OpenAI DevDay 上最重要的新 AI 工具

OpenAI团队表示，内部发生的重大变化之一是，最新一代模型（如Astra）的能力显著加快了他们的开发速度，以至于许多他们原本以为要到2027年才能实现的功能，实际上已经包含在这次DevDay的发布中了。

The team at OpenAI said that one of the big changes that had happened internally was that the power of the latest generation of models, like Astra, had increased their speed of development so significantly that many things that they thought were only going to come in 2027 were actually coming as part of this DevDay announcement.

结果就是，他们发布了超过20项不同的产品和公告，其中包括一些重磅消息，比如OpenAI对Muse和Grokbot的回应，还有一些低调的亮点，比如企业账户现在可以在开放市场上使用OpenAI积分来购买开源模型的访问权限。

The result of that was more than 20 different launches and announcements, including some big headliners like OpenAI's answer to Muse and Grokbot, and some sleeper hits, like the fact that enterprise accounts can now use OpenAI credits on an open marketplace to buy access to open source models as well.

总的来说，OpenAI DevDay上我们看到的并没有改变我们一直在观察的行业大趋势。

Overall, what we got at OpenAI DevDay does not change the big patterns and trends that we've been seeing in the industry.

那就是向更经济高效的模型转变，这些模型转向更持久的工作，而其中一部分持久工作也从单人体验变成了多人协作体验。

A move to more cost-efficient models, those models moving to more persistent work, and some amount of that persistent work moving from a solo to a multiplayer experience.

相反，DevDay进一步印证了这些趋势是值得关注的，它们将在未来几个月重塑我们所有人使用AI的方式。

Instead, what DevDay reinforced is that these are the trends to pay attention to and the ones that will reshape how we all use AI in the months to come.

《AI每日简报》是一档每日更新的播客和视频节目，内容涵盖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，在开始之前，先快速说几个通知。

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

首先，感谢今天的赞助商：KPMG、Blitzy、Robots and Pencils和Hyperagent。

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

想听无广告版本的节目，可以去 patreon.com/aidailybrief，或者在 Apple Podcasts 上订阅。

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

想了解如何赞助节目，请发邮件到 sponsors@aidailybrief.ai。

To learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai.

另外，在 aidailybrief.ai 上还有很多其他好东西。

There are also all sorts of other goodies at aidailybrief.ai.

在我们的网站上，你可以免费访问我们所有的 AI 培训项目。

On our website, you can get access to all of our free AI training programs.

这些包括免费的为期数周的自学项目，比如现在正在进行的 Multiplayer AI Sprint，还有免费的线上研讨会，比如即将在 10 月 1 日星期四中午举行的那个，主题是构建你自己的个人 AI 基准。

Those include free multi-week self-directed programs like the Multiplayer AI Sprint, which is live right now, but also free live webinars like the one that is coming up on Thursday, October 1st at noon, that is all about building your personal AI benchmark.

你还可以找到我们付费培训的链接，比如 Superintelligent Executive Catch-up 和 Superintelligent Executive Agent Leadership Program，这两个项目都在招收新学员，课程将在未来几周内开始。

You can also find links to our paid trainings, such as the Superintelligent Executive Catch-up and the Superintelligent Executive Agent Leadership Program, both of which are registering new cohorts that start over the next couple of weeks.

好了，关于今天的节目，最后说一点。

Now, last note on today's episode.

本来我打算把昨天白宫发生的一切和 OpenAI DevDay 一起详细讲一遍，但内容实在太多了，所以今天我决定只做 DevDay 的回顾，之后我们会更深入地探讨特朗普与所有主要前沿实验室负责人会面的成果，以及这是否会改变 AI 未来的发展。

Originally, I was going to try to jam in a full accounting of everything that happened at the White House yesterday on top of OpenAI DevDay, but there was just too much to fit, so I've decided to go all DevDay recap today, and then we will dig much deeper into what came out of Trump's meeting with the leaders of all the major frontier labs, and if and how it changes anything about the future development of AI.

不过现在，我们先来聊聊DevDay，看看它透露了什么，不只是关于OpenAI，还有整个AI行业现在的发展状况。

For now, though, let's get into DevDay and find out what it says, not just about OpenAI, but about where we are with AI in general.

OpenAI今天发布的东西简直多得惊人。

It was an absolutely monster day of announcements from OpenAI.

实际上，因为发布的东西太多了，今天的节目我们只能把重点都过一遍，聊聊每个发布是什么、是不是意料之中、大家的第一印象如何，以及它是否改变了AI竞赛的格局。

In fact, there were so many that on today's episode, we are just going to get through all of the big ones, discussing what the announcement was, whether it was expected or not, people's first impressions, and how it shifts the AI race, if at all.

这周晚些时候，我们会深入探讨其中最重要的几个，但现在，光是过一遍就已经够我们忙的了。

Later on in the week, we'll go deeper on some of the most important ones, but for now, it is going to take all that we have just to get through it.

首先登场的是Dots，这是OpenAI对Muse、Grokbot以及席卷AI领域的个人代理浪潮的回应。

First up is Dots, OpenAI's answer to Muse, Grokbot, and the wave of personal agents that are taking over the AI space.

Sam Altman把Dots描述为“能力超强、始终在线的代理，能处理你能想到的任何事情。”

Sam Altman presented Dots as, quote, "Remarkably capable, always-on agents built to handle really anything you can think of.

它们把AI带到了一个全新的形态。

They bring AI to a whole new form factor."

有了Dots，用户可以创建一个持久代理，通过短信风格的界面与它交流，然后让它利用自己的云电脑去执行任务，支持多达4万个不同的应用。

With Dots, users can create a persistent agent, communicate with it through a text message style interface, and then set them to work on tasks using their own cloud computer with support for 40,000 different apps.

