BONUS: How We Would Teach AI From Scratch in 2026

我刚看到你问了这个问题。

I just saw you ask that question.

它可能需要几分钟的时间。

It can take as long as a few minutes.

它也可能短一些。

It can be a little short.

这完全取决于你问的是什么。

It totally depends on what you ask.

它们在调整这个时间长度方面已经进步了很多。

They've gotten much more about kind of flexing that amount of time.

以前，思考意味着固定几秒钟，后来时间变短了，但人们又希望更长一些，因为他们面临更复杂的问题，给它更多时间能让它思考得更深入。

It used to be that thinking meant x number of seconds, and then it got a little shorter and then people wanted longer because they had harder problems and giving it more time lets it think more.

所以，你知道，如果你真的想运行它，就告诉它开始，如果它开始转圈，你可以去另一个聊天窗口做别的事，它完成后会通知你。

So, you know, if the truth is you can run it, tell it to go, and if it starts circling, you can go over to another chat and work on something else and it'll you can have it notify you when it's done.

所以你不必干坐着盯着它。

So you don't have to just sit and stare at it.

通常，我会在这里打开两三个ChatGPT窗口。

Usually, what I will do is I'll spin up two or three ChatGPT windows here.

如果我今天早上知道有两三个任务要处理，我会逐一启动它们，然后去喝杯咖啡，顺便处理邮件，让AI自己运行。

And if I have two or three things I know I need to work on this morning, I'll go start each of those, and then I'll go over and I'll drink my coffee while I go through my email and just let it work.

你不必一直盯着它看。

You don't have to just stare at it.

事实上，这才是真正提升效率的地方。

The truth is that's where the big productivity boost is.

它正在为我做一些基础工作，而这些工作我本来就要做。

It's it's doing some legwork for me while I'm going through something else I need to do right now anyways.

我的日程一直很满，而这种方式真的让我豁然开朗。

My day is my day is busy, and this has been such an unlock.

我之前看过，比如格兰特和我过去十四个月的产出数据。

I was looking at, like, even just Grant and I's production numbers from the last, like, fourteen months.

格兰特，我还没告诉过你这件事。

Grant, I haven't even told you this.

哦，真的吗？

Oh, really?

我的录音停止了。

My recording has stopped.

哦，不妙。

Oh, no.

你没事的。

You're good.

我还在

I still

看到是的。

see Yeah.

好的。

Okay.

让我在这里按一下。

Let me hit something here.

哦，看来他不小心被踢出了直播。

Oh, looks like he'd just accidentally kicked himself off the stream.

嗯，他马上就会重新加入。

Well, he'll rejoin in just a moment.

我很想知道他刚才在说什么，但我刚刚在聊天框里发了一条消息，总结一下我们刚才讨论的内容。

I'm curious to hear what he was talking about, but I just posted something in the chat just to recap what we covered.

所以我们讲了，你去访问你最喜欢的网站，也许你还没有最喜欢的网站。

So we covered, you go to your favorite website and maybe you don't have a favorite website yet.

所以最容易开始的是 chat.com。

So the easiest one to start with is chat.com.

我最近特别喜欢的是 claude.ai。

The one that I really like lately is claude.ai.

但这两个网站都不会让你失望。

But you can't go wrong with either of those.

Gemini 和 Grok 大概是第三和第四好的选择。

Gemini and Grok are the third and the fourth probably best.

而且 Gemini 的智能水平至少与最高水平相当，但我们发现它的表现不够稳定，而且应用程序体验也不够好。

And Gemini has higher intelligence than, is at least on par with the highest intelligence, but we find that it's not as consistent and actually the app is not as good.

所以我们首先推荐 Chat 和 Claude，然后是 Gemini，再然后是给完全新手的。

So that's why we would recommend Chat and Claude first and then Gemini second and then, yeah, for total beginners.

给完全新手的。

For total beginners.

是的。

Yeah.

对。

Yeah.

在你下线之前，不过呢。

Before you drop off, but yeah.

因为还有 Copilot。

Because there is Copilot as well.

是的。

Yeah.

我经常这么做，我知道有些人对Copilot有强烈的意见，甚至非常不喜欢它。

The thing the reason I often and I know I know there are people who have strong opinions about Copilot and and really dislike it.

但我有时推荐它给初学者的原因是，只需一个订阅，你就能获得ChatGPT最好的模型和Anthropic最好的Claude模型。

But the reason I recommend it sometimes to beginners is because for one subscription, you get access to ChatGPT's best models, Claude's best Anthropic's best models.

所以你可以获得Claude、ChatGPT和Claude，全部整合在一个账户里。

So Claude, you get ChatGPT and Claude in one account.

你可以在这几个模型之间切换，而且他们昨天刚推出一项新功能，让这些模型互相点评。

You can switch between them and there's a new feature they just dropped yesterday to where those models critique each other.

其他人没有这样做。

No one else is doing that.

至少没有大规模地这么做。

At least not on a broad scale.

我们在智能体中已经做过类似的事情，但还没有向大众广泛推出过。

It's a thing we've done in agents and such but not a thing that has been, like, rolled out to the masses.

而且这项功能很快就会推出，非常快。

And that's that's coming very, very soon.

它现在已经在前沿版本中了，到夏天就会向公众开放。

It's in Frontier now, and it'll be to the public by summer.

事实上，如果你只是普通用户，你会从中获得很多好处，尤其是如果你本来就习惯把文件保存在三月之类的文件夹里的话。

The truth is if if you're a casual user, you're gonna get a lot out of that, especially if you already save your files in in March or something anyway.

话虽如此，是的，我用的是ChatGPT和Claude。

With that said, yeah, I I'm ChatGPT, Claude.

对。

Yeah.

我用的是Gemini。

Those I mean, I use Gemini.

我们待会儿再聊这个。

We'll get into this.

还有Gemini。

And Gemini.

Gemini适合特定的使用场景，但不适合我的日常主力工具。

Gemini for specific use cases, but not for my daily driver.

