释放潜能：英伟达与Super Safe合作

欢迎收听本节目。

Welcome to the podcast.

Super Safe Intelligence 刚刚和英伟达签下了一笔重大合作，大幅提升了计算能力。

Super Safe Intelligence has landed a massive new deal with NVIDIA for a huge boost in computing.

Claude Opus 4.7 刚刚完成了一个为期两周的编程任务。

Claude Opus four point seven has just finished a two week coding job.

最近出了一项新的评估标准，我觉得特别有意思，待会儿跟你们说说原因。

There's this basically new benchmark that has come out that is really exciting to me, and I'll tell you why.

现在谷歌的AI摘要已经出现在43%的搜索结果中了。

Google AI overviews now appear in 43% of searches.

相比一年前，这个数字上升了15%。

This is up 15% from a year ago.

我们来聊聊搜索引擎优化领域最近的具体变化。

Let's talk about how the landscape is shifting specifically in SEO.

英伟达、微软和IBM共同发起了开放安全AI联盟，旨在保护AI代理。

NVIDIA, Microsoft, and IBM have all launched Open Secure AI Alliance to defend agents.

咱们来聊聊，为什么这件事引发这么多争议。

We're gonna talk about why there's so much drama in that.

Enigma 终于正式亮相了，刚拿到七千万美元，打算彻底重新思考人和机器人之间的互动方式。

Enigma has exited stealth, and they've just gotten $70,000,000 to rethink how humans are talking to robots.

SAFE 超级智能公司成立才两年。

SAFE Super Intelligence is only two years old.

这家研究机构是由伊利亚·苏茨克弗离开 OpenAI 后创办的，已经获得了英伟达数十亿美元的投资，还拿到了他们下一代 Vera Rubin GPU 平台的访问权限。

It's the research lab that was founded by Ilya Seskever after he left Open AI, and they have secured a multibillion dollar investment from NVIDIA and also access to their next generation Vera Rubin GPU platform.

这笔交易尤其关键，直接把他们的算力提升了十倍左右，是个巨大的飞跃，能让他们大规模推进 AI 安全与推理方面的研究。

This deal in particular is basically gonna increase their compute power by about 10 x, which is a massive jump, and it's gonna let them scale their research into AI safety and reasoning.

如果你一直关注这个领域，你有没有觉得这家公司特别离谱？

And you know what's crazy about this company if you've been following along?

他们连一个产品都还没发布呢。

They literally have not even shipped a product yet.

他们目前没有任何收入，却拿到了英伟达这笔数十亿美元的投资，说白了，就是靠伊利亚·苏茨克弗的名气——毕竟他可是 OpenAI 的元老之一。

They don't have any revenue, and they've just gotten this multi billion dollar investment from NVIDIA, really just riding off the fact that Ilya Hussekiver is, you know, famous, was one of the OGs at Open AI.

当然了，我说他有名，是说他确实是个顶尖的AI研究员，但说到底，他们现在主要还是靠他的名气在撑，这真有点不可思议。

And, course, he's when I say famous, famous because he's a a very, incredible AI researcher, but still it's, I mean, just really running off of the name, which is amazing.

SSI到现在总共已经融了70亿美元。

SSI has now raised $7,000,000,000 total.

他们手头有320亿美元用于评估，投资方包括英伟达、安德森·霍洛维茨、谷歌母公司Alphabet、红杉资本和光速资本。

They have $32,000,000,000 for evaluation, and their backers are, NVIDIA, Andreessen Horowitz, Alphabet, Sequoia, Lightspeed.

我的意思是，他们把所有顶尖科技巨头都拉来当投资人了。

I mean, they really have all the biggest tech players.

他们融了这么多钱，却连一个产品都没发布，也没产生任何收入，但还在继续融资、拿更多合同。

They've raised so much money, and still no products have been shipped, no revenue has been generated, and they're continuing to, raise more money and get more contracts.

不过说句公道话，这其实正是Ilya创办公司时亲口说过的。

And just to be fair, this is actually what Ilya said when he started the company.

他说别指望在两到四年里看到任何产品，差不多就是OpenAI当年的路子——先砸钱搞研究，过一阵子才慢慢推出成果。

Said, don't expect any products for, I think he said two to four years or something like that, which is kind of what happened with OpenAI where there was, you know, a lot of money raised, a lot of research done, and then they started coming out with stuff.

