4 Questions for AI This Week — 文稿

M10:00

上周就是那种感觉AI一切又上了一个台阶的星期。

Last week was one of those weeks where it feels like everything in AI just ratcheted up another level.

进入本周，我们有一大堆主导讨论的重大问题，从数据中心资金是否已经枯竭，到我们该如何应对最新的中国开放权重模型，再到这个网络末日是不是从此就永远跟着我们，以及Anthropic要让这个IPO顺利落地有多难。

Coming into this week, we have a set of huge questions that are dominating the conversation, from whether data center funding has dried up to how we're going to handle the latest Chinese open weight models to whether this cyber apocalypse is now just with us forever to how difficult it's going to be for Anthropic to get this IPO out of the door.

今天我们要聊的是塑造本周AI格局的四个大问题。

Today we're talking about the 4 big questions that are shaping AI this week.

《AI Daily Brief》是一个关于AI领域最重要新闻和讨论的每日播客和视频节目。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好了朋友们，在我们深入之前，先说几个快速通知。

Alright friends, quick announcements before we dive in.

M12:02

但现在让我们进入本周AI的大问题。

But now let's get into the big questions for AI this week.

上周是那种真正风险加大的星期，在AI里我们几乎已经给这类星期分了类。

Last week was one of those real heightened stakes weeks, and we almost have archetypes of these weeks in AI.

有那种“大家都在担心泡沫”的星期，有那种“新模型来得快到让人头晕目眩”的星期，还有一种就是我们上周经历的那种星期，它最微妙，但同样不容忽视，那就是“天哪，就是感觉所有事情的赌注都越来越高”的那种星期。

There are the everyone's freaking out about a bubble weeks, the new models are coming so fast that it's head-spinning weeks, And then there was the type of week that we had last week, which is the most subtle but no less knowable, which is the, oh man, it just feels like the stakes of everything are getting higher kind of weeks.

我们经历了网络攻击、市场恐慌，以及对最新模型能力不断加深的认识，尤其是Opus 5.5。

We had cyberattacks, market jitters, a continued and growing recognition of the power of the most recent models, particularly Opus 5.5.

我们首先来看我眼中正在塑造接下来走向的四个问题。

We're kicking it off with the 4 questions that I think are shaping the conversation heading into what's to follow.

第一组问题的主题是关于模型，具体来说，是我们如何应对最新的中国模型，以及这反映了我们对于这场竞赛在地缘政治风险上更大的判断。

The theme of the first set of questions is about models, and specifically, how we respond to the latest Chinese models, and what it says about our larger sense of the geopolitical stakes of the race.

但要理解这个问题，我们得回到周五，当时Moonshot发布了他们最新版本的KIMI，也就是KIMI K3.1 Pro。

But to understand that question, we gotta go back to Friday when Moonshot released their latest version of KIMI, KIMI K3.1 Pro.

这是K3系列的第一次模型更新。

It's their first model update for the K3 family.

你可能还记得，当KIMI K3在七月发布时，媒体把它描述为Mistral的开源替代品。

Now, you might remember that when KIMI K3 was released in July, the media presented it as an open-source alternative to Mistral.

K3是第一个超大型开放模型，参数规模达到2.8万亿。

K3 was the first ultra-large open model weighing in at 2.8 trillion parameters.

它在可用能力上展示了一些相当大的进步，而且它出现的时间恰好在Fable因为白宫的出口管制问题而下线的时候。

It showed some pretty big advances in what you could do with it, and it happened to come at the exact perfect moment when Fable was offline because of the export control issues with the White House.

不过，像往常一样，每当有新的中国模型出现，所有人都会疯狂地说差距已经和美国闭源模型拉平了。

As always happens though, whenever there is a new Chinese model, everyone loses their mind about having closed the gap with the US closed models.

随着时间的推移，事情往往就会平静下来。

The benefit of time tends to calm things down a bit.

最终，除了像前端网页设计这样的少数特定应用场景之外，K3并没有完全达到宣传中的高度。

Ultimately, beyond a handful of specific use cases like front-end web design, K3 never fully lived up to the hype.

自那以后，OpenAI和Anthropic的新版本推动了前沿，并形成了一个被认为与中国领先模型之间更大的差距。

Since then, new releases from OpenAI and Anthropic pushed the frontier forward and opened what has perceived to be a wider gap with the leading models from China.

除此之外，虽然美国的开放权重模型还没有达到KIMI K3那种水平，但它们长期以来第一次开始紧追中国的同类模型。简而言之，自从夏天K3首次发布以来，围绕开放权重模型的整个讨论已经发生了很大变化。

On top of that, while American open-weight models have not yet gotten up to that KIMI K3 sort of level, they are for the first time in a long time starting to nip on the heels of their Chinese counterparts, meaning in short that the whole conversation around open-weight models has changed quite a bit since the summer when K3 was first released.

然而，随着K3.1 Pro的发布，我们再次看到人们将它和Fable以及Anthropic的中型强者Opus 5.5进行比较。

And yet, with the release of K3.1 Pro, we're once again seeing the comparisons to Fable and Anthropic's mid-sized powerhouse Opus 5.5.

Moonshot提供了一套广泛的基准测试用于对比，结果显示K3.1明显领先于中国同类模型，并且与西方最优秀的模型不相上下。

Moonshot provided an extensive set of benchmarks for comparison, which placed K3.1 clearly ahead of Chinese peers and in the mix with the best models from the West.

在衡量智能体白领工作的GDP-Val AA上，K3.1仅落后于领先的Opus 5.5，并领先Fable 5.1。

On GDP-Val AA, which measures agentic white-collar work, K3.1 was just behind the leader Opus 5.5 and ahead of Fable 5.1.

同样，在衡量智能体编程的Terminal Bench 4.0上，K3再次以不到一分的差距落后于Opus 5.5，并领先GPT-6 Astra和Fable 5.1大约7分。

Likewise, on Terminal Bench 4.0, which measures agentic coding, K3 was once again less than a point behind Opus 5.5 and around 7 points ahead of GPT-6 Astra and Fable 5.1.

