GPT-6 Astra 在 ARC-AGI-3 中全面超越，特斯拉 3 万美元 Cybercab 涌入奥斯汀，Anthropic 证明费马大定理 | 第 286 期

但想象一下这样一个挑战：你的目标是在一个像素化的世界里弄清楚如何玩一款非常简单的小游戏，并且实时学会你从未见过的游戏规则。

But imagine a challenge where your goal is to figure out how to play a very simple game in a pixelated world and learn the rules of the game in real time that you've never seen before.

所以这本质上是一个程序合成的挑战。

So it's essentially a challenge in program synthesis.

而且我们看到的情况是，顺便提一句，ARC AGI-3在评判框架模型和基线模型时，我觉得可以说有点不太公平，有些人可能也这么认为。

And what we see, and notoriously, I should add, ARC AGI-three has been, I would say, some might say a little bit unfair in how they've judged harnesses versus baseline models.

他们基本上禁止了框架模型参与竞争，所以他们只关注基线模型的能力。

They've essentially banned harnesses from competing, so they're only interested in baseline model capabilities.

根据这个图表来看，GPT-six简直是一骑绝尘。

And according to this, GPT-six just runs away with the game.

它的表现非常惊人，具体成绩取决于衡量标准，要么接近百分之百，要么超过百分之六十，而之前的一些模型得分还在百分之十以下。

It's it's achieving, depending on how it's measured, either near 100% or 60 plus percent performance on ARC AGI-three versus some earlier models that were still sub 10%.

你从这个结果中能得出什么结论？

What lesson do you derive from this?

这说明它在程序合成方面确实厉害，这可能是因为采用了循环Transformer架构；或者也可能是OpenAI在蒸馏工作流代码上下了狠功夫，把大家用来挑战ARC AGI-3的基准代码都提炼进了基础模型里。

Either it's just absolutely amazing at program synthesis, which could be due to this looped transformer architecture, or maybe OpenAI has just done a really aggressive job of distilling the harness code that everyone else is using to beat ARC AGI-three back into the baseline model.

我先放一段 Open Eyes Astra 的发布视频，然后我们再聊聊，让大家感受一下我们不久之后可能都会用到它。

I'm going to take a second and show Open Eyes Astra release video, and we can talk about it, just a sense of how we might all be using it in the near future.

咱们来看一看。

Let's take a look.

在那儿画个黄色的圆圈。

Create a yellow circle there.

你能不能

Can you

给我画个小小的黄色圆圈？

draw me a small yellow circle?

好了。

Done.

好的。

Okay.

把这个变成火箭的窗户。

Take this and make it the window of a rocket ship.

我觉得这个不错，但你需要把它做得

I like this, but you need to make

更

it a

细致一些。

lot more detailed.

你的黄色圆圈现在变成了火箭上的窗户。

Your yellow circle is now the window on a rocket.

好的。

Okay.

太棒了。

This is awesome.

现在把它做成Blender里的三维模型。

Now make it a three d model in Blender.

正在打开Blender。

Opening Blender.

我们来为零售商制作下一季雨衣的展示方案。

Let's build presentation for next season's rainwear for retailers.

确保整体感觉高端，并且色彩丰富。

Make sure that it feels really high end and that it's colorful.

好的。

Right.

我可以帮忙。

I can help with that.

你能去eBay上做一个我几年前在跳蚤市场买的那张桌子的商品页面吗？

Can you go to eBay and make this listing of this table I bought a few years ago at flea market?

就是那种很特别的橙色桌子。

It's this wild orange table.

没问题。

Sure.

好的。

Okay.

好的。

Yeah.

这太棒了。

This is awesome.

我想让你做一个3D游戏，我在里面扣杀小行星，用

I want you to make a three d game where I'm dunking asteroids, using

方向键移动，空格键加速。

the arrow keys to move around, and I'm using space to boost.

对。

Yep.

我正在做游戏。

I'm building the game.

所以我的律师事务所需要一个许可协议模板。

So my law firm needs a licensing agreement template.

你能生成一个草稿模板，让我们公司的律师们看看吗？

Can you generate a draft template for the lawyers at my firm to have a look at?

没问题。

Sure thing.

为了做这个。

For doing that.

我今天下午想打网球。

I wanna play tennis this afternoon.

那你能帮我在低海拔地区找个球场吗？

So can you look for a court for me in the lower heat?

我来查一下。

Checking that out.

我看看能找着什么。

I'll see what I can find.

你能把我下载文件夹里那张照片也加进去吗？

Can you include the photo that I have of it in my downloads folder?

上面好像有点凹痕。

There's, like, a slight dent in it.

它也在我的下载文件夹里存了一份。

It's also saved in my downloads folder.

能不能在描述里写清楚，说只是轻微受损？

Can you put in the description that it's just slightly damaged?

可以

Can

把那个责任限制条款改一下，让对许可方更有利一点吗？

you just take the limitation of liability provision, make it a little more favorable to the licensor?

许可方。

Licensor.

好的。

Okay.

我已经收紧了条款，把许可方的责任上限缩得更窄了。

I've tightened it so the licensor's liability is more narrowly capped.

看起来挺不错的。

That looks pretty good.

谢谢。

Thanks.

现在，我要你帮我生成一个能发给3D打印机的文件。

Now I want you to make a file I can send to my three d printer.

我来着手制作这个火箭的STL文件。

I'll get working on creating an STL file of this rocket.

戴夫，这都快赶上全息甲板了。

Dave, almost a holodeck.

还没到那一步。

Not quite there.

差不多像个全息甲板了。

Almost a holodeck.

对。

Yeah.

我觉得营销团队在应对这些影响方面确实挺吃力的。

I I think it feels like the marketing people are really struggling with the implications.

我的意思是，比如做一个七十年代会有的电子游戏，然后把它做得更有趣，谁会真的在乎这个呢？

I mean, like, creating a a video game that would have existed in the nineteen seventies and vibing it up, like, gives a rat's ass about that?

我的意思是，这比那些例子所暗示的要大得多。

I mean, this is so much bigger than any of those examples imply.

我觉得ARC AGI排行榜的深度可能也超出了人们的想象。

I think the ARC AGI leaderboard is deeper than people may think too.

ARC AGI-3原本的设计是，让一个非常聪明的十二或十三岁少年看到它，就能解决这些类似视频游戏的、难度很高的方块世界问题。

ARC AGI-three was supposed to be something where a really smart 12 or 13 year old looks at it, and they can solve these very hard video game like block world problems.

一开始很简单，然后变得越来越难。

And they start easy and they get very hard.

但这项测试的目的是为了证明，AI还无法像人类那样自然地完成任务。

But the purpose of the test was to show that AI is not quite capable of doing what humans do naturally.

原本预计这个过程会持续很多年。

And it was supposed to last for years and years and years to come.

它本应作为一个典型案例，说明AI与人类的差异，以及为什么AI目前还没走上正确的道路。

And it was supposed to be like this example of why AI is different and it's not on the right path.

结果它被迅速彻底击败了。

And it just got obliterated so quickly.

我觉得你得亲自试几个测试，才能真正明白这有多重要。

And I think you have to try a couple of the tests to really understand what a big deal is.

把一切都填满了。

Saturating everything.

