GPT-6 Astra 太棒了——这是我构建的一切

今天我们直奔星辰大海，因为没错，今天GPT-6 Astra正式上线了。

We are going to the stars today because, yes, today, GPT-six Astra is live.

嗯，目前是对Daybreak客户开放。

Well, it's live for Daybreak customers.

其他人很快也都能用上了。

It will be live for the rest of us very soon.

我有幸提前拿到了测试权限，这六个月来一直在用这个模型。

I was lucky enough to get some early access to test the model, and this has been the model in the last six months.

你可能觉得我老调重弹，但这次真的让我更有干劲了。

And, you know, maybe I sound like a broken record, but this is the one that has made me actually feel more ambitious.

它解决了我一些一直搞不定的老任务，连Fable都搞不定的那种。

It has crushed some of my standing tasks that Fable couldn't figure out.

就连我心爱的5.6 Sol也搞不定的。

Even my beloved 5.6 Sol couldn't figure out.

这个GPT-6模型，一次就搞定了好几个。

This model, GPT-six, one shot a bunch of them.

我要告诉你这个模型是什么，OpenAI说它擅长什么，我还要给你看我用它做的一些非常厉害的东西，包括，没错，又黑进这台破电脑了。

I'm gonna tell you what this model is, what Open says it's great at, and I'm gonna show you some pretty incredible things I've built with it, including, yes, hacking this dumb computer again.

我们开始吧。

Let's get to it.

好的。

Okay.

我不想在博客文章上花太多时间。

I don't wanna spend too much time on the blog post.

我们可以在节目笔记里放个链接。

We can link to it in the show notes.

你们都可以去读一读。

You all can go read it.

简单来说，基本上在所有方面都是最先进的。

TLDR, state of the art on basically everything.

我们有科学、数学、编程、知识工作、不烦人、速度快、电脑使用，所有这些东西。

We got the science, math, coding, knowledge work, not being annoying, being fast, computer use, all the things.

所以OpenAI把它定位为最智能、最对齐的模型。

So OpenAI is positioning it as its most intelligent and aligned model.

特别是前沿数学方面，有巨大的飞跃，数学能力提升很大。

Huge jumps, particularly on frontier math, a big jump there in mathematical ability.

他们重点强调的是计算机使用和软件应用。

The big thing that they're calling out is computer use and the use of software.

所以你们会在今天的节目里看到，它用电脑真的非常厉害。

So you're gonna see this in today's episode, but it is so good at computer use.

不只是浏览器，还有Excel、Unity、Power BI、Blender、Jupyter、文档编辑器，几乎什么都能用。

Not just browsers, but Excel, Unity, Power BI, Blender, Jupyter, document editors, like kind of everything.

所以我们回来了。

So we're back.

SaaS回来了。

SaaS is back.

产品也回来了。

Product's back.

按钮又回来了，因为现在Astra能用了。

Buttons are back because now Astra can use it.

它在生成工作成果方面真的很擅长。

It's really good at producing work artifacts.

他们做了很多工作，让工作成果符合你的风格、品牌和语气等等。

They've done a lot of work to make work artifacts match your style, match your brand, match your voice, etcetera.

然后再说，自主性、操控性这些方面都做得很好。

And then again, autonomy, steering, all that kind of stuff is is very good.

他们真的花了很多精力在常见任务上节省时间，这又是因为电脑使用。

They've really focused a lot on kind of time saving on common tasks again because of computer use.

所以这有点像是个小预览，我觉得，我们都在花很多钱在令牌上，这些模型也感到压力，要拿出投资回报率的模式。

And so this is a little, like, sneak peek into, I think, like, we're all spending a lot of money on tokens, and these models are feeling pressured to come up with an ROI model.

而且我觉得你会听到很多关于Astra在电脑使用和节省时间方面的讨论，对吧？

And and I think you're gonna hear a lot about them talking about Astra in terms of computer use and time saving, right?

比如找房子可以从六个小时缩短到十分钟。

Like apartment hunting can go down to ten minutes from six hours.

我给你举个例子，看看怎么生成缩略图。

I'll show you an example of generating thumbnails.

这又是那种能从一个小时缩短到几分钟的活儿，而且完全不用动手。

Again, something that takes us like an hour down to a couple of minutes, completely hands free.

所以，速度快多了，效率也高多了，而且真能完成他们说的实际工作。

And so, a a lot faster, a lot more efficient, and kind of able to do what they say real work.

他们特别提到演示文稿、电子表格、文档和软件是Astra的绝佳应用场景，但咱们先看看一些数据。

They really call out presentations, spreadsheets, documents, and software as great use cases for Astra, but let's look a little bit at kinda some of the numbers.