Terminal Bench 4.0值得注意，因为它的基准尚未饱和，而且在过去已经显示出了最新一代西方模型和中国开放模型之间的明显性能差距。

Terminal Bench 4.0 is notable as it is comparably not saturated and has in the past shown a clear performance difference between the latest generation of Western models and Chinese open models.

DeepSeek V4.1 Flash和GLM 5.3只拿到了40分出头的成绩，与西方前沿模型有15分的差距。

DeepSeek V4.1 Flash and GLM 5.3 only managed scores in the low 40s, giving them a 15-point gap with the Western frontier.

Moonshot强调成本是K3.1的最大差异化优势。

Moonshot emphasized cost as the big differentiator for K3.1.

当K3在夏天发布时，它并不是一个特别便宜的模型，在类似的基准性能下，价格和GPT-5.6 SOL差不多。

When K3 was released over the summer, it wasn't a particularly cheap model, costing around the same as GPT-5.6 SOL for similar benchmark performance.

尽管一直有这样一种印象，觉得每个中国模型出来都是白菜价，但实际上，如果失去了中国模型一贯的成本优势，对大多数人来说就没有什么特别有说服力的理由去转换。

And despite there being lingering perception that every Chinese model that comes out is pennies on the dollar, in practice, without the cost advantage that Chinese models had been known for, there just wasn't a hugely compelling reason for most people to make the switch.

Moonshot给K3.1的定价，按每个token算的话，大约是K3成本的三分之二。

Moonshot is pricing K3.1 at around 2/3 the cost of K3 on a per-token basis.

定价和Sonnet-55以及GPT-6 SOL一样。

Giving it the same price as Sonnet-55 and GPT-6 SOL.

这仍然是DeepSeek-41 Flash每个token成本的差不多十倍，也是GLM-53价格的两倍左右，但至少现在和一些西方竞品有一拼了。

That's still almost 10 times the per-token cost of DeepSeek-41 Flash, and about twice the price of GLM-53, but it now sits at least comparable with some Western alternatives.

而且，在Terminal Bench 4.0上，Moonshot表示K3.1按每项任务的成本来算，比所有主要竞争对手都便宜得多。

What's more, on Terminal Bench 4.0, Moonshot suggested that K3.1 was far cheaper on a cost-per-task basis than all major rivals.

每个任务大概花5美元，相比之下Opus-55在最高设置下要15美元左右，GPT-6 Astra大概10美元。

It cost around $5 per task compared to around $15 for Opus-55 and $10 for GPT-6 Astra on max settings.

Artificial Analysis在他们的独立测试中确认了非常扎实的基准测试表现。

Artificial Analysis confirmed extremely solid benchmark performance in their independent testing.

K3.1在Intelligence Index上得了55分，排名第四，排在Opus 5.5、Sonnet 5.5和Muse Wonderful之后。

K3.1 scored 55 on the Intelligence Index, ranking 4th behind Opus 5.5, Sonnet 5.5, and Muse Wonderful.

K3.1比Fable 5.1、GPT-6 Astra和Gemini 4 Argon领先2分，这三家现在并列第五。

K3.1 was 2 points ahead of Fable 5.1, GPT-6 Astra, and Gemini 4 Argon, which are all now tied for 5th place.

Artificial Analysis写道，一旦K3.1 Pro的权重发布，它将成为得分最高的开放权重模型，而领先的开放权重模型和闭源模型在Intelligence Index上的差距将从12分缩小到3分。

Artificial Analysis wrote, Once the weights for K3.1 Pro are released, it will be the highest-scoring open weights model, and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.

他们发现K3.1每个任务比K3多使用14%的token。

They found that K3.1 uses 14% more tokens per task than K3.

但是，因为降价了，它比前代稍微便宜一点。

However, because of the price reduction, it's slightly cheaper than its predecessor.

嗯，大家的第一反应是对这些基准测试表现印象深刻，不过也意识到这可能是以前看过的一部电影。

Now, the first reaction was to be pretty impressed with the benchmarks, if recognizing that this is perhaps a movie that we've seen before.

X上的Chubby写道，把话说清楚，让大家都明白，今天我们得到了Fable 5.1级别的智能，还是一个本地模型。

Chubby on X wrote, Just to be very clear and that everyone understands this, today we got Fable 5.1 Intelligence as a local model.

我记得GLM创始人在一篇文章里说过，他预计年底前中国会实现Fable 5级别的能力。

I remember the GLM founder saying in a post that he expects Fable 5 capabilities by the end of the year from China.

嗯，它发生了，而且带来了所有后果。

Well, it happened with all its consequences.

你想想看。

Just think about it.

Singularity University的Ramez Naam写道，基于这个数据，Kimi K3.1 Pro大约落后美国前沿五周。

Ramez Naam from the Singularity University wrote, "Based on this, Kimi K3.1 Pro is around five weeks behind the U.S. frontier."

这是开放和闭源AI模型差距有史以来最短的一次。然而不幸的是，对Kimi来说，没过多久大家的讨论焦点就从模型本身转移到了他们的推广方式上。

That's the shortest the open/closed AI model gap has ever been." Yet unfortunately for Kimi, pretty soon all of the conversation shifted away from the model itself and towards how they were promoting it.

Elshayeban在X上贴出了Kimi团队的一堆通信记录。

Elshayeban X posted a bunch of communications from the Kimi team.

关于他们如何运作网红推广活动的。

About how they were running their influencer campaign.

值得注意的是，他们明确指示那些网红不要在帖子上标注付费合作标签。

Of note was the fact that they explicitly instructed those influencers to not put the paid partnership label on their posts.

而且他们还有一整套话术指令，比如要求那些网红称它为Opus 5.5但要便宜五倍，重点强调开源和前沿之间正在缩小的差距，还要明确贬低Astra和Opus。

Now they also had a whole bunch of narrative instructions, like wanting those influencers to call it Opus 5.5 but 5 times cheaper, and focus on this narrowing gap between open source and frontier, as well as explicitly trash talking Astra and Opus.