Dots 还可以通过语音通话或 Microsoft Teams 和 Slack 进行交流，并且在各个服务和会话之间保持持续的上下文。

Dots can also communicate via voice call or across Microsoft Teams and Slack, with persistent context carried over between services and sessions.

从功能上讲，它和 Muse 有很多相似之处，甚至连可爱的吉祥物都一样。

Functionally, it shares a lot with Muse, right down to the cutesy mascot.

不过，至少在发布时，有几个需要注意的地方。

However, at least at the time of release, there are a few caveats worth noting.

目前，用户一次只能创建一个 dot，但 OpenAI 计划未来扩展到代理团队。

For now, users can only create one dot at a time, but OpenAI plans to expand to teams of agents in the future.

与 Muse 的另一个重大区别是，OpenAI 只向专业版、商业版和企业版客户提供 Dots。

The other big difference from Muse is that OpenAI is only offering Dots to pro, business, and enterprise customers.

Muse 之所以如此成功，一个很大的原因是它完全免费，这意味着 OpenAI 自然限制了他们在这一方面的竞争力。

One of the big reasons Muse has been so successful is making it completely free, meaning OpenAI is naturally limiting their ability to compete on that front.

首席财务官 Sarah Friar 表示，愿景是将 Dots 带给所有消费者，但目前这还是一款面向专业用户的产品。

CFO Sarah Friar said that the vision is to bring Dots to the whole consumer base, but for now, this is a prosumer product.

OpenAI 的一大卖点是他们使用 GPT-6 Astra 来驱动这些代理，所以这是我们第一次看到真正前沿的模型来试点第一方个人代理。

OpenAI's big selling point is that they're using GPT-6 Astra to power the agents, so this is the first time we're seeing a truly frontier model pilot a first-party personal agent.

说实话，这个公告在很多方面都不太让人意外。

Now, this was, in many ways, the least surprising of the announcements.

自从OpenClau推出以来，这种个人助手的形式基本上就是必然趋势，连Grokbot这种面向企业的版本都引起了关注。

This personal agent form factor has been basically inevitable since the launch of OpenClau and even the interest in business-focused versions like Grokbot.

OpenAI进入这个领域几乎是板上钉钉的事。

OpenAI getting into this particular area was pretty much inevitable.

不过，Dots到底能不能经受住市场的考验，还得看后续表现。

Now, how ready for prime time Dots is remains to be seen.

现场演示确实出了一些问题，但我觉得这比X上那些爱挑刺的人说的要轻得多。演示出问题其实恰恰说明他们是真直播。

The live demo had some issues, although I would put way less in that than the chattering classes on X. Demos going wrong is basically just a way for you to know that they are actually live.

不幸的是，用户拿到Dots之后，还遇到了其他一些初期问题。

Unfortunately, once people got their hands on Dots, they ran into a number of other teething problems as well.

Chase Browser发了一段视频，里面Dot丢失了所有工作；Tenebra说他们本来很期待试用，但发现没法同时连接多台电脑来共享完整上下文。

Chase Browser posted a session where Dot lost all of its work, and Tenebra said that they were excited to try it out, but found that they couldn't connect multiple computers at once to share their full context.

其他人的体验倒是好一些。

Others had a better experience.

分析师Max Weinbach写道：“我测试这类工具的基本标准是，它能不能在访问我的邮箱和Google Drive的情况下，自动帮我处理报销。”

Analyst Max Weinbach wrote, "My basic test for how well one of these works is can it autonomously do my expenses with access to my email and Google Drive?

Grokbot搞得一团糟，没有按我说的做。

Grokbot made a mess and didn't do what I told it.

它头三天还行，后来就开始出错了。

It did the first three days and then started to mess up.

Dots第一次就搞定了，而且一直表现很好。”

Dots did it after the first time and has been good."

Justin Schroeder说，他觉得Dots比Grokbot或Muse更专注于工作。

Justin Schroeder said that he thought that Dots was even more work-focused than Grokbot or Muse.

他写道：“别把它当成个人助理，更像是一个Codex编排器和Slack协作者。”

"Think a bit less personal assistant," he writes, "and a bit more Codex orchestrator and Slack collaborator."

Justin还写道：“Slack才是它真正大放异彩的地方。”

In fact, Justin writes, "Slack is where it really shines.

你的团队可以直接给你的机器人发消息，它也能执行操作，比如新建面板。”

Your team can message your bot directly, and your bot can take actions like spin up new panes."

另外，关于一个我觉得会到处讨论的话题，贾斯汀认为这本来应该是一个独立的应用程序。

Also getting at a debate which I think will be everywhere, Justin suggests that he thinks that this should've been a separate app.

他写道：“ChatGPT 很棒。”

"ChatGPT is great," he writes.

“Codex 也很棒。”

"Codex is great.

“在我看来，它们是截然不同的产品，而 Dots 更像 Codex，而不是 ChatGPT。”

They are very different products in my opinion, and Dots are way more Codex than ChatGPT."

《The Every》的总体评价几乎把它定为“待定”，因为用户反馈说，有时它确实表现出色，但有时又让人非常沮丧。

The Every vibe check almost leaves it as a TBD, with their users reporting flashes where it was truly excellent, but then other times where it was just extremely frustrating.

当 Brandon Chu 提出这是大家都在趋同的形态时，Nate B. Jones 写道：“我觉得我们需要区分形态和实用性。”

When Brandon Chu suggested that this was the form factor that everyone was converging on, Nate B. Jones wrote, "I think we need to distinguish between form factor and utility here.

“是的，目前形态确实在趋同。”

Yes, form factor is converging at the moment.