在Automation Bench上，分数又是Sol报告的两倍。

So Automation Bench, again, double Sol's reported scores.

所以，你知道，在自动化任务方面要高得多。

So, you know, a lot higher on automating tasks.

再看看编码方面。

Let's look at coding.

在Deep Suite上，这里比Sol的5.6稍微高一点。

Deep Suite, a little bump up here against 5.6 Sol.

但再说一次，结合电脑使用，你能做更多真正有趣的事情。

But again, combined with computer use, you're gonna be able to do a lot more really interesting things.

数学在这里是个大话题。

Math is a big story here.

在前沿数学四级上有了很大的飞跃。

So big jump in frontier math tier four.

然后我们来谈谈价格和你能不能用上它。

And then let's talk about price and can you get it?

所以它首先推出。

So it's rolling out first.

它将提供给有Daybreak访问权限的企业客户。

It's going to go to enterprise customers with Daybreak access.

我提前获得了一点访问权限，这样我就能测试这个模型，并告诉你们所有相关情况。

I got a little bit early access so I could test the model and tell you all about it.

但在接下来的几天里，它将逐步向GPT Plus、Pro和企业用户推出，在API和AWS上也可以使用，定价是每百万输入token 10美元，每百万输出token 50美元。

But over the coming days, it's going to be rolling out to GPT plus Pro and enterprise in the API and in AWS and then pricing $10 per million input tokens and 50 per output tokens.

而且快速模式会很快。

And fast mode will be fast.

所以这是个很棒的模型。

So it's a great model.

你会想用它的。

You're going to want to use it.

我就想先把这点说清楚。

And I just want to get that out of the way.

你知道它是什么。

You know what it is.

然后我们来看看它有多厉害，先从电脑使用说起。

And then let's get into how amazing it is starting with computer use.

既然我已经简单介绍了Astra擅长什么，现在就来看看它的实际表现。

So now that I've told you a little bit about what Astra is great at, let's see it in action.

我在介绍这个模型的高层能力时，提到的第一点就是它在浏览器使用方面有多出色。

Now the number one thing I covered when talking about the high level of the model is how good it is at browser use.

我跟你说，我可是浏览器操作的高手。

And I am telling you, am a browser use maxer here.

我也是电脑操作的高手。

I'm a computer use maxer here.

我超爱浏览器操作。

I love browser use.

我超爱电脑操作。

I love computer use.

但是……

But...

而且我平时在Codex里也经常用这个功能。

And I've been using it a lot in Codex generally.

不过我得说，Astra把我的电脑操作水平提升到了一个新的高度。

But I will say Astra has taken my computer use to the next level.

我想给你看几个我用电脑操作完成的事情，我觉得这些真的非常非常有用。

I wanna show you a couple examples of some things that I've done with computer use that I think are really, really, really useful.

让我特别惊讶的是，它在处理复杂的网页界面、管理相当复杂的节点和交互方面表现得非常好。

The one I was really surprised by is how good it is at complex web UI and managing pretty complex nodes and interactions and doing it really well.

那我给你举个例子说明一下我的意思。

So let me show you an example of what I mean here.

这是我们的cxo。

So this is our cxo.

开发CRM，我们开始收到一些线索了。

Dev CRM, and we are starting to get leads.

我们上周刚推出这个业务，现在收到不少主动咨询。

We launched this business last week, and we're getting a bunch of inbound.

我们想筛选这些线索，然后决定是应该发给我还是发给我的合伙人扎克。

And we want to filter those leads and then determine whether or not they should be sent to me or sent to my business partner, Zach.

一旦确定是应该给我还是扎克，而且……

And once they are determined if it should be me or Zach and it's...

比如，我负责那些大型企业客户，他负责那些技术型初创公司，我们分工合作，一对一地回复线索。

Like, I get the big fancy enterprises, he gets the, like, technical startups, and we kinda divide and conquer responding to leads one on one.

我们还没把这个交给智能体，但我们现在要现场演示一下。

We haven't given that to the agents yet, but we're gonna do it live here.

所以我们用这个内置在CRM里的音频节点式工作流构建器来搭建分配逻辑，比如通过提示词判断是给克莱尔还是扎克，然后路由到不同的Slack消息和不同的CRM属性。

So we use this kinda built into our CRM audio, node based workflow builder to build out the assignment logic, do things like prompt to determine Claire or Zach, and then route it to different Slack messages and different CRM attributes.

我发现自己坐在这儿，一个个地摆弄这些按钮。

And I found myself sitting here and, like, futzing with these buttons one off.

我搞不清楚怎么从AI的响应中推断出正确的对象。

I couldn't figure out, like, how to infer the right objects from the responses, the AI responses.

