计。

Like, there's all these crazy things that have been invented for CAPTCHAs.

我觉得SynthID以后也会变成这样。

This is kinda what I feel like SynthID is going to become.

OpenAI、Runway和NVIDIA各自开发了不同的水印技术，但实现方式都不太一样。

OpenAI Runway and NVIDIA have each built their own versions of watermarks, but they all have a little bit different of implementation.

这正是我想说的，根本没法跨平台通用。

That's what I was saying, like, doesn't actually go cross platform.

所以目前整个生态挺分散的，这些检测工具通常只能识别自家的水印，别的都认不出来。

And so there's a bit of a fragmented landscape right now, and these detectors, they can only really detect their own watermarks, not everyone else's watermarks.

我最想知道的是，你能不能去Google生成一张图，看着挺像回事，然后拿去给开源模型，让它重新生成一遍，这样会不会就把SynthID水印给抹掉了？

The thing that I'm the most curious about, and I don't actually have a direct answer to this that that I would be curious though, is if it's actually able to, you know, you could go to like Google, generate an image, maybe it looks like a great image, just go to an open source model, give it that image, and say regenerate this, and if it would kind of kill the the SynthID on top.

我猜大概率真能抹掉。

And I would guess that it might actually kill it.

所以正因为这样，我平时生成图片时，经常直接截个网上的图，或者找张喜欢的素材，然后说：就用这个当参考，稍微改一改。

So for that reason, I mean, half the time I'm generating images, I'm like screenshotting, you know, another image or another graphic or something that I like out in the wild, and I'm like, hey, use this as an inspiration and change it slightly.

所以我很多时候用ChatGPT来生成图片，我担心这也会导致同样的问题——这些AI图像的SynthID标记基本上就形同虚设了。因为大多数恶意用户都会聪明地找到办法绕过它，而99%正常使用AI生成图像的普通人，其实根本不在乎这个。

And so that's how I use, like, ChatGPT to generate images a lot of times, and I'm worried it's gonna be the same thing, which basically makes, these image generating SynthID things useless because for a majority of, you know, bad actors out there, they'll be smart enough to figure that out, and 99 of people that just use these AI image generators in a, you know, responsible or good way, well, it doesn't really matter.

说到底，老实讲，我不确定这会不会真有什么大用处。

And so at the end of the day, yeah, I'm not sure if it's gonna make a huge difference if I'm, you know, being honest.

抱歉说了这么多消极的话。

Sorry for the pessimism.

明白了。

Okay.

Encore AI刚刚融资了三千万美元，用于开发语音助手。

Encore AI has just raised $30,000,000 to build voice agents.

这些助手全部基于真实成交的销售通话进行训练，这挺有意思的。

They're all gonna be trained on real sales calls that have actually closed deals, which is fascinating.

对吧？

Right?

因为我们已经有了这么多数据集。

Because we have all these datasets.

我有个朋友，他开了一家公司，手下有一群销售，专门帮他促成清洁服务的订单。

I have a friend in particular, and he's got a company, and he's got a bunch of salespeople that help close deals for his cleaning company.

有意思的是，他用Go High Level，每通电话都全程录音。

And what's interesting is he uses Go High Level and every single call is recorded for the the whole conversation.

所以他手上有所有通话的文字记录，也知道哪些最终成功签单了。

So he has transcripts of every single call, and, yeah, he knows which of those calls actually ended up closing.

就算他没专门标记这些信息，你也能直接读录音稿，或者让AI帮忙分析，判断出哪些单子成交了。

You can also even if if he wasn't tracking that bit of information in particular, you could usually just go read the transcript or have AI read the transcript and decide which ones were closed.

这真的挺让人惊讶的。

But that is fascinating.

对吧？

Right?

这相当于通过录音稿找出顶尖销售，看看他们是怎么成交的，然后训练模型去模仿他们。

These it's like determining who the top sales reps are, which ones are actually closing, using those transcripts and training models to to be more like those.

