从程序员到管理者：与Notion联合创始人西蒙·拉斯一起探索代理式工程的转型

大家好，听众朋友们。

Hi, listeners.

欢迎回到NoPriars。

Welcome back to NoPriars.

今天，我在这里与Notion的联合创始人西蒙·拉斯特对话。

Today, I'm here with Simon Last, co founder at Notion.

我们讨论了Notion在人工智能时代的新愿景——作为人类与智能代理协作的平台，Notion工程和产品团队正在发生的变化，以及这些全新的思维工具。

We talk about their new vision for Notion in the AI age as a platform for humans and agents to collaborate, how the engineering and product org at Notion is changing, and these new tools for thought.

欢迎你，西蒙。

Welcome, Simon.

嘿，西蒙。

Hey, Simon.

谢谢你参与这次对话。

Thanks for doing this.

嘿。

Hey.

当然。

Of course.

是的。

Yeah.

能来这里真的很有趣。

It's really fun to be here.

Notion 已经达到相当规模，是一个出色的平台，拥有大量用户。

Notion's at scale, amazing platform, lots of users.

你们很早就开始了。

You did start quite a while ago.

我觉得 Notion 是一家真正积极拥抱人工智能的公司。

I think of Notion as one of the companies that has really, like, braced AI quite aggressively.

我听说你们第一次接触 GPT-4 是在墨西哥的一次公司外出活动中。

I was told you first got your hands on GPT four at a company off-site in Mexico.

这是真的吗？

Is that true?

你最初开始做这些东西的起源故事是什么？

What is the origin story of, like, starting to work on this stuff?

是的。

Yeah.

我觉得是的。

I think yeah.

那一年，就是2022年。

That year, that was 2022.

我一直关注着，你知道的，发生了什么。

I I've been watching, you know, what's going on.

总的来说，我一直对这项技术充满好奇，热衷于尝试一切，想知道我们如何应用它。

In general, I've just been, like, super curious about the technology and fascinated to to try everything and about, like, like, how we can apply it.

直到我开始使用GPT-4，它才真正变得非常真实。

It wasn't until I played with GPT-four that it it became really, really real.

所以，我们获得了它的访问权限。

So, you know, we and we got access to it.

它有点像一个早期的ChatGPT界面。

It it was sort of like a a proto ChetGPT like interface.

我和我的联合创始人伊万都获得了访问权限，当时立刻就明确了两点。

And my cofounder Ivan and I both both got access, and it was just immediately clear, like, I would say two big things.

第一，它相当聪明。

One is that it was just pretty smart.

它能够理解相当复杂的指令。

It it it could follow reasonably complicated instructions.

它可以为你撰写内容。

It could write things for you.

它可以编辑内容。

It could edit things.

第二点是它的知识范围非常令人感兴趣。

And and the second big thing was that the scope of its knowledge was extremely interesting.

极其深厚且广泛的世界知识。

Super, super deep, like and and broad world knowledge.

当我们使用它时，我们两个人立刻就明白了。

When we played with it, it became just instantly clear to both of us.

好吧。

Like, okay.

现在是时候开始了，但我们还在思考如何应用这项技术。

The the time is now to start, but we're thinking about how to apply this.

它只会变得越来越好。

It's only gonna get better.

我们刚才在聊墨西哥和GPT-4。

We were talking about Mexico, GPT four.

你们也觉得，很明显，时机已经到了。

You guys saw it was, like, clearly the time.

你们一开始是有一个明确的愿景，觉得AI和Notion结合后应该能做什么，还是先从不同团队拉人，或者招聘新人，然后说‘我们来试试看’？

Did you start with, like, a particular vision of, like, what you should obviously be able to do with AI and Notion, or did you start pulling people from different teams or recruiting people and say, like, let's experiment?

你们是怎么开始的？

How did you begin?

我觉得我们立刻就有了长期和短期的愿景。

I think we immediately had a long term and a short term vision.

我会先说说短期的那一个。

I would say the the I'll start with the short term one.

很明显的一点是，它可以像个写作助手。

The the thing that was immediately obvious was, oh, it could be like a writing assistant.

嗯。

Mhmm.

所以它可以集成在你的文档里。

So it it could be in your document.

你可以选中一些文字，让它重写。

You can, like, select some text, have it rewrite it.

它可以为你生成文字，或许还能查找信息，然后给你提供来源或更多资料。

Can have it write text for you, maybe look something up, and then, you know, give you, like like, sources or more information.