我当时真的很沮丧。

I was just getting really frustrated.

我在这儿花了一个小时。

I spent an hour here.

然后我就想，电脑操作。

And then I was like, computer use.

当然，电脑操作能搞定这个。

Of course, computer use can do this.

所以我想给你看看实际操作效果。

And so I wanna show you what that looks like.

好的。

Okay.

所以，与其手动拖拽这些东西，我会打开我和Codex聊CRM的对话窗口。

So instead of dragging and dropping all this stuff, I'm gonna bring up my chat where I talk about the CRM with, Codex.

然后我会说，嘿。

And I would say, hey.

我有一个工作流，用来把线索分配给Zach或者我。

I have a workflow open to route leads to Zach or I.

我希望在这个流程中，用一个自定义的AI代理，根据线索提交生成一封定制邮件，邮件由Zach或我署名，取决于线索分配给谁，并邀请他们通过我们的日历链接预约时间，再加上一些我们为什么期待和他们交流的内容。

I want, as part of this flow, to generate with a custom agent, AI, an email that is custom to the lead submission authored from either Zach or I, depending on who it's routed to, and invite them to schedule with our calendar link, plus a bit about why we're excited to chat with them.

你能在Chrome里更新这个工作流来实现这个功能吗？

Can you update the workflow in Chrome to do this?

谢谢。

Thanks.

好的。

Okay.

所以我们应该看到Codex开始打开Chrome，接管Chrome，然后开始构建新的节点、新的逻辑、新的提示，并把所有这些重新连接起来，把正确的节点接好，然后它就开始运行了。

So what we should see Codex start to do is actually open up Chrome, take over Chrome, and start building out new nodes, new logic, new prompts, and route all of this stuff, reconnecting the right nodes, and here it goes.

所以你可以看到它正在为我们正在做的AI生成添加一个新的对象，包括邮件发送给谁、邮件主题是什么、邮件正文是什么。

So you can see it's adding a new object to the AI generation that we're doing, who the email is going to, what the email subject is, what the email body is.

我猜，你看，这里没有手动操作。

I'm guessing, look, there's no hands.

这些都不是我做的。

I'm not doing any of this.

它正在这个相当复杂的基于节点的应用里自动进行，找出在哪里做修改，添加内容。

And it's going ahead and going in this, like, fairly complex node based app and figuring out where to make the changes, include things.

而且我认为，理想情况下，它会显示一封来自Claire或Zach的邮件草稿，供我们审核，最终由我们发送。

And I think, ideally, it's going to show an email draft from Claire or from Zach to review and for us to ultimately send.

然后那会发送到我们的Slack频道，之后我们可以复制粘贴，然后通过邮件发送给潜在客户。

And so that's gonna send to our Slack channel, and then we can copy and paste that and email it to the prospect.

你知道，我可能还会做的一件事是，等我们觉得这个真的不错之后，让它实际把邮件发出去。

You know, the other thing what I'll probably do here is have it actually send the email after we feel like this is really good.

但话说回来，这种事情，就像打字和思考一样，特别费时间。

But, again, this kind of thing, just like typing and thinking takes forever.

在Astra之前……

And before Astra...

我的意思是，在GPT-5、GPT-6之前，但肯定是在Astra之前，我真的觉得浏览器操作和电脑操作没法很好地处理这些节点。

I mean, really before GPT-five, six, but certainly before Astra, I really just didn't feel like browser use and computer use can navigate these nodes as well.

现在我发现它真的可以。

And now I found that it really can.

我现在几乎一直不用手碰电脑了，这真的是一种全新的、很酷的与电脑和网站交互的方式。

And I just am hands off my computer all the time now, and it's a really fascinating new way to think about interacting with your computer, interacting with websites.

我在想这个问题。

And I think about this.

我们大家都知道，无界面就是下一个界面，我们也聊了很多关于MCP和命令行工具的话题。

We, you know, we all said no UI is the next UI, and we had all these conversations about MCPs and CLIs.

最让人兴奋的是，我们现在正处在这个时刻。

And what's so fascinating is now we're at this moment.

我们真的可以说，UI回来了，宝贝。

We're actually like, UI's back, baby.

比如SaaS，这可能是你的救星，因为人类确实喜欢按钮。

Like, SaaS, this might be your savior because humans really do like a button.

现在既然你能让AI帮你点击按钮、替你操作界面，如果产品更适合用UI来呈现，你就不一定非得去琢磨那些CLI和MCP了。

And now that you can get AI to click buttons for you and use a UI for you, you sort of don't have to think about these CLIs and MCPs necessarily if the right expression of your product is more of a UI based thing.