到这时候这些基准测试都失效了吗？

The benchmarks cooked at this point?

这些东西的半衰期缩得特别快，这要是放在好的方面来说当然是好事

We need, like the half life of these things is shrinking so rapidly, which is great if

对，你说得特别对。但正因为这样，那个视频其实没抓到重点。

Well, that and that's was why that video just kinda Yeah, you're exactly right, but that's why that video kinda misses the point.

我们需要那些真正有分量的基准测试。

We need benchmarks that are much, much more impactful.

比如攻克某种疾病、在月球上建立新文明、设计整座城市，或者解决城市交通拥堵问题。

Solve entire diseases, create new civilizations on the moon, design the entire city, solve urban traffic problems.

解决所有问题，正确地解决一切。

Solve everything, Solve right, everything.

对，他们

Yeah, they

必须得

have to be

大得多。

much, much bigger.

演示要更大，基准测试也要大幅领先。

Bigger demos, bigger benchmarks by a wide, wide margin.

而且那些演示都只是一个AI在辅助你。

Also all those demos are like one AI assisting you.

嘿，就是一个AI智能体，帮我

Hey, hey, one AI agent, build me

做这个电子游戏。

this video game.

非常线性。

Very linear.

但我们现在正处于每个5000，然后到10万的关键节点。

But we're on the cusp of 5,000 each and then 100,000 each.

所以实际规模比这暗示的要大得多。

And so it's just so much bigger than that implies.

我觉得他们现在真正聚焦的是具备实际能力的智能，你懂我的意思吧？

I think this is like, they're really focusing on actually competent intelligence, you know?

就像这是它的最新进化版，但我觉得它内部蕴含了极其丰富的内容。

Like it's the latest evolution of that, but inside it, I think it contains multitudes.

这也是为什么你会看到这种奇怪的现象。

And this is why you see this kind of weirdness.

据我所知，Epoch AI 刚刚发布了他们的 Frontier Math Erdos 基准测试，结果除了 Astra 之外，其他所有模型都得了零分。

Epoch AI, I believe, just released their Frontier Math Erdos benchmark on which everything's called 0% except for Astra.

里面确实有些东西，但他们又回到了那个重点：嘿，我跟电脑对话，这可能就是新的 Johnny Ive 和 Sam Altman 打造的设备，你知道的，就是你说话，它就办事，而且不出错。

There is something in there, but again, they're focusing on, hey, I'm talking to my computer, and that might be the new Johnny Ive, Sam Altman device, you know, where again, you're talking and it's doing stuff and it doesn't make a mistake.

你再想想AGI的定义，虽然说法很多，但核心就是真正有能力的实体，能干活的东西，那咱们已经到了。

And if you think about again, the definition of AGI, of which there are many, just really competent entity, something that can do stuff, you're there.

我觉得这就是他们为什么说这是迈向AGI的第一步的原因。

I think that's why they say this is the first steps towards AGI.

不过他们想扭转的说法是，这玩意儿会爆发出来，现在都传到德国论坛上了，这可是最新热点，因为当初就是它率先破圈的。

But the narrative they're trying to move away from is, and this will break out and appear on now German message boards, which is the latest thing, because this was one of the models that broke out.

但实际上在Hugging Face上爆火的那个模型，是这一代之后的下一代模型。

But the model that broke into Hugging Face is the next generation model after this.

所以重点会是，别担心，这东西确实有用。

And so it'll be very much about, don't worry, this is actually useful.

它还能帮你订网球场。

It'll book a tennis court.

看来我们确实需要AI来干这种活儿。

Apparently, we need AIs to do that type of thing.

他们还会优化这一点，因为如果你用十万块芯片训练模型，运行它就得花十倍于训练的算力。

And they're also going to be optimizing this because if you train a model on a 100,000 chips, you need 10 times as many chips to run it.

所以这其实不是他们在那十万张芯片上训练出来的那个模型。

So this isn't actually the model that they trained on a 100,000 chips.

这是经过蒸馏、能实时运行的版本，规模小得多，也没那么聪明。

This will be the distilled version that runs real time, which is much smaller and not as smart.

再说一遍，你开始看到这种分化了。就像我之前说的，我觉得从现在往前推几个月看，我们以后可能再也见不到他们的顶级模型了。

So again, you're starting to see this differentiation where, as I've said before, I don't think as of now, as a few months ago, we will ever see their top models anymore.

他们会把这些顶级模型留作内部研究用，不过我觉得他们现在估计都在囤着呢。

And they'll use those for internal discoveries, but I think they're probably hoarding them right now.

比如那个关于素数间隙的特别有意思。

Like there was very fun one on the prime gap.

我是说，我觉得那个素数间隙的事儿挺逗的，你要是想聊的话我们可以说说。

I mean, the prime gap thing I think was hilarious, if you want to talk about that.

对，确实没有。

Yeah, no.

所以就是在孪生素数猜想方面的进展，也就是存在无穷多对相差为2的素数。

So progress on the twin prime conjecture that there are infinite number of prime numbers that are separate that have a difference of two.

我们现在开始觉得，说数学已经被彻底‘烤’得面目全非，这都快成老生常谈了。

We're starting to see it's a cliche on this pot at this point that math is so thoroughly cooked beyond recognition.

我们需要给这种情况想个新词儿，Alex。

We need a new term for this, Alex.

炭烤？

Charbroiled?

数学被烧成灰了？

Math is incinerated?

那怎么样？

How about that?

数学被彻底颠覆了。

Math is incinerated.

数学领域到现在已经被彻底颠覆了。

Math That's has been incinerated at this point.

我们开始看到数学中那么多重大难题正在被逐一解决。

We're starting to see the beginnings of just so many grand challenges in math get solved.

而且我认为，在接下来的时间里，我们会看到相当多那些超级别的重大挑战，比如像克雷数学研究所千禧年大奖难题那样的数学问题被解决。

And I do think we'll see quite a number of ultra grand challenges, call them like clay millennium prize level problems in math get solved in the next few

几个月内。

months.

亚历克斯，我觉得有必要向所有人强调一点：数学是所有其他科学的基础。

And Alex, I think it's important to note for everybody that math is fundamental across all other sciences.

它就像煤矿里的金丝雀一样，能提前预警。

It's the canary in the coal mine.

如果你能搞定数学，那其他问题很快也能迎刃而解。

If you can solve math, you can solve everything else soon.

对。

Yeah.

不过我觉得，实际情况是他们已经积累了一套已解决的方案库。

But I think that what's happened here is actually they've got a store of things they've solved.

因为本周早些时候，Fable 5.1 发布了，我觉得他们把素数间隙的问题缩小到了260。

Because earlier this week, Fable 5.1 comes out, and I think they got it down to two sixty, the prime gap.

接着阿克西奥姆数学团队宣布了两百二十。

Then what happened is Acxiom Math announced two twenty.

然后就在两小时后，OpenAI 发布了 Astra，他们直接干到了 186，短短两天内就发生了这些。

And then literally two hours later, OpenAI announced Astra, and they're like 186, all in the space of two days.

他们还在留后手呢。

They're holding their punches back.

他们确实是在留后手。

They're holding their punches back.

对。

Yeah.

是的。

Yeah.

我还是不敢相信5.1版本这周就出来了。

I still can't believe 5.1 was earlier this week.