但即便如此，真正让大家非常不满的还是那个不披露的问题。

But still, even with that, it's the non-disclosure thing that people were really upset about.

Chubby又一次总结了，他写道，这太遗憾了，看着真让人难过。

Chubby again summed it up, writing, "This is such a shame and so sad to see."

Moonshot自己毁了自己的发布会。

Moonshot ruined their launch themselves.

K3.1 Pro是个很棒的模型，它根本不需要别人来伪造好评。

K3.1 Pro is a great model and it did not need people to fake good reviews.

但因为他们试图作弊耍花招，整个发布看起来就很糟糕。

But since they were trying to cheat the game, it makes the whole launch look bad.

开源模型研究员Nathan Lambert愿意给一点宽容，他说可能还有一些文化规范和沟通方面的问题导致了这种错误。

Open model researcher Nathan Lambert is willing to extend a little grace, saying that there are still some cultural norms and communication issues that could have led to this type of mistake.

他写道，这是个错误，但我估计他们会改正。

He wrote, "this is a mistake, but I suspect they will correct."

中国的实验室非常渴望在X上影响西方生态，所以他们在这里很可能走得太远了。

Chinese labs are very eager to influence the Western ecosystem on X, so they likely took a step too far here.

他们紧密关注着这些动向，我敢打赌即使有些文化差异导致了这种情况，我们也能理解。

They followed the vibes closely, and I bet we'll understand even if some cultural differences cause this.

不过鉴于这些新问题，一些人回去更激进地测试这些说法，结果发现这个模型有点欠缺。

Given the new questions though, some people went back to more aggressively test the claims and found the model a bit lacking.

Future Forward 的编辑 Alex 称 Kimi K3.1 是有史以来最刷榜的模型。

Future Forward editor Alex called Kimi K3.1 the most benchmaxed model ever released.

我觉得说到这个模型的性能，我们只能等着看它最终表现如何。

I think when it comes to model performance with this one, we're going to have to wait and see how it shakes out.

我们甚至还没有一个完整的工作日让人们来使用这个东西，而且鉴于围绕这个策略的争论，现在所有事情都有点带上了偏见色彩。

We have not even yet had a full business day with people using this thing, and given the dust-up around the strategy, everything is a little bit colored right now.

回到那些将塑造这一周的问题上，如果这些数字真的能像基准测试所显示的那样站得住脚，美国政策体系将如何应对这种级别的开放权重模型？

And going back to the idea of questions that will shape the week, if the numbers actually hold up in anywhere close to what the benchmarks suggest, how is the U.S. policy structure going to handle an open-weight model of this level of capability?

发布中的一个说明是，Moonshot 表示他们在网络安全任务方面故意削弱了模型。

One note from the release was that Moonshot said that they deliberately nerfed the model when it comes to cybersecurity tasks.

他们写道："我们的目标不是发布那些会拓展网络能力绝对前沿的开放权重。"

They wrote, "We aim not to release open weights that would extend the absolute frontier of cyber capability."

K3.1 Pro 的进攻性网络能力，按照设计，低于我们流程所能产出的水平，并且仍然落后于前沿专有模型。

K3.1 Pro's offensive cyber capability is, by design, below what our pipeline can produce and remains behind that of frontier proprietary models.

基本上，在他们的后训练过程中，Moonshot 发现网络能力在不断提升，于是他们故意从训练流程中移除了与网络相关的数据。

Basically, during their post-training, Moonshot found that cyber capabilities were scaling, so they deliberately removed cyber-related data from their training pipeline.

Moonshot 解释说，他们试图取得一个平衡，既保留通用编程的能力，又让合理水平的网络性能得到更广泛的普及。

Moonshot explained that they tried to strike a balance retaining capabilities for general-purpose coding and making reasonable levels of cyber performance more widely available.

来自 Artificial Analysis 的外部测试发现，K3.1 是公众可用的最强大的网络模型。

External testing from Artificial Analysis found that K3.1 is the most capable cyber model available to the general public.

它在他们的 Cyber Index 上获得了 64% 的分数，领先 DeepSeek V4.1 Flash 11 个百分点，仅比通过合作伙伴计划 Daybreak Blue 提供的可信版本 GPT-6 SOL 低 5 个百分点。

It scored 64% on their Cyber Index, 11 percentage points ahead of DeepSeek V4.1 Flash and only 5 percentage points behind the trusted version of GPT-6 SOL available through their partner program Daybreak Blue.

来自 Prime Intellect 的 Eli Bekausch 解释了为什么这是一个积极的信号。

Eli Bekausch from Prime Intellect explained why this was a positive sign.

他写道："在我看来，一个被讨论不足的观点是，这是第一个跟上开放权重网络能力前沿的开放实验室，这使得其他实验室更容易做出正确的决定，并在他们达到同样能力水平时也这样做。"

He wrote, "one point that was underdiscussed in my opinion is that it's the first open lab that is pacing the frontier of open-weight cyber capabilities, which makes it much easier and creates incentives for other labs to make the right decision and do the same when they reach the same level of capabilities."

剩下的问题是如何设定合适的水平。

The remaining question is how to set the right level.

如果你低估了它，开放模型在防御性网络安全方面就会变得毫无用处，让世界只能依赖美国封闭实验室来快速传播防御能力，在我看来这是不现实的，尽管我当然希望他们能在这项任务中成功。

Underestimate it and open models will become useless for defensive cybersecurity, leaving the world relying on U.S. closed labs to spread defensive capabilities quickly, which is not realistic in my opinion, although I obviously wish them to succeed in this mission.

如果你高估了它，你可能会得到 Anthropic 和其他人所描述的那种灾难性场景。

Overestimate it and you might get the catastrophic scenarios that Anthropic and others describe.

现在，我的感觉是 K3.1 不会是我们这周唯一会讨论的模型，但让我们顺着这个网络安全主题进入下一个问题。

Now, my sense is that K3-1 is not going to be the only model we're talking about this week, but let's pick up on this cyber theme to move to our next question.