所以我真的觉得，计算机使用能力简直太厉害了。

So I really do think computer use is kind of incredible.

我会让它继续跑着。

I'm gonna let this keep running.

我再举一个全新的例子，我还没试过，但我觉得它会表现得很棒，那就是我们经常在Flora里编辑播客的缩略图。

I'm gonna give another example of something kind of completely fresh that I haven't done yet, but I think it's gonna do really well, which is I'm gonna say we do a lot of thumbnail editing for our podcast in Flora.

我现在已经在Chrome里打开了Flora，里面存了我们所有播客缩略图放大的工作流程。

I have Flora open in Chrome right now with all of our podcast thumbnail upscaling workflows.

我现在就给你发几张我正在录的播客的照片。

I'm gonna drop you a couple images from the podcast I'm recording live right now.

你能生成我们的设计师Naomi真正需要用来制作这个播客缩略图的图片吗？

And can you generate the images that our designer Naomi would need to actually generate the thumbnails for this podcast?

然后我要拿...

And then I'm going to take...

我今天早些时候拍了一些照片。

I took some photos, earlier today.

所以我要发...

So I'm gonna drop...

就在我开始之前，我发一些高质量的照片，是你的朋友Claire，她看起来对Astra超级兴奋，因为我也很兴奋。

Right before I started this, I'm dropping really high quality photos of your friend Claire looking super excited about Astra because I am.

我要发送这个，然后看看它能不能打开Flora，Flora又是一个基于节点工作流的AI图像和视频编辑工具。

And I'm gonna send this, and we're gonna see if it opens up Flora, which again is sort of this, like, node node workflow based AI editing tool for images and videos.

我要看看它能不能操作那个界面，因为，再说一次，我以前没这么干过。

I'm gonna see if it can navigate that because, again, I haven't done this before.

所以我想看看它是不是真的能行，而且能做得很好。

And so I wanna see if it actually works and can do that well.

它会检查我已经在 Flora 里打开的缩略图工作流，然后用这三张照片来准备 Naomi 需要的图像素材。

So it's gonna inspect the thumbnail workflows that I already have open in Flora and then use these three photos to prepare the image assets Naomi needs.

你可以看到，这就是我们的工作流。

You can see here, this is our kind of like workflow.

看起来乱七八糟的。

It looks cuckoo bananas.

这个组织得不太好。

This is not super well organized.

就像你看到的，我们拍一些我或者嘉宾录播客的普通照片，然后让它们变得发光发亮、看起来很棒，因为不好意思，YouTube 说你们就喜欢这样。

As you can see, we take, like, kind of normal photos of me recording the podcast or a guest recording the podcast, And then we make them, like, glowy and look fabulous because I'm sorry, but that's what YouTube says you like.

它说你们喜欢这种发光的紫色头发，Claire。

It says you like this, like, glowing purple hair, Claire.

我希望你喜欢普通的办公室灯光，Claire，但你喜欢紫色发光头发，Claire。

I wish you liked normal office lighting, Claire, but you like purple glowy hair, Claire.

所以它要做的是，把我刚拖进来的这些新图片，操作这个应用，导入这些图片，然后希望能生成我们想要的视频，就是你今天来播客会看到的那些。

And so what it's gonna do is it's actually gonna take these new images that I dropped, navigate this app, import these images, and hopefully generate us the videos that we want that you'll see today, when you come to the podcast.

那我就让它先跑着。

So I'm gonna let that run.

等节点生成的时候我们再回来看。

We're gonna come back to it once the nodes are being generated.

好。

Okay.

那我们来看看它在做什么。

So let's watch what it's doing.

它拿了我刚才在Photo Booth里拍的那张照片。

It took this imported photo of me that I just took in photo booth.

它正在画一个节点，选择要用的图像生成模型。

It's going through and drawing a node, selecting the image generation model that's gonna use.

我们用GPT Image 2。

We use GPT Image two.

我们以前用的是Nano Banana，但其实我们真的很喜欢GPT Image 2。

We used to use Nano Banana, but we actually really like GPT Image two.

它正在选择宽高比。

It's selecting aspect ratio.

它根据我们已有的提示词填入提示，然后就会开始生成。

It's putting in the prompt based on other prompts we have, and it's gonna go ahead and generate that.

我觉得它会对这三张图片都这么做。

And I think it's gonna do it for all three of these images.

而且，这又是我们的设计师Naomi，她特别棒，每次都是她手动完成的。

And, again, this is something that our designer, Naomi, who's fabulous, does by hand every time.

这活儿有点机械，拖拽节点、整理好一切，确实需要费点功夫。

It's sort of mindless, and it takes a little bit of effort to drag the nodes, get it all organized.