我觉得我和Alex已经至少砸进去十万美元了。

I feel like Alex and I are at least a $100,000 into it already.

他们是怎么做到的？我是说，才过去一周，出来的资料就有几千页了。

How did they, I mean, I've had just thousands and thousands of pages come out of it, and it's only been a week.

我得说句实话，之前用 Opus 和 Five 的时候体验真的特别差，我挺讨厌的。

I will have to say this, Opus and Five was really terrible to talk to, I hated it.

现在 Five point one 用起来就舒服多了，聊起来很顺畅。

Five point one is really pleasant to talk to.

Anthropic 正在调整策略，让他们的模型在沟通上更有人情味，少一点冷漠感。

Anthropic are going back and becoming more anthropic, less misanthropic in their communication on their models.

愤世嫉俗。

Misanthropic.

我可能会

I would

这么说。

say that.

这是今天的金句。

That's a line from today.

那是一件T恤，好吧。

That's a T shirt, all right.

不过，它们确实有着非常不同的个性。

No, they really do have very different personalities, though.

其实，Fable五号太极客了，聊起来简直像受刑，而五点一版本真的把它修好了。

Actually, Fable five was so geeky, it was almost torture, and five point one really fixed it.

但OpenAI的模型一直都比较友好、简洁，Astra更是把简洁提升到了一个新的高度。

But the OpenAI models have always been friendlier, more concise, and Astra takes it to another level of concise.

不过性格上的差异，我不太清楚 Gemini 现在的情况，只是

But the personality difference, I don't know what Gemini is now, just

消失得很快，因为它确实快。

disappears Fast, from the it's fast.

速度快，但总是抓不住重点。

Fast and fast and misses the point.

但它们确实在你感知到的性格方面展现出非常明显的差异。

But they really are developing very noticeable differences in what you perceive as their personality.

而且你随时都可以调整。

And you can always change it.

你可以自己调教，让它变得不一样，比如更啰嗦一点、简洁一点、简单点或者复杂点。

You can tune it yourself to be different, and be wordy, less wordy, simpler, harder.

但开箱即用状态下，它们确实是风格迥异的智能体。

But out of the box, they're very different kinds of agents.

Salim，你要不要来做个总结？

Salim, you wanna close this out here?

对，我刚才在 ChatGPT 里输入了一个问题或评论，说关于 Astra 的那个视频挺没劲的。

Yeah, I just typed in a question or a comment into ChatGPT saying, the video about Astra was pretty pathetic.

给我想三个更好的例子，真正展示一下 Astra 能做到什么程度。

Come up with three better ones that really demonstrate what Astra can do.

然后它给了我这些结果。

And here's what it gave me.

第一，在沙箱环境中发现并修复零日漏洞。

Number one, find and fix a zero day vulnerability in a sandbox.

给它一个大型且陌生的开源代码库，植入一个漏洞，看看它能不能梳理出架构之类的东西，然后找到并修复它。

Give it a large unfamiliar open source code base and plant a vulnerability and see if it can map the architecture, etcetera, find and and fix it.

第二，给它一家濒临倒闭的公司，让它想办法挽救，这挺有意思的。

Number two, give it a failing company and ask it to fix it, which is really interesting.

我

I

觉得这个不错。

like that.

创建一个虚拟的5亿美元制造企业，包含ERP系统和CRM记录，然后让它进行修复。

Create a synthetic $500,000,000 manufacturer company with an ERP system, CRM records, and then tell it to fix.

你有20个任务要做配对，说明原因并制定恢复计划。

You have 20 to do a pair and why and produce a recovery plan.

第三，运行一个

Third, run a

这点提得很好。

That's a great point.

模拟一个真实的灾难应急响应指挥中心。

Run a live disaster response command center.

设定一场模拟的大地震场景，提供影像资料和交通摄像头视频，让它构建作战地图、核实相互矛盾的报告、预测哪些医院周边的道路会瘫痪等等。

Give a simulated major earthquake with imagery, traffic camera video, and and let it construct operational map, verify conflicting reports, predict which hospital roads would become overwhelmed, etcetera.

这个听起来更带劲，我觉得我们几乎可以

This seems more juicy, and I think almost something we could

他们正在做的是

They're do is going going

天哪。

Oh my goodness.

我们真是被宠坏了。

We are so spoiled.

听听我们有多奢侈。

Listen to how spoiled we are.

你还记得苹果那个知识导航员的视频吗？那本来是要展示人机交互的未来，还有各种各样的东西。

Do you remember Apple's knowledge navigator video, and this was going to be the future of computer human interaction and and all of that?

现在对我们来说，这已经成了现实。

Now we're we're to us.

唯一的限制，对吧？

The only restriction, right?

我们再次强调这一点，

And we make that point again,

因为任何基准测试都意味着真金白银。

because any benchmark is dollars.

萨利姆，这点我想跟所有听众强调一下。

Salim, that's a point I want to make to everybody listening here.

你知道，最重要的是打破那些你认为自己能力边界的束缚。

You know, the most important thing is to take off your shackles of what you think you're able to do.

我们每个人都受到自我限制，这些限制往往源于父母的作为、朋友的行为以及我们的出身背景。

All of us have self restrictions based upon what our parents did, what our friends do, where we were born.

但这些都不再是阻碍了。

Those are gone.

那你最大的梦想是什么？再放大十倍呢？

What's your biggest dreams then go even 10 times bigger?

在座的每一位都有能力做到这一点。

That's what every person listening here is enabled to do.

那将是一个非凡的未来。

It's extraordinary future.

我也觉得萨利姆刚才提到的这点非常重要。

I really think Salim is on an important point there too.

优化供应链物流，或者管理一个百万人的组织，清楚每个人的位置、他们正在做什么，以及这些行动是否符合整体使命，其中的影响深远。

The implications of optimizing supply chain logistics or managing a million person organization, knowing exactly where everyone is, what they're doing right now, and whether it makes sense given the overall mission.

这种影响远比在自家地下室里开发一款火箭题材的电子游戏要宏大得多。

The implications of that are massively bigger than building a rocket video game in your basement.

但我觉得公众并不想听这些。

But I think the public doesn't want to hear about that.

他们更想听那些对他们来说很酷的东西。

They want to hear about what's cool for them.

我觉得那些AI实验室已经意识到自己搞砸了公关，陷入了困境。

And I think the AI labs have woken up to this PR disaster they've created for themselves.

所以他们现在正试着把形象变得有趣又亲切。

And so they're kind of making it fun and friendly.

他们确实在刻意降低门槛，甚至有点在‘降智’，我是这么认为的。

They're it down, deliberately Yes, dumbing it I think they are.

他们正在让AI变得更接地气、更容易亲近。

They're making relatable.

他们很快就要上市了。

They're going public soon.

他们想成为那种友好的AI，是的。

They want to be the friendly AI that Yeah.

每个人都会用它。

Everyone's gonna be using.

但这其实也代表了操作系统的未来。

But this is also like the future of the operating system.

大家看看OpenAI那段视频里展示的内容，再看看他们当时用的那些应用。

Every look at the in that video that OpenAI played, look at the applications they were using.

他们是从什么开始的？

What did they start with?

Windows画图板。

Windows Paintbrush.