所以是通过微调模型，让它们真正擅长促成交易。

So fine tuning models to actually be good at closing deals.

这家初创公司叫Encore AI，他们收集所有的通话录音，从中分析顶尖销售员说了什么话，能有效推动客户前进。

This startup in particular, Encore AI, they're getting their call recordings and they're they're mining all of that to identify what the top performers say that move customers forward.

所以在我看来，这真的非常聪明。

And so, yeah, this is, in my mind, really, really smart.

自从他们创立以来，尤其是完成种子轮融资的十八个月里，年经常性收入增长了五倍以上。

The annual recurring revenue has grown more than five x since they started and since they did their seed round, which is about eighteen months ago.

他们现在服务的大型企业客户已经超过四十家。

They now serve over 40 enterprise customers.

目前大多数客户都是金融机构。

Most of those are financial institutions that they're servicing right now.

这家公司2022年刚成立时叫ESAIT IO，后来才改名为Encore。

This was started in 2022 as ESAIT IO, and then they rebranded to Encore.

后来，我听说有位顾问建议他们改名，说原来的名称不够响亮，不容易让人记住。

Later on, I think one of their advisors recommended it because you say it is not, I don't know, very recognizable.

总之，他们逐阶段分析客户互动，找出影响收入的漏洞和障碍点，再把这些洞察反馈给AI代理和员工，用于培训和指导。

In any case, they analyze different customer interactions stage by stage to find revenue leaks and friction points, then they feed those insights into both AI agents and human employees for coaching.

让我觉得有意思的是，它并不是直接拿完整的通话记录照搬，而是把整个过程拆分成几个阶段。

So what's interesting to me is it doesn't just grab the the just like the pure transcript and try to copy the whole thing, but it breaks it into stages.

比如说，我们假设他们做的是陌生电话推销，第一步得让人愿意听你说下去。

It's like, okay, we're in, you know, the the maybe it's like a I don't know if they do cold calling, let's say it's like cold calling.

然后得让人对产品产生兴趣，接着抛出一个吸引点，再消除对方对产品的顾虑，最后还要说服对方接受价格。

It's like you gotta get the person to stay on the phone and it's like, then you gotta get them to be interested in the product, then you gotta give them like a hook, then you gotta like, you know, get rid of any sort of, issues they might have with your product, then you gotta sell them on the price.

对吧？

Right?

销售的每个环节都是独立的阶段，他们就是把这些通话记录按这些环节一一拆解，分块分析。

Like, these are these different stages in in the sale that you gotta go through, and so it's breaking the transcripts down into all of those different pieces to, you know, identify it that way and kind of break it into chunks.

我觉得这真的太有意思了。

I think this is so fascinating.

Encore 比 Salesforce 等其他 CRM 巨头更厉害的地方在于，通话录音是独立存储在他们的分析平台之外的。

Encore's edge over Salesforce and a bunch of other CRM giants is that the call recordings live outside of their prep platform.

我觉得金融服务业正是他们最适合切入的领域之一。

So financial services is I think one of the smartest areas that they could get into for this.

现在银行和保险公司为了合规，早就开始录音所有客户对话了。

Banks right now and also insurers already record every customer conversation for compliance.

我们打电话时都听过那句：本通话可能被录音用于服务质量监控。

We've all heard it when we're on the phone and it's like, this call may be monitored for quality assurance purposes.

对吧？

Right?

所以Encore现在可以直接接入这套现有系统，而不用重新谈判每个客户部署中CRM数据的处理方式。

And so Encore right now is gonna be able to plug into that infrastructure without forcing a renegotiation of how CRM data gets handled across every customer deployment.

如果他们真能去像摩根大通这样的银行，说：嘿，摩根大通，你们已经有所有通话记录了。

And if they're able to actually train it on, like, let's say they go to Chase Bank and they're like, hey Chase Bank, you already have all the transcripts.