另外，我们很喜欢Flora，但这里简直是一团乱麻。

Also, we love Flora, but it's kind of like chaos reigns here.

所以任何能帮我们重新整理这些节点、让它们井井有条的工具，都会非常非常有帮助。

So anything that can help us reorganize these nodes and keep them in check can be very, very helpful.

所以我们接下来可能会让 Codex 来生成所有这些图片。

And so what we're probably gonna do is have Codex now generate all these images.

然后 Naomi 的工作，就像所有人类一样，就是要有好的审美，确保模型不会把我的脸弄得很奇怪，或者给我多几根手指，或者让我看起来像恐怖谷那种感觉。

And then Naomi's job, like all humans, is just to have good taste and make sure that the model doesn't make my face look insane or give me extra fingers or, give me, like, an uncanny valley look.

但你可以看到，这里已经开始生成一张图片了。

But you can see here, it started generating one image.

它会继续生成下一张。

It's gonna go ahead and generate the next.

所以你可以把这两个例子看作，一个是这种复杂的 CRM 工作流节点构建器，另一个是用 Flora 来生成素材和写提示词，这些都非常受益于这个界面，但不需要人类用手指去输入提示词、拖拽东西来构建界面。

And so you can just think of this as two examples, this, like, complex CRM workflow node builder and, using Flora to generate assets and do prompting as something that really benefits from this UI, but does not benefit from humans using their fingers and typing prompts and dragging things to make the UI.

这正是 Astra 真正大放异彩的地方。

And this is where Astra really, really shines.

所以我一天大概有 90% 的时间都坐在这里，看着 Astra 用我的电脑，特别是用 Chrome 浏览我非常喜欢的网站和应用程序。

And so I do spend about 90% of my day sitting here watching Astra use my computer and specifically use Chrome to navigate websites and applications that I really love.

顺便提一下，我使用 Astra 和浏览器操作的另一个用例，我觉得对软件开发者普遍很有帮助，我们一会儿会讲到编程，那就是 QA 测试。

As a side note, another use case that I use for Astra and browser use that I think is really helpful for software developers in general, and we'll get to coding in a minute, is just QA.

所以最近，我在用Astra来稳定ChatPRD主聊天体验。

And so recently, I was working on stabilizing the main ChatPRD chat experience with with Astra.

我们一直在追查一些竞态条件和流式持久化中的一些奇怪的边界情况之类的问题。

And we were chasing down some, like, race conditions and some weird edge case things in streaming persistence and all this kind of stuff.

当我们终于把代码和PR都搞定后，我就说，你能不能就在Chrome里测试一下这个？

When we finally got the code shipped and the PR shipped, and I was like, can you just, like, test this in Chrome?

最棒的是，它直接打开了预览分支，然后开始点击测试、发送聊天消息。

And what was really awesome is it just pulled up the preview branch, and then it started clicking through testing, sending chats.

在这里，浏览器使用真正帮到忙的是，它检查了控制台，查看了错误日志。

And what was really helpful about browser use here is it was inspecting the console, checking for error logs.

它处理了像竞态条件这类问题。

It was doing things like race conditions.

它会刷新Chrome。

It would, like, refresh Chrome.

它会做所有这些作为人类执行起来非常繁琐和困难的事情，但作为QA却非常有用。

It would do all this stuff that would be very tedious and hard to execute as a human, but was super useful as a QA.

你可以看到，它实际上跑了差不多一个小时四十五分钟。

You can see here it actually ran for an hour and forty five minutes, basically.

在测试这个分支的时候，它发现了一些导航时的问题。

QA ing this branch, it found a couple issues when you were navigating.

它修复了一些东西。

It fixed some things.

它特别检查了一些竞态条件，效果真的很棒。

It specifically looked at some, race conditions and was really awesome.

所以如果你还没有用浏览器工具做QA，请一定要用。

So if you are not using browser use for QA, please do.

我准备用“发现”和“制作链接”来结束我们的计算机操作演示，效果简直不可思议。

I'm just gonna go wrap our computer use action by looking at found and made the links kind of incredible.

我现在就给Naomi发Slack消息，告诉她好消息。

I'm actually just going to Slack Naomi right now and say, good news.

Astra在Flora里自己生成了缩略图。

Astra made the thumb photos itself in Flora.

超级可爱。

Super cute.

然后我会给她发一个示例。

And then I'm going to send her an example of it.

而且，这真的很有趣，因为想想一个完美的例子。

And, like, this is so funny because think of a perfect example.

她有个预约。

She's got an appointment.

我们需要完成这件事，因为你们都喜欢尽快看到模型评论，而Astra为我们的缩略图制作了一张很棒的图片。

We need this thing done because you all like your model reviews ASAP, and Astra made a great image for our thumbnail.