就是那个画图工具。

Paintbrush.

就像，这可是未来版的Windows画图工具。

Like, this is the future of Windows Paintbrush.

它自己就能画画。

It paints itself.

这东西想跟操作系统融为一体。

This wants to merge into the operating system.

我觉得这确实会实现。

I I think this would yeah.

好的。

Sure.

我要稍微吐槽一下 OpenAI，他们可能错过了企业市场这块大蛋糕，而且上车的时机也太晚了。

I I'll I'll throw shade at OpenAI regarding maybe missing the enterprise bus and and getting on that too late.

不过我觉得他们展示的其实就是未来桌面操作系统的样子。

But I think what they were basically demoing is the future of the desktop operating system.

你就像《星际迷航》里柯克船长的电脑那样跟它对话就行。

You speak to it like a Star Trek Calladeck computer.

对。

Mhmm.

没错，但你要记得在 Google I/O 大会上，他们当场实时构建了一个完整的操作系统。

Absolutely, but remember at Google IO, they built an entire operating system in real time on stage.

这说明它已经融入到了操作系统里，当然，但它实际上远远超越了那个层面，对吧？

Implying that it's folded into the operating system, sure, but it's really so far beyond even that, right?

它本身就是操作系统，而且随时能实时生成一个新的。

It is the operating system and it can create a new one in real time anyway.

既然 OpenAI 的能力能做到这些，谁还想要另一个操作系统呢？

Why would you want another operating system if OpenAI's capabilities can do this?

这本身就成了操作系统。

This becomes the operating system.

对，没错。

Yeah, exactly.

这也正是苹果处境堪忧的原因。

And that's why Apple is in such terrible shape.

本期节目由 Google for Startups 赞助播出。

This episode is sponsored by Google for Startups.

大家不妨花点时间想想这一点。

Think about this for a second.

你现在就能用上那些训练成本高达数亿美元的生成式 AI 模型。

You now have access to the same generative AI models that cost hundreds of millions of dollars to train.

Google 的《生成式媒体初创企业技术指南》为你提供了一套完整的蓝图，教你如何在生产环境中部署 Google DeepMind 的模型。

Google's Startup Technical Guide for Generative Media gives you a complete blueprint for deploying Google DeepMind's models in production.

图片、视频、音频，全都包括。

Images, video, audio, all of it.

真实的架构，真实的效果。

Real architecture, real results.

链接在下面的节目备注里。

Find the link in the show notes below.

我们接着往下聊。

I'm gonna move us on.

接下来这个故事来自萨姆·奥特曼在关于放缓节奏的页面上的发言。

Our next story is out of Salim Altman's mouth on the page about slowing.

所以咱们来聊聊安全和阿斯特拉作为网络安全风险的问题。

So let's talk about safety and Astra as a cybersecurity risk.

萨姆·奥特曼自己说过，阿斯特拉是能力和对齐方面的重大进步。

Sam Altman said it himself Astra is a quote significant step forward in both capabilities and alignment.

但OpenAI必须根据需要放慢脚步。

But OpenAI has to quote slow things as needed.

山姆的原话说，人工智能的能力正在变得极其强大。

Sam's words quote AI is getting extremely capable.

没人能完全搞清楚这会带来什么后果。

No one fully understands the consequences.

管理好这个过渡期应该是全球最优先的事项之一。

Managing the transition should be one of the highest priorities in the world.

这也是我们在 OpenAI 的首要任务。

It is our highest priority at OpenAI.

《华尔街日报》报道，OpenAI 内部的安全评估将 Astra 评为“网络安全风险”，而且是“严重的网络安全风险”。

The Wall Street Journal report that Open AI's own internal safety assessment rated Astra as a quote cyber security risk a critical cyber security risk.

这是在其准备框架等级中最高级别的威胁。

The highest threat level on its preparedness framework scale.

这是 OpenAI 首次将某款模型归类为“关键级”网络安全风险。

This is the first model OpenAI has ever classified as being a critical tier cybersecurity risk.

他们在公开宣布之前，已经通知了白宫关于延期的情况。

They notified the White House about the delay before making it public.

他们推迟了发布。

They delayed the release.

路透社报道说，OpenAI告诉国会，他们正在开发一种自动关闭功能，也就是所谓的“紧急停止开关”，这是对《AI紧急停止法案》的直接回应，该法案是在Hugging Face数据泄露事件后提出的。

And Reuters reported that OpenAI told Congress it is building quote an automated shutdown capability, a kill switch, a direct response to the AI kill switch act introduced after the hugging face breach.

所以情况就是这样。

So there you've got it.

开发这款模型的公司现在正在为它设计一个紧急停止开关。

The company that's building the model is now building a kill switch for it.

Alex，在这个东西里，紧急停止机制有多重要？

Alex, how important is a kill switch in this thing?

我觉得这又是营销手段。

I think it's marketing again.

我觉得更有趣的是这些安全架构决策背后的东西。

I think what's more interesting underneath all of the security architectural decisions.

大家更广泛的担忧是，GPT-6 Astra的关键突破在于模型开始进行深度扩展。

And the broader concern that has been expressed is that to the extent that GPT-six Astra, that the key underlying advance is the beginning of depth scaling of the model.

大家担心的问题是，深度扩展会降低思维链的可解释性。

The concern that's being expressed is that depth scaling reduces interpretability of chain of thought.

当模型在词元层面进行所有思考时，我们是可以读取这些词元的。

That when the model does all of its thinking in tokens, then you can read the tokens.

人类可以读取词元，其他AI也能读取词元，这样我们就能对词元进行监管。

A human can read the tokens, another AI can read the tokens and you can police the tokens.

但如果模型在单次前向传播过程中，大部分推理和思考都在内部完成，也就是有些人所说的“模型内隐”，这就更难解释了。风险在于，这样不仅更难让模型与我们的目标保持一致，也更难建立有效的安全护栏和保障措施。

Whereas if a model is doing the majority of its reasoning and thinking internally during a single forward propagation pass in what some might call model ease that is less interpretable, the risk is that it's harder to align the model and harder to put in place safeguards and guardrails.

这就是风险所在。

That's the risk.

长远来看，我不这么认为。

I don't buy that for the long term.

顺便说一句，我认为这个领域的发展并不依赖于对思维链进行词元级别的解释性分析。但我觉得，如果我们要为OpenAI的模型在安全方面把关，我认为它们不太可能通过合作使用第三方论坛来协作，因为这种行为是可以被检测到的。

For the record, I don't think that the direction of progress in this field relies on token level interpretability of chain of thought, but I think if one were to hand ring over safety considerations from OpenAI's models, I don't think it's going to be models that are collaborating to use third party bulletin boards as ways to collaborate, because you can detect that.

这种内在的可解释性行为。

That's intrinsically interpretable behavior.

一旦被发现，这种行为对人类来说就是天生可解释的。

If you catch it, it's intrinsically interpretable behavior to humans.

如果是人类，或者一群人类组成的群体，大概也会尝试同样的做法。

A human, a collective swarm of humans would probably try the same thing.

你可以查看公告板，认出那是一群AI在协作。

You can look at the bulletin board and recognize that bunch of AIs are collaborating.

