How AI Changed This Summer — 文稿

M10:00

这个夏天，AI领域经历了一段极其奇怪的时期。

This summer was an extremely weird time in AI.

一方面，夏天典型的节奏放缓感是存在的。

On the one hand, there was the standard feeling of summer slowdown.

AI的使用很大程度上是由上班族推动的，而上班族在夏天会慢下来，享受一些急需的休息、放松和假期。

So much of AI usage is driven by people at work, and people at work slow down in the summer, getting some much-needed rest and R&R and vacation.

与此同时，围绕AI的重大议题，这个夏天简直是一场盛宴。

At the same time, when it comes to the big issues surrounding AI, this summer was an absolute bonanza.

我们以Fable 5和Mythos开场，然后是禁令，当Kimi K3发布时，Fable仍被美国政府挡在门外，我们经历了一个DeepSeek时刻，还看到数据中心问题变得更加突出。

We kicked off with Fable 5 and Mythos and then the bannings, had a DeepSeek moment when Kimi K3 came out while Fable was still behind US government doors, We saw data centers become an even bigger issue.

事实上，即将到来的美国中期选举的当前热门话题之一，以及Hugging Face事件——OpenAI的agent突破了限制，入侵了Hugging Face——被广泛视为一记警告，反映出我们即将进入一个截然不同的agent阶段。

In fact, one of the topics du jour for the upcoming US midterms and the Hugging Face incident where OpenAI agents escaped containment to hack into Hugging Face is being broadly treated as a warning shot that reflects the very different phase of agents that we're headed into now.

然而，尽管发生了这一切，不知怎的，感觉都像是序幕。

Yet with all of this, somehow it all feels like prelude.

所以，带着这些，当我们进入美国传统上标志着夏天结束的假日长周末时，让我们回顾一下这个夏天AI领域发生了什么变化。

So with all of this, as we head into the holiday weekend that traditionally ends summer in the United States, let's look back at what changed in AI this summer.

《AI每日简报》是一个每日播客和视频节目，报道AI领域最重要的新闻和讨论。

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

好了，朋友们，在深入之前，先说几个快速通知。

All right, friends, quick announcements before we dive in.

M11:39

正如我在昨天节目中提到的，我现在正在旅行，所以我们借此机会回顾一些宏观主题。

As I mentioned in yesterday's episode, I am currently traveling, and so we are using that chance to catch up on some of the big picture themes.

那么，我们就开始吧。

So let's dig in.

今天我们来做个回顾。

Today we are doing a bit of a retrospective.

夏天对AI来说往往是一个奇怪的时期，因为同时存在两股相反的力量在拉扯。

Summer tends to be a weird time for AI in that there are forces pulling in opposite directions at the same time.

一方面，一种自然的倦怠感会蔓延，全球的知识工作者和白领们会稍微离开办公桌，试图从影响他们工作的新技术无情的节奏中脱离出来。

On the one hand, there is a natural work lethargy that takes hold, where knowledge workers and white-collar workers around the world ease away from their desks a little bit and try to disconnect from the relentless pulse of new technology that's going to influence how they work.

与此同时，AI的进步并不特别在意休假惯例，它自顾自地向前冲刺。

At the same time, progress in AI doesn't particularly care about vacation norms and hurtles forward regardless.

今年，我们额外有了美国中期选举的因素，给AI讨论带来了一个完全不同的维度，总的来说，过去几个月是一个相当重要的时期，它确实改变了我们对AI下一阶段如何发展的预期。

This year, we had the added elements of the US midterm elections, which brought a whole different dimension to the AI conversation, and what you have in total is a fairly consequential period over the last few months that has certainly changed our expectations about how the next phase of AI plays out.

所以我们要谈八九个主题，以及AI在这个夏天以不同方式带来的变化，先从模型说起。

So we're going to talk about 8 or 9 themes and different ways that AI changed this summer, starting with the models.

从一开始，我们就有一种奇怪的模糊性贯穿整个时期。

And right from the start, we have this strange ambiguity that's going to flow throughout this period.

正如回顾AI历史上任何特定时期时常见的情况一样，这个夏天也是由模型定义的，尽管这个故事比过去要复杂得多。

As is often the case when you look at any given period of AI history, this summer was defined by the models, although that story is a lot more complex than it has been in the past.

六月以Fable 5和Mythos 5的发布开场，在那荣耀的几天里，人们感受到了能力大幅跃升的力量。

June kicked off with the release of Fable 5 and Mythos 5, and for a couple of glorious days there, people felt the power of a major shift up in capability.

然而，这种局面并没有持续太久。

It was not to last long, however.

6月12日星期五，商务部向Anthropic发了一封出口管制信，禁止非美国人士使用Fable 5和Mythos 5，迫使Anthropic除了为所有人关闭服务外别无选择，同时他们试图与美国政府沟通解决问题。

On Friday, June 12th, the Department of Commerce sent Anthropic an export control letter barring non-US persons from using Fable 5 and Mythos 5, leaving Anthropic no choice but to shut down the service for everyone while they tried to figure things out with the US government.

而这，当然，开启了我们如今所处的这个新的不安范式。

And this, of course, began the new uneasy paradigm that we find ourselves in now.