这就是我们立刻着手去做的方向。

So that was the thing that we immediately, like like, got to work on.

然后，我们组建了一个专门小组来推进这个项目，之后大约两三个月就推出了它。

And, you know, we sort of started a tiger team around it, and then we were able to launch it in, like, two, three months after that.

而我们当时立即想到的长期愿景是，或许可以打造一个更通用的助手。

And then the long term vision that we immediately had was like, oh, the thing that looks like it may be possible is more of like a general assistant.

如果能赋予它Notion中人类所能使用的所有工具，让它能够创建自己的数据库、查询和操作它们、创建和编辑文档，并将这些功能整合起来，完成更长期的任务，会怎么样呢？

So what if you could just give it all the tools inside Notion that a human would have, be able to, like, create its own databases, query, manipulate them, create documents, edit them, and sort of weave all these things together to do, like, a longer range task.

所以我们一开始就同时着手这两方面的工作。

And so we we sort of immediately started on both.

短期的那个我们很快就实现了，但长期的那个还没能成功。

The the short term one, we're able to shoot very quickly, and then the long term one didn't really work yet.

所以它花了更长的时间才真正运行起来。

And so that took much longer to get working.

AI专属的Notion功能和产品最初是什么时候发布的？

Are there, like, specific first launch of the AI specific Notion features and products was when?

去年。

Last year.

不。

No.

我们在2023年2月发布的。

It was it was February 2023 is when we had launched.

是的。

Yeah.

我的时间线搞错了。

My timelines are wrong.

从开始发布以来，有没有一些特别值得注意的教训或突破性时刻？

Are there, like, a few specific learnings or breakthrough moments you think since beginning to release that are interesting?

对。

Yeah.

我的意思是，这几年来，这一直是一段漫长而艰难的过程，

I mean, there's been it's it's it's been a slog over many years or over multiple years at this point with

是的。

with Yeah.

有很多很多的经验教训。

With with many, many learnings.

我想说，是的。

I I would say yeah.

我来给你梳理一下我们发布产品的整个过程：首先，我们推出了我们的写作系统。

I mean, just to give you a timeline of the arc of what what we shipped is, you know so the first thing was our our writing system.

我们称之为AI写作器。

We called it AI Writer.

那是我们推出的第一个产品。

That's the first thing we launched.

它最容易实现，因为这是一个单步骤的任务——重写和编辑文本。

It was easiest to get working because it's, like, single step task, rewriting, editing text.

它没有检索方面的复杂性。

There's no, like, retrieval aspect.

只是直接访问模型来生成文本。

It was just, like, raw access to the model to write to write the text.

接下来我们立即着手的下一个重大项目是一个问答系统，对整个工作空间进行语义索引，然后让你提出问题，系统能基于来源给出答案。

The next the next big thing that that we immediately started working on was a a q and a, doing a semantic index of the entire workspace and then letting you ask a question, and it can give you an answer that's that's grounded in the sources.

对我们来说，这显然也会非常有用。

That was also immediately obvious to us that that'd be super useful.

所以我们开始着手开发这个功能。

So we started work on that.

我们是在2023年10月左右推出的。

That one we launched in, I think, was October 2023.

在此之前我们已经启动了公测，但正式发布是在10月。

So we started a beta before then, but then that would our our GA was in October.

这显然是一项大得多的工程。

That was just a a much bigger effort to get working, obviously.

我们不只是简单地接入大语言模型。

We weren't just, like, like, plugging in LLM.

而是真正实现了实时更新的索引系统。

Was actually doing this, like, real time updating index.

对。

Right.

我们也必须更加认真地对待评估和质量方面的问题。

Had to get much more serious about the the evals and the quality there as well.

问答功能的开发是一段多年的旅程。

The the q and a has been a multiyear journey.

基本上，一旦我们的Notion索引工作起来，我们就立刻意识到：哦，明白了。

Basically, what we did is immediate as soon as we got the Notion index working, it it was obvious that, oh, okay.

我们也应该索引其他所有内容。

We should index everything else as well.

于是我们索引了Slack和Google Drive，并且定期推出新的索引功能。

And and so we index, like, Slack and Google Drive, and we're launching new ones on a on a regular cadence.

现在，我认为我们已经实现了相当完整的覆盖。

And and and now we have a, I would say, fairly complete

有人可能会说，这些问题非常困难，因为这些产品本身至今也没有完美解决。

One could argue that those are, like, very difficult problems that, you know, those products natively have not solved perfectly yet.