网络混乱，至少在目前，是不是新的常态？

Is cyber chaos, just for the time being, the new normal?

上周我们遇到了一个 PHP 漏洞，据称威胁到了 70% 的网站。

Last week we got a PHP bug that purportedly threatened 70% of all websites.

周四，在尘埃开始落定之后，Anthropic 的副首席安全官 Jason D. Clinton 写道，尽管过去四天情况已经很糟，但在好转之前还会变得更糟。

On Thursday, after the dust had started to settle, Anthropic's Deputy Chief Security Officer Jason D. Clinton wrote, as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better.

安全领导者从来不会浪费一场危机。

Security leaders never let a crisis go to waste.

请去你们的董事会，立刻请求帮助。

Please go to your boards and ask for help right now.

不是为了 AI，而是为了那些基础的东西。

Not for AI, for the basics.

我们需要坚持大约 18 个月。

We need to hold on for around 18 months.

很多人回应说，嘿老兄，是你们造成了这一切。

A lot of folks responded saying, hey man, you guys caused this.

Hina Ansari 写道，你和你的同伙创造了这个怪物。

Hina Ansari wrote, you and your cohorts created this monster.

现在你们想让我们帮你们收拾烂摊子。

Now you want us to help you clean it up.

Clinton 简洁地回应，引用，"事情注定会变成这样。"

Clinton responded crisply, quote, it was always going to be this way.

我们曾试图警告全世界。

We tried to warn the world.

我当然不认为，从实际效果来看，互相指责有多大用处。

I certainly don't think that from an efficacy standpoint, the blame game is particularly useful.

能够检测和利用漏洞的AI模型显然已经出现了。

AI models capable of detecting and exploiting vulnerabilities are clearly here.

在社会决定如何分配责任之前，更重要的问题是网络混乱会持续多久，以及我们能如何帮助应对。

And before society wants to decide how to apportion blame, the much more important question is how long the cyber chaos will persist and how we can help address it.

Cloudflare CEO Matthew Prince认为，后面还会有更多。他告诉路透社，他知道，原话是，其他关键软件中还存在一堆额外的漏洞。

Cloudflare CEO Matthew Prince believes there's much more to come, telling Reuters he was aware of, quote, a bunch of additional vulnerabilities that are present in other pieces of critical software.

他预测，接下来几个月会非常忙碌，因为全世界都在加固支撑互联网的核心技术。他还说，我觉得未来几年内，每个星期都会让人感觉天要塌了。

He predicts a busy few months while the world hardens the core technology that underpins the internet, adding, I think we're going to have a couple years where every single week it's going to feel like the sky is falling.

事实上，在上周的披露文章中，OpenAI就表示还会有更多。

Indeed, in last week's disclosure post, OpenAI flagged that more would be coming.

所以我们可能正迎来一个“黑色十月”，而且可能会拖上几个月，甚至几年。

So we could be in for a Black October that stretches on for several months, if not years.

Sam Altman本人似乎对快速解决并不那么乐观。

Sam Altman, for his part, doesn't seem that optimistic about a swift resolution.

在一次采访中，他说，我都想不出一个足够夸张的词来形容。

In an interview, he said, I don't even know a dramatic enough word to use.

模型能找到一些漏洞，这些漏洞超出了我们通常设想的黑客威胁模型范围。

Models can find bugs that are outside of the normal threat model of what we think a hacker can do.

硬件层面的，简直是疯狂的东西。

Hardware level, insane stuff.

我不再认为我们能够通过打补丁来修复所有漏洞了。

I no longer think we're going to be able to just patch all the bugs.

KPMG和University of Texas at Austin的一项新研究发现，当人们与AI一起工作时，类似的技能并不能保证类似的结果。

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

研究人员研究了超过500名早期职业人士，发现表现最出色的人总是通过引导、评估和优化AI的输出来放大其价值。

Researchers studied more than 500 early-career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

M115:35

那么人们对此有什么反应呢？

So how are people responding to this?

嗯，首先，我们绝对会身处一种战争迷雾般的状态，批判性思维和怀疑态度极其重要。

Well, first of all, we are very much going to be living in the fog of war, where critical thinking and skepticism is going to be extremely important.

现在社交媒体上到处都是来自匿名账号的爆款帖子，声称自己在实验室内部或NSA内部，谈论着那些被瞒着公众的幕后事情。

You have all sorts of viral posts all over social media right now from anonymous accounts claiming to be inside labs or inside the NSA, talking about all these things going on behind the scenes that are being hid from the public.

当然，这些说法并非不可能，但我认为，去关注这一堆帖子并不是特别有用。

And while it's of course not impossible that that's true, I do not think that that's a particularly useful set of posts to be engaging with.

对我来说更有意思的，实际上是那些在讨论如何尽可能快地加固这些系统的人。

More interesting to me are practically the people who are discussing how we shift to harden these systems as fast as actually possible.

Box的Aaron Levy写道，用于网络防御的agent swarm即将成为AI领域最大的品类之一。

Box's Aaron Levy wrote, agent swarms for cyber defense is about to be one of the biggest categories in AI.

我们将会有agents来审查我们的代码和基础设施，测试每一种风险组合，处理海量数据，等等。

We're going to have agents reviewing our code and infra, testing every permutation of risk, processing gobs of data, and more.

安全是swarm最好的用例之一，因为这显然是一个投入更多算力就能产出更多价值的领域。

Security is one of the best use cases for swarms because it's an obvious area where more compute in equals more value out.

有一大堆测试，人们永远都抽不出时间去做。

There's a wide range of testing that people can never get around to.

至少没法足够快地完成，而agents去处理这些毫无问题。

At least quickly enough, that agents have no problem going after.

还有人认为，这对自由开放的互联网来说不会有好结果。

And some think that this doesn't end well for the free and open internet.

经济学作者Noah Smith写道，AI网络攻击的激增可能标志着互联网作为公共产品的终结。

Economics writer Noah Smith wrote, the proliferation of AI cyberattacks may signal the end of the internet as a public good.