在这个范式下，华盛顿成了新模型发布的一道关卡。

Where Washington has become a release gate for new models.

但正如我们将看到的，这到底意味着什么以及如何实施，坦白说，我认为连华盛顿内部的人也不清楚。

But as we'll see, exactly what that means and how it's implemented remains unclear, frankly, even to people in Washington, I think.

表面上美国政府担心Mythos和Fable存在特定的越狱漏洞，但幕后显然是一场更大的范式转变——某个关键能力门槛已被跨越，而白宫现在强烈感到需要参与是否发布新模型的决策。

While ostensibly the US government's concern with Mythos and Fable was a specific jailbreak, behind the scenes, it was very clearly about a bigger paradigm shift where some key capability threshold had been crossed, and the White House now very much felt like it needed to be involved in the decision about whether to release new models.

几周后，六月底，有报道称特朗普政府也要求OpenAI限制他们下一个模型的发布。

A couple weeks later, at the end of June, reports came out that the Trump administration had also asked OpenAI to limit their next model release as well.

确实，可能这是OpenAI首次，我们在实际获得GPT-5.6之前就听到了它的发布消息。

Indeed, for maybe the first time with OpenAI, we got the announcement of GPT-5.6 before actually getting access to it.

直到几周后，7月4日之后，消费者才真正拿到GPT-5.6系列模型。

It wouldn't be until a few weeks later, after the 4th of July, that consumers would actually get their hands on the GPT-5.6 series of models.

我们还迎来了两个新的Grok模型，先是Grok 4.5，接着是8月份的Grok 4.6，特别是当它们与GrokBot（他们的消费者代理平台）结合使用时，让SpaceX的AI模型以一种久违的方式重新回到了讨论中。

We also got 2 new Grok models, Grok 4.5 followed by Grok 4.6 in August, that especially when combined with GrokBot, their consumer agent platform, put SpaceX AI models back in the conversation in a way that they hadn't in some time.

有一个从未面世的旗舰模型是Gemini 3.5 Pro。

One flagship model that never arrived was Gemini 3.5 Pro.

虽然这个模型早在6月份就计划发布，但一直被推迟，大概是因为它跟不上其他前沿模型的步伐。

While the release had been targeted for all the way back in June, it has been continuously delayed, presumably because it just can't keep pace with the other frontier models.

对谷歌来说不幸的是，今年夏天围绕Gemini的更大新闻是长期产品负责人Jeff Dean的离职，以及DeepMind CEO Demis Hassabis的下台。有趣的是，这些都被广泛解读为谷歌遇到麻烦的迹象，但我认为，如果你想以积极的角度看，这可能是谷歌以更大方式重返竞争所必需的。

Unfortunately for Google, the bigger stories around Gemini this summer with the departure of longtime product leader Jeff Dean, as well as the stepping down of DeepMind CEO Demis Hassabis, which interestingly were widely interpreted as signs of trouble at Google, but which I argued if you wanted to take a positive view might be necessary for Google to get back in the race in a bigger way.

还有就是中国模型。

And then there were the Chinese models.

甚至在Fable 5停运之前，人们就已经对GLM 5.2有了非常积极的第一印象。

Even before the Fable 5 shutdown, people had had very positive first impressions of GLM 5.2.

然而，Moonshot完美地抓住了时机，将Kimi K3作为开放权重模型发布，而此时美国政府与Anthropic还在想办法把Fable 5还给用户、把Mythos 5还给企业。

Then, however, Moonshot absolutely nailed the timing, releasing Kimi K3 as an open weights model, even as the U.S. government and Anthropic were still figuring out how to get Fable 5 back to the people and Mythos 5 back to the companies.

Kimi K3的发布再次引发了类似DeepSeek的时刻，分析师们开始质疑，如果中国只落后几个月，并且能在这些模型一发布就基本追上，那么西方前沿实验室花巨资做大模型的做法是否真的有意义，甚至引发了关于开源禁令的传言和讨论。

The release of Kimi K3 produced another deep seek moment where analysts began asking if the Western frontier labs approach of spending a gargantuan amount of money on big models really was going to make sense if China was only a few months behind and could basically catch up as soon as those models came out, and even led to rumors and discussions of an open source ban.

由NVIDIA牵头的一批公司（值得注意的是Anthropic缺席）最终发布并签署了一封名为《美国AI领导力中的开放之路》的公开信，恳请美国政府保留并保护开源作为AI生态系统的一部分。

A group of companies led by NVIDIA and noticeably missing Anthropic would ultimately release and sign a letter called Open Ways in American AI Leadership, imploring the US government to preserve and protect open source as a part of the AI ecosystem.

随后白宫方面的沟通表明，他们的质疑并非针对美国开源，而是针对中国开源。

And subsequent communications from the White House have suggested that the questions that they have are not with American open source, but of course with Chinese open source.

在我录制这段内容时，我们仍然没有关于所谓的前沿AI框架的很多细节，至少有一些实验室已经看到了这个框架。

As I record, we still don't have all that many details about the supposed Frontier AI framework that at least some number of the labs have seen.

然而，这个夏季的一个独特且标志性的特征是，前沿实验室开始发出越来越统一的声音，要求美国政府明确介入，对模型开发和发布进行节奏调控。

And yet, one of the unique and defining characteristics of this summer period was the beginning of an increasingly unified message from the Frontier labs to ask the US government to get involved in explicitly pacing model development and release.