我稍后会解释为什么。

And I'll explain why in a minute.

是的。

Yeah.

格兰特和

Grant and

你应该知道，我们的日常使用工具是相反的。

I have opposing daily drivers is what you should know.

这并没有什么真正的原因。

And and it's not for any real reason.

事实上，这两个工具的大部分功能都是相同的。

The truth is most of the same features exist in both tools.

在任何给定时刻，它们都拥有该领域中排名前二的模型。

Both of them have, at any given moment, the top two models in in in the space.

但我想说的是，一旦你习惯了某个工具，有时候就没必要每次出现新模型或小改动就更换。

But what I will say is that once you get a tool you're comfortable with, sometimes it's not worth changing every time there's a new model or a little tweak.

我的意思是，如果有重要的事情，你当然可以换。

I mean, if there's something significant and you want, absolutely.

但我强烈建议你找一个你真正喜欢界面、用着顺手的工具，可以试试它们的免费版，看看：我真的不喜欢这个的使用方式。

But I highly recommend finding one you're you just like the interface on, you're comfortable, you can go and try out their free versions and see, like, I just don't like how this work.

这个用起来不错，或者我真的很喜欢这个的使用方式。

This works or I I really like how this works.

如果是这种情况，就坚持用下去吧，因为这些公司现在每隔几个月就会推出新模型。

And if that's the case, stick with it because these these companies are putting out models every couple of months right now.

只要你用的是其中任何一个，你都能随时拥有接近甚至达到最前沿的模型。

And if you're with either one of these, you're always going to have something that's right near state of the art, if not the state of the art.

所以我强烈推荐你选其中一个，或者两个都用，只要你有余钱——因为我每天都用这两个。

So I would highly recommend either one of them or both if you just have the expendable money because I use them both daily.

所以我想再补充一点。

So I just wanna cover one thing.

有人问到了Perplexity。

So Perplex someone asked about Perplexity.

我目前不打算订阅Perplexity的原因是，除非你打算购买他们的高价产品Perplexity Computer（这是另一个产品），因为Perplexity的大部分功能，ChatGPT、Gemini都能做到，Claude也能部分做到。

The reason why I would maybe not subscribe to Perplexity right now, unless you're gonna pay for their expensive version for Perplexity Computer, which is a different product, is because most of what Perplexity does, ChatGPT does and Gemini can do and Claude can kind of do.

我们会在第二部分进一步讨论这个话题。

And we'll get into that more in part two.

但如果你喜欢Perplexity，并且觉得它在搜索场景中对你特别有价值，那完全可以用它。

But if if you like Perplexity and you find it really, really valuable for you specifically for search use cases, totally fine to use it.

是的。

Yeah.

在我看来，它并不一定能够替代ChatGPT或Claude这样的日常主力工具。

It's not it's not, in my opinion, necessarily a replacement for like a daily driver like Chat or Claude.

对。

Yeah.

如果你是把它当作主力工具来用，我很想了解你是怎么使用的。

If you're using it that way, I'd love to learn more about how you're using it.

对我来说，它更多是一种补充。

For me, it's more of a supplement too.

我非常喜欢Perplexity的发现功能，它会根据我的兴趣推送新闻。

I really love Perplexity's discover feature where it caters news to me.

它会根据它认为我今天想看的新闻内容，自动生成新闻文章。

Essentially, writes news articles based on what it thinks I wanna read the news about today.

是的。

Yeah.

而且很多时候，它推送的内容比真实新闻更积极、更有趣。

And and a lot of times, it's a lot more positive and interesting than the real news are.

没错，Perplexity在金融领域有很多非常高级的应用场景。

And yeah, perplexity has a lot of really sophisticated financial use cases.

正如我所说，Perplexity Computer是完全不同的产品。

Like I said, perplexity computer is like a totally different product.

它有点像是OpenAI的替代品，而且非常出色。

And that would be kind of like a replacement for OpenCLaw and that's really good.

所以，如果你愿意为Perplexity Computer花钱，那就是完全不同的话题了，而且很可能值得你投入时间。

So I think if you're willing to spend the money for a Perplexity Computer, that's a completely different conversation and that probably would be worth your time.

这并不是说你不能在ChatGPT或Claude中做Perplexity Computer能做的事情。

Now it's not to say that you can't do what Perplexity Computer does in ChatGPT or Claude.

只是你无法达到Perplexity Computer那样的水平。

You just can't do it to the level of Perplexity Computer.

是的，完全正确。

Yes, exactly.

是的，Perplexity用于研究，马库斯·斯通说，嗯，确实很不错。

Yeah, Perplexity for Research, Marcus Stone says, yeah, I mean, it's pretty good.

我发现它能访问的网站数量可能仅次于ChatGPT。

I find that it can access probably the most websites besides ChatGPT.

我觉得ChatGPT几乎能访问你需要去的每一个网站。

Think ChatGPT can access pretty much every website that you need to go to.

但话说回来，我们还有没有其他想聊的初学者内容？

But yeah, so let's get into, are there any more absolute beginner things we wanna cover?

我觉得应该聊聊如何使用提示词和界面。

I I think hitting prompts, how to use the interface.

我唯一没展示出来的东西是，由于我的网络连接问题，我不太想再提了：如果你进入ChatGPT的设置，会有一个个性化选项，那是一个很大的菜单。

The only thing I could have shown that I didn't, and I'm hesitant to bring it back up because of my internet access, is if you go to your settings in ChatGPT, there's a personalization option and it's a big menu.

它并不让人畏惧，但你会想：你希望它怎么跟你对话？

It's not intimidating but it's like, how do you want it to talk to it?

你希望它友好一点、书呆子气一点，还是务实一点？

Do you want it do you do you want it to be friendly, nerdy, pragmatic?

你希望它非常专业吗？

Do you want it to be, you know, very professional?