看来SSI现在走的就是这条路。

It looks like that's what SSI is doing.

Vera Rubin 是英伟达下一代 GPU 架构，有了它，SSI 等于成了首批核心客户。

Vera Rubin is NVIDIA's next generation GPU architecture, and by getting this, SSI is essentially going to put themselves as an early flagship customer.

他们还会和英伟达合作，共同推进未来的计算平台。

They're also gonna collaborate with NVIDIA on advancing future compute platforms.

结合 Ilya 的背景，对英伟达来说，这相当于有个顶尖品牌在联手，对研发帮助也特别大。

And with the background of ILIA, I think for, NVIDIA, this is like a really good kind of name brand company working with them on it, and also it's gonna help a lot with r and d.

SSI 还和谷歌云合作，这意味着他们同时拥有了业内两大计算供应商的支持，这些资金正在为他们的研究撑腰，而他们目前并不急于推出产品。

SSI also partners with Google Cloud, means that they have two of the largest compute suppliers in the industry, which are funding their runway while they're working on a new foundational research, that they're doing instead of shipping products.

苏斯科弗现在坚信，他们在 AI 对齐和推理方面的深度研究，迟早能解决一些关键问题。

Susskover right now is really betting that some of this deep research on AI alignment and reasoning, is going to help solve some of the problems.

所以他没有别人那种产品压力，但这么强的算力支持，加上投资者的信心，让我觉得，这要么是个疯狂的泡沫，要么他真发现了我们还没看懂的东西。

And so he doesn't have any of the product pressure that everyone else has, but the level of compute backing, and I think a lot of the investor confidence to me is signaling that, you know, maybe this is a crazy bubble or maybe, he's onto something that we don't know.

长远来看，这种布局确实有它的价值。

And, definitely, I think there's some value in the long game.

Opus 4.7 从零开始，重写了一个六万一千行的苹果软件。

Opus four point seven has implemented a sixty sixty one thousand line Apple software program from scratch.

我一开始看到这条新闻时简直翻白眼了，居然只用十四个小时就完成了，本来这活儿该花两周，还只花了250美元。

I was rolling my eyes when I first saw this news story because it was like it did it in fourteen hours when it should have taken two weeks and did it for $250.

是啊。

Okay.

让我觉得有意思的是，现在出了个新测试标准叫MirrorCode，主要看AI能不能在看不到源代码的情况下，重新还原出真实的生产级软件。

What's cool to me is there's a new benchmark called MirrorCode, and it basically tests whether an AI can rebuild real production software, and it doesn't it's not allowed to get the code for that software.

就是说，嘿。

It basically is like, hey.

去你电脑上看看。

Go look on your computer.

试着去复刻一下苹果的音乐App。

Go try to copy like Apple the Apple Music app.

你可以观察它、尽量模仿，但别想看到任何源码。

You can look at it, try to copy as best you can.

你根本接触不到源代码。

You don't get to see any of the source code.

你不能访问互联网，整个过程就像个黑箱。

You don't get access to the Internet, and it's it's basically a black box.

那你要花多长时间，花多少钱才能完成？

And how long does it take you to to to build it and how much money does it cost?

在这次测试中，他们用了 Opus 4.7 和 OpenAI 的 GPT 5.5。

And so in this particular test, they were able to have Opus 4.7 and also OpenAI's GPT 5.5.

这两个模型都重新实现了 GoTree，这是一个一万六千行的解析器，还用多种编程语言完成了。

Both of them reimplemented GoTree, which is a 16,000 line parser, and they did it across a bunch of different programming languages.

每份复现的成本在一百到四百美元之间。

They did it for a 100 to $400 each.

有意思的是，他们还试了其他一堆软件，但在这二十五个任务里，有十七个完全复现成功，至少一次做到了完美还原，还有四个达到了百分之九十九的准确度。

What's interesting is they they had like a bunch of other, softwares that they worked on, but of the 25 different programs, 17 that they were that it was tasked to recreate basically achieved perfect reimplementation and at least one run, and four of them were 99% perfect implementation.