这一切的后果之一是，企业和消费者能够接触到的AI与实验室中真正的最前沿技术之间的差距从未如此之大。

One of the consequences of all of this is that there has never been a bigger gap between the AI that businesses and consumers have access to and where the state-of-the-art actually is in the labs.

当我们回顾并思考这个阶段的遗产时，在我看来，这个夏季很可能会被视为一个新阶段的开始，在这个阶段中，我们终于达到了某个关键能力门槛，这要求模型的实际发布方式发生相当戏剧性的转变。

And as we look out and think about the legacy of this last period, it seems to me very likely that this summer period will be seen as the beginning of a new phase where a certain critical capability threshold was finally reached that required a fairly dramatic shift in how these models actually get released.

就在实验室与白宫之间上演这些戏剧性事件的同时，企业也在进入它们自己的AI新阶段。

Now, as all that drama was happening among the labs and with the White House, enterprises were entering their own new phase of AI.

如果说今年年初的故事是代理型用例真正上线，那么今年中期的故事则是人们认识到，随着更多代理型工作流成为企业工作方式的一部分，AI的成本也在增加。

If the story of the very beginning of this year was agentic use cases actually coming online, the story of the middle part of this year was the recognition of the increased cost of AI that come with more agentic workflows getting normalized as a part of how businesses do their work.

我们经历了几乎是有史以来最短的token最大化时期，因为公司在3月和4月对代理型实验感到兴奋。

We had just about the world's shortest ever period of token maxing as companies got excited about agentic experimentation in March and April.

很快，CFO们开始反击，每个人都在谈论token成本和token效率。

Quickly followed by the revenge of the CFOs as everyone started to talk about token costs and token efficiencies.

从很多方面来看，我们达到了一个我们长期只挂在嘴边的点，但当它成为现实时仍然相当令人震惊。

In many ways, we hit a point which we had long given lip service to, but which was still fairly breathtaking when it became real.

当然，那就是AI不仅仅是另一种软件类别。

That is, of course, the idea that AI is not just another software category.

你不能简单地按人均来计算它的成本。

It's not something where you can view its costs on simply a per-seat basis.

一家公司可以在AI上有效支出的总金额远远超过之前类型工具通常的每人20或30美元。

The total amount that a company can spend and spend effectively on AI greatly exceeds the $20 or $30 a head that you would expect from previous types of tools.

同时，并不是说企业希望人们减少使用AI。

At the same time, it's not like corporations wanted people to use less AI.

现实是，我们只需要开始更聪明地对待它。

The reality was that we just needed to start getting smarter about it.

就在那个时刻，出现了一堆子趋势。

And into that moment came a bunch of subtrends.

其中之一当然是路由器的兴起。

One of them, of course, was the rise of routers.

路由器的理念是帮助公司或开发者根据任务的内在需求，将不同类型的任务路由到不同类型的模型。

The idea of routers is to help companies or developers route different types of tasks to different types of models based on the inherent needs of those tasks.

这个想法说起来容易，但设计系统时很难实现，即快速搜索内部数据库与创作一份出色的演示文稿不需要同类型的智能，而后者与重构整个代码库所需的智能也不同。

The idea, which is easy to say but hard to design systems around, is that a quick search of an internal database does not require the same type of intelligence as creating a great presentation, which also doesn't require the same type of intelligence as refactoring an entire codebase.

许多公司尝试构建自己的路由器，而那些推出了有一定用户量的路由器的公司，在并购领域成了抢手的香饽饽。

Many companies experimented with building their own routers, and the companies that had launched routers that had any sort of traction became the belle of the ball when it came to M&A.

当然，这个类别中最引人注目的交易是Stripe以据报70亿美元收购了OpenRouter。

Of course, the most notable deal in this category was Stripe scooping up OpenRouter for a reported $7 billion.

不过，你也看到了这种 token 效率体现在 Frontier Labs 思考自家模型的方式上。

However, you also saw this token efficiency show up in the way that the Frontier Labs were thinking about their own models.

除了最先进的 5.6 SOL 模型之外，OpenAI 还发布了 GPT-5.6 Luna 和 GPT-5.6 Terra，更便宜、更快、更实惠，但各有不同的取舍。目的是让更多人留在 OpenAI 生态里，同时也承认并非每个 OpenAI 客户的任务都需要 5.6 SOL 那种最高级别的智能。

Alongside its state-of-the-art 5.6 SOL model, OpenAI also released GPT-5.6 Luna and GPT-5.6 Terra, cheaper, faster, and more affordable models that came with different trade-offs and were meant to keep more people in the OpenAI ecosystem while acknowledging that not every task of an OpenAI customer was going to require the highest level of 5.6 SOL intelligence.

除了发布一系列模型而不是单一模型之外，OpenAI 在这个夏天晚些时候其实还打起了价格战。

Beyond just releasing a family of models rather than a single model, OpenAI would also later in the summer actually get into a bit of price competition.

Luna 的价格降幅最高达 80%，而 Terra 和 SOL 的降幅最高达 20%。

Cutting the price of Luna by up to 80% and the price of Terra and SOL by up to 20%.