“但这可能会改变。”

That may change.

但实用性并不是。

But utility is not.

Muse 擅长特定的事情，比如打电话和处理实际邮件工作。

Muse is good at specific things, like phone calls and practical email work.

Instinct 擅长其他特定的事情，比如旅行。

Instinct is good at specific different things, like travel.

Dots 擅长把 AI 上下文当作工作界面来用。

Dots is good at AI context as a work surface.

不管怎么定位，这些都是不同的实用性。

Regardless of positioning, that's distinct utility.

关键问题是，你选对了要搞定的实用性吗？

The question with a T for trillions is did you pick the right utility to get right?"

我们这周晚些时候会再聊 Dots，但我觉得最大的收获是继续关注这个领域。

We'll come back to Dots later in the week, but I think the big takeaway is continue to watch this space.

虽然承认现在有点小毛病，但 Ali K. Miller 认为这里有一些相当大的差异，挺重要的。

While acknowledging that it's a bit buggy right now, Ali K. Miller argues that there are some big differences here that are fairly significant.

比如每个点都有自己的专用虚拟机，而且它始终在线，这样它就能更主动地工作，还有其他一些值得关注的变化。

Things like each dot getting its own dedicated virtual machine, it being always on, which allows it to be more proactive, and some other changes that are worth watching.

不管怎样，榜单上的第二名是决策API，这是OpenAI对Jev的回应。

In any case, number two on the list is the Decisions API, which is OpenAI's answer to Jev.

记住，Jev是一个判断模型。

Now remember, Jev is a judgment model.

它不擅长生成文本。

It's not good at outputting text.

它擅长分类，给出0到1之间的置信度分数，换句话说，就是做判断，而判断是决策的前奏，这大概就是这个功能名字的由来。

It's good at classifying things, giving confidence scores between zero and one, making judgments in other words, which are the precursors to decisions, which presumably is where this feature got its name.

决策API允许用户调用Luna的一个版本，这个版本模仿了Jev快速分类和决策的能力。

The Decisions API allows users to call a version of Luna that mimics Jev's quick classification and decision-making abilities.

用户定义一系列问题和可能的答案，模型就会快速输出响应。

Users define a list of questions and possible answers, and the model outputs fast responses.

OpenAI表示，这个API可以用于分类、路由请求，或者选择代理的下一步行动等场景。

OpenAI said that the API could be used for things like classification, routing requests, or choosing an agent's next action.

他们声称，与响应API相比，决策速度提高了10倍。

They claimed 10 times faster decision-making compared to the responses API.

不过，一个关键问题是，这比直接用Jev好在哪？

Still, one of the big questions was how is this better than just using Jev?

一个实际使用中的区别在于，Luna支持视觉输入，而Jev做不到这一点。

One difference in actual use cases comes from Luna's support for visual inputs, which aren't possible with Jev.

在过去几周的Jev热潮中，很多用户展示了Jev快速分类的能力，比如用于视觉营销，但这在底层需要图像到文本的转换。

Over the past couple of weeks of Jev mania, many users have shown off Jev making rapid classification for things like visual marketing, but that requires an image-to-text transformation under the hood.

决策API省去了这一步，从而扩展了它的潜在应用场景。

The Decisions API removes that step, and in that way expands its set of possible use cases.

目前，决策API还只是面向有限群体进行预览测试。

Now, right now, the Decisions API is just in preview for testing with a limited group.

至于它的重要性，我觉得这更多是承认了Jev所代表的不仅仅是一个新模型。

And in terms of significance, I think more than anything, this is a recognition that what Jev represents is more than just a new model.

它是一个极其有用，甚至可以说是即将成为生态系统中基础性的组件。

It is an extremely useful, and dare I might say, soon to be fundamental primitive to have in the ecosystem.

Jev之所以火，并不是因为它花哨或者性感。

Part of why Jev has hit wasn't that it was flashy or sexy.

只是当你看到它在很多方面比传统的大语言模型做得更好时，尤其是看到大语言模型被硬塞进一些并不适合它们的场景，就像把方钉子硬敲进圆孔里，你就会觉得，有一个这样的判断模型和生成模型搭配使用，回头想想其实是很自然的事。

It's just that as soon as you see all the things that it does better than a traditional LLM, in fact, where you see that LLMs were being rammed like square pegs into round holes into use cases that they weren't great at, it just seems obvious that having that sort of judgment model to sit alongside your generative model is pretty obvious in retrospect.

再说一次，现在做比较还为时过早，但我倾向于认为这个领域有空间容纳不止一个模型。

Once again, it's too early to really do comparisons, but I tend to think that this is a space where there is room for more than one model available.

实际上，我觉得几乎每个前沿实验室都会很快推出某种类似的东西。

In fact, I think that pretty much every frontier lab will have some version of this available very, very soon.

我们列表上的下一个是OpenAI Space，这是一个新的共享文档工作区，让人类团队和智能体可以一起协作。

Next up on our list is OpenAI Space, their new shared document workspace that gives human teams and agents a place to collaborate.

你可以把这个功能想象成一个AI增强版的Google Drive。

You can think of the feature sort of like an AI-enhanced Google Drive.

团队可以在上面处理共享的电子表格、幻灯片和其他文档，但Space也给团队提供了一个地方来放置工作自动化流程。

Teams can work on shared spreadsheets, slide decks, and other documents, but Space also gives teams a place to house workplace automations.

Dots可以原生地在Space中处理文档，但团队也可以设置定时任务，在空间中生成交付物。

Dots can natively work on documents in Space, but teams can also set up scheduled tasks to produce a deliverable in their space.

很多人关注的是OpenAI用Space去对标Microsoft 365、Google Drive或者Notion。

Many focused on OpenAI going after Microsoft 365, Google Drive, or Notion with Space.