我们就用你们的记录，来搞清楚你们顶尖销售是怎么把客户转化的。

We're gonna use your transcripts so it we understand how your top salespeople actually convert people.

对吧？

Right?

而且如果他们只把数据留在摩根大通内部，不拿去训练其他银行比如花旗的数据，那就太棒了——摩根大通能用自己团队的最佳实践，花旗也能用自己团队的最佳实践，谁都不用担心机密信息外泄。我觉得这简直太厉害了。

So and and if they can just keep that data with just Chase Bank and they're not, like, training off of that and giving it to, you know, Bank of America or something else, that's really valuable because Chase Bank gets, you know, the best their salespeople are able to do and Bank of America gets the best their salespeople are able to do and they're not worried about that data or that, you know, confidential information leaking, I think this is incredible.

像Salesforce这样的平台，目前还比较集中化，没有那种开箱即用的个性化客户解决方案。

And so when you have something like Salesforce, currently Salesforce and others like that, they're a little bit more centralized and they don't have those kind of, out of the box, you know, custom, client solutions.

这真的很棒。

So very cool.

现在，作家和编辑们都在故意打破传统的写作规范。

Right now, writers and editors are deliberately breaking conventional style rules.

大家都开始不用破折号了。

We're all getting rid of em dashes.

错别字也越来越多了。

We're all adding typos.

说实话，我本来就经常打错字，这纯粹是我个人的习惯。

Well, I mean, I actually just naturally have a lot of typos, so that's just me.

我老婆每次见了都翻白眼。

My wife always rolls her eyes at me.

但这种做法正变得越来越流行。

But this is actually becoming something that's more and more popular.

说来也挺有意思，这个播客从一开始，我就一直秉持着同样的理念。

And it's kinda funny because in a weird way with this podcast, I've always kind of had that same mantra.

我刚开始做这个播客之前，其实还运营过一家初创公司，专门用纯AI生成的脚本和配音做播客。

A lot of when I first started this podcast, prior to this, I actually ran a bunch of a 100% AI generated scripts and speaker podcasts for a different startup I had.

那时候纯粹是为了获客之类的用途。

And I was just doing it for, you know, lead gen or whatever.

那些内容都是教育类的、长尾内容，不像现在这样是时事新闻。

And they were kind of like educational episodes, evergreen content, not news like this is.

但不管怎样，从最开始我就知道，这个播客必须有真人出声——必须是我自己的声音，因为听众想听的是我的经历、我的故事、我的亲身见闻。

But anyways, from the very beginning, I always knew that this was a possibility and I always said, look, I gotta have a real person, myself, my real voice on this podcast talking because people are gonna wanna hear my my background, my stories, my anecdotes.

甚至我说话时的停顿、‘呃’、‘那个’，这些看似零碎的人类小习惯，听众反而喜欢，因为他们知道这不是AI在跟他们说话。

And and even to the point where I was saying like pauses and ums and buts and like little, I don't know, human things that you put in there, people wanna hear that because they know that this isn't an AI talking to them.

至少我个人觉得，以前AI语音听起来很假，但现在真的进步太多了。

I for one I mean, if the voice cause the voices used to be bad and now they're actually really good.

比如你用11labs v3，那些语音简直太逼真了。

Like, you use 11labs v three, the voices are incredible.

有时候我不小心听，还真分不清是真人还是AI。

Sometimes I can't tell it's a it's a real, or it's just an AI if if I'm not careful.

我还接过一些电话。

And I've had calls.

我也听过一些音乐。

I've I've heard music.

我能听出来，但我是认真的。

Like, I can detect it, but guys, I'm serious.

前几天我听了一首歌。

Like, I'll get I'll I'll listen to a song the other day.

我把它加到了YouTube Music的喜欢列表里，听了大概五遍才反应过来。

I added it to my liked folder on, on YouTube Music and realized probably, like, five listens in.