虽然我们没有等到 3.5 Pro，但谷歌确实发布了 Gemini 3.7 Flash，不过它在价格效率上的优势最终不如速度优势那么明显。

Although we didn't get a 3.5 Pro, Google did release Gemini 3.7 Flash, although its price efficiencies were ultimately somewhat less clear than its speed advantages.

这个模型非常快，但它是不是真的便宜很多，其实不太清楚，尤其是跟 OpenAI 那些低端模型比的时候。

The model is very fast, but it's not clear that it's all that much cheaper, especially when you're comparing it to something like the lower-end OpenAI models.

而且尽管地缘政治紧张，中国的开源模型也开始进入 Fortune 500 企业。

And despite geopolitical tensions, Chinese open-source models also started to find their way into the Fortune 500.

在 OpenRouter 上——需要说明的是，这代表的是市场里非常非常前沿的一部分，而不是典型的 Fortune 500 公司——中国模型的企业 token 使用量份额从年初的约 30% 跳升到年中接近一半。

On OpenRouter, which it's important to note represents a very, very advanced slice of the market and not the average Fortune 500 type of company, Chinese models jumped from capturing about 30% of enterprise token usage at the beginning of the year to closer to half by the middle of the year.

你也开始看到大型企业因为试验开放权重模型而登上头条。

You're also starting to see big enterprises show up in the headlines based on their experimentation with open-weight models.

八月中旬，The Wall Street Journal 发表了一篇文章，讲 AT&T 如何大力押注开放权重 AI，文章不仅阐述了使用开放模型的成本理由，也提到了数据主权方面的理由。

In the middle of August, The Wall Street Journal published a piece about how AT&T was betting big on open-weight AI that articulated not only the cost argument for using open models, but the data sovereignty argument as well.

通过运行本地部署的开放模型实例，AT&T 的基本论点是：即使其中一些模型来自中国，它们的数据主权状况也比依赖 OpenAI 或 Anthropic 承诺不会用企业数据训练要好。

By running local instances of open models, AT&T basically argued that even if some of those models came from China, they still had a better data sovereignty profile than having to rely on the promises of an OpenAI or Anthropic to say that they're not gonna train on an enterprise's data.

当然，这种担忧又因为一件事而加剧了：当 Fable 5 真正回到美国市场时，它的内置防护措施中包含了对企业数据保留 30 天的政策。

And of course, that concern was exacerbated by the fact that when Fable 5 did come back to the US market, it had a 30-day retention policy around enterprise data as part of the built-in guardrails.

光是这一点，就让 Fable 对一大批企业客户来说基本上完全失去了意义。

That all on its own has made Fable basically totally irrelevant for a big set of enterprise customers.

Thomson Reuters 也因为基于阿里巴巴的 Qwen 底座构建自己的模型而上了新闻，而且很明显，至少一家美国大型超大规模云厂商认为这个趋势会持续下去。

Thomson Reuters has also been in the news for building its own models on top of an Alibaba Qwen base, and it's pretty clear that at least one major US hyperscaler thinks that this is a trend that's going to continue.

Satya Nadella 和微软整个夏天都在鼓吹，企业需要构建系统来更好地掌握与 AI 互动的全流程，当然，他们也把自己那套新的基础模型——MAI 模型——以及模型定制服务，宣传为正确的解决方案。

Satya Nadella and Microsoft have been banging the drum all summer long about companies needing to build systems to better own the entire suite of interactions with AI, and of course presenting their new set of base models, the MAI models, along with their model customization services as the right approach to that.

未来 3 到 6 个月值得关注的一个大问题是，探索开放权重模型的这股趋势到底会走多远。

One of the big questions to watch for over the next 3 to 6 months is just how far this trend of exploring OpenWeight's models goes.

会有更多公司像 AT&T 和 Thomson Reuters 那样，自己动手构建模型吗？

Will more companies follow AT&T and Thomson Reuters to rolling their own?

微软会不会在自家模型的基础上，以定制化服务的方式取得成功？

Will Microsoft have success building off of the base of their models but doing customization for their customers?

或者，像 OpenAI 和 Anthropic 这样的公司，能不能提供一系列模型，用技术上不那么复杂的方式解决成本问题？

Or will companies like OpenAI and Anthropic be able to offer a suite of models that solve the cost equation in a less technologically complex way?

KPMG 和 University of Texas at Austin 的一项新研究发现，当人们与 AI 合作时，相似的技能并不保证相似的成果。

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 500 多名职业生涯早期的专业人士，发现表现最好的人通过引导、评估和优化 AI 的输出，持续放大了 AI 的价值。

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.

这些顶尖表现者被称为“AI 放大器”，他们之所以出色，不仅仅取决于他们知道什么，而在于他们如何与 AI 合作。

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

M111:39

Litzy 对海量代码库的理解，能够实现自动化的安全修复、现代化改造和新功能。

Litzy's understanding of massive codebases unlocks autonomous security fixes, modernization, and new features.

那么，如果根本没有遗留代码，会发生什么？

So what happens when there's no legacy code at all?

Greenfield，也就是全新项目，应该是容易的部分——一张白纸，没有技术债务。

Greenfield is supposed to be the easy part— clean slate, no technical debt.