所以，这真的是一个很好的例子，说明模型和浏览器通过计算机使用，即使在创意领域，也能减轻人类的负担，让你工作得更快一点，让你的客户印象深刻。

And so, this is just a really great example of a model and browser using computer use just taking toil off humans, even in the creative field and letting you do your work a little bit faster and letting you impress your customers.

你们都是我的客户。

You all are my customers and my clients.

你们是我的观众。

You're my audience.

我更爱你们一点了。

I love you a little bit more.

所以，这真是一个非常酷的例子。

So very, very cool example.

我爱Astra。

I love Astra.

我爱浏览器使用。

I love browser use.

但让我们来看看其他有趣的东西。

But let's get to some other fun stuff.

好的。

Okay.

我们来谈谈编程吧。

Let's talk about coding.

是的。

Yes.

显然，它在编程方面很厉害。

Obviously, it's great at coding.

但具体来说，它突破了一些软件工程上的难题，这些难题我之前用GPT-5.6、Fable或者Opus都没能解决。

But specifically, it has broken through a couple software engineering challenges that I have not been able to crack thus far with GPT-five six or with Fable or with Opus.

具体来说，我……

Specifically, I...

对于ChatPRD，我一直在开发一个功能，一个产品智能功能，它能把来自Intercom、Granola、Linear和GitHub的所有产品信息汇总起来，然后告诉你发生了什么、什么趋势、你需要做什么。

For ChatPRD, I have been working on this feature, this product intelligence feature that, like, sort of, like, sucks up all your product intelligence from, like, Intercom and Granola and Linear and GitHub and kinda tells you what's going on, what's trending, what you need to do.

除了生成这些洞察和趋势之外，它还会自动生成一个关于你产品状态的维基。

And in addition to creating those insights and trends, it generates this sort of, like, auto Wiki of the state of your product.

我把所有模型都试了一遍。

And I have thrown every model at this.

我挺聪明的。

I'm pretty smart.

我大概知道它在架构上应该怎么运作，但它很复杂。

I kind of, like, architecturally know how it needs to work, but it is complex.

它得先接收数据，然后决定哪些要长期保存。

It has to ingest data, has to decide what it durably stores.

还得处理数据，把重复的见解去掉。

It has to process it, has to deduplicate insights.

还要生成一个知识库。

It has to generate a wiki.

并且要对照外部来源来验证这个知识库。

It has to verify that wiki against external sources.

然后，我……

And, like, I...

Fable在架构上做得特别厉害。

Fable did insane things with the architecture.

5.6就是在见解质量上卡住了。

5.6 just, like, got stuck on quality of insights.

我跟你们说，Astra一次就搞定了。

And I am telling you all, Astra one shot it.

我当时就想……

And I was like...

就好像……

It was like...

这就是我整个模型使用过程中的体验，做这些非常具体的任务一直特别难，我反复试了六个月。

This has been my experience this entire model, which has been like, I've been so hard to do these, like, very specific tasks, I've tried them over and over and over again for six months.

然后突然之间，Astra 一下子就搞定了。

And then all of a sudden Astra started to one shot it.

所以我一直在做这个产品图谱和产品智能的项目，做了很久，最后干脆交给 Astra 了。

So I have been working on this product graph, product intelligence forever, and finally just gave it to Astra.

我就说，你能不能……

I'm like, can you...

我开了一个分支。

I have a branch open.

你能修一下吗？

Can you fix it?

然后它就把它建好了。

And it built it.

所以它洞察……

So it insight...

它，你知道的，从各种不同的来源摄取数据，比如Intercom、Granola、Linear等等。

It, you know, ingests data from a bunch of different sources, intercom, granola, linear, etcetera.

它把这些洞察拉进来。

It pulls in those insights.

它告诉我作为产品负责人，我的三大优先事项应该是什么。

It tells me what my top three priorities as a product person should be.

AI卡住了，那是我们一直在解决的那个竞态条件bug。

AI stalling, that was that race condition bug that we were figuring out.

自助取消功能坏了，然后人们还想要本地部署的Jira。

Self-service cancellation is broken, and then people want on prem Jira.

所以Jira的朋友们，请赞助这个播客，人们想要本地部署的Jira集成。

So Jira friends, please, sponsors of the pod, people want on prem Jira integration.

它会随着时间生成这类图表。

It generates these kind of graphs over time.

它会告诉你有哪些机会。

It it tells you what opportunities.

我跟你说，我试过其他模型，都没法从一组数据里得出这么高质量的洞察。

And I am telling you, I could not get any other model to get to this quality of insight out of a set of data.