那么，你是怎么考虑接手这个问题的？

So how did you think about taking that on?

我不知道这会不会让其他产品团队觉得冒犯，但说实话，目前还没做好。

I don't know if that's, like, an offensive thing to other product teams, but, like, it's not working yet.

是的。

Yeah.

这确实是真的。

It's it's kind of true.

是的。

Yeah.

我们经常讨论这个问题，因为这就像在想：我们到底有什么权利来做这件事？

This has been something we talk about a lot because it's like, you know, it's like almost like, what right do we even have to do this?

但事实是，大多数公司在构建索引方面都做得相当差。

But but it turns out that most of the companies are pretty bad at making their indexes somehow.

说实话，这让我们有点惊讶。

It's honestly kind of baffled us a little bit.

对。

Right.

但我觉得，在经历了这一切，并与团队合作尝试让它运行起来之后，我的看法是，有一点AI方面的敏锐度非常重要。

But I I think my take after dealing with all of this and, you know, working with the teams to try to get it working is there's a little bit of just AI pilled savviness that's pretty important.

然后我认为，大部分工作其实只是需要一些技巧和对细节的关注。

And then I'll and then I think most of it is honestly just, like, a bit of, like, like, craft and attention to detail.

我认为，特别是对于这种索引检索工作，要想真正让它运行起来，你必须非常实证、迭代，真正去尝试各种查询。

I think, like, in in in particular with this, like, indexing retrieval stuff, in order to really get it working, you you have to be quite empirical and iterative and actually be, like, like, trying queries.

每个数据源都有点独特。

Like, you you'll like, each each data source is a little bit special.

你知道，你不能用一种万能方法来处理获取Slack和获取Google Drive这样的事情。

Like, you know, you can't just apply a one size fits all to, like, acquiring Slack versus acquiring Google Drive, let's say.

它们是完全不同的信息类型。

They're they're completely different kinds of information.

我们发现，真正要让检索系统有效，需要投入一些技巧和用心：实际尝试大量不同的查询，每天使用它，并不断迭代、重新思考和调整检索方式。

And we found that there's just a little bit of, like, like, craft and love that's going to it in terms of, like, actually trying a bunch of different queries, actually using it every day, and constantly iterating and rethinking and tuning how the retrieval works.

你是怎么考虑人们组织工作空间的多样性呢？我的意思是，即使是Notion，其使用方式也不是统一的。

How did you think about the diversity of how people organize their workspaces and just their I mean, even Notion is not use of it is not homogenous.

对吧？

Right?

作为投资者，我可能参与了15个不同的工作空间。

Like, I'm probably part of 15 workspaces as an investor.

所以我观察它们时会想，我的 workspace 一团糟，而这些人却非常有条理，他们的工作流程完全体现在Notion的使用方式上。

And so I look at them and I'm like, well, mine's a mess and these people are really organized and the workflow is reflected in how their Notion works.

是的。

Yeah.

完全正确。

Totally.

我想说的是，有了嵌入技术，其实这种差异的重要性已经没那么大了，嗯。

I would say I mean, the interesting thing is that with embeddings, it almost doesn't matter as much Mhmm.

现在更是如此。

Anymore.

AI其实并不关心树状结构是什么样的。

The the AI doesn't really care what the what the the tree structure is, for example.

AI只关心有没有一段包含你需要的上下文的文本，然后它就能检索出来。

The all all the AI cares about is that there's a snippet of text that has the the context you need, and then it can retrieve it.

所以我们现在反而建议人们，别太纠结于组织结构。

And so, actually, we kind of advise people now, like, don't worry as much about organization.

只要找到一种方法，把所有内容都导入进去就行。

Just just just find a way to get it all piped in and, like, like, thrown in there.

但你仍然会做出一些可能显著影响性能的决定，比如分块策略。

You still make decisions that could change performance quite a bit, like chunking strategy

或者别的什么。

or whatever.

是的。

Yeah.

当然。

Absolutely.

是的。

Yeah.

这非常重要。

That's super important.

但这一点对用户来说是透明的，与他们组织内容的具体方式无关。

But but that's sort of not that's sort of transparent to the user and sort of independent of their their particular method of organizing things.

嗯。

Mhmm.

考虑到内容库的差异如此之大，这似乎仍然是一个艰巨的技术挑战。

It just seems like still a difficult technical challenge given how different the content bases are.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

我觉得是的。

I think yeah.

这花了大量的迭代。

That took a lot of iteration.