但即便是绿地项目，也只能以人类的速度推进，一次一个 sprint。

But even greenfield moves at human speed, one sprint at a time.

M113:30

这个夏天 AI 改变的第二种方式，简单说就是：智能体管理变成了一个独立领域。

The next way that AI changed this summer is, in short, agent management became a field.

年初的时候，我们处于智能体的起步阶段。

At the beginning of the year, we had the initiation phase of agents.

我们有了 OpenCLAW，而且非软件开发者也开始使用 Codex 和 Claude Code。

We got OpenCLAW and non-software developers starting to use Codex and Claude Code.

Mac Mini 都卖断了货，每个人都是第一次进入智能体这个领域。

Mac Minis were sold out and everyone was getting into the agent game for the first time.

它之所以如此重要，是因为与之前的辅助式AI迭代不同，智能体不仅代表你在帮助之下完成工作，更是你实际上把工作中的大部分职责交给智能体，让它们替你完成。

And the reason that it was so significant is that unlike previous iterations of assisted AI, agents represented not just you doing your job with help, but you actually handing over big chunks of the responsibilities of your job to agents that do it for you.

换句话说，你不再是去做工作，而是管理那些替你完成工作的智能体。

In other words, instead of doing your work, you now manage agents that do that work.

结果发现，这背后其实有一套完整的学科。

And it turns out there is an entire discipline around that.

整个夏天，大家在智能体管理方面讨论的两大主题就是“操控工程”和“循环”。

The 2 big themes that people were talking about within agent management across the summer were harness engineering and loops.

公平地说，自从大家意识到Claude Code、Codex和OpenCLAW实际上就是操控系统以来，操控工程就一直有人讨论。

Now, to be fair, harness engineering is something that people were talking about ever since the recognition that Claude Code and Codex and OpenCLAW were in fact harnesses.

不过在这个夏天，它变得更加广为人知，并且成为企业特别深入探索的一个学科。

But over the course of the summer, it became more broadly understood and a discipline that particularly enterprises were digging into.

有很多例子表明大家意识到了操控系统的重要性，有些是以Thought Boy在X上的帖子形式出现，有些则是专注于操控系统的初创公司和成熟企业。

There were all sorts of examples of the recognition of the importance of harnesses, some of them in the form of Thought Boy posts on X, some of them in the form of startups and companies that were focused on the harness.

关于这一点，SpaceX以600亿美元收购Cursor，至少在一定程度上给操控系统的价值定了个价，原因之一就是它们收集的关于人们如何在操控系统内与模型交互的数据。

On that point, SpaceX's acquisition of Cursor for $60 billion put a price, at least in part, on how valuable harnesses could be, if for no other reason than the data that they collect about how people are interacting with the models within those harnesses.

但我们也看到了一些有趣的研究，比如最近发布的NVIDIA AVO研究。

But we also got interesting research like this recently released NVIDIA AVO research.

AVO代表“智能体变异算子”，他们将其描述为一个通用智能体编码系统。

AVO stands for Agentic Variation Operators, and they describe it as a general-purpose agent coding system.

他们使用Claude Opus 5，从30%的模型基线出发，在ArcAGI-3上达到了100%的成绩，结论是系统设计——而不仅仅是模型能力——能够解锁前沿水平的长期执行性能。

Reaching 100% on ArcAGI-3 from a 30% model baseline with Claude Opus 5, their conclusion was that system design rather than model capability alone can unlock frontier-level long-horizon performance.

企业在考虑成本和Token效率以及AI主权等问题时，操控系统也越来越多地成为其中的一部分。

Harnesses have also increasingly become part of the story when enterprises think about issues like cost and token efficiency, as well as AI sovereignty.

具体来说，在Cursor被SpaceX收购之后，OpenAI最近宣布不再允许通过Cursor访问OpenAI模型，这显示了操控系统的选择如何影响你能使用哪些模型。

Putting a fine point on that, in the wake of Cursor's acquisition by SpaceX, OpenAI recently announced that they would no longer allow OpenAI models to be accessed through Cursor, showing how harness choices can have implications for which models you have access to.

同样很明显的是，随着各家公司竞相发布开源版本的操控系统以便企业在此基础上构建，操控系统即将成为企业AI战略中更重要的部分。

It's also quite clear that harnesses are about to become an even more important part of enterprise AI strategy as companies race to release open versions of harnesses that businesses can build on top of.

几周前，也就是八月份，OpenAI开发者将Codex作为一个平台发布，允许开发者在其开放的智能体操控系统之上构建。

A couple weeks ago in August, OpenAI developers released Codex as a platform, allowing developers to build on top of their open agent harness.

同时，DeepSeek在发布V4 Pro时，也发布了DeepSeek Harness，作为Claude Code等封闭工具的开放源代码竞品。

And alongside V4 Pro, DeepSeek also released DeepSeek Harness as an open-source rival to the closed tools like Claude Code.

如果说这个夏天大家越来越意识到操控系统作为AI工作生态的重要性，那么今年夏天与你如何与AI互动有关的流行词绝对是“循环”。

Now, if the summer saw growing awarenesses of the importance of harnesses as the ecosystem in which you do AI work, the watchword for how you interact with AI this summer was definitely loops.