你可以自定义所有这些设置，还可以告诉它一些关于你的信息，从而决定你们互动的方式。

You know, you can customize all of that and you can tell it things about yourself and determine, like, how your interactions go.

这真的是一次巨大的突破

And that was a really big unlock

对我而言也是如此。

for me.

我们可以在我这边展示一下。

We can show it on mine.

你能不能给我演示一下怎么操作？

Why why don't you walk me through how to do that?

好的。

Okay.

往下滚动到左下角，那里有你的信息和设置选项。

Go down to the bottom left where your information is there, your settings.

点击确定。

Click Okay.

往上找到个性化设置。

Go up to set personalization.

好的。

Okay.

你看到这个了吗？

So you see this?

我之前是点击了左下角的个人资料。

So what I did is I clicked on my profile in the bottom left corner.

那会是你的名字、邮箱之类的。

That'll be your name or your email or something.

然后我去设置了个性化。

And I went to personalization.

对，没错。

And yeah.

当你来到这里时，比如那个顶部的下拉菜单，就是一个非常好的选项。

And when you get over here, like, that top drop down, for example, is a really good one.

你看，你有这么多不同的选项。

You know, you've got all these different options.

你可以让系统以你想要的任何方式与你互动。

You can have it communicate with you however you want.

我通常用的是古怪模式，我想是这样。

I usually use Quirky, I believe.

好像有个很技术宅的选项，不过我也可能记错了。

Seems like there was a nerdy, but I could be wrong.

你可以通过多种方式微调它的语气，如果有什么让你烦躁，或者你特别需要它为你解锁更多功能的话。

You can adjust its tone in a variety of ways up and down a little bit if something's getting on your nerves or you've got something you really feel like you need from it to to better unlock capabilities for you.

这真的很酷。

This is really cool.

是的。

Yeah.

确实如此。

It really is.

这是一个很棒的个性化功能，但使用的人还远远不够。

It's a great personalization feature and not nearly enough people use it.

在我的自定义指令中，我让它做一些事情，比如质疑我的假设，当我把某些未经证实的事情当作真理时提醒我，还有其他各种类似的功能，目的是让它更好地满足我的需求。

In custom instructions like mine is told to do things like challenge my assumptions, double check me if I've taken anything for granted as true that is not, and a variety of things like that that are designed to get it to give me what I need out of my AI.

而我需要的，是一个能在我冲动时拉住我的缓冲器。

And what I need is a backstop to I rush.

我每天都会在极短的时间内做很多事情，所以我需要一个能提醒我：‘慢下来，科里。’的东西。

I do a lot of things in a small amount of time every single day of my life, and I need something that can be like, slow down, Corey.

你把钥匙留在门里了，就像昨天克劳德做的那样。

You've for you've left the keys in the door much like Claude did yesterday.

对。

And Yeah.

我们稍后再谈这个吧。

We'll talk about that maybe later.

如果没有，我看聊天里有一大堆问题。

If not, I I see there are a a million questions in the chat here.

是的。

Yeah.

我们来回答这些问题吧。

Let's answer these.

所有的

All of

这些都整理出来，之后可以写成一篇博客文章，因为很明显今天我们不可能回答每一个问题。

these out, and we can put them in a blog article after because I've I've it's pretty clear we're not going to get every single one of them today.

但无论我们没讲到什么，都会写进博客文章，并在Neuron中分享。

But whatever we don't, we'll put in a blog article and share in the Neuron as well.

关于马克汉姆MVP的问题，如何最好地使用AI工具融入求职过程，我想说一下。

I will say for the question about Markham MVP, best way to use AI tools to integrate into a job search.

我们会介绍一些能帮助你做到这一点的内容。

We're gonna cover something that will help you do that.

事实上，也许我们可以把这个作为我们情景规划的一部分。

And in fact, maybe we can use that as part of our scenario planning here.

我另外想说的是，我在聊天中发布了我们涵盖的内容，去你最喜欢的平台看看。

The one other thing I wanted to say, so I posted in the chat, okay, things we covered, go to your favorite platform.

你可能还没有最喜欢的平台。

You might not have a favorite platform yet.

所以不妨试试，比如试试chat.com、claude.ai、gemini.google.com或者drop.com。

So just try, maybe try chat.com, claude.ai, gemini.google.com, or drop.com.

向每个平台提几个问题，看看你最喜欢哪一个。

And just ask a couple questions to each of them and see which one you like the best.

然后就选其中一个吧。

And then just go with that one.

因为目前这些工具的智能水平都差不多。

Because right now the intelligence is pretty on par.

我觉得目前前三个要更好一些。

I would say the first three are the better of the three at the moment.

而接下来我们要讲的所有内容中，前两个是最合适的。

And the first two are the better of the two for all of the things we're gonna cover next.

然后你可以不断优化。

Then you can iterate.

所以你先问一个问题。

So you ask a question.

但重要的是，你第一次问不会得到想要的答案。

Then the important thing to remember is you're not gonna get what you want the first time.

所以你要和它不断互动，这正是科里向我们展示的方法。

So you iterate with it, which Corey showed us how to do.

然后对问题进行迭代，这样你可以搜索之前的聊天记录。

And then iterate on the question and then you can search to find previous chats.

另外一件事我之前在那条消息里没提到，就是科里提到的模型选择器，他谈到了

And then the other thing that we I didn't mention in that message is Corey talked about this here, which is the the model picker, and he talked about

是的。

Yeah.

为什么你需要这个功能。

Why you would want this.

Instant 适用于日常聊天。

Instant is for just everyday chats.

Thinking 适用于你希望获得更高智能水平的回答，并且愿意等待结果的情况。

Thinking is if you want it to use a higher degree of intelligence and you're okay waiting for an answer.

而 Pro，这个

And pro, this

从五秒到一两分钟不等。

anywhere from five seconds to a minute or two.

是的。

Yeah.

对。

Right.

这可能需要一些时间。

It can take some time.

你会在这里看到它已启用。

And you'll know that that's turned on right here.