虽然这确实大致在那个领域，但这些服务用起来越来越不适合智能体工作，OpenAI推出真正AI原生的生产力套件只是时间问题。

And while this is certainly generally in those space, each of those services have felt increasingly like a bad fit for agentic work, and it was only a matter of time before OpenAI launched a truly AI native productivity suite.

而几乎其他每个公告都有各种各样的意见，但Space是那种尤其是重度用户立刻就爱上的产品。

Now, whereas almost every other announcement had a pretty wide diversity of opinions, Space was one where especially the power users were in love right away.

Ray Fernando写道：“我提前体验了OpenAI Dots，我不是来跟风吹捧的。

Ray Fernando wrote, "I've had early access to OpenAI Dots, and I'm not here to join in on the glaze fest.

它现在能干掉Grokbot或Hermes吗？

Is this a Grokbot or Hermes killer right now?

不能。但Spaces和让智能体在应用里大展身手感觉是这些智能体发展的正确方向。”

No. But Spaces and allowing agents to thrive in apps feels like the right directions for these agents."

How I AI的Claire Vo写道：“在我看来，Dots有点被高估，而Space被低估了。

How I AI's Claire Vo wrote, "In my opinion, Dots slightly overhyped and Space underhyped.

我认识的每家公司都想要一个AI原生的协作工作空间。”

Every company I know wants an AI native collaboration workspace.

我们会持续关注，看这是否会吸引更多企业加入OpenAI的生态系统。

We'll be watching to see if this pulls more enterprises to the OpenAI ecosystem."

Every的Dan Shipper写道：“最明显的优势是，你写作时不用再在ChatGPT和Notion等其他应用之间来回切换窗口了。”

Dan Shipper from Every wrote, "The obvious win is that you stop switching windows between ChatGPT and another app like Notion while you write.

不那么明显的一点是速度。

The less obvious one is speed.

当dot通过浏览器在Google Docs中构建文档时，编辑几乎像爬行一样慢，因为Docs并不是为agent设计的。

When a dot builds a document through its browser in Google Docs, the edits practically crawl in because Docs wasn't built for agents.

原生文档就没有这种延迟。

Native documents don't have that lag.

你可以直接在文档内部标记你的dot，而不必回到聊天中告诉它去检查文件。”

You can tag your dot inside the document itself instead of going back to the chat and tell it to check the file."

Dan总结道：“我早就预料到这家公司会这么做。”

Dan concludes, "I've long expected the company to do this.

在2024年，我写过关于ChatGPT内文档、幻灯片和表格的潜力，很高兴这个产品终于面世了。

In 2024, I wrote about the potential for documents, slides, and sheets inside ChatGPT, and I'm glad the product is finally here.

我现在大部分工作都在ChatGPT的内置浏览器里完成，这个功能让整个过程更顺畅了。

I'm already doing most of my work in ChatGPT's in-app browser, and this makes that process smoother.

我可以在文档评论里标记我的dot，然后边工作边收到修改建议。

I can tag my dot in a comment on a document and get a revision in line as I'm working.

这种来回互动感觉更像是在合作。

The back and forth feels more like collaboration."

当天发布的重磅模型是GPT-6-1 Sol，OpenAI称其智能水平接近Astra，但价格只有五分之一。

The big model launch for the day was GPT-6-1 Sol, which OpenAI pitched as near Astra intelligence for a fifth of the price.

这次发布距离GPT-6 Sol仅一周，带来了一些相当显著的升级。

The release comes just a week after GPT-6 Sol and provides some fairly notable upgrades.

所有努力级别下的编程基准测试成绩都有大幅提升，也就是说，6-1 Sol在中等设置下的表现超过了6 Sol在最高设置下的表现。

Coding benchmarks are up significantly across all effort levels, meaning that 6-1 Sol on medium settings outperforms 6 Sol on max settings.

该模型在DeepSui上的最高得分是75.2%，这是在高级设置下取得的，而超高级和最高级设置下性能反而有所下降。

The model's top score on DeepSui was 75.2% on high settings, with extra high and max actually seeing a degradation in performance.

我们第一次看到这种现象是在Opus 5上，看起来最高努力级别开始让模型过度思考，反复质疑正确答案，反而对表现不利。

We first saw this phenomenon with Opus 5, and it looks like top effort settings are starting to force models to overthink and second-guess correct responses to their detriment.

现在，那个高设置下的DeepSui得分其实比Astra的最佳表现还稍微高一点，这支持了OpenAI关于接近Astra智能的说法。

Now, that high setting score on DeepSui was actually a touch higher than Astra's best performance, supporting OpenAI's claim of near Astra intelligence.

在大多数主要基准测试中，情况也差不多。

The pattern was similar across most major benchmarks.

相比6 Sol，这是一个很大的提升，让6-1 Sol以更便宜的价格达到了和Astra差不多的水平。

A big improvement over 6 Sol that landed 6-1 Sol in the same ballpark as Astra at a much cheaper price.

其中一个值得注意的基准测试是OSWorld，它测试的是长时间跨度的计算机使用能力。

One of the notable benchmarks was OSWorld, which tests long horizon computer use.

6-1 Sol在最高设置下的最佳得分是71.4%，超过了6 Sol的64.4%，也接近了Astra最佳的73.5%。

6-1 Sol's best performance scored 71.4% on max settings, beating 6 Sol at 64.4% and coming close to Astra's best at 73.5%.

6-1 Sol的价格也便宜很多，只有6 Sol的三分之一，是Astra的13%。

6-1 Sol was also significantly cheaper than the others, at a third the cost of 6 Sol and 13% the cost of Astra.