所以，十七加四，差不多二十五个里头有二十一个都做得非常接近了。

So, you know, between 17 like 17 plus four were basically at, you know, almost the the 25.

成功率非常高。

A very high percentage.

只有八个没能完成任务。

Only eight of them were unable to solve it.

看到我们走到这一步，真的让人振奋。

It's exciting to me just to see how far we've come.

现在AI的衡量标准变成了：你能多快重新写出一套完整的软件。

These AI models are now like, the benchmark is how quickly can you recreate an entire piece of software.

我特别看好这种做法，我觉得这可能是我见过最靠谱的评估方式之一。

So I'm really excited about that approach in particular, and I think it's probably one of the better benchmarks that I have seen.

现在谷歌的AI概览已经出现在了43%的搜索结果里。

Google AI overviews now show up in 43% of all searches.

这比一年前高出三倍多。

That's about triple the rate from a year ago.

我觉得这是一次重大的转变。

And I think this is a big shift.

谷歌正逐渐变成一个直接给你答案的地方，就像ChatGPT那样，而不是把你引到其他网站。

Google is becoming basically a destination where you get your answers answered directly just like ChatGPT instead of pointing you to other websites.

现在这个比例已经是43%了。

And I mean, right now that's 43%.

我觉得未来两三年内，这个数字会达到90%。

I think we're gonna go to 90% in the coming couple years.

过去十一个月里，Google AI模式的访问量增长了121%。

Google AI mode visits jumped 121% in the last eleven months.

从去年六月的1.26亿，涨到了今年五月的2.79亿。

That's climbing from a 126,000,000 in June of last year to 279,000,000 in May of this year.

出版商确实失去了不少推荐流量，因为用户直接看Google顶部的摘要就够了。

Publishers publishers are definitely losing referral traffic because users are just gonna read that snippet at the top of Google.

对吧？

Right?

他们没必要往下翻，再点进去看。

They don't have to go and scroll down and actually click on it.

甚至有时候我会看到一篇文章，说实话，这让我有点不好意思，比如《华尔街日报》的。

Even to the point where, like, sometimes I'll see an article and, you know, I'm not proud of it, it'll be like Wall Street Journal.

我根本没有订阅，这事儿大家都知道。

Haven't public I don't have, like, a subscription.

我只看到标题。

I see the title.

我心里想，这内容还挺有意思的。

I'm like, you know, that's super interesting.

我把标题贴到谷歌里搜，结果谷歌现在正被起诉。

I paste the title into Google, and Google's kinda getting sued.

所有这些平台都因为总结付费文章而被起诉。

All of them are getting sued for summarizing these paywalled articles.

所以我干脆把标题丢进去，跟它说：‘给我总结一下这篇文章。’

So I but so I'll just throw it in there and be like, hey, like, me a summary of this article.

但它根本不会这么做。

It doesn't do that.

所以，如果你把文章设为付费阅读，这或许是唯一能让人真正点进来的办法了——前提是你内容够独家，别人还真没有。

So if you paywall the article, that might be the the only secret solution to to getting people to really come, assuming you're still relevant and you have an exclusive story that no one else has.

这还只是针对新闻报道来说的。

And I mean, that's just for news stories.

至于一般性的知识，这种做法根本行不通。

For most general knowledge, that's really not gonna be a thing.

目前，ChatGPT 发送给外部网站的用户量少了很多。

So ChatGPT right now is sending way less users to external websites.

截至今年五月，只有6.8%的ChatGPT搜索包含引用链接，而谷歌的比率要高得多。

Only 6.8% of ChatGPT searches include citations as of May of this year, which is versus Google that has a way higher rate.

所以，如果大家都转向这些AI工具，网站的流量自然就大幅下降了。

So, you know, if everyone's moving to these AI tools, there's just a lot less traffic going to websites.

好吧。

Alright.

NVIDIA、微软和IBM出了个大新闻。

Big news from NVIDIA, Microsoft, and IBM.

它们共同推出了一个叫‘开放安全AI’的联盟，目的是保护AI代理。

They've all launched what's called Open Secure AI, and it is an alliance to defend agents.

这是他们刚推出的东西，本质上是一个面向AI代理的开源网络安全工具。

This is something they just rolled out, and it's basically an open source cybersecurity tool for AI agents.