是的。

Yeah.

那里还有一个扩展的思考选项。

And you've got an extended thinking option right there as well.

是下拉菜单吗？

Is that the dropdown?

哦，是的。

Oh yeah.

你可以选择标准模式和扩展模式。

You can do standard and extended.

如果你需要它为你提供最高的智能水平，那么最好一直开启扩展思考功能。

If what you need for it is to give you the highest intelligence possible, then it's probably a good idea to just have extended thinking on all the time.

我使用Claude时就是这样做的。

That's what I do when I use Claude.

这里还有其他工具，但我们会稍后在相关时再详细介绍。

And then there's other tools here, but we'll get into this slightly later as they're relevant.

但请记住，如果你看到这里的加号，就可以使用这些工具，这里有很多不同的工具。

But just remember that if you have this plus sign here, you can use these tools and there's lots of different tools on here.

好的。

Alright.

是的。

Yeah.

所以，好的。

So Alright.

现在我们要进入第二部分。

Now we're gonna get into the second part of this.

我先暂停一下共享屏幕。

I'm gonna stop sharing my screen for a second.

我们之前讨论过，如果你是完全的新手，你是应该先 setup 一个项目，还是先学习如何高效地进行提示？

So we were talking about whether or not if you're a total beginner, if you start by setting up a project or setting or learn how to prompt efficiently.

科里，你学习基本提示礼仪的原因是什么？

Corey, what's your reason for learning basic prompting etiquette?

然后我会分享我为什么在真正开始之前先 setup 一个项目。

And then I'll share my reasoning for setting up a project before you really get involved into anything.

哦，我来告诉你我这么做的原因。

Oh, I'll tell you why I do.

我这么说是因为，你知道，几年前提示工程是个热门话题。

And and and the reason I say this is is, you know, prompt engineering was a big subject a couple of years ago.

它当时被技术圈广泛嘲笑，却被非技术人员广泛采用。

It was largely mocked by the technical community and heavily adopted by non technicals.

在与AI互动时，这仍然是一个有价值的技能，但它并不像一个独立的职业。

It is still a valuable skill in interacting with AI, but it is not like a profession of its own.

所以，我们早期认为的最终目标是，世界上将充满提示工程师。

So, something I see as what we thought was the end game early on was like, world's gonna be full of prompt engineers.

我们都会去设计提示，那就是全部了。

We're all gonna be engineering prompts, and that's what it's gonna be.

相反，我认为提示工程就像是AI入门课程。

Instead, I think of prompt engineering as AI one zero one.

有很多短期课程。

There are a ton of short classes.

你可以看一个十五分钟的YouTube视频，获得大量想法和可以使用的不同模式，这能为你提供一个小工具箱，让你参加派对时有所准备。

You can watch a fifteen minute YouTube video and get a lot of ideas and different patterns that you can use to help you out and what that does is it gives you a little toolbox to show up to the party with.

所以，你知道，啊，我想做这个。

So, you know like, ah, I wanna do this.

我怎样才能对这个问题获得最好的结果呢？

How can I get the best results to this question possible?

你完全可以直接说一些话，它很可能就会接近你的需求。

You can absolutely just say some words and it'll probably get you close.

但如果你需要那额外的3%、5%来提升准确性或确保得到你想要的结果，那么理解如何精心设计问题、如何避免在提示中自相矛盾就非常重要，而自相矛盾是一个大问题。

But if you need that extra 3%, that extra 5% that that increases your accuracy or ensures what you want, the ability to understand how to how to carefully craft a question, how to ensure you're not doing things like contradicting yourself in a prompt, which is a big issue.

理解如何清晰地表达。

Understanding how to just just be clear.

事实上，你希望的是AI能清晰地理解你。

The truth is what you want is the AI to be clear.

它会回答你提出的任何问题。

It's gonna answer whatever you ask it.

大多数时候，当它回答错误时，很可能是因为提问的人问得不对。

Most of the time when it answers it wrong, it's likely because someone asked it wrong.

只要多花一点时间，看个视频。

And just taking a little extra time there, watch a video.

这其实就是我想说的全部。

That's that's really all I would do.

我会看一两个视频，在思考时多花点时间，尤其是在处理难题时。

I would watch a video or two, take a little extra time when you're thinking about it, especially on hard problems.

事实上，格兰特和我大部分时间都根本不会去想这个问题。

The truth is most of Grant and I's day, it gets no thought.

但对于一个绝对的新手来说，这对我来说是一个巨大的突破，我强烈推荐。

But as a as an absolute beginner, I found it to be a major unlock for me personally, and I highly recommend it.

好的。

Alright.

现在让我解释一下项目背后的逻辑以及它们是什么。

Now let me explain the logic behind projects and and what they are.

为了做到这一点，我会共享屏幕，直接进入解释。

And and to do that, I'll I'll share my screen and I'll just kinda get into the explanation.

是的。

Yeah.

是的。

Yeah.

给我一点时间。

So give me one moment.

我们现在要切换到Claude吗？

Are we bouncing over to Claude now?

是的。

Yeah.

接下来这部分我们要用Claude。

We're gonna we're gonna go to Claude for this next part.

你可以简要介绍一下，虽然一切都没变，但看起来有点不同。

You might give the twenty second rundown of how everything's the same, but it looks a little different.

至于你目前所处的位置，你知道的，你在哪儿，你

As far as like, you know, where you're where you're where you're at, where you

因为我觉得你在这里已经掌控得很好了。

Because I think that's a you you've got that under control here.

是的。

Yeah.

当你使用桌面应用时，你并不在。

When you're in the desktop app, you're not in.

是的。

Yeah.

所以我们来谈谈这个。

So we're going to talk about that.

这些工具的另一个混淆点是，你可以通过访问 chat.com 或 claud.ai（你偏好哪个都行）来使用网页版。

So another layer of confusion, for these tools is that there's a web based version that you can access by going to chat.com or, claud.ai, whatever your preference is.