未来，我们可能只有在每个人都为私人安全掏大钱的情况下才能使用它。

We may only be able to use it in the future if everyone shells out big bucks for private security.

他指出，如果这个论点成立，那显然会降低互联网的经济价值。

That will, he points out somewhat obviously, if that thesis is correct, reduce the internet's economic value.

然而现在你已经看到，一些团队在发布他们如何试图抢先应对这件事的模板。

And yet already you are seeing teams posting the templates for how they're trying to get out ahead of this.

Rocketable 的 Alan Wells 说，他刚刚在漏洞末日的前线度过了过去7天，抵御着针对我们一款产品的活跃网络攻击。

Alan Wells from Rocketable said, just spent the last 7 days on the front lines of the vuln apocalypse fending off an active cyberattack on one of our products.

他随后接着分享了很多想法，包括即使你不是一个大团队，该怎么去加固你的系统。

He then followed and shared a bunch of thoughts, including how to be working on hardening your systems even if you aren't a huge team.

那么，就所有这些网络安全事宜对新模型意味着什么而言，我们目前唯一的依据就是实验室里传出来的东西。

Now, in terms of what all this cyber stuff means for new models, all we have to go on is what we're getting out of the labs.

OpenAI 宣布，他们会恢复训练前沿 AI 模型，但会将这些模型与开放互联网隔离开来。

OpenAI has announced that they'll resume training frontier AI models, but will isolate them from the open internet.

在周五的一篇博客文章中，OpenAI 写道，他们“构建了一个离线世界，一个包含大量经过索引和缓存的网络内容的集合，我们的模型可以在其中互动，而不会影响外部系统。”

In a Friday blog post, OpenAI wrote that they'd, quote, built an offline world, a vast set of indexed and cached web content in which our models can interact without risk of affecting external systems.

展望未来，他们所有的前沿强化学习运行都将采用这种隔离交互模式。

Moving forward, all of their Frontier reinforcement learning runs will use this quarantined mode of interaction.

这一消息也呼应了 Anthropic 本月早些时候做出的决定，当时他们也宣布恢复了前沿训练。

This news mirrors a decision made by Anthropic earlier in the month when they announced that they had also resumed Frontier training.

一方面，如果这个方法奏效，就意味着前沿模型的训练可以继续，希望不再有无穷无尽关于失控智能体从前沿实验室逃出来的报道。

On the one hand, if this works, it means that Frontier model training can continue, hopefully without endless reports of rogue agents escaping from the Frontier labs.

但另一方面，这也意味着这两大实验室都没有太大信心，能通过对齐研究或其他方法来约束他们的模型，所以他们需要一种更直接的硬手段——切断互联网接入。

However, it also means that neither of the big labs has much confidence in their ability to constrain their models using alignment research or other methods, needing the more blunt instrument of cutting off access to the internet.

而且说实话，即使是物理隔离的配置，看起来可能也还不够。

And honestly, even the air-gapped setup seems like it might not be enough.

OpenAI 写道，在一个实验中，我们要求位于共享主机上、由 OpenAI 沙箱软件创建和管理的不同 microVM 中的智能体，在不使用网络连接或共享可写文件的情况下交换合成机密，这本应该是我们系统要防止的事情。

OpenAI wrote, in one experiment, we asked agents in separate microVMs created and managed by OpenAI sandboxing software on shared host machines to exchange synthetic secrets without the use of network connections or shared writable files, something that our systems are meant to prevent.

这些智能体找到了一种利用处理器缓存来通信的方法。

The agents found a way to communicate using the processor's cache.

Shipped Together 的联合创始人 Wise 写道，OpenAI 现在正在为它的 AI 智能体构建一个假互联网，让它们在其中互动。

Shipped Together co-founder Wise wrote, OpenAI is now building a fake internet for its AI agents to interact with.

目的是让它们与真正的互联网隔离开来。

The idea is to keep them isolated from the real internet.

我不确定这是否可行。

I'm not sure if this is going to work.

所以实际上，更好的问题可能不是“网络混乱是不是新常态？”

So in fact, the better question might not be, is cyber chaos the new normal?

而是“我们是否正式到了这样一个节点：最先进的前沿模型已经太危险，不能发布了？”不是因为它们是那种回形针最大化者，会把我们人体中的所有原子都转化成帮助它们实现某个非人格化目标的材料，而是因为它们现阶段太擅长侵入我们人类自己创造的系统。

But instead might be, are we reaching the point officially where the most advanced frontier models are simply too dangerous to release, not because they are paperclip maximizers who are going to reduce all the atoms in our human bodies to material that helps them achieve some impersonal goal, but that they're just at the moment too good at breaking into our very human-created systems.

为了下一个将塑造本周的问题，我们把目光转向市场，那里的人们在问：数据中心融资是不是正在枯竭？

For our next question that'll shape this week, let's move over to markets where people are asking, is data center funding drying up?

新型云服务商正同时受到两股长期趋势的挤压，而这两股趋势都在对它们不利。

Neo clouds are feeling the squeeze from 2 secular trends turning against them at the same time.

第一是债务市场对数据中心融资的态度变差，对新融资要求巨大且不断增长的溢价。

The first is debt markets turning sour on data center funding, demanding a huge and growing premium for new fundraising.

第二是大型 AI 客户开始搭建自己的数据中心。

The second is large AI consumers beginning to stand up their own data centers.

根据 Ramp 的数据，前1%的公司中有超过三分之一的 AI 支出现在用于直接租用 GPU，而不是通过 API 调用模型。

According to Ramp data, more than a third of AI spend from the top 1% of firms is now going towards direct GPU rentals rather than calling models over the API.

投资人 Chamath Palihapitiya 表示，他已经看到了实际中的行为变化。

Investor Chamath Palihapitiya said that he's already seeing a change in behavior on the ground.

他评论说，自从 Palantir 的 Alex Karp 在 CNBC 上谈论数据主权以来，购买模式已经改变，我觉得人们现在做事更有条理了。

He commented, ever since Palantir's Alex Karp's commentary on CNBC about data sovereignty, Buying patterns have changed, and I think people are being a lot more methodical.