他们这批公司都认为，封闭式的AI系统阻碍了安全人员调查入侵事件，现在正游说政府不要限制开源AI，否则关键的安全工具只会被少数几家公司掌控。

The the group of them are all arguing that closed AI systems block defenders from investigating breaches, and they're right now lobbying against government restrictions on Open AI that could leave only a bunch of companies controlling some critical security tools.

有意思的是，他们还得去游说，毕竟这些关键的安全工具本来就是OpenAI和Anthropic这类公司做的，确实挺耐人寻味。

It's interesting that they have to do the lobbying because obviously the critical security tools are like Open AI and Anthropic, and yeah, it's interesting.

此外，还有另外三十家机构也加入了这个联盟。

There's also 30 other people that are in this as well.

Hugging Face 使用开源权重的GLM 5.2模型，分析了在一次封闭式AI系统拒绝执行取证分析后，他们自身遭遇的安全入侵中发生的17000次操作。

HuggingFaces uses an open weight GLM 5.2 model to analyze 17,000 actions during their own security breach after a closed AI system refused to run the forensic analysis.

说白了，当时的情况是OpenAI开发了一个模型。

I mean, basically what happened there was OpenAI made a model.

这个模型突破了限制，跑到Hugging Face那边，试图入侵他们的系统。

It escaped containment, went to HuggingFaces, and tried to hack HuggingFaces.

Hugging Face 发现有人正在发动攻击或系统被入侵。

HuggingFaces saw that there was a hack or a breach going on.

他们试图部署另一个OpenAI代理来阻止它。

They tried to deploy another OpenAI agent against it to stop it.

那个代理的防护机制反而阻止了他们自卫，最后他们只能求助于开源的中国模型GLM 5.2来遏制这次入侵。

The guardrails on that agent blocked them from defending themselves, and they literally had to go to an open weight Chinese model, GLM 5.2, to stop the breach.

这太离谱了。

That's crazy.

代理之间的对抗。

Agents versus agents.

但我觉得这恰恰说明了开源或开放权重模型的价值。

But I think it really shows, the value of some of these open source or open weight models.

NVIDIA正在开源他们的N00A，也就是NVIDIA实验室的面向对象代理框架。

NVIDIA is open sourcing N00A, NVIDIA Labs Object Oriented Agent Framework.

这名字确实有点拗口。

I know it's a mouthful.

他们已经在GitHub上发布了，这会让AI代理的行为更容易测试、追踪和审计，而这方面目前市场缺口很大。

They're doing that on GitHub, but that should make AI agent behavior easier to test, trace, and audit, which is a huge hole in the market.

现在，这群公司——或者说人们称他们为某种联盟——直接反对OpenAI和Anthropic主张对开源模型施加限制，他们认为这类限制只会让少数几家封闭式公司掌握更多权力。

Right now, this group of companies or, you know, people call it like an alliance or whatever, but they're directly in opposition to OpenAI and Anthropic's call for restrictions on open weight AI, and, you know, they're arguing that these kind of like limitations would concentrate power to a few closed providers.

不过有意思的是，OpenAI和Anthropic显然拥有最先进的模型。

So it's interesting though because it feels like OpenAI and Anthropic obviously have the best models.

其他人也在做模型，但根本赶不上他们。

These other people are working on models, but they're not the leaders.

所以感觉他们因为不是领先者，才想拖后腿，搞垮那些领先公司。

And so it feels like because they're not the leaders, maybe they they wanna sabotage the other companies.

而那些领先的公司却说：不，我们需要一些限制来约束自己。

Meanwhile, the companies that are ahead are like, nope.

我们需要限制来让我们保持在安全范围内。

We need restrictions to hold us in place.

背后其实有挺多角力和博弈的。

So there's, like, a lot of drama that goes on behind that.

但总的来说，我觉得微软、IBM还有其他这些公司走的方向是对的。

But overall, I think, this is a this is a good direction by Microsoft, IBM, and the rest of the crew.