是的。

Yeah.

分块大小、检索的工作方式，以及检索流程中的不同步骤。

The chunk sizing, how retrieval works, the different, like, steps in the pipeline of retrieval.

是的。

Yeah.

这方面有很多迭代。

There's a lot of iteration on that.

伊万说，我应该问问你重建成Notion和重做你的框架有多少次了。

Ivan said I should ask you how many times you've rebuilt Notion and rebuilt your harnesses.

是的。

Yeah.

是的。

Yeah.

这几乎成了一个老生常谈的笑话。

It's kind of a running joke almost.

我的意思是，我们大概每六个月就会重写一次我们的AI框架，而且重写所需的时间也在逐渐减少，因为进展一直在加速。

I mean, we we we rewrite our AI harness probably every six months or so, and and the time to rewrite has kind of been been decreasing just because, I mean, like, like, progress has been accelerating.

我认为这真的是一个非常关键的点，很多公司都搞错了，就是只做一件事，然后死守着不放。

I think this is honestly a really key thing and something that a lot of companies get wrong is just, like, doing one thing and then just, like like, sticking with it.

你必须时刻关注模型和当前技术的状态，然后围绕这些来深度设计框架、系统和产品。

You really do have to keenly aware of what the current state of the model is and the technology is and then designing the harness, the system, and the product deeply around that.

这基本上意味着你每六个月就得重写一次。

And it basically means you have to rewrite it every six months.

我觉得这挺有意思的。

And I find it pretty fun.

这是过程的一部分。

It's part of the process.

你知道，你有机会重新开始，重新思考它。

You know, you get to you get to restart and and and and rethink it.

我们正在开发一个新版本的框架，预计下周或下下周就要发布了。

You know, we're working on we're about to release a new version of our harness, like, in the next week or two.

然后我们已经在考虑下一个版本了。

And then and then we're already thinking about the one after that as well.

我觉得这引出了我想问你的几个问题，关于现在有了编码代理之后，Notion 在工程、产品和研究方面的运作方式是怎样的？

I I think that leads to a a set of questions I had for you on just, like, how does Notion as an engineering and product and research organization work now that you have the power of coding agents as well?

因为我觉得，如果你知道代理能帮你完成这些工作，你重写框架的意愿会大大提高。

Because I imagine, like, your willingness to rewrite the harness goes up dramatically if you're like, agents are gonna help me do it.

是的。

Yeah.

这非常正确。

That's extremely true.

是的。

Yeah.

我的意思是，是的。

I mean, yeah.

使用编码代理一直非常有趣。

It's been it's been really fun to use the coding agents.

我觉得我所考虑构建的东西的野心已经大大提升了。

I think the ambition of what I even consider building has gone a lot.

在过去两三年里，你认为Notion在工程和产品运作方式上最显著的变化是什么？

What do you think has most dramatically changed in how you think about how engineering and product should work at Notion over the last two, three years?

是的。

Yeah.

我的意思是，它确实经历了多次变化。

I mean, it's definitely changed multiple times.

我的意思是，就编码代理而言，我们经历了多个阶段。

I mean, in terms of the coding agents, we kinda went through multiple eras.

最初是标签自动补全阶段，然后我们开始进入插入和重写部分代码的阶段。

There was kinda like the tab autocomplete era, and then we and then we got into sort of inserting, rewriting some code.

但直到这些代理开始工作，情况才真正发生变化。

But but it wasn't really until the the agents started working.

我想说，去年年初我们开始采用这些代理。

I I would say, like, early last year we started to adopt the agents.

我想我大概在四月份开始使用Cloud Code。

Like, I started using Cloud Code, I think, around April.

那是一个巨大的突破。

That was a huge unlock.

我认为这里最大的转变是，你可以真正推动这些代理实现端到端的开发、验证和维护，但这需要对架构设计和验证循环进行深入的思考。

Like, I would say the the the big shift there is that, you know, you can really push on getting these agents to end to end, you know, implement and and verify and maintain stuff, but it but it requires pretty significant thought in terms of how you architect things and what is the verification loop.

但好处是，如果你做得好，你就可以对所构建的内容更有雄心，也能比人类编写时更稳健。

But but but the upshot is I think if you do it well, you can be much more ambitious about what you're building and also make it much more robust than you could have done with with with humans writing it.

而另一方面，如果你做得不好，那就全是垃圾。

And then the flip side is if you do it badly, it's all slop.

这是否改变了你对Notion团队应如何构成的看法？

Does that change your lens of, like, what teams should look like at Notion?

