本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
总的来说，我对CLI非常看好。
Broadly speaking, I'm really bullish on CLIs.
在特定环境下，我仍然看好MCP。
I'm still bullish on MCPs in a certain environment.
我认为当你需要一个狭窄且轻量级的代理时，这会非常好。
I think it'd be really great for when you want, like, a narrow lightweight agent.
我觉得
I think
确实有很多应用场景，你
there's definitely a lot of use cases where you
不希望有一个带有计算运行时的完整编码代理。
don't want, like, a full coding agent with the compute runtime.
而且，你还希望它的权限更加严格。
And, also, you want it to be, like, more tightly permissioned.
MCP本身就具有非常强大的权限模型。
MCP inherently has a really strong permission model.
你只能调用这些工具。
Like, all you can do is call the tools.
MCP 就是个简单直接又管用的东西，效果相当不错。
MCP is just, like, the dumb simple thing that works, and it that is pretty good.
Notion 致力于成为记录企业人员工作活动的最佳系统。
Notion is dedicated to being the best system of record for where people do their enterprise work.
只要其他人还在使用 MCP，我们就会持续支持我们的 MCP。
So we will always support our MCP insofar as other people are using MCPs.
无论我们的观点如何，我们已经为 MCP 投入了大量精力，并正在组建一支出色的团队。
Regardless of our perspective, we've put a lot of effort into our MCP, and we have a fantastic team that we're building.
大家好。
Hey, everyone.
欢迎收听《Laid in Space》播客。
Welcome to the Laid in Space podcast.
我是 Alessio，And 的创始人，今天和我一起的是《Laid in Space》的编辑 CWIC。
This is Alessio, founder of And I'm joined by CWICs, editor of Alayed in Space.
你好。
Hello.
你好。
Hello.
我们回到了Alessio为我们布置的漂亮演播室，这里有Notion的Simon和Sarah。
We're back in the beautiful studio that Alessio has set up for us with Simon and Sarah from Notion.
欢迎。
Welcome.
谢谢你们的邀请。
Thanks for having us.
谢谢你们的邀请。
Thanks for having us.
是的。
Yeah.
祝贺你们最近的发布。
Congrats on the launch recently.
自定义代理，终于来了。
Custom Agents, finally, it's here.
感觉怎么样？
How's it feel?
我们发布东西很慢。
We ship things slowly.
所以它之前已经处于测试阶段一段时间了。
So it had been in alpha for a little bit.
当它处于测试阶段时，有一群人确保它已准备好上线。
And at the point at which is it's in alpha, there's a group of people that are making sure it's ready for prod.
还有一群人正在开发下一个功能。
And then there's a group of people working on the next thing.
所以有时候这些发布会带来延迟的满足感。
So sometimes some of these launches are a bit delayed satisfaction.
所以提醒自己他所做的一切工作是很不错的，因为我们确实有习惯性地领先两到三个里程碑，因为你必须如此。
So it's quite nice to remind yourself all the work he did because we do have a habit of being two or three milestones ahead just because you have to be.
你不能自满。
You can't get complacent.
但很高兴人们理解了这有多有用。
But it's been great that people understood how this is helpful.
而且我认为，与两三年前相比，如今构建AI工具普遍更容易了。
And I think that's just easier in general building AI tools today than it was two, three years ago.
人们都懂。
People get it.
因此，在用户教育方面，这是我们迄今为止最成功的发布，无论是在免费试用、转化率等方面。
And so that user education, there's just it was our most successful launch in terms of free trials and converting people and things like that.
非常成功。
It was really successful.
是的。
So yeah.
但还有很多东西要构建。
But there's a lot to build.
免费提供三个月很有帮助。
Making it free for three months helps.
你知道吗？
You know?
是的。
Yep.
这对我来说确实非常令人兴奋，因为这可能是我们第四次或第五次重建它了。
It was definitely super exciting for me because it's probably the fourth or fifth time that we rebuilt that.
是的。
Yes.
你从2022年就开始构建这个了。
And You've been building this since 2022.
是的。
Yeah.
甚至就在2022年底我们获得GPT-4访问权限的时候。
It was even right when we got access to G P G 4 in late twenty twenty two.
好的。
Okay.
我们来创建一个智能体吧，我当时用的词是助手。
Let's make an agent that I mean, I would use the word assistant at the time.
但这还不是我的第三个智能体。
That wasn't my third agent yet.
不过，我们会让它拥有Notion的所有功能，并在后台运行。
But, oh, we'll give it access to all the tools that Notion can do, and then it will run-in the background.