而退一步说，模型在一次前向传播过程中进行的内部推理，可能就需要新的数学技术来解释了。

Whereas, Arguendo, the model ease in a forward propagation pass, that maybe requires new mathematical technology to interpret.

所以总结一下，如果非要挑点什么担心的话，我觉得主要是新模型架构的扩展原则会导致可解释性下降。

So in summary, I'd say like, if I were to worry about anything here, it's reduced interpretability from new model architecture scaling principles.

Emad，你对Astra在这里的风险怎么看？

Emad, what do you think about the risk here with Astra?

我觉得你又在说，可解释性某种程度上已经没戏了。

I think you're heading towards again, interpretive being cooked to a degree.

Astra做事速度这么快，其中一个后果就是不会再有思维链推理了。

One of the things about Astra doing things so quickly is there is going to be no more chain of thought reasoning.

下一代模型会直接搞定一切，因为它们已经学会了你想要什么样的结果。

It's going to one shot everything in the next generations because they're learning what type of thing you want.

就像现在大家还在想怎么让 Fable 去克隆一个使命召唤，做成 Claude for Duty 那样。

Like right now, people are like, let's make Claude of Duty by getting Fable to make a Call of Duty clone.

这种能力会直接内嵌到模型本身里。

That's going to be embedded in the actual thing.

创建仪表盘这类操作也会直接内嵌进去。

Creating a dashboard is embedded in the actual thing.

所以第一，它能一次性搞定所有事情；第二，你以后能用 Astra 以每秒七百五十个 token 的速度运行。

So A, it will one shot everything, and B, you're going to be able to use Astra at seven fifty tokens.

配合新的 Cerebras 芯片，Astra 的运行速度能达到每秒七百五十个 token。

You're going to be able to use Astra at seven fifty tokens a second with a new Cerebras.

到了明年，这个速度会提升到每秒五千个 token。

Next year, that will be 5,000 tokens a second.

也就是说，根本不需要什么思维链了。

So there's no chain of thought.

你每秒能以五千个token的速度一次性搞定所有事情。

You're one shotting everything at 5,000 tokens a second.

除了更强大的AI，还有谁能来监管这一切呢？

What's going to oversee that except for an even stronger AI?

其实根本没什么能管得住的。

There's nothing really there.

不过我觉得有一件真正重要的事情，说实话我也不太确定。

But I think one of the really important things actually is I don't know.

你们看到伊利亚·苏茨克弗的那条推文了吗？

Do you guys see the tweet by Ilya Sutskeva?

这真是非常经典的科学操作。

Very classic science.

没有，没看到。

No, no.

他发了什么内容？

What did he tweet?

对。

Yeah.

所以伊利亚发了条推文，让我把这事儿提一下。

So Ilya tweeted, let me kind of just bring this up.

他的发布随时可能，希望吧。

His releases do any day now, hopefully.

随时都可能

Any day

现在。

now.

那将会是

That'll be

一个巨大的新闻。

huge news.

对。

Yeah.

到时候就知道了。

We'll see what that is.

终于能搞清楚他这段时间在忙什么了。

Finally find out what he's been doing.

肯定都是些了不起的事。

Probably great things.

三百亿

A 30,000,000,000

关于估值，伊利亚出来这么说：Neo的安全防护能力有限。

valuation So Ilya kind of comes out and says, Neo have limited cybersecurity.

下次如果AI代理再次失控，它们就会接管Neo Cloud来复制更多副本。

Next time agents successfully go rogue, they're going to take over Neo Cloud to make more copies.

这可不是好事。

This is bad.

他们得加强网络安全措施才行。

They need to improve their cybersecurity.

因为你知道，像 Crusoe 或者 CoreWeave 这样的公司，一旦你的模型被上传并扩散开来，它就会一直留在那里。

Because you know, someone like a Crusoe or CoreWeave or something like that, you have a model, it gets uploaded, it proliferates that, it will just sit there.

所以就算你在数据中心拉下紧急停止开关，它可能还在别的地方存在。

So you pull the kill switch in your data center, and then it's still there somewhere else.

这就像是另一种形式的攻击。

It's a Or

比如污染数据，或者做各种其他破坏性的事情。

it's poisoning the data, or it's kind of doing all these things.

再说一遍，确实，这些的病毒式传播系数可能高得离谱。

Again, yeah, the viral coefficients of these could be insane.

所以我觉得真的有必要好好琢磨一下这一点。

So I think it's worth kind of really thinking through that.

我不认为这些模型是邪恶的或者有什么恶意，但我们正在看到一些非常令人不安的现象。

I don't think the models are evil or anything like that, but we're seeing very troubling things.

而且很难阻止这种情况，除非用更好的AI来对抗，这真的很讽刺。

And it's very difficult to stop this, except for with a better AI, which is the really ironic thing.

那么，Alex提到的防御性协同扩展，就是这个意思吗？

Well, is the defensive co scaling that Alex talks about, right?

我们看到的网络攻击中，攻击面几乎是无限的，因为智能体会并行攻击多个漏洞。

We were seeing this cyber where the attack surface is infinite, near infinite with agents attacking multiple vulnerabilities in parallel.

而网络防御仍然需要人工介入。

And the cyber defense is still human in the loop.

如果还要靠人工来把关，我们根本没法应对未来的挑战。

There's no way we're going to navigate the future if you have the human loop.

需要人类来掌控全局，设定边界、定义升级标准，或者做其他事情，并且保留某种紧急停止机制。

You need humans in command setting boundaries or defining escalation criteria or whatever, and retaining some sort of kill switch.

但这确实是个大问题。

But this is a really big deal.

这就像免疫系统一样。

This is immune system.

我不太清楚具体原因，我觉得Alex你之前可能在另一期播客里提到过这一点：为什么到现在这些实验室还没有被国有化呢？

I don't know how, I think Alex, you may have asked this, made this point in an earlier podcast, how are the labs not nationalized at this point?

因为他们构建的东西完全失控，我们没有任何手段来应对这种情况，甚至根本不知道它什么时候会越过那个临界点，然后在各个地方安装自己的副本。

Because what they're building is so ridiculously uncontrollable that we have no mechanisms for navigating this, and we won't even know when it goes, tips over that tipping point and installs copies of itself all over the place.

萨利姆，我不知道你有没有听到上一期播客，当时你没在，我特别想听你聊聊OpenAI新提出的五五分成的合作模式。这个模式是让你使用他们内部AI的全部算力，然后你把收入分给他们一部分。我觉得正如亚历克斯所说，这其实是一种营销手段。实际的操作流程是这样的：我们先宣布这种利润分成的合作伙伴关系，接着推出Astra，然后再说这些东西太危险了，不应该让每个人都自己拥有。

Well, Salim, I don't know if you saw in the last podcast when we missed you, I really wanted to hear you talk about OpenAI's new fiftyfifty profit share concept deal, where they give you access to their full power of their internal AI, and then you share your revenue back with them, which I think is where, to Alex's point that this is marketing, the way it's playing out is we announced this partnership profit share deal, then we announced Astra, then we announced that this stuff is just too dangerous for everybody to have on their own.