此外，还有一个可以下载的桌面版本。

And then there's also web or sorry, there's also a desktop version you can download.

两者都有吗？

Of both?

是的。

Yeah.

两者都有。

Of both of them.

我实际上建议下载桌面版，特别是如果你打算用它来做一些正式的工作。

And I actually would recommend downloading the desktop version, especially if you want to be using this for any sort of serious work.

我推荐它的原因是，桌面版能提供更多的功能。

And the reason I recommend that is because there's just more capabilities that you can do.

所以我要给你演示一下。

So I'm gonna show you

我推荐桌面版，因为这样你的AI就不会藏在你那1200个打开的标签页中的一个里。

I recommend it because then your AI isn't hiding in one of the 1,200 tabs you have open.

它是一个独立的程序。

It's its own program.

是的。

Yeah.

所以我来给你举个例子。

So I'll give you an example here.

这是Claude的网页版本。

So this is the web version of Claude.

对吧？

Right?

你可以看到这是一个聊天窗口。

And you can see it's a a chat window.

它看起来和ChatGPT非常相似。

It looks very similar to ChatGPT.

侧边栏这里有几个不同的选项，但有趣的是，它没有桌面版的一些功能。

It has a couple different items on the side panel here, but then it's interestingly, it it doesn't have some of the same features that the desktop version does.

而桌面版（不展开太多细节）有一个叫做Cowork的工具，就在这里。

And the desktop version without getting too complicated has this tool called Cowork, which is where this lives.

它还有另一个叫做Code的工具，如果你听说过Claude写代码，我们在通讯中经常提到这个。

And then it has another tool called Code, which if you ever hear Claude code, we write about that a lot in our newsletter.

这里指的就是这个功能。

That's what this is referring to.

正如你在这里看到的，我本来想问Claude：使用MCP服务器安全吗？

And as you can see here, I was going to ask Claude, is this safe to use an MCP server?

不过对的，我们后面会讲到这个话题。

But yeah, we'll get into that.

今天就先不聊这个了。

That's not for today.

我们这次不讨论编程相关的内容。

We're not talking about coding.

只聊聊聊天功能，以及这个工具是怎么运作的。

We're just talking about chat and how this works.

所以这就是我推荐桌面版的原因。

So that's why I recommend the desktop.

现在ChatGPT也有桌面应用程序了。

Now ChatGPT also has a desktop app.

我目前没有登录那个账号，但它也有一个叫Codex的功能。

I'm not logged into that right now, but it also has one called Codex.

而我现在是登录着Codex的。

And I am logged into Codex.

我们稍后可能会在Codex中演示科里的网络行为。

And we might do some demos of Corey's internet is behaving inside Codex later.

但我将向你展示我在Claude中展示的相同内容。

But I'll show you the same thing I'm showing you in Claude.

我会在Codex中展示，这相当于ChatGPT的版本。

I will show you in Codex, which is sort of like ChatGPT's equivalent.

是的。

Yeah.

那么我们来谈谈这个叫做项目的东西。

So let's talk about this thing called projects.

所以我现在要共享我的屏幕。

So I'm gonna share my screen here.

把这缩小一点。

Shrink this a little bit.

好的。

Okay.

那么，什么是项目呢？

So what is a project?

项目就是你用来组织文件夹的文件夹。

Projects are folders that you are basically like organizing folders.

就像你会在Google Drive中整理所有文档一样，这是一种整理所有聊天记录的方式，但它强大得多，我会解释一下。

So in the same way that you would have a Google Drive where you organize all of your documents, this is a way of organizing all of your chats, but it's a lot more powerful and I'll explain Yeah.

它是如何运作的。

How that works.

科里，如果聊天中有任何关于这个的问题，请告诉我，我可以回答，因为我看不到。

Corey, also let me know if there's any questions in the chat related to this and I can answer that because I can't

我会在你讲解时留意聊天内容，先生。

see I'll my keep an eye on them while you're going, sir.

好的。

Okay.

太好了。

Great.

所以我认为，在开始任何工作或聊天之前，你应该先设置一个项目，因为这是整理这些杂乱无章的聊天记录的绝佳方式。

So the reason why I think you should set up a project before you do any sort of work or or chatting is because it's a really good way to keep this like really messy, like all chats log organized first and foremost.

但更重要的是，通过使用项目，你可以为自己创造成功的机会，而这是仅依赖单个聊天所无法实现的。

But also you can do you can set yourself up for success by using projects in a way that you can't if you just rely on a standalone chat.

那我们一起来设置一个测试项目吧。

So let's let's set up a test project together.

所有有想要解决的问题的人，请在聊天中提出这些问题。

Everybody who has a problem that you're trying to solve, go ahead and ask those questions in the chat.

比如，你希望在AI领域为自己设置什么样的项目？

Like, like, what type of project would you like to set up for yourself in AI?

你希望AI帮你完成哪些事情？

Like, what what do you wanna have that AI help you do?

给我们一个好例子。

Give us a good example.

格兰特会当场展示我们是怎么做的。

Grant will show you how we do it right there.

Copilot 中项目对应的术语是什么？

What is the Copilot equivalent of a project?

我本该知道这个的。

I should know this.

我觉得他们也有项目。

I think they have projects.

是的。

Yeah.

他们称之为项目。

They have called projects.

但让我先说一下

But let me let me

我以为它叫项目，但是

I was thinking it was called project, but

我可能错了。

I could be wrong.

它可能被称为文件夹点点。

It could be called folders Dot dot.

有可能。

Potentially.

你正在屏幕上查看吗？

Are you pulling that up on your screen?

点点点。

Dot dot dot.

是的。

Yeah.

我在看。

I am

在查看。

looking.

我们来看看。

Let's see.

天啊，今天电脑跟我过不去。

Man, the computer is not happy with me today.

我们会给你一个答复的。

We'll get you an answer to that.