但我最喜欢的部分是它居然自动生成了维基页面。

But my favorite part is it actually generated the auto wiki well.

它的作用就是把你能找到的关于你产品的所有信息都抓取过来，这次是针对 ChatPRD，它还去查了 GitHub。

So what this does is it sucks up all the information you can find about your product, in this case, about ChatPRD, and it went and looked at GitHub.

它看了我们的文档，基本上把产品当前的功能、最近的更新、客户背景、需要修复的问题这些都整理出来，然后通过 MCP 接口暴露出去。

It looked at our documents, and it just basically explains current function of all the product, recent changes, customer context, things that we need to fix, and then it exposes this via MCP.

我跟你说，这感觉简直就像……

And I am telling you, this was like...

我管这叫三样本提示（three-shot）。

I would call it three shot.

一次提示就达到了90%，然后又试了两三次其他提示。

One shot got it to 90%, and it was like two or three other prompts.

最后终于跑通了。

And finally, it got working.

我还稍微美化了一下界面。

I had to polish the UI a little bit.

但这种事情，其他模型一开始都做不好。

But this is one of those things that just none none of these models could do really well out of the out of the gate.

不是它们……

It wasn't that they...

我没法从技术上说明白我想要怎么实现。

I couldn't technically explain how I wanted executed.

它自己做的架构和决策，结果质量很差，而且真的……

The architecture and whatever decisions it made on how to do it just resulted in poor quality, and it really...

Astra这次真的搞定了。

Astra really figured this one out.

所以它大大加速了我在ChatPRD上的个人产品开发和软件开发。

So it has super accelerated my personal product development, software development on ChatPRD.

但更重要的是，它让我能做一些有趣的东西。

But more important, it's let me build some fun stuff.

那我们就来聊聊编程，但要让编程变得有趣。

So let's talk about coding, but make it fun.

你们都知道，我被这个东西坑过。

So you all know that I have been victimized by this thing.

就是那个Divoom Mini 2。

It's a, Divoom Mini two.

它其实就是一个蓝牙音箱。

It is really just like a Bluetooth speaker.

它预装了专有软件。

It comes preinstalled with proprietary software.

你不能……

You can't...

没有API。

There's no API.

你破解不了它。

You can't crack it.

你也黑不进去。

You can't hack it.

而GPT-5是第一个让我在这台小电脑上做最基础操作的东西。

And GPT-five five was the first one that even let me do the most basic thing on this little computer.

这就像是……

This has been...

这是我的珠穆朗玛峰。

This is my my Everest.

而且，我已经能在这台小设备上刷入的东西，质量已经非常惊人了。

And already, I mean, the quality of what I've been able to flash onto this little device is amazing.

你看。

Look.

这就像克莱尔的小实验室。

It's like Claire's little lab.

它已经弄明白了像素算法，这个算法我之前还得自己逆向工程。

It's figured out the pixel algorithm, which I had to to reverse engineer.

那可真不容易。

It was not easy.

所以它现在已经在画可爱的画了。

So it's it's already doing cute art.

但让我给你看看Astra在这里能做什么。

But let me show you what Astra was able to do here.

Astra建了一个小应用，让我能实时连接到这个设备。

So Astra built a little app that allows me to connect live to, this device.

我可以用鼠标在这个设备上画画。

I can draw on the device using my mouse.

我还能改这个东西。

I can change the thing.

我可以打字，我爱我的订阅者。

I can type, I love my subscribers.

它不会帮我检查拼写。

It's not gonna spell check for me.

我可以在直播中做所有这些。

I can do all that live, live streaming.

但更酷的是，我让Astra给我建了一个命令行工具，现在它已经完全搞懂了这个东西，所以它可以随时跟我实时聊天，想聊什么就聊什么。

But even cooler, I had Astra build me a CLI, now that it's totally figured out this thing, so that it can talk to me live about anything it wants.

所以我可以对它说，再用命令行工具告诉我们的观众最新一期的《How AI》节目。

And so I can say, use the CLI again to tell our audience about the most recent episode of How AI.

然后它现在就可以去查……

And so it can now go look...

现在它说Codex正在工作。

Now it's saying Codex is working.

真可爱。

Very cute.

它会非常认真地思考自己在做什么。

It's going to think really hard about what it's doing.

它会找到最新一期的《How I AI》播客，然后向观众直播一个友好的CLI介绍。

It's gonna find the latest How I AI published episode, and it's going to stream an audience friendly introduction to the CLI.

我是说，以前那些模型，我花了老半天才让它打个招呼，现在我真的想做什么都能做。

I mean, again, it took me forever to get this thing to say hi with these previous models, and now I am truly able to do anything I want on this.

所以它会说Codex正在工作。

So it's gonna say Codex is working.