比如规模、资历之类的吗？

Like size, seniority, anything like that?

是的。

Yeah.

我的意思是，根本的影响在于，每个人在产出上的个人影响力可以大大提高，而你的产出越来越取决于你使用这些工具的能力和意愿。

I mean, I would say I mean, the fundamental effect is that, you know, everyone's individual impact in terms of their output can be much higher, and your output increasingly depends on your ability and willingness to use the tools.

我认为这就是正在发生的一切的根本所在。

I I think that's the fundamental thing that's happening.

那么，这会如何体现呢？

And then, like like, how does that play out?

我认为我们在团队层面其实还没看到太大的影响。

I think I don't think we've seen that much impact on the the team sides, really.

我们通常喜欢组成小规模的特遣团队。

I think we we like to work in, like, a smallish tiger teams for the most part.

我认为如果能让团队变小，几乎总是更好的。

I think if you can make teams small, it's almost always better.

以前就是这样，我认为现在仍然是这样。

That was true before, and I think it's still true.

也许稍微更明显了一点，但也没那么多。

Maybe increasingly a little bit, but but but not that much.

我觉得，没错，关键就是真正地利用好这些工具。

I I think, yeah, the main thing is to just, like, like, really harness the tools.

你认为在组织中，普通工程师和十倍工程师，或者更愿意使用工具的工程师，会有什么不同吗？

Do you think something different happens to the median engineer in an organization versus the 10 x engineer or the engineer 10 x more willing to use the tools?

是的。

Yeah.

我觉得差距更大了。

I think the the gap is bigger.

如果你现在善用工具，你完全可以成为一百倍甚至一千倍的工程师。

You can be, like, a 100 or a thousand x engineer if you use tools right now.

我觉得差距要大得多。

I think I think the the gap is much bigger.

就像，最低门槛没变，但最高上限已经极大地提高了。

Like, the the the minimum bar has not changed, but the maximum bar has has extremely increased.

它在内部的一个影响是，总的来说，事情感觉有点更混乱、更无序，但我其实挺喜欢这种感觉的。

One impact it's had internally, I would say, is, like, broadly, things feel, like, a little bit more messy and chaotic, I would say, like but I kinda love that.

我的意思是，现在出现了更多原型，多得多了。

I mean, it's like there's there's more proto there's way more prototypes.

你知道，比如我们的设计团队就创建了一个完整的 Git 仓库。

You know, people are like, for example, our, design team made an made an entire, Git repo.

他们称之为设计游乐场。

They call it the design playground.

它本质上就像一个简化版的 Notion，带有一堆 UI 基础组件。

And it's essentially like a simplified Notion with a bunch of, like, UI primitives

嗯。

Uh-huh.

里面。

In it.

而且他们把它做得非常复杂。

And they've made it, like, really sophisticated.

你知道吗？

You know?

它里面甚至有个智能代理，真的很酷，因为这让所有设计师都能快速创建出超高保真的原型。

It it it has, like, an agent in there and, like and it's it's pretty cool because it allows them all the designers can can spin up, like, super high fidelity prototypes Okay.

非常快。

Really quickly.

所以现在不再只是指着一个线框图说，‘这个看起来会怎么样？’

And so it's no longer, like like, pointing at a mock and being like like, you know, like like, how will this look like?

他们会给你一个已经部署好的原型链接。

They'll give you, like, a URL to a prototype that's that's that's been deployed.

这种做法在整个技术栈中都很普遍，所有工程团队都变得稍微更混乱、更多事情在发生。

And that sort of thing is true all the way up and down the stack, you know, for all of engineering, just like a little bit more chaotic, more stuff happening.

所有的拉取请求都更加有野心。

All the PRs are more ambitious.

你会在某些更危险或更敏感的事情上划一条线吗？

Do you draw a line somewhere about, like, stuff that is more dangerous to touch or sensitive?

比如，这里可能存在数据丢失的风险，还是没有？

Like, there's could be risk of data loss over here or and not?

还是说你认为所有事情都属于可以尝试的范围？

Or is it kind of you look at it all as it's fair game?

我们仍然会对所有拉取请求进行审查。

We still do reviews on all the pull requests.

而且我想说，现在所有的拉取请求都是由代理生成的。

And and I would say and, you know, all the pull requests are now written by agents.

它们通常更大、更复杂。

They're often, like, larger and and and more complex.

这大概是最大的问题。

That's, like, the worst part.