我们会我们会。

We'll we'll we'll

我用的不是最好的电脑。

I'm not on, like, my best computer.

我用的是，嗯，这个。

I'm on, like Yeah.

我会确保把内容发到聊天窗口里。

I'll make sure you have it in chat.

聊天GPT的等价物也叫项目。

The chat GBT equivalent is also called projects.

那我来打开chat.com的那个。

So let me pull up the chat.com one.

所以项目就在侧边栏的顶部这里。

So projects is right here in the top in the sidebar corner.

你只要点击那个小项目图标，就可以开始一个项目了。

So you just hit the little project thing and you can start a project.

所以无论你是用这个还是用这个——它去哪儿了？

So whether you are using this or this where'd it go?

消失了。

Disappeared.

你好。

Hello.

好了，来了。

Here we go.

好的。

Okay.

实际上，我点到了。

Actually, I hit it.

好的。

Okay.

所以无论你是使用Claude应用中的项目功能，还是ChatGPT应用或桌面端，或者抱歉，是网页版。

So whether you're using projects in Claude on the app or ChatGPT on the app or on the desktop or sorry, on the web version.

就是这样操作的。

That's how you do it.

那我们来创建一个新项目吧。

So let's go ahead and make a new project.

科里，有人给我们提过想创建什么类型的项目吗？

Corey, did anyone give us any ideas for what type of project we wanna create?

如果没有，我就用之前那个求职的例子。

If not, I'll use the job search example from before.

我们看看。

Let's see.

项目是一样的。

Project's the same.

使用量化Excel。

Use quant Excel.

求职管理器，寻找企业，联系人员，定制联系人，采访职业转型者。

Job search manager, find businesses, find people, customize contacts, interviews of career pivoters.

我们之前提到了两个求职相关的点。

We've had two mentioned job search.

我们来看看。

Let's see that.

另一个提到的是线索开发。

Another another one mentioned lead gen.

对。

Yeah.

我快速查了一下求职相关的内容。

I I hit job search real quick.

我们就做求职项目吧。

Let's do job search.

所以我们接下来要做的，是把所有与求职相关的聊天记录整理到一个个人项目中。

So what we're gonna do is we're gonna organize all of our chats related to job searching into a personal project.

之前我们提到过AI效率的五个层级，我说过实际上建立一个项目属于第一级。

So before we mentioned the five levels of AI efficiency, and I said that actually setting up a project is level one.

因此，无论你想要完成什么任务，或者需要任何组织结构来帮助你实现使用AI的目标，我都建议你创建一个项目。

So any task that you're trying to do or any organizing structure to help you accomplish whatever you're trying to do with AI, I recommend making a project.

是的。

Yes.

我也是这么做的。

I do too.

我有一个通用项目，用于记录我们在The Neuron做的所有事情。

I have one that's like a general purpose project for everything we do at The Neuron.

基本上，我会点击创建项目，这里有几个关键部分需要注意。

And basically what I do is I'm going hit create project and there's a couple different components here to keep in mind.

Claude和ChatGPT的界面大致相同，但我们会展示它们之间的差异。

And this is pretty much the same between Claude and Chat, but we'll show the differences.

如果你要做超过两次，那就应该放进一个项目里。

If you're gonna do it more than twice, it belongs in a project.

是的。

Yeah.

对。

Yeah.

我觉得没错。

I think that's right.

而且，这样做也很有帮助。

And also, it's just helpful.

比如，你可以创建一个项目，我也可以拿这个当个例子。

Like, you can create a project, and maybe I'll do this as an example too.

我会创建一个叫‘格兰特的工作区’的项目。

I'll create a project that's just like Grant's work zone.

只是为了整理我所有的想法和项目。

And I'm just gonna say for organizing all my ideas and projects.

比如，这可能不是最高效的做法，但如果你只想有一个集中区域来存放所有东西，也可以这样操作。

Like, this is probably not the most efficient way to do it, but if you want just like one staging area to put all of your stuff in, you could do it that way as well.

让我来解释一下这里的各个组成部分。

So let me explain the different components here.

首先，你有一个聊天窗口，就像你访问 chat.com 并开始提问一样。

So first, you have a chat window, which is just like if you went to chat.com and you started asking questions.

但在这里，你还有这些东西。

But then you have some of this stuff over here.

你有记忆、指令和文件。

You have memory, you have instructions, and you have files.

你为什么要使用这些功能呢？

Why do you want to use any of this stuff?

这些功能对你有什么帮助？

Like, how is this helpful to you?

文件功能允许你上传文档、添加文本内容，或连接 Google Drive 或 GitHub 仓库到你的数据库中，这样每次在这个项目中开启聊天时，系统都能访问这些信息并从中学习。

Files allows you to either upload documents, add text content, add Google Drive or GitHub repos to your database so that anytime you start a chat in this project, it will be able to access that information and learn from it.

指令的作用基本上和为项目中的所有聊天设置一组固定提示一样。

Instructions basically works the same way as having a set prompt that you apply to all chats in a project.

所以，让我们回到我们的求职管理器。

So basically, let's go back to our job search manager.

那么，我们想给求职管理器哪些指令呢？

So what are some instructions that we'd want to give our job search manager?

我们可以做的一件事是，把我们的简历上传到这个项目中。

One thing that we could do is we could give we could upload our resume to the project.

这样，每次你开始一个新的聊天时，你所对话的AI都会知道你的简历，并能参考它。

So anytime you start a new chat, your this AI that you're talking to will know your resume and it will be able to reference it.

我通常的做法是，你可以直接从你的设备上传，或者直接添加文本并粘贴进去。

And the way that I would do this is I would, you can upload it directly from your device or you could just add text and paste it in.

你可以写上，比如‘当前简历，4月2日’。

And you could say like current resume, April 2.

然后把简历内容粘贴在这里。

And then resume content pasted here.

对吧？

Right?

然后你可以说：帮我根据这份职位描述更新我项目知识中的简历。

Then you could say, help me update my resume in your project knowledge for this job description.