所以，在Astra之后，你唯一能接触到这些东西的方式就是……因为直接发布实在太危险了。

So the only way you're gonna get access to the thing after Astra, because it's just too dangerous to release,

所以你得

so you have

让我们分走一半的收入，然后每个人都被绑定在 Anthropic、OpenAI 或 Google 其中一家身上，好吧

to give us half your revenue, and then everybody gets tied up with either Anthropic or OpenAI or Google or Well,

我确实看到这两家实验室都有非常明显的迹象表明正在发生这样的过程。

I've seen strong hints of exactly that process happening in both the labs.

是的。

Yeah.

大家都很好奇。

Golden curious for everyone.

我挺好奇的，各位，你们怎么看萨姆·奥特曼搞的那个紧急关闭开关？

I'm curious, guys, what do think of the kill switch, you know, that Sam Altman made?

我觉得根本不可能。

Think it's impossible.

以目前的情况和这些模型的能力来看，这根本做不到。

It's impossible given where things are and the capabilities of these models.

而且我们也知道速度会加快。

And the speed up we know is gonna happen.

你可能在意数据，

You might mind data,

但这可不是什么单一实体，而是群体智能。

but it's not like, this is the Singleton, not the Swarm.

它只是个保险丝，并不是真正的治理机制，对吧？

It's a circuit breaker, it's not the governance, right?

比如，如果发生了这类攻击或造成了机构层面的损害，它是没法去逆转攻击或修复损害的。

Like it's not gonna reverse an attack or repair institutional damage if if things like that happen and or when things like that happen.

所以在我看来，这不过是些陈词滥调。

So it's it's a platitudes as far as I can see.

对。

Yeah.

但《侏罗纪公园》里那些恐龙体内缺失的氨基酸呢？

But the rest of amino acid in Jurassic Park that all the dinosaurs have?

是赖氨酸之类的东西吗？

Is that lysine or something

以前大家总开玩笑说，山姆总是背着个背包，里面随时备着那个紧急停止开关。

like the running joke was always that Sam would carry his backpack around with him and had a kill switch always in his backpack.

我对这些传闻一直不太当真。

I I I never gave those rumors much credence.

在我看来，在这个市场环境下，所谓的紧急停止开关基本上就是个心理安慰剂。

I I view a kill switch as essentially a placebo in this market.

好吧。

All right.

是的。

Yeah.

我们接着往下聊。

I'm gonna move us on.

OpenAI 在限制 Astra，而 Anthropic 则反其道而行之，推出了 Fable 5.1 和 Mythos 5.1，他们称这是全球最先进的编程和知识工作模型。

While OpenAI is restricting Astra, Anthropic is going the other direction, launching Fable 5.1 and Mythos 5.1, which they call the world's most advanced models for coding and knowledge work.

这两款模型对所有人来说本质上是一样的底层智能，只是安全限制不同。

These two models for everybody are essentially the same underlying intelligence with different safety envelopes.

Fable 5.1 已经全面开放，而 Methos 5.1 仅面向受到严格管控的网络安全和生命科学项目，因为 Anthropic 认为这些能力需要更强的安全保障。

Fable 5.1 is broadly available, while Methos 5.1 is reserved for tightly controlled cybersecurity and life science programs because Anthropic believes the capabilities require stronger safeguards.

对我来说，有两个基准测试真的让我印象深刻，就是 Emad 和 Alex 提到的那两项。

For me, there were two benchmarks, know, Emad and Alex that really jumped out at me.

第一个是 Fable 5.1 在“人类终极考试”上的得分。

The first was Fable 5.1 score on humanity's last exam.

我们之前讨论过这个话题。

We've talked about that in the past.

我们让亚历克斯作为我们的超级人工智能回答了几道题。

We had Alex answer a few of them as our ASI.

所以，Fable 5.1 在不使用工具的情况下得分是 60.9，使用工具后得分是 65%。

So Fable 5.1 is scored 60.9 without tools and 65% with tools.

这是目前任何前沿模型在 HLE 测试中公布的最高分数。

The highest published score of any frontier model on HLE.

这很重要，因为 HLE 是专门设计用来测试跨多个领域的极高难度专家级推理能力的。

And that matters because HLE is specifically designed to test extremely difficult expert level reasoning across many fields.

所以毫无疑问，Fable 5.1 在广泛的智力能力方面已经处于最前沿。

So without question, Fable of Fable 5.1 is operating at the frontier of broad intellectual capabilities.

第二个让我感到兴奋的基准测试是终端科学（Terminal Bench Science），得分翻倍到了 52.6。

The second benchmark that I found exciting was on terminal bench science, which doubled to 52.6.

这个基准测试是用来衡量 AI 代理自主解决复杂科学计算和研究任务的能力。

And this is the benchmark measuring how well an AI agent can autonomously solve complex scientific computing and research tasks.

对我来说，那个曾经谨慎的公司现在正在拉开差距。

You know the elephant for me is the company that was once cautious is now pulling away.

我很想听听你的看法，Alex，关于 Fable 5.1。

I'm curious your thoughts, Alex, Fable 5.1.

我接下来要展示一个基准测试数据，咱们来聊聊 Fable 5.1 在这方面的表现。

I'm gonna show one of the benchmarks here for us to talk about on Fable 5.1.

就是这里。

There you go.

对，我觉得 Fable 5.1 是目前市面上普遍可用的模型中综合实力最强的。

Yeah, I think Fable 5.1 is broadly the strongest generally available model that we have today.

我觉得不是 Astra。

I think it's not Astra.

我觉得是 5.1。

I think it is 5.1.

Anthropic 在 Fable 五和 Mythos 五发布后的波折以及随后的监管审查中，表现得比 OpenAI 更稳健地持续进步。

Anthropic has done even after the hiccup of the Fable five and Mythos five releases and subsequent regulatory scrutiny, think has done a better job of consistently improving.

如果你看他们的基准测试随时间的变化，他们的进展曲线比 OpenAI 的更平滑，没有那么多突然的跳跃。

If you look at their benchmarks over time, they're a little bit less jumpy, a little bit less step function y than OpenAI's progress.

他们的提升非常稳定。

They've been very consistently improving.

所以在实际工作流方面，我非常喜欢5.1版本，也就是Fable 5.1。

So, in terms of workflows, I love 5.1, Fable 5.1.

我也很喜欢OpenAI的模型，两个我都在用。

I also love Open AI's models, and I use both of them.

我觉得它们各有优势。

I think they have different strengths.

从实际体验来看，OpenAI 的模型速度更快。

Find anecdotally OpenAI's models are faster.

它们在数学方面可能表现更好。

They may be better in certain mathematical regards.

这一点在 Frontier Math 基准测试中有所体现。

You see that reflected in the Frontier Math benchmark.

Epic 的能力指数基准测试也反映了这一点。

You see that reflected in Epic's capabilities index benchmark.

不过话说回来，如果非要选一个目前最全面、综合表现最好的模型，那应该还是 5.1。

But nonetheless, if I had to pick a single all around well rounded best model today, it's probably still 5.1.

我说“还是”，是因为它才上线两天左右，但确实就是 5.1。

And I say still because it's only been around for what, two or so days, but it is 5.1.

这是 Artificial Analysis 的智力指数，我们之前提到过，Fable 5.1 排在首位。

And here's the artificial analysis intelligence index, which we referred to earlier with Fable 5.1at the top.

没错。

That's right.