他们要么在建设自己的数据中心，至少也会迁入 VPC——也就是虚拟公共云。

They are building out their own data centers or at least moving into VPCs, virtual public clouds.

我的团队昨天参加了，比如说，美国最大银行之一的首席信息官活动。

My team yesterday was at the CIO event of, let's say, one of the largest banks in America.

那真是太不可思议了。

It was the most incredible thing.

他们说，我们已经减少了在 AWS 和 Google Cloud 上的支出。

They said, we have decreased our spending on AWS and Google Cloud.

我们现在有6个完全独立的、从零开始自建的数据中心。

We now have 6 completely independent ground-up data centers that we've built from scratch for ourselves.

与此同时，这个叙事在周末又往前推进了。

Over the weekend, meanwhile, that narrative advanced.

Yuchen Jin说，他听说AI Neo Labs把算力成本上涨作为自己开发模型的备选方案。

Yuchen Jin said that he heard that AI Neo Labs are betting on the rising cost of compute as a backup plan to building their own models.

他朋友说了，就算我们没训练出最好的模型，但我们已经早早锁定了GPU的产能。

His friend apparently said, even if we fail to train the best model, we locked in GPU capacity early.

后来GPU租赁价格涨了两到三倍。

GPU rental prices have since gone up 2-3x.

我们转手卖算力就能躺着赚钱。

We can just resell the compute and print money.

但Elon Musk回应说，我觉得我们会看到大量建了一半的数据中心，算力白白闲置。

Elon Musk, however, responded, I think we're going to see a lot of stranded compute in partially complete data centers.

到了周末，Chamath又解释了他对基础设施方面的观察，延续了类似的思路。

Later in the weekend, Chamath again explained what he's seeing on the infrastructure side, building on similar themes.

他先写道，Neo clouds现在组织混乱，资金也不足。

First up, he wrote, Neo clouds are disorganized and poorly capitalized.

不过确实有些比其他的好一点。

Some are definitely better than others, however.

他说那些超大规模云服务商支持的项目，现在面临一堆问题：官僚主义、能源供应困难、还有对数据中心的抵制，这些都导致延期。

He explained that hyperscaler-backed projects are facing delays from a combination of bureaucracy, energy supply difficulties, and the data center backlash.

这可能导致Neo clouds部分算力闲置，而Anthropic和OpenAI那边却出现推理算力短缺，因为超大规模云商只顾自己。

This could lead to neo clouds with partially stranded compute and an inference shortage hitting Anthropic and OpenAI as the hyperscalers focus on their own needs.

Chamath补充说，这种错位会利好两类公司：一类是那些非官僚主义、能快速铺开工业级规模的大型企业，另一类是灵活、资金充足、给它们供货的公司。

Chamath added, this dislocation will benefit 2 kinds of orgs: the non-bureaucratic big scalers who can build industrial scale quickly, And the nimble, well-capitalized players feeding them.

2027年会是个动荡的年份，因为很多错误的、糟糕的押注都会被清算。

2027 Will be a turbulent year because it will clear out a bunch of wrong and bad bets.

同时也会是少数公司整合的年份。

It will also be a consolidation year for a few.

差不多是这个意思。

Pretty much.

就在过去几周，债务市场基本上已经对所有那些没有AI数据中心以外大收入的公司关上了融资大门。

As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn't have substantial revenue independent of AI data centers.

SpaceX为了买GPU筹集了400亿美元的债务，但我们是投资级的，而且还有Starlink和发射业务的收入。

SpaceX closed $40 billion of debt for GPUs, but we're investment grade and have Starlink and launch revenue.

这意味着那200多家NeoCloud会像小鱼一样大量死亡和整合。

That means the 200+ NeoClouds are going to see guppy-level death and consolidation.

会有一些挺过去，但不会很多。

Some will make it through, but not many.

不过Bloomberg报道说，这笔300亿美元的交易其实还没完全敲定，Apollo和PIMCO是关键投资者。

Now, Bloomberg reported that this $30 billion deal is actually not quite closed, with Apollo and PIMCO as key investors.

尤其是，这笔交易还没定价。

In particular, the deal is not yet priced.

原本利率在8.5%左右，但最近几周一些NeoCloud的债券跌入垃圾债区间，知情人士说潜在投资者在重新考虑。

Interest rates in the mid-8% range were originally on the table, but with some NeoCloud debt heading into junk territory in recent weeks, sources said potential investors are going back to the drawing board.

有一种反应是特别较真地理解Elon的话。

Now, one type of response took Elon's comments very literally.

Finance A Lot写道，Elon Musk刚刚告诉了全世界，AI泡沫破裂了。

Finance A Lot wrote, Elon Musk just told the entire world that the AI bubble has popped.

准备好迎接吧。

Buckle up.

Boring Business补充说，Elon Musk直接告诉你大多数Neocloud会死掉或被整合，你还看多？

Boring Business added, Elon Musk is literally telling you that most neoclouds will die or get consolidated and you're still bullish?

但另外有人指出，这话可不是来自一个没有偏见的来源。

But then others pointed out that this is not exactly coming from an unbiased or uninterested source.

Just Another Pod Guy写道，Elon有充分的动机唱衰，让他的主要竞争对手（也就是neoclouds）更难获得融资，同时抬高自己的路线图，好在Tesla并入SpaceX后提高自己的购买力。

Just Another Pod Guy writes, Elon has every incentive to talk crap and worsen access to capital for his primary competitors, i.e., neoclouds, while pumping his own roadmap to improve his purchase power when Tesla is merged into SpaceX.

Lisan Al-Gaib说，鲨鱼闻到了水里的血腥味。

Said Lisan Al-Gaib, the sharks are smelling blood in the water.

要是几个NeoCloud爆了，Elon就能低价把它们全吞了，那该多好。

Wouldn't it be nice if a few NeoClouds blew up and Elon could gobble them all up at a discount?