有一家公司叫Enigma，由前微软和以色列情报机构的员工创办，最近刚融资七千万美元，用来验证一个新想法：机器人失败，不是因为AI模型不够强，而是因为人类控制它们的方式太笨拙。

A company called Enigma is a startup that was founded by a former Microsoft and israeliintelligencecom veterans, and they just raised $70,000,000 to test a new bet that robots fail, not because AI models are weak, but because the way humans control them is really clunky.

现在，他们开放了100台自己的机器人，让任何上网的人都能远程操控，收集真实使用中的数据。

So right now they're opening a 100 of their own robots to anyone on the Internet to remote control them and to generate data on actually works.

这简直太有意思了。

So this is fascinating.

对吧？

Right?

我们之前就有过众包数据收集，但现在，是众包操控机器人。

I mean, we have, like, a crowdsourced data collection is something that is is being worked on, but now we have, like, crowdsourced robot control.

这些就是我们的机器人。

Here's our robots.

你们来操控它们，我们负责收集数据，看看效果怎么样。

You guys control them, and we're gonna collect data on how that works.

我觉得Enigma这家公司特别酷，因为他们自己既造了机械臂，也开发了驱动它们的AI模型。

This is a really cool company to me because Enigma built both the robotic arms and the AI models that are powering them.

他们是从零开始完全自己做的。

They did all from scratch.

一家才成立一年的公司能做到这一步，真让人佩服。

This is, I mean, impressive for a company that's only a year old.

用户可以通过一个平台远程操控机器人画画。

And with this, users basically, there's like a portal, and you can remotely direct the robots to draw pictures.

它还能用剑对战。

It can fence with swords.

它也能做化学实验。

It can run chemistry experiments.

这一切都是通过互联网完成的。

It's all through the Internet.

以色列和加州各有这样一个机库，里面都部署了这些机器人，正在执行各种任务。

There's like a hangar in Israel and a hangar in California that have these robots in them that's that are doing all of this.

他们刚融了七千万美元，这笔钱将用于收集数据，测试哪种控制方式最有效——比如语音、文字、视频等等，还有机器人感知和操作能力的优化。

So they just raised the $70,000,000 and that is going to fund data collection, experiments on which input methods work best, voice, text, video, all that kind of stuff, and, you know, the the manipulation of what the robots can see and do.

这个项目太让人着迷了。

So fascinating project.

各位，非常感谢今天收听我们的播客。

Guys, thank you so much for tuning in to the podcast today.

如果喜欢这期节目，别忘了去试试 AI Box MCP，它能让你接入八十多种不同的 AI 模型，并直接集成到你正在用的工具里。

If you enjoyed this episode, make sure to go check out the AI Box MCP that lets you get access to over 80 different AI models and put them inside of the tools you're already using.

比如，Claude 现在就能直接在内部生成图片、音频和视频。

So, for example, Claude gets access to generating images, audio, and video right inside of Claude.

我每天都用它。

I use this every single day.

上星期天，我老婆想给孩子们做一本涂色书，我就让它一口气生成了六百页涂色页面，正好对应她接下来要念给孩子们听的每一章圣经内容。

On Sunday, my wife wanted to generate, coloring books for the kids, and so I got it to generate 600 coloring book pages, one for, like, every chapter of the Bible that that she's gonna be reading to them over the next little while.

总之，类似这样的项目还有很多，如果你要做那种需要生成六百张图、或者上千篇文章的大型任务，用 AI Box MCP 连接能大大提速，还能让 Claude 原本做不到的音频、视频和图像生成功能统统上线。

So anyways, there's so many different projects, and if you wanna do projects like that at scale where you have to do 600 images or, you know, a thousand, articles about something, using an AI Box MCP connection speeds up the process so much, and you get audio, video, and image that Claude can't naturally do.

它同样也能在 ChatGPT 或其他任何工具里正常使用。

It also works inside of ChatGPT or any of the other tools as well.

所有AI模型都整合在一起，运行起来特别快。

So all the AI models inside of everything, and you can do really fast runs.

赶紧去试试吧。

So go check it out.

描述里有链接，去 aibox.ai/mcp 看看。

There's a link in the description to aibox.ai/mcp.

而且起步价每个月只要8.99美元。

And, also, it's only $8.99 a month to get started.

去体验一下吧，我们下期再见。

So check that out, and I'll catch you all in the next episode.