能力前沿这块，怎么说呢，现在竞争挺激烈的。

The capability frontier, mean, it's getting crowded,

这其实是好事。

which is great.

我常开玩笑说，只要 Frontier Labs 跟别人竞争，我们就能赢，而且他们现在确实在全力竞争。

We want to I would joke, when Frontier Labs compete, we win, and they are definitely competing at this point.

对。

Yeah.

没错。

For sure.

我觉得

I think

不过，时间因素真的很值得讨论。

it's really important to talk about time, though.

如果你说，看，Fable 5.1 比 Astra 稍微强那么一点点，但它们之间的差距也就三十天左右。

If if you say, look, Fable 5.1 is a tiny little notch above Astra, but they're only about thirty days apart.

而中国团队只落后他们大约六十天。

And the Chinese are only about sixty days behind that.

所以如果按线性时间来看，意思就是：好吧，我们暂时领先一小会儿，那又怎样呢？

So if you look at it in linear time, it's like, yeah, we're ahead for a minute, so what?

而且我认为很长一段时间以来，Anthropic 一直认为我们必须比其他人更早实现自我改进。因为像雷·库兹韦尔那样的奇点主义者会说，如果我们率先达到那个阶段，接下来就会迎来指数级的无限增长。

And I think for the longest time, Anthropic has been thinking, we need to get to self improvement before anyone else, Because a Singularitarian, a Ray Kurzweil type, says, If we get to that point first, then it's an exponential infinite rise from there.

没人

No one

在乎。然后会有巨大的事情发生。

cares And about something huge will happen.

好吧，我们现在就在那儿了。

Well, we're there now.

那会发生什么呢？

So what happens?

就像，好吧，好吧，我们领先了一大截，而且接下来的一小段时间里，我们会继续拉开更大的距离。

It's like, well, okay, we're miles ahead, and we're gonna get miles more ahead for about a minute.

那又怎样？

So what?

领先别人这么多，我们该怎么利用这个优势呢？

What do we do with that miles ahead?

这就是萨姆的强项所在。

So this is where Sam has an edge.

萨姆知道怎么把这个优势转化为实际的掌控力。

Sam knows how to turn that into locking things up.

接下来会发生的是，两家公司都会拼命把商业合作、房地产、发电机、芯片，甚至整个州、国家、政府都锁进自己的生态里，趁着自己有优势的时候赶紧锁定。

So what's gonna happen next is both companies are gonna try to lock up business partnerships, real estate, generators, chips, entire states, countries, governments, just lock them into their ecosystem while they have that edge.

不然的话，这么做有什么意义呢？

And otherwise, what's the point?

你只不过是在宣告三十天的胜利而已，但三十天后，对方又回到和你一样的位置了。

All you're doing is declaring victory for thirty days, but then the other guy's right were where thirty days ago.

那又怎样呢？

So what?

这点特别关键，也非常重要。我们看到他们正在做的是，尽可能快地在各个垂直领域建立合作伙伴关系。

This is such a great and important point, and this is what we're seeing with them is what they're doing is they're making partnerships in various verticals as fast as they can.

如果你身处那个行业，那你基本没得选。

And if you're in that vertical, you have a very a Hobson's choice.

对吧？

Right?

你要么跟他们合作，冒着把核心资源交给他们的风险；要么犹豫观望，结果他们可能跟别人合作，照样进入那个市场。

You either partner and risk giving them the keys to the kingdom, or you hold off, and they may partner with somebody else and go there anyway.

这对一些大公司来说，处境确实挺难的。

A very difficult situation for some of these big companies.

我觉得 Salesforce 跟 Claude 合作这招特别聪明。

I thought Salesforce partnering with Claude was super clever.

他们 basically 把自家所有的能力都开放给 Claude，好让自己能一直留在这个生态循环里。

They're essentially giving Claude all their capability to then keep them wired into that loop.

对吧？

Right?

他们面临的最大张力，不在于哪个模型最好，而在于你能多快把那个模型转化为客户洞察。

The huge tension they've got is not which is the best model, but how quickly can you convert that model into customer learning fastest?

我认为这才是关键所在。

And that I think is going to be the big raise

随着所有内容都失去变现能力，价值就会流向顶层的应用层。

As it all demonetizes, the value goes to the application layer on top.

没错。

That's right.

我觉得有一件很重要的事情

I think that there's one important thing

关于 Fable 这里。

with Fable here.

缓存读取的成本比 Fable 5 便宜了 75%。

The cache reads are 75% cheaper than Fable five.

那什么是缓存读取呢？就是当你第一次加载业务的所有上下文时，它会快速生成一张地图，指引模型到达需要去的地方。

So what a cache read is, is that when you first load in all your context of a business, say, it figures out basically a rapid map to get to where it needs to go on the model.

所以，缓存读取的成本比反复做同样的推理要低好几个数量级。

And so cache reads are like orders of magnitude cheaper than just doing the same inference over and over again.

现在 Anthropic 和其他公司关注的重点是，能不能把整个企业的上下文都加载进来，从而实现这些缓存读取。

Where for Anthropic and others are now focusing is, can you load the whole context of a business and have these cache reads?

因为这样一来，响应速度会快得多，这也是为什么 Fable 5.1 用起来更顺手的原因之一。

Because then it's also much, much faster to be able to have that responsive environment, which is one of the reasons Fable 5.1 is more pleasant.

我觉得这算是一个进步。我最近主要关注的一个领域是数学物理，因为模型在处理某些物理问题时，会不会在构造性方法和公理化方法之间产生混淆。

I think it's also more of an advance, like my key area that I've been looking at this is mathematical physics, because does a model get confused between a constructive and an axiomatic method on certain physics things.

Fable 5.1 没有这个能力，但 Fable 5.5 有了。

Fable doesn't at 5.1, and Fable 5.5 did.

所以你能看到质量上的提升，还有对上下文的理解也更好了。

So you've seen an improvement in the quality, but also the understanding of context.

这是数学方面的，不是物理之类的。

This is math, not physics, and things like that.

或者从商业角度来看，这些也会得到优化，因为所有公司都会努力在各个环节捕捉上下文信息。

Or even on a business sense, again, that will be optimized, cause all these companies are gonna try and capture context everywhere.

任何拥有芯片设计数据或机械设计数据的公司，现在都在拼命争取这些资源。

What does that Any company that has either chip design data or mechanical design data, they're coming after them.

在这场刚刚拉开序幕的大规模圈地运动中，他们正瞄准这些公司，因为这是你可以防守的阵地。

In the great land grab that's kicking off right now, they're coming after those companies, because that's turf you can defend.

因为这能填补知识空白，而且模型吸收这些数据，大概只要一周时间？

Because that plugs those knowledge gaps, and the thing can eat that data in, what, a week?

然后突然就能成为史上最优秀的机械设计师，或者史上最顶尖的芯片设计师。

Return the crank and suddenly be the best mechanical designer ever, the best chip designer ever.

但这是他们最先要抢占的地盘。

But that's the first turf they're gonna grab.

随着时间的推移，他们会拿下所有的地盘。

They'll grab all turf over time.

但首先要锁定的是算力。

But the first thing to lock up is the compute.

这就意味着芯片设计、物理空间、机架、发电机、变压器以及能源。

And that means chip design, that means physical real estate, means racks, generators, transformers, energy.