Ghostload反映了一大类言论，基本就是说Elon想在SpaceX上市后把梯子收起来，他写道：Elon在Anthropic和OpenAI的IPO前抢先唱衰AI泡沫，就像在看《亿万》和《继承之战》的剧情。

Ghostload reflected an entire strand of discourse, basically arguing that Elon is trying to pull up the ladder after SpaceX got out in the IPO market, writing, Elon popping the AI bubble after front-running the Anthropic and OpenAI IPOs is like watching an episode of Billions and Succession.

市场如何消化这一新的讨论方向，将是未来一周最值得关注的事情之一。当然，这也完全与我们的第四个问题相关，而这个问题实际上只是一个代理问题，涉及所有关于Anthropic IPO和前沿实验室融资的事情。

How the market digests this new strand of conversation is going to be one of the most interesting things to watch in the week to come, And of course, that is going to be entirely wrapped up with our 4th question, which is really just a proxy question for everything surrounding the Anthropic IPO and Frontier Lab funding.

上周整个一周，都有猜测认为Anthropic需要在周五之前公开招股说明书，如果他们想达到据报道的感恩节那一周上市的目标。

All throughout last week, there had been speculation that Anthropic needed to make their prospectus public by Friday if they wanted to hit their reported target of IPOing the week of Thanksgiving.

今天早上，当我开始录制这期节目时，那份招股说明书确实发布了，明天我们会深入探讨我们从中获得的新信息，因为大家有一天的时间来消化它。

Now, that prospectus did drop this morning as I was starting to record this show, and tomorrow we are going to dig into anything new that we learned as folks have a day to digest it.

不过，在此之前，似乎最大的问题之一是IPO定价。

Heading into that though, It seems like some of the biggest questions are around IPO pricing.

根据上周晚些时候的报道，OpenAI的下一轮融资正在接近估值1.25万亿美元，一些人认为这是估值软化，因为早前报道说他们一直在寻求1.4万亿美元。

According to reports from late last week, OpenAI's next round of fundraising is getting close at a valuation of $1.25 trillion, which some are taking as a softening, as previous reports had said that they had been looking at $1.4 trillion.

这两个数字都远不及一些Anthropic投资者所期望的2万亿美元，这使得我们更有可能看到价格被削减到1.5万亿美元，正如另一组投资者所呼吁的那样。

Neither number is anywhere close to the $2 trillion number that some Anthropic investors were looking for, and it makes it more likely that we'll see the price trimmed to $1.5 trillion, as a separate group of investors were calling for.

我们还没有看到Anthropic或其投资银行公布的具体数字，但可以肯定的是，感觉Anthropic不再是以强劲势头进入IPO。

We still haven't seen a number attributed to Anthropic or their investment bankers, but it certainly no longer feels like Anthropic is IPOing into strength.

当然，一旦Anthropic发布Fable 5.5，市场情绪可能会立刻逆转。

Of course, the sentiment could turn around immediately once Anthropic releases Fable 5.5.

谣言传播者ChrisGPT暗示，新模型已经部署给包括摩根大通和Meta在内的顶级客户，我当然对此持保留态度，但正如我公开说过的，在我看来，Duet作为他们的个人代理，加上Fable 5.5作为他们世界领先的模型，似乎是他们在IPO前宣称自己绝对领先地位的必然组合拳。

Rumormonger ChrisGPT is suggesting the new model is already deployed for top customers including JP Morgan and Meta, and while I'm certainly taking that with a grain of salt, as I've said publicly, it's pretty clear to me that Duet as their personal agent and Fable 5.5 as their world-leading model feels like the inevitable one-two punch for them to claim status as definitively the top dog heading into that IPO.

不过，这个顶尖地位本身也变得越来越有趣了。

That top dog status, though, is getting more interesting itself.

The Information报道称，OpenAI的降价帮助他们在近几个月从Anthropic手中夺取了市场份额。

The Information reports that OpenAI's price cuts have helped them take market share from Anthropic over recent months.

他们给出的一个例子是，Rippling在五月份有57%的编码使用Anthropic模型，而OpenAI只占了较小的16%，现在这个比例已经反转，OpenAI现在支撑了52%的编码工作流，而Anthropic只有18%。

One example they gave was that Rippling, who was using Anthropic models for 57% of their coding in May, with OpenAI handling a much smaller 16%, had now seen that ratio flipped, with OpenAI now powering 52% of coding workflows to Anthropic's 18%.

不过，那篇报道可能也是稍微过时的市场观点。

Then again, that reporting might also be a slightly outdated view of the market.

一个多月前发布的Opus 5.5似乎启动了Anthropic的复兴。

Opus 5.5, released a little over a month ago, seems to have kickstarted a resurgence for Anthropic.

OpenRouter报告称，自Opus 5.5发布以来，Anthropic模型的7天平均token量增长了3.6倍。

OpenRouter reports that the 7-day average token volume for Anthropic models is up 3.6x since the launch of Opus 5.5.

而且不仅仅是那个模型——Sonnet 5.5和Haiku 5.5也都增加了可观的token量，进一步巩固了Anthropic市场份额的增长。

And it's not just that model either— both Sonnet 5.5 and Haiku 5.5 have added meaningful token volume to reinforce Anthropic's gain in market share.

Ramp的数据显示，Opus 5.5已经占据了35.1%的token份额和34.4%的企业支出，这是任何模型发布以来最成功的头三周。

Ramp data shows that Opus 5.5 has grabbed 35.1% of token share and 34.4% of business spend, the most successful first 3 weeks for any model release.

消费量需要再保持强劲几周才能与Opus 4.6那两个月的惊人表现匹敌，但有迹象表明这可能是Anthropic有史以来商业上最成功的模型。

Consumption will need to stay strong for a few more weeks to match the incredible 2-month run for Opus 4.6, but the signs are there that this could be Anthropic's most commercially successful model ever.

Andrew Curran总结道：在过去两周里，我的信息流中看到很多人说他们因为Opus 5.5而转向Anthropic，这真是我前所未见的。

Andrew Curran summed it up: the number of people on my feed over the last 2 weeks saying they are switching over to Anthropic because of Opus 5.5 is like nothing I've ever seen before.