为我们做事。
Do work for us.
然后我们试了很多次，但当时时机太早了。
And then we just tried that many times, and it just was too early.
我得逼你好好想想这一点。
I need to force you to, like, double click on that.
什么叫太早了？
What is too early?
什么没有成功？
What didn't work?
在函数调用功能推出之前，我们一直在尝试使用Frontier Labs和Fireworks对Notion功能进行微调，做一个函数调用模型。
We were fine like, before function calling came out, we were trying to fine tune with the Frontier Labs and with Fireworks, like a function calling model on Notion functions.
这正好是我加入的时候。
This is right when I joined.
我加入是因为我们需要一位经理。
I joined because we needed a manager.
西蒙需要能够去度假。
Simon was needed to be able to go on vacation.
我加入的时候差不多就是这个时候，你可以更详细地讲讲。
That's around when I joined, you so can speak much more to it.
是的。
Yeah.
我们曾先后与Anthropic和OpenAI合作，试图在当时——我们第一次尝试的时候，甚至还没有‘工具’这个概念。
We did partnerships with both Anthropic and OpenAI at different times to try to at the time, the when we first tried, there wasn't even a concept of, like, tools yet.
我们自己设计了一个工具调用框架，然后尝试对模型进行微调，使其能在多轮对话中使用它。
We designed our own, like, tool calling framework, and then we tried to fine tune the models to to use it over multiple turns.
因为这个功能一开始根本没法正常工作。
And because it didn't work well out of the box.
我觉得，是的，模型太笨了，而且上下文也太短了。
I think yeah, the models are just too dumb and the context, like, was also way too short.
是的。
Yeah.
而且，我们确实为此撞了很长时间的南墙。
And, yeah, we just banged our head against it for a long time.
不幸的是，虽然总是有一些迹象表明它在起作用，但始终不够稳定，无法成为真正有用且令人愉悦的体验，直到我认为去年年初Sonic 3.6或7版本的突破出现。
Unfortunately, it was always, like, there was always like glimmers that it was working, but it never felt quite robust enough to be like a useful, delightful thing until I would say the big unlock was probably like Sonic 3.6 or seven early last year.
从那时起，我们开始开发我们的智能体，并在去年上线了它。
And that's when we started working on our agent, which we shipped last year.
然后是自定义智能体，具备类似的功能。
And then Custom Agents, kind of a similar capability.
而这一项花费了更长时间，因为我们只是想把可靠性提升得更高，因为它实际上是在后台运行的。
And that one just took longer because we just wanted to get the reliability up a lot higher because it's actually running in the background.
而自定义代理的权限和理解产品界面，是通过Slack频道与一组人员共享，并允许访问呈现给另一组人员的文档。
And the product interface of, like, permissions and understanding this Custom Agent is shared in a Slack channel with x group of people and has access to documents that are surfaced to y group of people.
而x组与y组的交集可能并不完全重合。
And the intersect of x versus y might not be whole.
那么，你该如何设计产品，以确保管理员理解权限设置需要多次调整？
And so how do you build the product around making sure administrators understand that permissioning took multiple swings.
归根结底，一切都很困难。
Everything is hard back at the end
一天结束时。
of the day.
是的。
Yeah.
我很好奇，当模型无法正常工作时，你们如何将这些情况反馈到产品路线图中？比如，我们应当预期模型会在某个合理的时间内有所改进，但与此同时，你们在2022年已经拥有了大量客户。
I'm curious, like, when the models are not working, how do you inform the product road map of, okay, we should probably build expecting the models to be better at some reasonable pace, but at the same time, we need to you had a lot of customers in 2022.
你并不是一家没有用户基础的初创公司。
It's not like you were a new company with no user base.
是的。
Yeah.
我的意思是，你总是要在成为AGI导向、前瞻性思考和构建未来方向之间找到平衡。
I mean, I think there's always the balance of, like, you wanna be AGI pilled and thinking ahead and building for where things are going.
但同时，你也想推出真正有用的产品。
But also, you wanna be, like, shipping useful things.
所以我们总是努力保持这种平衡。
And so we always try to, like, keep a balance there.
我们倾向于采取投资组合的方式。
We try to take, like, a portfolio approach.
我们一直在同时进行多个项目，并且始终试图
We're always working on multiple projects, and we're always trying
在维护我们已经发布的产品上投入精力，让那些已经运行良好的新功能变得更好。
to work on maintaining things where that that we've already shipped, like shipping new things that are, like, eminently working well and make them really good.