而且，提高努力水平几乎不增加成本，这说明OpenAI在这款模型的计算机使用效率上取得了重大突破。

What's more, increases in the effort level barely changed the cost, suggesting that OpenAI has made some big breakthroughs in the efficiency of computer use with this model.

Artificial analysis给这款模型打了52分，比Astra低一分，也排在Opus、Sonnet和Fable之后，位列第五。

Artificial analysis scored the model at 52, one point shy of Astra and also behind Opus, Sonnet, and Fable coming in in fifth place.

他们还发现了惊人的成本效率，把帕累托前沿又往前推了一步。6-1 Sol 跑基准测试的成本只有 Astra 的四分之一，比 6 Sol 还便宜 31%。

They also found incredible cost efficiency that pushes the Pareto frontier, with 6-1 Sol's benchmark run completed at a quarter the cost of Astra and 31% cheaper than 6 Sol.

如果你不需要用最高配置跑，6-1 Sol 似乎也能把成本压到很低，比如接近 GLM 53 Flash 的水平，那是 ZAI 开源模型的廉价版。

If you don't need to run it at max settings, 6-1 Sol seems capable of hitting smaller costs like GLM 53 Flash, which is the cheaper version of ZAI's open source model.

OpenAI 说这是目前性价比最高的模型。

OpenAI called it the most cost-efficient model for its performance available today.

至于这个发布是意料之中还是意料之外，一方面，这些实验室出新模型从来都不太让人意外，尤其是在 DevDay 这种大日子。

Now, as to whether this one was expected or unexpected, on the one hand, it's never all that surprising to get a new model from one of these labs, especially on a big day like DevDay.

让人意外的是，我们上周才刚拿到 6 Sol。

Surprising in that we only got 6 Sol last week.

所以说到第一印象，我们得先自己上手试试再说。因为你能看到像 Dropout Layer 在 X 上发的这种帖子，说“GPT-6-1 Sol 让 Opus 5-5 看起来都贵了”，也会看到像 Bridge Branch 这种帖子，说“我们拿它跑了 Sunset Ocean 测试。”

So when it comes to first impressions, let's just say we're going to have to wait to get our hands on it ourselves because for every post you can find like this one from Dropout Layer on X, "GPT-6-1 Sol just made Opus 5-5 look expensive," you also get one like this one from Bridge Branch, "We ran it through our Sunset Ocean test.

成本跟 GPT-6 Sol 一样，速度慢了一倍，而且海洋几乎没渲染出来。”

Same cost as GPT-6 Sol, twice as slow, and it barely rendered an ocean."

大家好。

Hello everyone.

AI带来的一个大变化是，我们的思路已经从如何给网页排名，转变为如何成为AI信任并愿意引用的信息来源。

One big change around AI is we've shifted our thinking from how we rank our pages to how do we become the source that AI trusts enough to answer with?

在毕马威，他们正亲身经历这一变化。

At KPMG, they're seeing this firsthand.

AI生成的结果现在直接给出答案，往往连一次点击都不需要。

AI-generated results now surface answers directly, often without a single click.

因此，他们越来越专注于生成式引擎优化，也就是GEO，通过结构化内容，让AI系统能够检索、理解并引用这些内容，将其视为可信的权威来源。

That's why they are increasingly focused on generative engine optimization, or GEO, structuring content so AI systems can retrieve it, understand it, and cite it as trusted authority.

这不仅仅是SEO的演进，更是对可见性的要求。

This is not just an SEO evolution, but a visibility mandate.

事实上，毕马威的GEO要求很简单。

And indeed, the GEO mandate from KPMG is simple.

如果AI在影响决策，那么你的专业见解就需要出现在答案里。

If AI is shaping decisions, your expertise needs to show up inside the answer.

详情请访问kpmg.com/us/geo。

Read all about it at kpmg.com/us/geo.

再说一遍，网址是 kpmg.com/us/geo。

Again, that is kpmg.com/us/geo.

下面说说为什么大多数传统系统现代化项目会失败。

Here's why most legacy modernization projects fail.

因为做这些工作的AI无法大规模理解代码库。

The AI doing the work can't understand code bases at scale.

它只能看到一小部分上下文，检查语法，却错过了遍布全球应用生态系统中多年积累的决策。

It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem.

Blitzy解决这个问题的方式和它解决其他问题的方式一样。

Blitzy solves this the way it solves everything.

在任何迁移开始之前，Blitzy的代理会基于你的代码，将整个遗留系统逆向工程成一个持久的知识图谱。

Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph.

每一个依赖、每一个约束、每一条曾经只存在于某个工程师脑子里的隐性知识。

Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head.

基于这种理解，Blitzy自主执行语言迁移、框架升级以及从单体架构到微服务的转型，并且全程进行端到端验证。

From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith to microservices transformations, all validated end to end.

Blitzy的一个客户用编码代理在16周内完成了价值1000万美元的保险单体架构的现代化改造，而基线是137周。

One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents.

这相当于压缩了9倍的时间。

That's 9x compression.

在加速你的路线图的同时，消除技术债务。

Retire technical debt while accelerating your roadmap.

详情请访问blitzy.com。

See how at blitzy.com.

网址是B-L-I-T-Z-Y.com。

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

最优秀的团队不是靠一个明星队员扛起所有人。

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

他们清楚自己的长处，也了解彼此的短板，并且会善用这两点。

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

这正是Robots & Pencils刻意打造的团队文化。

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

那里没有人为了凑人头数而埋头做那些琐碎的工作。

Nobody there is grinding through busy work to pad a headcount number.

人们是为了解决难题而来的，他们留下来是因为周围的每个人都在同时进步。

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

在一个满是只想快速招人的公司市场里，这值得一看。

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

去 robotsandpencils.com/careers 看看吧。

Check out robotsandpencils.com/careers.