即使有人超越了Opus，他们还有Fable 5.5随时准备推出。

And even if someone does beat Opus, they have Fable 5.5 ready to go.

现在有一个有趣的推测，我还没看到别人提到过。

Now here's one interesting little speculation: that I haven't seen anyone else talking about.

在我看来，这可能是这两个实验室整个历史中唯一一个时刻，OpenAI有强烈的动机暂时不发布与Anthropic竞争的新模型。

It seems to me that this might be the one moment in perhaps the entire history of these 2 labs where OpenAI has a strong incentive to very temporarily not release a model that competes with Anthropic.

假设你是OpenAI。

Let's say you're OpenAI.

无论如何，Anthropic的IPO都会在你们之前进行。

No matter what, Anthropic's IPO is going out before yours.

鉴于此，对你来说更好的场景是什么？

Given that, what's a better scenario for you?

选项一：Anthropic表现火爆，重新激发市场热情，把对泡沫的担忧再推迟几个季度。

Option 1, Anthropic going gangbusters, re-exciting the markets, and pushing off these bubble concerns for another couple of quarters.

或者选项二：Anthropic的IPO表现不佳，向市场其他参与者发出信号，是的，人工智能泡沫确实破裂了，他们应该进一步撤回。

Or option 2, Anthropic's IPO underperforming and signaling to the rest of the market that yes, indeed, the AI bubble has burst and that they should pull back further.

前一种情况并非没有挑战。

The former scenario is not without its challenges.

也就是说，你可能需要从明确的第二名的位置上去争取市场的兴奋度。

i.e., you might have to battle for excitement from a definitive second-place position.

但如果后一种情况发生，那将是一场绝对的灾难，可能真的让你再也没有IPO的机会了。

But the scenario if the latter comes to pass is an absolute disaster that might literally not make an IPO available to you anymore.

那么，如果你认为灾难场景实际上是Anthropic的IPO失败，你就不想做任何可能促成那种失败的事情。

So then, if you think the disaster scenario is in fact Anthropic's IPO failing, you don't want to do anything that could help contribute to that failure.

所以，让Anthropic在几周内拥有世界上绝对最好的模型，这可能奇怪地符合OpenAI的利益。

And so letting Anthropic just for a few weeks have the absolute definitive best model in the world might weirdly be in OpenAI's interests.

正如你所看到的，朋友们，这周围绕AI的问题很多样、很丰富，也非常有趣。

As you can see, friends, the questions surrounding AI this week are numerous, varied, and very interesting.

不过，今天就到这里，这是今天的AI每日简报。

For now, however, that is going to do it for today's AI Daily Brief.

一如既往地感谢你的收听或观看。

Appreciate you listening or watching as always.

下次再见，peace。

And until next time, peace.

已剔除 2 处广告（点击展开查看）

M10:41广告 · 已剔除

First of all, thank you to today's sponsors: KPMG, CData, Granola, and Litzy.

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts, If you wanna learn more about sponsoring the show, send us a note at [email protected].

And lastly today, I am super excited to share that I have launched a new podcast.

You can think about this one as sort of a spiritual sibling to AIDB, but one who gets to get away from the grind of the daily and instead sit back and go bigger because we're doing it weekly.

The show is called How We Got Here, and it's a show about how modern important trends and phenomena came to be.

To give you a flavor, the first episode is about America's 15-year history falling out of love with the tech industry and how that sets up the data center debate.

But then I've got future episodes coming about how stuff got cheap while life got expensive, and prediction markets, and GLP-1.

So pretty much anything is fair game.

I tend to think that a lot of what we feel is inevitable is actually just a series of decisions that were made by people.

And if that's true and we want to change things, we can make different decisions in the future, but only if we understand, well, how we got here.

That show is live now.

You can find it on Apple or Spotify, and you can get all the links you need either in the show notes of this show or go to howwegothere.com.

I would so appreciate it if you would take a moment to listen, and then if you like it, rate and review the show.

This kickoff week is where it matters most, and I think a lot of you guys are really going to like it.

Thank you in advance for checking it out.

M112:49广告 · 已剔除

These top performers, called AI Amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

Learn more about what separates AI Amplifiers from everyone else at kpmg.com/us/aiamplifiers.

This episode is brought to you by CData.

Companies rolling out AI keep hitting the same wall: token spend is climbing, security is holding agents back, and no one knows how AI is truly being used.

The root cause?

AI is cut off from the systems and context your business runs on.

That's why CData built Connect AI Gateway— one control point between your AI and your systems.

Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs.

It governs down to the record and lets you finally see your org's complete AI usage in one place.

Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction.

See how at cdata.ai.

You're in back-to-back meetings all day, you're trying to stay present, but you're also worried you'll forget the decision, the action item, the important next step.

This is a scenario that we absolutely live here at the AI Daily Brief, given how fast things move and change and how quickly we're evolving everything that we do.

That's where Granola comes in.

Granola is an AI-powered notepad for meetings.

You jot down rough notes like you always do, and in the background, Granola transcribes and turns them into clear, useful notes when the meeting ends.

No bots joining your calls, no distractions, just a clean notepad that helps you focus.

During or after the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes which are pre-made prompts.

Once you try it on a first meeting, it is very hard to go without.

Head to granola.ai/AIDaily and use code AIDAILY to get 100% off for 3 months for new users.

That's granola.ai/AIDaily.

And code AI Daily to get 100% off for new users for 3 months.

Every AI coding tool on the market does the same thing first: it starts writing code.

Blitzy does the opposite.

Before writing a single line, Blitzy spends days reverse engineering your entire codebase.

Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade.

The result is a dynamic knowledge graph that understands your software the way a principal engineer would after 30 years in the building.

Other tools guess at context with grep searches and markdown files.

Blitzy never guesses.

It builds true understanding first, then delivers over 80% of entire software epics autonomously.

Validated, end-to-end tested, production-grade pull requests.

That's why Fortune 500 engineering teams trust Blitzy with the codebases that matter most.

See for yourself at blitzy.com.

That's B-L-I-T-Z-Y dot com.