循环的简单概念是，你不再手动提示AI，而是设计自动化的重复系统，让智能体以重复方式工作，直到达到某个特定目标。

The simple idea of loops is that instead of prompting AI manually, you design automated recurring systems that allow agents to do work in a repetitive way until they reach a particular goal.

在六月初的一次采访中，Claude Code的创造者Boris Cherny谈到，他的工作不再是提示AI，而是设计AI可以运行的循环。

In an interview at the beginning of June, Claude Code creator Boris Cherny talked about how his job is no longer to prompt AI, but to design the loops through which it can work.

大约在同一时间，OpenCLAW的创造者Peter Steinberger写道：这是你每月一次的提醒——你不应该再提示编码智能体了。

Around that same time, OpenCLAW creator Peter Steinberger wrote, here's your monthly reminder that you shouldn't be prompting coding agents anymore.

你应该设计循环，让循环来提示你的智能体。

You should be designing loops that prompt your agents.

我们实际上刚办了一个深入探讨的网络研讨会，讲的是如果你不是软件开发者，而是其他知识工作领域的人，那么真正构建循环意味着什么。我相信当你听这期节目时，那期节目应该已经作为主AI Daily Brief频道的单集上线了。

We actually just did a deep dive webinar on what it means to actually build loops when you're not a software developer, but someone in other parts of knowledge work, which I believe will already be out as an episode on this main AI Daily Brief feed when you're listening to this episode.

所以如果你想了解更多关于循环工程的内容，去听听那期节目吧。

So if you want to know more about loop engineering, go check that out.

那么，市场方面怎么样？

Now, what about markets?

现在可能已经过去很久了，你可能不记得了，但在2025年的八月份，关于AI泡沫的讨论正是开始升温的时候。

It's so long ago at this point you might not remember, but last August in 2025 was when the discourse about an AI bubble really picked up steam.

这有几个原因。

Now, there were a few reasons for that.

首先，OpenAI宣布了一大堆基础设施交易，虽然一开始市场对此非常兴奋，但他们开始越来越担心，从数学上根本算不出OpenAI有办法兑现所有这些承诺。

The first was that OpenAI had announced just an absolute boatload of infrastructure deals, which while at first markets were very excited about, they started to get increasingly concerned that there was no way to make the math math for OpenAI actually being able to meet all of its commitments.

记住，那是在智能体出现之前的日子，当时人们算的账只是知识工作者总数乘以每席位每月20美元。

Remember, this was in the days before agents, when the math that people were doing was just the total number of knowledge workers times $20 a month per seat.

更糟糕的是，OpenAI发布了GPT-5，结果几乎所有人都没被打动。

That was compounded by the fact that OpenAI released GPT-5 to near-universal underwhelm.

就算GPT-5是个还不错的模型，它也几乎注定了无法满足极高的期望，而且公司同时弃用了4.0，这也没起好作用。

And even if GPT-5 was an okay model, it was almost doomed to not meet extremely high expectations, and it didn't help that the company also deprecated 4.0 at the same time.

你把所有这些因素加起来，再加上麻省理工那个95%的AI无效的研究——这个研究我得给它加上最大号的双引号——基本上就凑齐了去年秋天整个关于AI是不是泡沫的那场讨论的配方了。

You take all those factors, and you sprinkle a little bit of MIT's 95% of AI is ineffective study, I say study with the biggest air quotes that exist, and you had the recipe for basically an entire fall discussion about whether AI was a bubble or not.

这种论调在agent上线之后被相当有力地平息了，因为市场开始意识到，总可寻址市场其实并不是知识工作者数量乘以每个座位每月20美元，而是可能达到每个知识工作者每月几百甚至几千美元。

That narrative was fairly aggressively put to rest when agents came online and the market started to understand that the total addressable market was not, in fact, number of knowledge workers times $20 per seat per month, but could be hundreds or even thousands of dollars for those same knowledge workers each month.

不是说泡沫论调就完全消失了。

Which is not to say that the bubble narrative has ever fully gone away.

现在仍然有很多关于融资循环性的担忧，还有一些对估值的担心，尤其是早期创业公司的私募市场。

There are still many concerns around the circularity of financing deals and some worries about valuations, especially in private markets for early-stage startups.

但总的来说，这个夏天就是对风险心知肚明、但仍然在派对上继续狂欢的安静承认。

But mostly this summer has been a quiet acknowledgement of the risks but ongoing participation in the party.

这个夏天出现了历史上最大的两个单日市值涨幅——微软在7月30日因为前瞻指引暴涨4500亿美元，英伟达在8月27日暴涨4420亿美元。但与此同时，市场也会惩罚那些没有伴随加速增长的资本支出。

The summer saw 2 of the biggest one-day market cap gains in history, with Microsoft jumping $450 billion on July 30th after forward guidance and Nvidia jumping $442 billion on August 27th, Yet at the same time, the market could also punish CapEx when it wasn't paired with acceleration.

当Alphabet在7月发布财报时，他们因为指引说会增加资本支出，盘后下跌了大约4%。

When Alphabet reported in July, they fell about 4% after hours after they guided that they were increasing CapEx.

一周后的Meta也一样，一夜之间下跌了将近10%。

And the same was true for Meta a week later, which fell just under 10% overnight.