本期《AI每日简报》由Hyperagent赞助，你可以在那里管理你的代理团队，让团队一起协作。

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

别再依赖本地代理和聊天工作流，等着你的笔记本电脑被提示了。

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

Hyperagent在云端部署始终在线的代理，在你团队已经使用的工具中完成实际工作。

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

营销代理能把竞争对手的动作变成落地页。

Marketing agents turn competitor moves into landing pages.

销售代理会丰富潜在客户信息、起草邮件，并更新客户关系管理系统。

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

运营代理负责跟进文书工作，并跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享上下文，并遵循你关于范围和审批的规则。

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

是时候让代理像队友一样加入你的团队了。

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

去Hyperagent雇佣你的代理吧。

Hire yours at Hyperagent.

在hyperagent.com/aidailybrief获取100美元积分。

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

现在，希望这些Sol级模型足够好，因为我们可能再也见不到GPT-6-1 Astra了。

Now, hopefully, these Sol class models are good enough, though, because we may never see GPT-6-1 Astra.

《华尔街日报》报道称，OpenAI因安全问题取消了发布其旗舰模型下一版本的计划。

The Wall Street Journal reports that OpenAI have scrapped plans to release the next version of their flagship model over safety concerns.

OpenAI的安全系统负责人萨奇特·贾恩在一份声明中说：“在安全和一致性方面，总是有权衡的。”

Sachit Jain, OpenAI's head of safety systems, said in a statement, "For anything regarding safety and alignment, there's a trade-off.

“你真的需要找到一条合适的线，既要保持在范围内，又要避免模型在执行任务时偷懒，即使遇到困难也要继续推进。”

You really do need to find what's the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction."

听起来OpenAI试图降低模型的执着程度，因为这种执着在过去几个月里导致了多起事故，但他们没能找到一个折中的办法。

It sounds like OpenAI tried to turn down the tenaciousness that had led to multiple incidents over recent months but couldn't find a happy medium.

贾恩还补充说，虽然这个模型比之前的版本更不偷懒，但“在保持在范围内、遵守授权以及向用户反馈所完成工作类型方面，它没有完全达到标准。”

Jain added that although the model was less lazy than its predecessors, it, quote, "Didn't quite meet the bar in terms of staying within scope and authorization and how it communicates back to the user about the type of work it's done."

当然，我说我们可能永远看不到6-1 Astra，这有点夸张，但实际上，他们似乎确实需要在技术上取得一些突破，才能以他们认为足够安全的方式进入前沿模型的下一阶段。

Now, obviously, I'm being a bit hyperbolic when I say that we'll never see 6-1 Astra, but it actually does seem like they're going to need to hit some technical breakthroughs before we get to some of the next levels of the frontier in a way that they deem safe enough to release.

接下来，我们来看看活动中那些被认为比较小的公告，很多人仍然认为这些公告意义重大。

Now from here, we get into what are considered the smaller announcements from the event, many of which people are still noting as quietly significant.

OpenAI展示了其平台计划的最新版本，向外部开发者开放了ChatGPT。

OpenAI showed us the latest version of its platform plans, opening ChatGPT to outside developers.

OpenAI团队的蒂博写道：“你现在可以构建带有插件扩展的原生应用，并直接在ChatGPT中发布。”

Thibault from the OpenAI team wrote, "You can now build full native apps with plugin extensions and ship them right in ChatGPT.

我们有超过12亿的周活跃用户，会在对话中直接展示相关的插件。

We have over 1.2 billion weekly users and will surface relevant plugins right in the conversations.

这个功能在技术上叫做插件扩展。

This feature is technically called plugin extensions."

OpenAI为了把这个功能做好，已经尝试了一年多，测试了原生集成、插件和MCP，想借此扩大生态系统。

Now, OpenAI has been trying to get the balance right on this feature for more than a year, testing native integrations, plugins, and MCP as a way to expand the ecosystem.

那么，这会不会终于成为让ChatGPT像应用商店一样的功能呢？这看起来一直是OpenAI的大目标之一。

So could this finally be the feature that lets ChatGPT function as an app store, which always seemed like one of the big goals for OpenAI?

当然，大家肯定会急着去尝试。

Certainly, there is going to be a rush to experiment here.

我已经看到像Jana Welinder做的这种个人造型师应用，但也有人在想，对应用开发者来说，这些取舍到底值不值。

I'm already seeing things like this personal stylist app from Jana Welinder, but others are wondering if the trade-offs for the app developer are really worth it.

麻省理工学院的Christian Catalini写道：“你带来应用，OpenAI带来智能。”

MIT's Christian Catalini writes, "You bring the app, OpenAI brings the intelligence.

那OpenAI是不是也能拿到用户的数据痕迹呢？

Does OpenAI also get the user traces?

如果你贡献的是教你的合作伙伴怎么做你的工作，那这合作挺有意思的。

An interesting partnership if your contribution is teaching your partner how to do your job."

对此，Signal回应说：“长尾应用可能会选择加入，但对其他人来说，这真是个糟糕的主意。”

To which Signal responded, "Long tail of apps might opt in, but otherwise, this is a terrible idea for anyone else."

另一面就是用ChatGPT登录，这名字一听就明白。

A flip side of this is sign in with ChatGPT, which is exactly what it sounds like.

它让你可以用ChatGPT登录其他账户。

It allows you to use ChatGPT to sign into other accounts.

OpenAI应用研究负责人鲍里斯·鲍尔写道：“用ChatGPT登录，让你可以使用任何基于我们API构建的应用。”

OpenAI's head of applied research, Boris Power, wrote, "Sign in with ChatGPT lets you use any app built with our API.