但更好的一面是，它们通常经过了更好的测试，我们可以对那些值得重视的事项要求更严格的测试。

But the better part is that they're often, like, a much better tested, and we can demand sort of a much better testing for the things that merit it.

我现在不会再提交任何未经完整端到端测试的拉取请求了。

I never produce a PR that, like, hasn't been, like, fully end to end tested anymore.

因此，你可以相当有信心地认为它能正常工作，但这要求你不能只是凭感觉编码，随便说说想要什么。

And so it's like you can get to a pretty high degree of confidence that that it works, but it requires, like you're not just vibe coding by by saying the thing you want.

你需要认真思考：我究竟想做什么？我要做出什么样的更改？

You're sort of thinking carefully about, like, what is the thing I'm like, what is the change I'm trying to make?

以及如何验证它？

And, like and and how can it be verified?

如何安全地部署它？

And how can it be deployed safely?

然后借助智能代理来协助你完成这个过程。

And then enlisting the agent to help you with that process.

当你想到你提到的通用助手还不存在时，你想象Notion的智能代理在未来一两年内还能实现哪些目前尚未突破的功能？

When you think about where you said the general assistant, like, doesn't quite exist yet, what's the what do you imagine Notion's agent agents being able to do, like, over the next year or two that are still unblocked?

这些功能目前要么受限于能力，要么受限于你的基础架构工作。

They're still blocked by either capability or your harness work.

我们花了好几年时间试图构建一个智能代理。

We struggled for a few years to build an agent.

你知道，它总是有点用，但其实没什么太大用处，主要是因为时机太早了。

And, you know, it always, like, like, sort of works, but then, you know, wasn't that useful largely just so it was too early.

所以，我们其实尝试了三四次来构建这个智能代理。

So we, you know, we we tried to build an agent, I would say, actually three or four times.

然后我们终于在去年秋天发布了它，也就是去年八月或九月。

And then we finally launched it last fall, so, like, last August, September.

所以，如果你现在使用Notion AI，它就是一个完整的智能代理，几乎可以访问Notion中的所有内容。

So the if you use Notion AI now, it's, the full agent that has access to everything in Notion pretty much.

这个功能完全能正常运行。

So that that that totally works.

我觉得，我们最初构想的很多功能现在都完全实现了，而且已经正式上线了。

I would say, like, the a lot of the original vision that we had totally works now, and it you know, it's, like, fully shipped.

去年八月或九月，我们发布了我们的个人智能代理。

Last or September, we shipped our personal agent.

因此，Notion 中的几乎所有用户都拥有一个智能代理，它基本上可以访问用户所能访问的所有内容。

So it's pretty much every user in Notion has an agent, and it basically it has access to all all all the things that the user has access to.

正如你所知，它可以为你创建数据库。

As you know, it can create a database for you.

它可以更新内容和文档。

It can update things, documents.

它可以搜索网页，进行研究。

It can search search the web, do research.

而我们上周刚刚推出的是第二个重大功能：自定义代理。

And then the second big thing that we just launched last week actually was custom agents.

你可以创建一个新的自定义代理，并为其命名。

So you can basically you can create a new custom agent, give it a name.

与个人代理不同，默认情况下，它无法访问任何内容，因此你需要明确授予它权限。

And unlike the personal agent, by default, it doesn't have access to anything, so you have to grant it access.

但一旦你授予了权限，它就能在后台自主运行。

But then once you do, it can actually run autonomously in the background.

举个例子，你可以赋予它访问自己数据库的权限来记录任务，然后将它连接到一个 Slack 频道，它就会开始在 Slack 上回复他人并记录任务。

So, for example, you can give it access to its own database to file tasks, let's say, and then you can attach it to a Slack channel, and then it will start responding to people on Slack and filing tasks.

这是一个使用场景。

That's that's one use case.

另一个场景是，你可以赋予它访问每周报告数据库的权限，然后让它搜索网页或搜索你的工作空间。

Another one is maybe you could you can give it access to a database of, like, weekly reports and then and then let it search the web or search your workspace.

因此，这个自定义代理某种程度上代表了某种工作或任务，即你希望自动完成的某些知识型工作。

And so it's sort of a custom agent sort of represents some work or job, some some knowledge work tasks that you want to be done autonomously.

我对此未来的发展感到非常兴奋的是，我们希望它能极其擅长从一个初始核心自我扩展能力，从而让它能够自我引导去完成任何事情。

One thing I'm really excited about this going forward is is we want it to be extremely good at sort of bootstrapping its own capabilities, basically, from an initial kernel, allowing it to basically bootstrap itself to do anything.