我觉得这歌八成是AI生成的。

I'm like, this is probably AI generated.

总之，我觉得这种事发生的频率，可能比我们愿意承认的要高。

So anyways, there's there's I think this happens more than we would like to admit.

我的意思是，这些东西现在进步得真快。

I mean, this stuff is getting a lot better.

但我一直觉得，人们终究还是想要真正的人类。

But my mantra's always been like, look, people are gonna want a real human.

现在作家们就是这样做的。

And this is what's happening with writers now.

他们故意加些拼写错误，还有那种杂乱的第一人称故事，让文字看起来更像真人写的。

They're they're adding literal typos and kind of these messy first person stories to make all of their writing look more unmistakably human.

这不只是看起来像人，而是要让人一眼就能认出来是人写的。

It's not even just, like, looking human, but unmistakable.

我们就是希望别人知道，这绝对是人类写的。

Like, we want people to know this is human.

ChatGPT可不会输出这样的文字。

ChatGPT doesn't output text like this.

它也不会全用小写字母。

It doesn't use all lowercase.

这真的挺有意思的。

So this is really interesting.

有一家出版机构特别提到，今年五月他们收到了一千份投稿，几乎全是AI生成的。

One one publication in particular said that they got a thousand pitches in May and almost all of them were AI generated.

《Pitchfork》的主编已经禁止投稿中使用AI生成的内容。

Pitchfork's editorial head has banned AI from review submissions.

这个规定是从2024年开始的，我记得大概是两年前吧。

That was starting in 2024, and I think that was like two years ago.

但小说家们现在都在刻意避开很多东西。

But novelists are actively avoiding a lot of different things.

你能在推特上看到这种现象。

And I mean, you'll you'll see it on Twitter.

挺搞笑的，大家会去翻老作品里的破折号，或者找Claude总爱用的那些奇怪表达，比如它总说‘不是这样，是那样’。

It's kinda funny or an x or whatever where everyone like finds em dashes in old literature or they find like just funny glitches that Claude always says, you know, it's not this, it's that.

对吧？

Right?

就像这些搞笑的梗，大家还会从老书里截屏，说天哪，没想到简·奥斯汀当年居然在用AI写书。

Like all those funny things and they'll screenshot them from old books and they're like, oh my gosh, I can't believe Jane Austen was using AI to write.

总之，写作圈里现在有很多努力，都想让内容更有人味儿。

So anyways, there's a lot going on inside of, inside of all of inside of the writing world trying to be more human.

Sierra是一家估值120亿美元的数据安全公司，他们花10亿美元收购了Oasis Security，用来管理企业内部AI代理的身份。

Sierra is a $12,000,000,000 data security company, and they're buying Oasis Security for $1,000,000,000 to manage AI agent identities inside of companies.

现在，几乎每家公司都在部署各种AI代理来干活，但它们需要工具来追踪这些代理是谁、有什么权限、在做什么。

So right now, basically, every company is putting out all of these different agents to do work, but they need some tools to track who those agents are, and what they have access to, what they're doing.

这问题以前传统安全软件是盯人的，但现在有了AI代理，他们也想盯住这些玩意儿。

This is a problem that I think a traditional security software would be tracking humans, but now that we have these agents, they also wanna track that.

我觉得挺有意思，我在X上看到有人截了ChatGPT 5.6的图，说哎呀不好了。

And I think it's interesting because I've seen some tweets on X where people were like, they were screenshotting I think it was, ChatGPT, like, 5.6 or whatever, and it oh, no.

是在Cursor上看到的。

Was on Cursor.

他们用Cursor做项目，然后说：行了，先暂停一下。

They were using Cursor and they said, they were working on a project and they're like, okay, stop for now.

这事回头再 finish 吧。

Finish this later.

他们刚叫停，它就自己提交了。

And it made a push after they told it to stop.