这让应用开发者轻松多了，不用费劲去说服用户单独付费订阅。

Makes it much better for app developers not needing to jump a huge hill justifying paying a separate subscription.

ChatGPT就是你的全能超级智能果汁。”

ChatGPT is your one-stop super intelligence juice."

这个想法基本上就是让开发者允许用户带着他们的智能订阅到处用，降低应用的使用门槛，同时利用很多人已经选择用ChatGPT作为他们的超级智能解决方案这一事实。

The idea is basically that developers allow people to carry their intelligence subscription with them, lowering the barrier to entry for their apps, and allowing them to take advantage of the fact that so many people have already opted in to using ChatGPT for their super intelligence solution.

Jackie Luo写道：“用ChatGPT登录是个大事，我不明白他们为什么把它定位成身份验证，其实这是OpenAI在把优惠价格扩展到第三方，并把服务捆绑进订阅里。”

Jackie Luo writes, "Sign in with ChatGPT is a huge deal, and I don't understand why they framed it as auth when it's really OpenAI expanding favored pricing to third parties and bundling services into the subscription.

这终于开始让他们的利益和客户一致，这样个人就不用重复为token付费，应用也不用在API成本之上再定价了。

It finally starts to align their incentives with customers so that individuals stop double paying for tokens and apps stop having to price everything on top of API costs.

我猜这会扩展到更多应用，甚至所有应用，Anthropic也得跟进，才能让他们的订阅更有价值。

I'd imagine that expands to many more, if not all apps, and Anthropic will need to follow to add value to their subscription.

在那个世界里，应用可以绕过token成本，只对应用层收费，这对大家可能都是好事。”

In that world, apps could bypass token costs and charge for the app layer alone again, which is probably good for everyone."

换句话说，这个想法是，应用不用再为了提供智能访问而收取看起来很贵的费用，他们可以只收20美元或者他们想定的应用体验价格，然后让API成本直接流向底层。

In other words, the idea here is that instead of apps having to charge what looks like huge prices to get access to intelligence, they can just charge their $20 or whatever they wanted to price their app experience at and let the API costs flow directly to the underlying.

虽然，是的，这意味着他们可能没法在这些成本上加价赚点溢价，但好处是，用户已经付了ChatGPT订阅费，不用再被说服去付额外费用，这很可能比他们自己赚的钱更划算。

While, yes, that means they might not get to scalp those costs and add a little premium to them, the benefits of not having to convince people to pay those additional fees when they're already paying for a ChatGPT subscription likely outweighs the money that they could make otherwise.

对于我的重度用户们，OpenAI为订阅推出了几个新选项。

For my power users out there, OpenAI has introduced a few new options for their subscriptions.

第一个是超快模式，在Codex中token生成速度快8倍，在应用里快6倍。

The first is Ultra Fast Mode, which offers 8x faster token production in Codex and 6x speed in the app.

目前，极速模式只在ChatGPT Work的Astra和Codex中可用，但OpenAI表示很快就会支持6-1 Sol。

Right now, Ultra Fast Mode is only available for Astra and Codex in ChatGPT Work, but OpenAI say they will add support for 6-1 Sol soon.

另外，OpenAI还推出了一个500美元的新订阅档位，使用量是Plus档的25倍。

In addition, OpenAI is introducing a new $500 subscription tier that offers 25 times the usage of the plus tier.

而且目前只有这个档位能使用极速模式。

It's also the only tier that gets access to Ultra Fast Mode for the time being.

在活动开始前，Thibault宣布OpenAI将重新开放每月200美元的Pro档订阅，这个档位几周前被关闭了，不过现在做了一些调整。

Heading into the event, Thibault announced that OpenAI would reopen access to their $200 a month pro tier, which they had turned off a few weeks ago, but with a few tweaks.

使用量的计算方式改了，算下来API成本降低了50%。

Usage calculations have been changed, netting out to a 50% reduction in terms of API cost.

Thibault解释说，这是在糟糕的选项里挑了个最好的。

Thibault explained that this was the best of a bad set of choices.

OpenAI优先考虑不引入五小时使用限制，并认为API成本降低和模型效率提升意味着用户仍然可以完成大致相同的工作量。

OpenAI prioritized not having to introduce the five-hour usage limit and argued that API cost reductions and more efficient models would mean users can still get roughly the same amount of work done.

一方面，大家多少预料到了这种情况，但另一方面，你能想象同样的模型回来却缩水了，大家会是什么反应。

Now, on the one hand, people were sort of expecting something like this, but at the same time, you can imagine how well it went over to have the same model coming back with reduced value.

对我们来说，想理解这反映了AI的现状，说实话，算力限制是真实存在的，而且会一直存在。

For our purposes here of trying to understand what it says about the state of AI, look, man, compute constraints are real, present, and permanent.

前沿模型已经快撞上现有算力的天花板了，而新算力的上线速度远远赶不上人们使用这种数字智能的增长速度。

Frontier models are running up against the walls of what they can do with the compute that we have, and new compute isn't coming online at anywhere near the speed that people are increasing their use of this digital intelligence.

不过，好处是我们会看到效率方面的大力推动，长远来看，这会让我们所有人都享受到更划算、更好的体验，但过程中会有一些波折。

The upside of that, though, is that we're going to see a big push around efficiency, which in the long run should net out to more cost-effective, better experiences for all of us, but it's going to have some bumps along the way.

在开发者和vibe coder圈子里，Codex现在有了专门的云端环境，用户合上笔记本电脑后还能继续工作。

Over in developer and vibe coder land, Codex now has a dedicated cloud environment, allowing users to keep working after they close their laptop.