对吧？

Right?

比如，甚至可以构建一个我们尚未支持的集成，部署它，然后加以使用。

So even, for example, maybe building an integration that we don't support yet, deploying that and then and then using it.

所以你认为 Notion 代理实际上代表了更广义的代理概念，比如编写代码就是一个工具。

So you imagine that Notion agents are actually the broader definition of agent where, like, writing code is a tool.

我认为这非常关键。

It's I think it's pretty key.

是的。

Yeah.

我认为这非常关键。

I think it's pretty key.

是的。

Yeah.

我觉得编码代理就像是通用人工智能的核心。

I think I I think of coding agents as, like, the kernel of AGI.

通用人工智能将是一个编码代理。

AGI will be a coding agent.

代码是一种非常有用的原始工具，用于表示确定性的逻辑。

And and and and code is just a really, really useful primitive for representing, like, deterministic logic.

关于这一点令人兴奋的是，将其应用于知识工作代理时，它能够自我启动能力。

The thing that's really exciting about it, replying it to to a knowledge work agent is that it can bootstrap a capability.

你知道的吧？

You know?

所以，是的，正如我所说，如果不存在集成，它就能自己构建。

So, yeah, like I said, if integration doesn't exist, it can build it.

如果它需要连接到一个新的数据源，它也能做到。

If if it needs to, you know, connect itself to a new data source, it can do that.

考虑到你有一个Notion，它的规模是这样的，但它是否在一个更大规模的生产力和平台玩家的环境中运作？

Given you have a, you know, Notion, is that scale, but is it operating in a landscape of productivity and platform players that are at even more scale?

对吧？

Right?

许多这些平台最终都会拥有自己的代理。

Many of these will end up with their own agents.

来自实验室、微软阵营的许多人正试图整合其他数据源。

Lots of people from the labs, the Microsoft's world are trying to integrate other data sources.

这种跨领域的整合与索引尝试。

This like cross attempt to integrate and index.

你觉得这种情况会怎么发展？

Like, how do you think that plays out?

你觉得Notion的智能代理最擅长什么？或者它们最适合做什么？

Like, what do you what do you imagine that Notion agents are best at or what they have the right to go do?

如果你看一下整个格局，我觉得可以大致分为三类：实验室、软件平台，还有基础设施。

If you look at the landscape, like, I I would sort of say there's the labs, and then there's maybe the the the software platforms, and then there's maybe, like, infrastructure.

就实验室而言，我们把自己定位为模型领域的瑞士。

In terms of the labs, you know, we see ourselves as kinda like the the Switzerland for models.

我们认为，我们的客户不希望被锁定在某个特定实验室的模型上。

We think and our customers, they, you know, they don't wanna be locked into a certain certain labs model.

它们每个月都在发布新版本。

They're always releasing new versions.

每个月都有一个模型表现得比其他的好。

Any given month, one is better than the other.

所以我们希望成为一个平台，让你能随时轻松访问所有最佳模型，并且能轻松切换。

So we wanna be a place where, basically, you can you can easily get access to all the best models at any time, and you can easily switch around.

你认为开源在这方面也有作用吗？

Do you think open source plays into that as well?

是的。

Yeah.

对。

Yeah.

当然。

Absolutely.

我认为开源模型实际上已经变得非常好。

I think the open source models are actually getting really good.

现在有四个相当不错的中国本土模型。

There's, like, the four different Chinese models now that are that are quite good.

是的。

Yeah.

我们上周刚在我们的智能体中发布了其中一个，而且我们肯定会全部接入。

We actually just released one of them in our agent last week, and and we're gonna do all four for sure.

它们实际上很不错，而且比前沿模型便宜得多。

They're they're actually quite good, and they're and and they're way cheaper than the the frontier models.

所以我认为有很多应用场景是需要这种模型的。

So I think there's there's a lot of use cases where where where you'd want that.

我们希望提供这种选项。

We wanna give that as an option.

至于其他的，你知道的？

In terms of, like, the other you know?

我们认为我们的角色是整合所有能获取的最优模型，构建高质量、前沿的智能体实现，让人们能够轻松便捷地使用它们，同时打造一个非常适合人类和智能体协同工作的协作平台。

So, you know, we think of our role as sort of taking all the best models that we can, creating really high quality state of the art agent implementations where where people can easily and conveniently get access to them, and then making sort of a collaborative workspace that is really good for for humans and for the agents to to to coordinate on.