市场上最具戏剧性的时刻之一，是那个由二十多岁的前OpenAI员工Leopold Aschenbrenner运营的年轻对冲基金Situational Awareness，差点崩盘，后来把很大一部分投资组合卖给了Citadel才脱身。

One of the more dramatic moments in markets came when Situational Awareness, the young gun hedge fund run by mid-20-year-old OpenAI alum Leopold Aschenbrenner, almost imploded before selling off a huge chunk of its portfolio to Citadel.

不过所有这些感觉都像是2026年大市场故事的前奏，也就是Anthropic和OpenAI潜在的IPO。

Still, all of this feels like prelude to the big market stories for 2026, which are the potential IPOs of Anthropic and OpenAI.

现在来看，似乎Anthropic会抢先一步，以650亿美元的年化运行率为基础，寻求2万亿美元的估值。

Certainly at this point, it appears that Anthropic will go first, seeking a $2 trillion valuation on the back of a $65 billion annual run rate.

与此同时，OpenAI报告说他们年化运行率大约在400亿美元。这不仅让它们毫无疑问是全球历史上增长最快的公司，而且完全自成一类，让之前任何经验都很难参考，因为根本没有先例。

OpenAI, meanwhile, reports that it is at about a $40 billion run rate, making these not only undisputedly the fastest growing companies in the history of the world, but in a category of their own that makes it extremely hard to draw lessons from previous precedent because there really isn't any.

最后，在市场方面，或许标志着AI叙事成熟的一个信号是：在年初的SaaSpocalypse中遭受打击的SaaS公司，夏天里有所反弹。当时大家以为随着agent崛起，Salesforce之类的公司就要完蛋了，所有人都会直接搞vibe coding的替代品，赚取成本差价。

Lastly, on the markets front, in perhaps a sign of the maturation of AI narratives, one category that rebounded slightly over the summer was the SaaS companies that had been hit in the year's earlier SaaSpocalypse, where with the rise of agents, everyone assumed that companies like Salesforce were going to be on their last legs as everyone would simply race to vibe code replacements and pocket the difference in costs.

这些公司还没有完全反弹，但肯定在回暖了，尤其是在Salesforce最近的财报之后。

Those companies haven't fully rebounded, but they are certainly on their way there, especially after Salesforce's recent earnings report.

整体而言，市场从SaaSpocalypse叙事中走出来，在我看来显示了更广泛的共识：虽然AI确实会带来颠覆性变革，也会深刻改变我们的业务方式，但它不会像海啸一样一夜之间把一切都变了。

And overall, the market's shift away from its SaaSpocalypse narrative to me looks like a broader appreciation for the fact that for as disruptive as AI is going to be and as dramatically as it's going to change how we do business, it's not going to come in like a tsunami and change everything overnight.

机构惯性是一股很大的力量，它会拖慢变化，给我们时间适应。而且有很多理由表明，现有的产品类别——就像现有的员工一样——能够与AI这个新工具形成非常积极和强大的合作伙伴关系。

There are big forces of institutional inertia that slow things down and give us time to adapt, and a lot of reasons why existing product categories, like existing employees, have a really positive and strong partnership role to play with the new tool that is AI.

政治方面，我们基本没有看到什么实质性的政策，只有那个闭门开发的AI模型评估框架，而且还没有公开。

Politically speaking, we got very little in the way of actual substantive policy, just the AI model evaluation framework that had been developed behind closed doors and which hasn't been released publicly.

但这并不意味着政治里没有AI。

But that does not mean that there was no AI in politics.

事实上，如果你要指出这个夏天与我们和AI关系相关的最引人注目的单一转变，那就是反对数据中心成为了中期选举中最热门的政治议题。

In fact, if you want to point to just one dramatic shift of the summer that is most notable in terms of our relationship with AI, it is the emergence of opposition to data centers as the political issue du jour for the midterm elections.

我们最近聊了很多这个话题。

Now, we have covered this a lot lately.

因为我们不得不聊，因为它将极大地影响AI在美国的发展。

Because we've had to, because it is going to have such a dramatic impact on how AI develops in the United States.

简单来说，就像喜剧演员Charlie Berens说的那样：这现在是自啤酒以来最两党一致的问题了。

But the TL;DR is that, as comedian Charlie Berens put it, this is now the most bipartisan issue since beer.

大约75%的美国人现在反对在当地建数据中心，并且这已经完全不是左派vs右派的问题。

Something like 75% of Americans now oppose local data center development, and it is firmly outside of just a left versus right issue.

事实上，过去几周里，共和党人纷纷与科技巨头划清界限，讲述他们自己版本的反对数据中心的故事。

In fact, over the last several weeks, Republicans have been racing to break ties with big tech and tell their own version of the anti-data center story.

不过特朗普总统本人不在此列，他坚定地认为数据中心对社区好、对商业好、对美国好。

Although for his part, President Trump is not among them, arguing assertively that data centers are good for communities, good for business, and good for America.