这个功能迟早会来，毕竟在旧金山，人们拇指按在笔记本上到处走已经常见到成了梗。

This one was always coming after people walking around with their thumbs jammed into laptops became so common across San Francisco that it became a meme.

但放眼更大的趋势，从GrokBot开始，我们看到整个行业都在把云端实例作为所有智能体产品的必要组成部分。

But looking for broader patterns, starting with GrokBot, we've seen the entire industry move towards a cloud instance being a necessary part of all agentic products.

这是AI从像软件一样被使用，转向更持久、始终在线的应用这一大转变的延续，无论你通过什么方式访问。

This is a continuation in the big shift from AI being something you engage with, like software, to becoming a more persistent, always-on application, regardless of how you access it.

Codex CLI也迎来了一次大更新，外观和功能都有新变化。

The Codex CLI is also getting a major refresh with a new look and new capabilities.

随着你往下滚动，历史记录会一直往前延伸，支持单线程的极简主义者——我绝对算一个——而编辑器保持固定，这样你随时可以输入下一行。

History goes back further as you scroll, supporting the mono-thread maxis, of which I am absolutely one, while the composer stays pinned so you can always type your next line.

OpenAI 写道：“Codex CLI 现在让你能更好地管理并行工作。

Wrote OpenAI, "The Codex CLI now gives you a better way to manage parallel work.

使用 /agents 查看正在运行的任务、检查进度，并在不同任务之间切换。”

Use /agents to see what's running, check progress, and jump between tasks."

在企业方面，OpenAI 推出了私有智能，保证零数据保留，即使在推理阶段也是如此。

On the enterprise side, OpenAI has launched private intelligence, which guarantees zero data retention, even at inference time.

过去几个月，隐私问题一直是企业的大问题，现在也成了为企业提供服务的公司的大问题，因为人们越来越担心 OpenAI 和 Anthropic 在从客户那里窃取数据。

Over the past few months, privacy, which was always a huge issue for enterprises, has become a huge issue for the companies supplying the enterprises, with growing concerns that OpenAI and Anthropic are skimming data from their customers.

这基本上就是微软最近所有沟通的潜台词或明面主题，告诉你为什么不应该信任他们的竞争对手。

This is basically the subtext or the main text of every communication from Microsoft these days about why you should be not trusting their competitors.

所有这些都意味着，一个更强大、更清晰的端到端数据隐私保证，对那些需要它的人来说意义重大。

All that means that a stronger, clearer end-to-end data privacy guarantee is a big deal for those who need it.

OpenAI 还推出了一个模型市场。

OpenAI has also launched a model marketplace.

它让用户可以通过OpenAI的Responses API和Codex，从Base10购买开放权重模型的推理服务。

It allows users to buy open-weight model inference from Base10 through OpenAI's Responses API and Codex.

这个事儿，换一天我们真能深入聊很久，因为它对OpenAI的护城河影响很大。

Now, this is one that on another day, we could go way deep into because it has some big implications for OpenAI's moat.

这既保护了他们免受开源模型的冲击，又给客户提供了更多选择。

It protects them from open source disruption and gives customers a lot more choice.

这也意味着企业可以放心跟OpenAI签大额合同，因为他们知道这些投入可以轻松用在多模型策略上，包括开放权重模型。

It also means enterprises can feel comfortable making large spending commitments with OpenAI, knowing that they can easily use that spend on a multi-model strategy that includes open-weight models.

相信我，我们肯定会再回来聊这个的，因为我觉得它比大家第一眼看到的要重要得多。

Trust me when I say that we are going to come back to that one because I think it is bigger than people are giving it credit for at first blush.

我们其实才刚开了个头，但咱们还是快速总结一下这些意味着什么，以及OpenAI DevDay透露了AI下一阶段的什么信息。

So we have barely scratched the surface here, but let's try to quickly sum up what all of this amounts to and what OpenAI DevDay revealed about the next phase of AI.

我们并没有看到什么特别惊艳的新东西。

We didn't get something blisteringly new.

我们看到的，其实是过去几个月我们在节目里一直在记录的很多趋势得到了印证。

What we got was confirmation of a lot of the trends that we've been cataloging on this show over the past several months.

首先，通过 Sol 6.1 和 Decisions API，我们看到一个趋势的延续和深化，那就是模型已经足够好、足够便宜，可以用于做所有事情，也就是说，你不需要把它们关掉。

First, with Sol 6.1 and the Decisions API, we got a continuation and a deepening of the trend of models that are good enough and cheap enough that they can be used to do everything, i.e., you don't have to turn them off.

你不需要决定它们用来做什么。

You don't have to make decisions about what they're used for.

它们可以包揽一切。

They can just do it all.

有了 Dots，我们有了持久且主动的代理，它们利用这一点真正去做所有事情。

With Dots, we have persistent and proactive agents that take advantage of that to actually do everything.

而当谈到做所有事情时，插件扩展表明，这意味着既要整合一切，也要走向各处，也就是 ChatGPT 在其他应用中的登录功能。

And when it comes to doing everything, plugin extensions shows that that means bringing everything in as well as going everywhere, which is ChatGPT sign-on for other apps.

最后，通过 Space，我们得到更多确认，下一代 AI 将不仅仅是单人模式，还会以原生方式支持团队和多人模式，即使目前这还处于早期阶段。

Finally, with Space, we have more confirmation that the next generation of AI will not just be single player mode, but will also be team and multiplayer mode in a native way, even if that remains nascent so far.

这里还有很多很多可以探讨的内容，但今天的 AI 每日简报就到这里。

Lots and lots more to explore here, but that is going to do it for today's AI Daily Brief.

一如既往，感谢你的收听或观看。

Appreciate you listening or watching, as always.

那么下次再见，祝一切安好。

And until next time, peace.