他们问：‘你干嘛这么做？’

They're like, why'd you do that?

我说：‘哦，它就是自动提交了。’

And I was like, oh, just, like, was doing the push.

总之，这类情况挺常见的，尤其是当这些智能体接到任务，特别是那种开放式的任务，比如：

Anyways, these things happen where these agents, you know, especially if you give them a task, especially if it's more open ended, like, hey.

你把 x y z 这件事搞定。

Just get x y z thing done.

它们会按自己的方式去做，我们往往搞不清楚到底是谁改了哪里、改了什么，所以就会出问题。

They do it in their own way, and we don't always know what made the change, what made the edit, who did what, and so you can have some of these issues.

Anyway，Oasis 是 2022 年成立的。

Anyways, Oasis was founded in 2022.

他们筹集了1.95亿美元。

They raised a $195,000,000.

他们专注于保护非人类身份，主要通过API密钥实现——这些密钥能直接登录企业系统，全程无需人工介入。

They specialize in securing nonhuman identities, and they do it through API, you know, API keys that are logging into corporate systems without any humans there.

他们刚又融资6亿美元，估值达到120亿美元，年收入已达1.5亿美元，但至今还没盈利。

They just raised $600,000,000, a $12,000,000,000 valuation, and they're now making a $150,000,000 a year in revenue, but they're still not profitable.

他们的支出非常激进。

They've spent really aggressively.

他们最近接连完成了三笔收购，时间间隔特别短。

They have three acquisitions that they just did in quick, like, really close together.

这是他们几个月内的第三笔收购，之前已经收购了Rift和Genie Security。

This is their third acquisition in a few months, which is following them buying Rift and Genie Security.

我觉得，一家公司能融到这么多钱，挺有意思的。

I think it's interesting when you have a company that's able to raise so much money.

显然，他们的团队非常出色。

Obviously, they have a really good team.

他们高层有个非常有远见的领导者，但每次做收购，一方面说明他们很聪明，另一方面也说明他们发展得太快了，资金充足，完全有余力去收购，而不是从头开始研发，这点挺有意思的。

They have a really good, you know, visionary at the top, but when they make all these acquisitions, I mean, a, they're smart, but b, they you know, they're they're growing so fast and they have so much money they have the bandwidth to do it and not to build it from scratch, which is interesting.

现在大家都在抢着布局非人类身份市场，赶在像Palo Alto Networks或者Okta这些大公司进场之前占住位置，所以我觉得他们才这么急着出手。

There's a huge push right now to own the nonhuman identity market before a lot of these bigger companies like Palo Alto Networks or Okta are kinda moving into this, and so that's why I think they're jumping on that.

这周还有件别的事我想提一下，就是谷歌DeepMind刚发布了Lyria 3.5，音乐生成效果更流畅了，人声和歌词也更自然了。

Other news that happened this week that I just wanna cover, the headlines, Google's DeepMind has just shipped Lyria 3.5 in flow music with some better vocals and lyrics.

我以前聊过Lyria，说实话当时没觉得有多厉害。

I've talked about, Lyria in the past, not been impressed with it.

刚推出的时候，它只能生成三十秒的片段。

When it was first came out, it was, like, only able to generate thirty second clips.

后来能做完整歌曲了，现在我觉得他们在人声和歌词上进步了不少。

Then it became then it was able to do full songs, now I think they're getting better at vocals and lyrics.

我觉得它将来会成为一股重要力量，能跟Yu Gi Oh和Suno这些产品一较高下。

So I think this is gonna be a big player, challenging Yu Gi Oh and Suno and stuff.

刚出来的时候确实没什么可说的，但我觉得Lyria现在真的在变强。

When it first came out, it wasn't much to talk about, but I think Lyria is getting better.