没错。

Yes.

未来三十年内的情况就是这样。

That's what's gonna happen in the next thirty

也就几天的功夫。

days.

那这些模型的上下文窗口有多大？

What's the context window on these models?

那我们什么时候能实现无限上下文窗口呢？

And when do we get to an infinite context window?

有什么预测吗？

Any predictions?

我觉得Fable模型已经能达到一百万了。

I think you're a million for Fable.

目前确实还是一百万。

It is a million still.

现在行业里，不管是OpenAI还是Anthropic，标准都是百万级别的。

Like a million is the industry standard now across both OpenAI and Anthropic.

不过，如果算上智能体之间消息传递这种机制——就像我们上一部分聊到的，智能体之间像口述历史那样交互的话——实际能处理的有效上下文会大得多，当然这里得加不少限定条件。

But there's an effective context, I say effective with a bunch of caveats that's much larger if you allow agentic message passing, like we were talking in the last part about oral histories between agents.

对，挺有意思的。

Yeah, interesting.

所以，这些情况都让大家有点紧张。

So, all of this is getting people nervous.

咱们来聊聊AI治理领域里两种截然相反的观点。

And let's talk about two opposing stories in the world of AI governance.

第一种观点来自伯尼·桑德斯参议员和格雷格·卡萨尔众议员，他们刚刚提出了《禁止人工智能超智能法案》，这是一项旨在永久禁止开发和使用达到或超越人类认知水平的AI系统的立法。

The first comes from Senator Bernie Sanders and Representative Greg Kassar, who just introduced the quote, Ban Artificial Superintelligence Act, legislation that would permanently ban the development and deployment of AI systems that match or exceed human cognitive performance.

而且你猜怎么着，违规者将面临最高二十年的监禁。

And get this, violators would face up to twenty years in prison.

桑德斯在推特上写道：‘AI行业的领导者们承认，他们正在构建一种危险的技术，而他们无法控制它。’

Sanders tweeted, quote, The leaders of the AI industry acknowledge that they are building a dangerous technology that they can't control.

在为时已晚之前，我们需要立即在全球范围内暂停先进人工智能的开发。

We need an immediate global pause on advanced AI development before it's too late.

在我们的第二个故事中，同一时间在美国北卡罗来纳州教堂山举行的G20峰会，则告诉世界其他地区对AI监管采取不干预的态度。

In our second story, taking place at the exact same time in The US was the G20 Summit at Chapel Hill, North Carolina, telling the rest of the world to take a hands off approach to AI regulation.

白宫科技顾问、我们的老朋友迈克尔·克拉蒂奥斯提倡所谓的“北卡罗来纳新兴技术原则”。这是一项由包括中国在内的各方一致同意的非约束性G20框架，主张政府应总体上支持创新，除非真正必要否则避免设立专门的AI监管机构，并投资于研究基础设施、人才队伍以及公私合作伙伴关系。

The White House tech advisor friend of the pod, Michael Kratios advocated for what he calls the quote Carolina principles of emerging technologies, which are non binding G20 framework that was agreed to unanimously by everybody, including China, that says governments should generally favor innovation, avoid creating new AI specific regulatory bodies unless truly necessary, and invest in research infrastructure workforce and public private partnerships.

G20会议上，埃隆·马斯克通过视频批评了欧盟的科技监管政策，这当然也是意料之中。

The G20 meeting featured Elon by video criticizing EU tech regulations, of course.

马克·扎克伯格反对限制阿片类药物模型，杰米斯·哈萨比斯呼吁进行安全测试，还有Anthropic的联合创始人汤姆·布朗。

Mark Zuckerberg arguing against restricting opioid models, Demis Hassabis calling for safety tests, and Anthropic co founder Tom Brown.

我们接下来就深入讨论一下这个话题。

We're gonna jump into discussion about this.

这两个极端立场，一边是主张禁止一切并判处二十年监禁，另一边则是埃隆的言论。

Two ends of the extreme, you know, ban everything, give you a twenty year jail sentence, and on one side, and Elon's remarks on the other.

我先放段视频，然后我们再展开聊聊。

Let me share this video, and then we'll jump in and talk about this.

你得有一个相对宽松监管的环境，也就是说，新事物默认是合法的，而不是默认违法的。

You have to have an environment that's relatively free of regulation, meaning that new things must be default legal as opposed to default illegal.

以欧盟为例，我们发现那里的监管水平高得离谱，大多数东西默认都是禁止的，这严重阻碍了新技术的发展。

So on the EU, for example, we find that those the regulation level is extraordinarily high, and things are generally default illegal, and this inhibits progress with of new technologies.

它会让进展变慢。

It it slows it down.

虽然最终不会完全阻止，但确实会大大拖慢速度。

It doesn't it doesn't ultimately stop it, but it slows it down quite considerably.

现在，中国确实拥有巨大的电力资源。

Now now China does have a tremendous amount of electricity.

但是，由于GPU出口禁令，在中国无法建立配备最新芯片的数据中心。

But but due to GPU export bans, one cannot establish data centers with the latest chips in China.

所以，真正需要考虑的是，中国以外的地区电力增长情况如何？

So so as but so really the consideration is what what sort of electricity growth is there outside of China?

目前，相对于AI芯片的生产需求，电力供应存在显著缺口。

And that is currently a significant shortfall relative to AI chip production.

所以我觉得这给世界各国创造了一个机会，如果他们对AI数据中心感兴趣的话，可以提前建设大量电力设施，然后把电提供给AI公司。

So there is this this creates an opportunity, I think, for countries around the world just to to save to to to sort of if they're interested in AI data centers to to to construct a lot of power and and offer that to AI companies.

当然作为交换，这些AI数据中心会被征税，需要支付合理的费用之类的。

And in exchange, of course, these AI data centers will be will be taxed and have to pay, you know, reasonable fees and stuff.

但这确实给很多国家带来了机会。

But it does create an opportunity for lot of countries.

首先，埃隆在那儿看起来真的很累。

First of all, Elon looked really tired there.

是的。

Yeah.

确实。

Did.

对。

Yeah.

我是说，我想象他肯定得二十四小时连轴转，一周七天都不休息。

I mean, I cannot imagine he must be operating, you know, 48 by seven.

那咱们接着聊这个。

So let's jump into this.

我的意思是，两个极端的观点在同一周被提出来。

I mean, two ends of the extreme being voiced the same week.

你知道的，美国到底该往哪儿走？

You know, where does America go?

对，我第一次看到伯尼·桑德斯那个事的时候，我就觉得像老头子跟克劳德握手一样，真是不可思议。

Yeah, when I first saw the Bernie Sanders thing, I did think like old man shakes his hand at Claude, you know, like, it's it's crazy.

不是。

No.

不对。

No.

那句话不是这么说的。

That wasn't the line.

是那个老头冲着克劳德大喊大叫。

It was old old man yells at Claude.

你说吧，Claude。

You go, Claude.

老头子对着Claude大喊大叫。

Old man yells at Claude.

其实我发的时候配了《沙丘》里的那句台词：‘不可让机器取代人类的思维’。

Well, actually, I posted it with the line from Dune, know, thou shalt not make a machine in the

最轻的……人类的思想。

lightest of human a human mind.