然后你可以把职位描述粘贴到下面。

And then you can actually paste the job description in down below.

如果你想要非常具体和精确，可以使用所谓的 XML 标签。

If you want to be really, really specific and really precise, you can use what's called XML tags.

这表示：请根据我项目知识中的当前简历（4月2日）来更新这份职位描述。

And this just says like, help me update my resume in current resume April 2 in your project knowledge for this job description.

你之所以想用这种方式，是因为假设你在项目知识中上传了多个简历。

And that the reason you would want to use that is because let's say you have multiple resumes uploaded in the file in the project knowledge over here.

你希望它参考的是这个最新版本。

You want it to reference this specific one that's the most current.

但你可能不想删除其他所有简历，因为保留它们或许会有帮助。

But maybe you don't wanna delete all those other ones because maybe it's helpful to have them.

你知道，也许你有不同的简历类型。

Know, maybe you have different types of resumes.

也许你会根据申请的不同职位类型来制作简历，比如你有两种不同的技能组合，有些过往经历对这个职位更相关，而其他的则更适用于其他职位，等等。

Maybe you create your resume a certain way for a certain type of job that you go out for and you create it differently for another type of job you got for because different let's say you have two sets of skills and maybe your some jobs and previous positions were more relevant for that past for this role and other ones are more relevant, etcetera.

所以，你使用XML标签的原因是为了直接引导系统关注某个特定文件。

So that's why you would use XML tags is if you wanna direct its attention directly to a specific file.

这样理解吗，科里？

Does that make sense, Corey?

我有没有漏掉什么？

Did I miss anything there?

是的。

Yeah.

这说得通。

That makes sense.

这说得通。

That makes sense.

太好了。

Cool.

现在，项目说明可能是：在撰写求职信之前，始终使用网络搜索来查询当前的日期和时间。

Now project instructions would be like, always use web search to look up the current date and time before drafting a cover letter.

这和我在ChatGPT中使用它们的方式完全不同。

This is so different to how I use them in ChatGPT.

好的。

Okay.

我们也得听听你的看法。

We gotta get your take as well.

但我也想说，你是一个简历撰写专家，你的工作是帮助我更新简历，并为我申请的具体职位撰写求职信。

But I would also say you are a resume builder and your job is to help me update my resume and write cover letters for specific for specific jobs I'm going out for.

现在，假设你甚至不知道这里该写什么。

Now let's say you don't even know what to write here.

你可以做的一件事是说：嘿，我正在尝试建立一个简历撰写或求职管理项目，我需要编写一些自定义指令，让我的代理完全按照我的要求行事。

One thing you could do is you could say, hey, I'm trying to set up a resume build or job search manager manager project, and I need to write some custom instructions to get my agent to do things exactly the way I want.

你能帮我根据 Claude 目前的最佳实践和文档，写出一套最优秀的自定义指令吗？

Can you help me write, the best set of custom instructions possible based on, Claude's current best practice, documentation.

那么，我们可以让 AI 实际上查找并为我们生成这些自定义指令。

Then what we can do here is we can have the AI actually look up and write the custom instructions for us.

它正在搜索网络，进行思考，并从 Claude 的最佳实践中获取信息。

And what it's doing is it's searching the web and it's thinking and it's fetching the information from Claude's best practices.

它现在会更新这些指令，以便在你任何聊天或项目文件夹内需要完成的任务中都应用这些规则。

And it's gonna now update them so that that applies whenever you have any chat or anything, any tasks that you need to do inside the project folder.

科里，这和你平时用它的方式有什么不同？

Corey, how is this different from what you use it for?

其实差别不大。

Well, it's not that different.

你一上来就说‘一定要这样做的时候’，让我有点措手不及。

You threw me off when you went in and started at always do it this way.

我当时想，等等，你这不是在给它一个系统提示吗？

I was like, wait, you're not giving it like a a system prompt.

我当时想，哦。

I was like, oh.

系统提示本质上就是一个案例，因为我们之前没解释过，而且我也不打算深入讲。

A system prompt would basically be a case because we haven't explained that before, and and I'm not gonna get into it too deep.

但你可以把它想象成每个AI都基于一个核心提示运行。

But think of it as every AI runs on one core prompt.

比如ChatGPT、Claude，它们都有一个主系统提示，以及一些其他关联的提示。

So, like, ChatGPT, Claude, they all have a single system prompt that and a variety of others they connect to.

但你可以把它理解为一个单一的系统提示，告诉它：你是一个乐于助人的助手。

But think of it as a single system prompt that tells it you are a helpful assistant.

你的任务是为用户提供他们需要的信息。

Your job is to provide users with the information they need.

你要安全地完成这项工作。

You're gonna do this safely.

你不会给它们危险的建议，诸如此类的东西。

You're not going to give them dangerous advice, that kind of stuff.

这就是系统提示的作用。

That's what a system prompt does.

本质上是对的。

And essentially Yeah.

在项目中，你是在给它一个专属的系统提示。

With a project, you're giving it its own little system prompt.

绝对要照他说的做，让它帮你写出来。

And absolutely do what he said and have it write it for you.

这提供了很好的背景信息。

So that's great context.

而且这一点是我之前忽略的，因为我每天都接触这些，没想过要把系统提示当作一种说明方式，但其实它就是这么回事。

So and that's something that, you know, I missed because I'm, you know, in this every day and I and I don't think about, you know, using the system's prompt as explaining it, but that's basically what it is.

当你设置自定义指令时，就是在使用一个系统提示，帮助这个项目尽可能直接地协助你。

When you're setting up custom instructions, it's a system prompt to help this project basically direct directly help you as much as possible.

关于这一点，我想补充的是，系统提示可以像你所需的一样宽泛或具体。

So one of the points about that, that I would add is they can be as broad or as specific as you need.

我以前有一个项目文件夹，专门用来删减字数，你们可以看到这里。

I used to have previous, you can see over here, I used to have a project folder that was just for cutting word counts.