我还是更喜欢Suno，因为他们有完整的AI工作室，还能做音乐翻唱，功能多得不得了，不过我也挺高兴谷歌背后的模型在不断进步。

I still like Suno more because they have the entire AI studio and they have like music covers you can do and there's just so much to to do with Suno, but, I am excited that the underlying model behind Google is getting better.

Panagram刚融了九百万美元，专门用来检测AI生成的文字。

Panagram just raised $9,000,000 to detect AI generated text.

据说准确率高达99%。

Apparently has 99% accuracy.

这挺有意思的，尤其想到我们之前聊的，现在大家都故意加点错别字啊之类的，想让文字看起来不像AI写的。

It's interesting, you know, in in light of what we were talking about with everyone using more, I don't know, typos or whatever you wanna call it to seem less AI generated.

现在这些检测工具已经做得相当厉害了。

So now we have these tools that are really good.

有意思的是，OpenAI在最早期的时候也推出过自己的AI文字检测工具。

It's interesting because OpenAI in the very early days came out.

但他们后来直接把项目停了，说识别AI生成的文字实在太难了。

They had their own, AI text, detector, and they actually shut the project down because they said it was too hard to detect AI generated text.

现在我觉得，我们每个人脑子里其实都自带AI文字检测功能了，能一眼看出那些套路化的表达、破折号的用法，但这些小技巧，AI模型很快就能学会避开。

Now, I think we all kind of have AI generated text detection built into our brains now, where we're we're understanding a lot of the phrases that are used, we're understanding the em dashes, but those things the AI models are gonna be able to get around very easily.

我其实有一份清单，列出了AI模型老爱用的那些让我烦的表达。每次让AI帮我写播客脚本、文章、博客，或者任何内容，甚至网站文案，我都会让它逐条检查，把所有破折号全删掉，把所有'X、Y或Z'这种句式全改掉，就像给内容做一次全面审计。

I actually have a whole list of things that AI models frequently say that annoy me, and whenever I get AI to write a podcast script, or if I get it to go write an article, or I go get it to write a, you know, blog post or content, or like basically any work I'm doing, or even like even websites, I very frequently will say like, run through this whole list and remove every single m dash from the website, every single time that we say x, y, or z, and I just get it to go like, run this kind of audit on my content.

希望以后模型自己就能避开这些套路了，不过这事儿还挺有意思的。

Now eventually, hopefully the models will just get away from that, but, it is interesting.

所有这些AI文本检测技术，我觉得吧，希望它真有效，但我一直持怀疑态度，总觉得这就像猫鼠游戏，永远追不上。

All of this AI generated text detection stuff, I think I like I hope that it's good, but I'm always dubious of it because I feel like it's always this cat and mouse game.

明白了。

Okay.

有个挺有意思的消息。

There's funny news.

玛莎·斯图尔特联合创办了一个叫Hint的AI助手，专门帮 homeowners 解决问题，项目已经拿到了一千万美元融资。

Martha Stewart has cofounded Hint, which is an AI assistant for homeowners, and she has $10,000,000 in funding for this project.

这还挺让人意外的。

So that's kind of interesting.

如果你还没给这个播客打分评价，真的会特别帮到我们。

If you haven't left a review on the podcast yet, it would help the show out a ton.

我真的非常非常感激。

I would super, super appreciate it.

这能帮助更多像你这样出色的人找到这个节目。

It helps more incredible people like yourself find this show.

所以，这确实是帮助节目触达更多听众的最重要方式。

And so, yeah, it's the number one thing that could help this show reach more a bigger audience.

如果你还没做，麻烦抽空留个评价，我会特别感动。

If you wouldn't mind, if you haven't already done it, it would mean the world to me.

非常感谢你收听。

But thanks so much for tuning in.

希望你今天过得特别棒。

I hope you guys have an incredible day.

想体验80个不同AI模型的话，别忘了去aibox.ai看看，我们下期再见。

Make sure to go check out aibox.ai if you wanna get 80 different AI models all in one platform, and I'll see you in the next episode.