它正在读取作为脚本的Python代码，然后希望，祈祷一下，它会向你们这些可爱的订阅者直播我们最新一期的《How AI》节目，让我们看看它能不能做对。

It's reading the Python that is the script, and then hopefully, fingers crossed, it will stream a message to you all beautiful subscribers about our most recent episode of Howe AI, and let's just see if it gets it right.

它做对了。

It got it right.

最新一期是《Grokbot对战Openclaw》，但让我们看看它能不能真的把它直播到那台小电脑上。

The most recent episode is Grokbot versus Openclaw, but let's see if it can actually stream it to the little computer.

我会展示给你们看，我们会看到它进来。

I'll show it to you all, and we'll see it come in.

来了来了。

Here it goes.

正在直播。

It's streaming.

它说克莱尔给我们展示了格罗克机器人。

So it's saying Claire shows us Grock Bot.

它正在直播节目信息，然后应该会给我们展示一张超级超级可爱的小图片，关于我的很多格罗克机器人。

It's streaming in the episode information, and then it should show us a super, super cute little image, about my many Grock bots.

看看那个。

But look at that.

它完全开箱即用。

It just works totally out of the box.

看看多可爱。

Look how cute that is.

它甚至还会动。

It even animates.

好的。

Okay.

这又是我的一座大山。

This, again, is my Everest.

里面还写着，在PR支持SOC2监控，以及从Open Clot迁移代理。

It also says inside PR support SOC two monitoring and moving agents from Open Clot.

三十六分钟，真人代理导览。

Thirty six minutes, a real agent tour.

今天就在你的播客应用里找到它吧。

Find it in your podcast app today.

这个东西。

This thing.

我们做到了，各位。

We did it, folks.

接下来，我得给它root一下。

Next, I gotta root it.

所以这是下一步。

So that's next.

另外一件我做的很酷的事，你们还没见过的，就是我把这个接入了我的 Codex 钩子。

The other really cool thing that I did with this that you all haven't seen is I hooked it into my Codex hooks.

现在它能在 Codex 需要东西的时候告诉我。

So now it can tell me when Codex needs something.

这就引出了我的下一个用例。

And that brings me to my next use case.

非常有趣。

Very fun.

还是那句话，就是做一些我以前做不到、现在能做到的编程事情。

Again, just coding things that I could not do before that I can do now.

我们来看看。

Let's look at it.

它正在给我做一个 Mac 应用，能把我的 Codex 线程变成九十年代风格的 AIM 即时通讯工具。

It is building me a Mac app that turns my Codex threads into a nineties style AIM a messenger.

对于那些年轻漂亮的你们……

For those of you who are young and beautiful and...

或者可能在AIM存在时还没出生的人。

Or maybe not even born when AIM was around.

AOL即时通讯，就是我们这些老一辈以前在网上聊天用的，它长这样。

AOL Instant Messenger, it's how we olds used to chitchat with each other on the Internet, and it looked like this.

它长这样。

It looked like this.

看看Astra给我做的这个东西。

And look at this thing that Astra made for me.

好。

Okay.

它把我所有活跃的Codex线程包装成了九十年代AIM即时通讯的风格。

So it made a nineties AIM instant messenger style wrapper around all my live Codex threads.

所以我的所有好友就是不同的线程。

So all my buddies are different threads.

这是Divoom的小迷你应用，我可以打开我们正在聊天的那个对话。

Here's the Divoom little mini app, so I can open up the chat that we're chatting there.

它说它直播了9月2号那期关于Grock Bot对战Ohenclaw的节目。

It said it streamed the September 2 episode on Grock Bot versus Ohenclaw.

我可以回去再给它发一条消息。

I can go back and send it another message.

我可以看看不同的东西。

I can look at different things.

比如，我做了这个Astra播客来回顾。

Like, I built this Astra podcast for review.

所以我可以跟Codex聊天，就像它是我的AIM小好友一样。

So I can chat with Codex as if it's my little AIM buddy.

它真的太可爱了。

It's so stinking cute.

它让我保留我的屏幕名。

It lets me have my screen name.

这是我的离开消息，马上回来，让我的智能体们继续干活。

Here's my away message, BRB, letting my agents cook.

然后你可以看到，我可以对我刚才在做的那个线程说，干得好。

And then you can see here, I can say to that thread I was just working on, great job.

那效果太棒了。

That worked awesome.

我太喜欢了。

I love it so much.

而且它还能……

And it can...

哦，它已经有一个正在工作的写手了。

Oh, it's already got an active active writer.

所以它知道什么时候真的在工作，就像现在这样。

So it knows when it's actually working, which it is right now.

它不让我写。

It won't let me write.