我还有另一个文件夹，用来帮我撰写‘环形bullet’和‘智能洞察bullet’，我们每天都会写这些内容。

I had another one that was for helping me write the around the horn bullets and the intelligent insight bullets that we would write every day.

我还有一个专门写主题行的文件夹，不过这个我们会在第三级再讲。

I had one for writing subject lines, but which we're gonna get to in level three.

实际上，我现在正用技能来取代这些文件夹，这一点我们稍后会讲到。

That's actually something that I am now replacing with skills, which we'll get to that.

我想在这里评论一点：我不认为ChatGPT能做到这一点，但Claude可以。

One thing I wanna comment here, don't I believe ChatGPT does this, but Claude does.

所以它正在向我提出一系列后续问题。

So it's asking me a series of follow-up questions.

你看到这里了吗？

Do you see this here?

那么这个代理会在哪里运行呢？

So where will this agent run?

是的

Yep.

这会影响我们如何构建指令。

This affects how we structure the instructions.

所以我会说，Claude 项目。

So I'll say, Claude projects.

哦，它也会这样。

Oh, it does this too.

ChatGPT 也会吗？

ChatGPT does?

很好。

Great.

对

Yeah.

只在特定情况下，比如它不会在每个查询中都这样做，但如果你问的是类似这样的问题，是的。

Only on specific things like it doesn't do it on every query, but if you're asking for something like this, yes.

代理人主要应该在哪些方面帮助你？

Where should the agent primarily help you with?

寻找和跟踪工作？

Finding and tracking job?

我们就说以上所有方面吧。

Let's just say all of the above.

因为我觉得我们希望这个功能涵盖范围广一些。

Cause I think we want this to be broad.

什么最重要，以及它如何行为？

What matters most and how it behaves?

不询问就采取行动，在重大步骤前询问，遵循严格的工作流程。

Takes action without asking, asks before major steps, follows a strict workflow.

我认为在重大步骤前询问是最关键的。

I would say asks before major steps is like the really important one.

好的。

Okay.

我看到聊天里有一些问题。

So I'm seeing some questions in the chat.

科里，你在回答这些问题吗？

Corey, are you answering those?

我在回答几个。

I'm answering a few.

你能不能花一分钟快速挑几个上来？

Do you wanna take a minute and bring a couple up real fast?

好的。

Yeah.

你能作为主持人把它们调出来吗？

Can you bring up them on your side as the host?

我从错误的一侧在看。

I'm I'm looking at it from the wrong side.

是的。

Yes.

我可以。

I can.

我在这边看的是YouTube上的评论，而不是公共聊天窗口里的。

I'm over here looking at them in YouTube instead of looking at them in the public chat.

是的。

Yeah.

对。

Yeah.

我其实也经常这样。

I've I do that too, actually.

我就在这儿一直点，但评论就是不显示在屏幕上，我搞不懂怎么回事。

I'm like I'm like, I'm over here clicking and they're not popping up on screen and I can't figure it out.

是的。

Yeah.

没错。

Right.

好的。

Alright.

任何与项目相关的问题都可以提。

And any any and all questions related to projects.

因为我觉得，如果你打算在应用中认真使用AI，理解项目是非常重要的。

Because I think projects are, like, really important to understand if you're gonna be doing anything serious with AI inside the applications.

所以尽管全部问出来吧。

So ask them all the way.

公共聊天消息看不到。

Public chat is not coming through.

我这里看不到，无法调出来。

I can't see it in here to bring them up.

有意思。

Interesting.

这非常奇怪。

It's very, very strange.

我以前从没遇到过这种情况。

I've never seen this happen before.

那我先开始了。

Well, I'll go ahead

然后把它们大声读出来。

and read them out loud.

所以。

So

好的。

Okay.

是的。

Yeah.

我也有几个问题可以读一下，如果你想的话，可以顺便跟上。

Let's I have a couple I could read too if you wanna if you wanna just catch

你为什么不读问题呢？我来回答，然后你再给出你的看法。

Why don't you why don't you read the questions and I'll respond and then you can give me your answer

也一样。

as well.

好的。

Do okay.

所以我会搞错大家的名字。

So from I'm gonna mess up people's names.

所以请提前原谅我。

So please please accept my apology in advance.

来自Kartik Suri的问题：在为编程项目创建多个脚本文件时，项目会保留上下文吗？

From Kartik Suri, does do projects retain context as we build out several script files for a coding project?

是的。

Yes.

不过并不完美。

Now it's not perfect.

对于这个特定的使用场景，你最好使用Claude Code。

Like for that specific use case, you'd wanna use Claude code.

是的。

Yep.

你希望使用你本地的代码仓库来存放你编写的所有代码。

And you would want to use your local like repo for where you're writing all the code.

那就是你应当使用的工具，也是最适合这个场景的。

That's where you would want, that's what you'd want to use for that.

但没错，如果你希望在聊天窗口中完成，它也能做到。

But yes, if you're, if you wanted to do it in your chat window, it could do that.

基本上，它的实现方式是你可以问它：‘我们在这个聊天中上一次对话内容是什么？’

And basically the way that it does that is you can ask it, hey, what was the last conversation we had in this chat?

它实际上可以读取项目内的之前聊天记录。

And it can actually like read the the previous chats inside the project.

有人在这里提到了一点，我想快速说明一下。

Something someone pointed out here, I wanna call out real fast.

哦，抱歉。

Oh, I'm sorry.

嗯嗯。

Mhmm.

说吧。

Go ahead.

我看到有几个人问到，是否可以把项目用作归档。

Is that, you know, I've seen a couple questions about using it using projects as an archive as well.

是的。

Like Yeah.

你可以这么做。

Where, like, you know, you can do that.

有人提到，你还可以把旧的聊天记录移到项目中。

And someone pointed out that you can also grab old chats and move them to your project.

这很关键。

That's key.

所以这就是

So That's that's