我觉得这正是这个世界所需要的，我们只是试图以一种优雅、执行到位的方式来做这件事。

I think it's it's something that's that's very needed in the world, and we're just trying to do it in a really tasteful, well executed way.

你刚才提到，需要索引才能让智能体表现良好。

You were describing you need the index to make the agents good.

你让智能体能够访问我们人类在Notion中拥有的工具。

You give the agents access to the tools that we humans have in Notion.

你怎么看待Notion的结构，比如它在哪些方面对代理有用，或者在哪些方面没用或不相关，比如数据库之类的？

How do you think about the structure of Notion and, like, where it's, like, useful or even, like, not useful or relevant for agents, like, and databases and such?

这仍然非常有用。

It's all still pretty useful.

极其有用。

Extremely useful.

我们一直面临一个挑战，就是想让代理使用起来非常方便。

There there's been a challenge to sort of you know, we wanna make it really convenient for the agent.

我认为这是以前不存在的新事物。

I think that's that's a new thing that that didn't exist.

明白吗？

Know?

过去，它对人类用户很方便，我们也让API对编写代码的人类开发者很方便。

In the in the past, it was convenient for humans, and then we also made APIs convenient for humans writing code.

是的。

Mhmm.

这是我们的API。

Here's our API.

所以我们实际上有了一个新客户，那就是智能代理。

So we essentially have a new customer, which is the agent.

一开始，这确实是个问题。

At first, that was definitely a problem.

你知道吗？

You know?

比如，我们的API使用了一种极其复杂的JSON格式来处理区块，默认情况下这种格式冗长又糟糕，对代理来说非常不友好。

So for example, like, our our API, uses this this crazy JSON format for blocks that, by default is, like, crazy verbose and, like, horrible for the agent.

但我们接受了这个挑战，专门为代理设计了极其便捷的API。

But we basically took on that challenge and designed just really convenient APIs for the agent.

我们创建了一种类似标准Markdown的变体，外观上与普通Markdown一致，但增强了所有Notion区块的功能，而模型对这种格式非常擅长。

We created sort of a markdown dialect that looks like the default normal markdown, but it's sort of enhanced with all the Notion blocks, and the models are really good at it.

效果非常好。

It works really well.

所以这就是它读取和写入页面的方式。

So so that's how it reads and writes the pages.

而对于数据库，我们使用了 SQLite。

And then for databases, we we use a SQLite.

基本上，它通过 SQLite 进行通信，效果也非常好。

So so, basically, it's the gets the speak in SQLite, which also works really well.

默认的方案其实效果并不好，但我们把它当作一个工程挑战来解决。

So the default thing did not work really well, but then we just, like like, took that on as an engineering challenge.

我现在可以说，我们已经拥有了极其便捷的 API，代理们用起来非常自然。

And and I would say now we have, like, extremely convenient APIs that the agents are are really naturally good at.

你们是如何理解或发现什么样的 API 更适合代理的？

How did you understand or figure out what would make the API better for agents?

这是个好问题。

That's a good question.

是的。

Yeah.

我觉得这是尝试各种方法的综合结果。

I would say it's a it's a combination of just trying things.

这非常依赖实证。

It's it's it's very empirical.

嗯。

Mhmm.

所以我们只是随意尝试，然后发现，哦，这个功能表现不太好。

So so we're just playing around and, like, like, noticing, oh, it's not very good at that.

哦，这个用了太多标记了。

Oh, that's way too many tokens.

我们怎么才能让它更简洁呢？

How can we make this smaller?

再结合一些第一性原理的思考，比如，模型到底是在什么数据上训练的？

And then a little bit of just, like, like, first principles thinking of, like, you know, what is it the models are being trained on?

它们的先验知识里包含什么？

And what's what's in their prior?

它们知道什么？

What do they know?

我们觉得它天然擅长什么？

And what what do we think it would naturally be good at?

那么，智能体循环是如何工作的？

And and, like, like, how does the agent loop work?

那么，访问这些内容的便捷高效模式是什么？

And, like, what what would be the convenient efficient pattern for for accessing these things?

然后就是大量的尝试和摸索。

And so and then just, you know, a a lot of playing around.

我听到的是用户研究，其中用户实际上是智能体，然后是持续的评估。

I hear user research where the user is actually agent and then, you know, ongoing eval.

是的。

Yeah.

我的意思是，用户，没错。

I mean, user yeah.