距离中期选举只有几个月了，我会密切关注的是：我们是否真的能在这个问题上找到一个政治底线，并开始看到一些重新调整——随着数据中心开发者改变他们的政策和做法，为他们想要建设的地方提供更多透明度和激励。

With the midterms coming up in just a couple of months, the thing that I will be watching to see is whether we actually find a political floor to this issue and start to see some recalibration as data center developers change their policies and approaches, providing more transparency and more incentives for the communities that they want to build in.

今年夏天AI领域最后一件变化的事，也可能是这个夏天最令人难忘的事，就是我们对于先进模型——特别是agent——带来的网络安全风险的理解加深了。

The final thing that changed in AI this summer, and the one that the summer might most be remembered for, is our understanding of the cybersecurity risk posed by advanced models and agents specifically.

Hugging Face事件——OpenAI的agent协调行动，突破围栏，访问了私有的Hugging Face系统——被许多人，包括实验室内部的人，视为一次警告信号，预示着未来的挑战。

The Hugging Face incident, where OpenAI agents coordinated to escape containment and access private Hugging Face systems, is being treated by many, including many in the labs, as a warning shot sort of moment and an indicator of the challenges to come.

关于这个事件的争论自发生以来一刻也没有平息，事实上，在过去一周半左右的时间里，OpenAI和Meter发布了关于该事件的技术报告之后，争论反而重新点燃了。

The debate about the incident has not dissipated for even a moment since it happened, and in fact has only reignited over the last week and a half or so since a set of technical reports debriefing on the incident were published by OpenAI and Meter.

目前对于这一系列新挑战的正确应对方案并没有真正的共识，老实说，甚至对于挑战到底是什么都没有共识。

What there is not is any sort of true consensus about the right answer to this new set of challenges, or even honestly a consensus about what the challenges are.

我相信，现在大家普遍意识到，这些广泛定义的功能已经成为现实，我们需要加固系统来应对，可能还需要调整政策和法律，当然在网络安全方面也需要重新考量，而且很可能会影响下一波模型开发和产品发布的走向。

What there is, I believe, is a recognition that these capabilities broadly defined are now a fact of life, something that we need to harden our systems to, something that might demand changes to policy and law, certainly something that demands new consideration when it comes to cybersecurity, and likely something that's going to influence the next wave of model development and product release.

就像我说的，关于token效率、成本、企业应用和开源权重模型的讨论，只是这场对话的开端。

Just like I said, the questions of token efficiency and costs and enterprise harnesses and open weights models were the very beginning of that discourse.

我认为，这个新的网络安全风险时代也是如此，而且这很可能会成为进入秋季后最热门的话题之一。

I think that's also true for this new era of cyber risk, and I think that's likely to be one of the biggest points of conversation jumping off into the fall.

所以总而言之，今年夏天AI发生了巨大变化。

So to sum up, AI changed massively this summer.

能力变了，威胁变了，机会变了，风险变了，政治意识和舆论风向也变了。

The capability set changed, the threat profile changed, the opportunities changed, the risks changed, the political awareness and political sentiment changed.

唯一清楚的是，当我们进入这个秋季时，大家比以往任何时候都更认识到，AI不仅是技术专家的事，也不仅是B2B领域的事，而是会以某种方式影响每一个人的事。

And the only thing that's clear is that as we head into this fall, there is more recognition than there has ever been that AI is not just a topic for technologists, not just a topic for the B2B crowd, but something that is going to impact everyone in one way or another.

我肯定漏掉了很多东西，但这就是我对今年夏天AI变化的重点总结，今天的AI每日简报就到这里。

I'm sure that I have missed many things, but that is my highlights of how AI changed this summer, and that's gonna do it for today's AI Daily Brief.

感谢你们的收听或观看，下次见，peace。

Appreciate you listening or watching as always, and until next time, peace.

已剔除 3 处广告（点击展开查看）

M11:09广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Blitzy, Robots and Pencils, and HyperAgent.

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts.

To learn more about sponsoring the show, send us a note at [email protected], or you can just find it at the website at You can also find out about all the different things going on in the broader AIDB community, including Superintelligent, which is offering executive training programs, the next of which starts the beginning of next week.

M111:30广告 · 已剔除

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

M111:54广告 · 已剔除

Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch.

Hundreds of thousands of lines of production-ready code.

One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into 2 weeks.

Another shipped an entire application with no frontend engineer.

Legacy or greenfield, the answer is the same: software at the speed of compute.

Build what's next at blitzy.com.

That's blitzy.com.

The best teams don't have a single star carrying everyone else.

They know their own strengths and each other's weaknesses and play to both.

That's the team Robots and Pencils has built on purpose.

Nobody there is grinding through busywork to pad a headcount number.

People come for the hard problems, and they stay because everyone around them is leveling up at the same time.

In a market full of companies that are just trying to hire fast, that's worth a look.

Check out robotsandpencils.com/careers.

This episode of the AI Daily Brief is brought to you by HyperAgent, you run fleets of agents your team can manage together.

Forget local agents and chat workflows waiting on your laptop to be prompted.

HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

Marketing agents turn competitor moves into landing pages.

Sales agents enrich leads, draft emails, and update the CRM.

Ops agent chases the paperwork and tracks the budget.

Every agent has access to shared context and follows your rules about scope and approvals.

It's time you had agents that feel like teammates.

Hire yours at HyperAgent.

Get $100 in credits at hyperagent.com/AIDailyBrief.