而且我们总是希望有一些稍微疯狂的项目。
And then we wanna always have a few projects that are a little bit crazy.
你们现在有哪些AGI导向的项目呢？
And what are the AGI pill projects that you have today?
我很好奇你们不方便透露的具体工作内容，但我想知道，现在有哪些项目，可能在十八个月后，人们会说：‘显然，这个一定会成功。’
I'm curious what you don't have to share exactly what you're working on, but I'm curious what are things today that maybe in eighteen months people will be like, oh, obviously, this was gonna work.
十八个月？
Eighteen months?
是的。
Yeah.
十八个月是很长的。
Eighteen months is.
这是一段很长的时间。
It's a long time.
是的。
Yeah.
是的。
Yeah.
目前有许多事情正在发生。
There's a number of things happening.
我认为越来越清楚的一点是，编码代理是EGI的核心。
I think one thing that's becoming more clear is I think the coding agents are the kernel of EGI.
一切都是编码代理。
Everything is a coding agent.
我认为这是一个方向。
I think that's one one direction.
而令人兴奋的是，你的代理可以自我启动软件和能力，并实际进行调试和维护。
And then, yeah, the exciting thing about that is your agent can bootstrap its own software and capabilities and actually debug and maintain them.
所以，是的，我们正在深入思考这一点。
And so, yep, we're thinking a lot about that.
另外，我还非常兴奋的一个领域是我们所谓的软件工厂。
And then, yeah, like, another category of things that I'm really excited about is what we call the software factory.
很多人在使用这种说法。
Lots of people are using this this sort of word.
基本上，它只是意味着：你能否创建一个尽可能自动化的流程，用于开发、调试、合并、审查和维护一个由多个代理协同工作的服务代码库？
Basically, it just means, can you create, like, a as automated as possible, a workflow for developing, debugging, merging, reviewing, and maintaining a code base of a service where there's a bunch of agents working together inside?
那么，这到底是如何运作的呢？
And, like, how does that work?
如果你回想起你最初的问题，比如，为什么这件事花了这么长时间？
If you think back to your initial question, like, why did this this take so long?
我认为Notion的
I think something Notion's
我没这么说。
I didn't say that.
但是，是的。
But, yes.
好的。
Okay.
去吧
Go
继续。
ahead.
在三年的尝试过程中，有什么变化？
What changed over the three and a years of trying?
因为大多数人总是说还没成功，直到推理模型出现，才成功了。
Because most people always say it didn't work yet, then Reasoning Models came, then it worked.
我当时想，好吧。
I was like, okay.
我们再往前走一点。
Let's go a little bit.
我的意思是，这是其中一部分。
I mean, that's part of it.
但我认为另一部分，也是我认为真正能让Notion在每一项前沿能力上脱颖而出的，是我们拥有两项至关重要的技能。
But I think the other part of it that I actually think is really what will set Notion apart for every new capability is we have two skills that are crucial when it comes to frontier capabilities.
一个是不要逆流而上。
One is not letting yourself swim upstream.
所以要快速意识到，如果你只是在硬碰模型的能力，而不是给模型提供正确信息，或者没有搭建好正确的基础设施，这本身就需要一种直觉能力。
So like quickly realizing if you're just pressing against model capabilities versus not exposing the model to the right information, not having the right infrastructure set up, that in of itself is a skill of intuition.
第二个是看清，你没有逆流而上。
And the second is to see, okay, you're not swimming upstream.
河流到底朝哪个方向流？
Which direction is the river flowing?
我们该如何提前思考产品，即使它还不成熟，也要开始构建，以便当它真正出现时，我们已经做好准备。
And what is like how do we think ahead about the product and start building it even if it's not great yet so that when it is there, we're ready for it.
这些有时候看起来可能违背直觉。
And like those can sometimes feel like counterintuitive things.
比如，我们可能在尝试优化一个根本还不存在的工具调用模型。
Like we can be trying to fine tune a tool calling model when they don't exist yet.
关键是不要长时间这样做，但要意识到其中确实存在某种可能性。
And the trick is to not do that for too long, but realize that there was something there.
我们之前做过很多事，其实根本没顺着正确的方向走。
And we've had a lot of things which like we're just like not swimming in the right direction with the streams.
在我们做出会议纪要之前，已经迭代过好几个版本的转录功能了。
I think we had multiple versions of transcription before we got meeting notes.
对吧？
Right?
天哪。
Oh my god.
我经常提到这个。
I talk about that.
是的。
Yeah.
对。
Yeah.
我觉得我们真的和前沿实验室在能力上紧密合作。
I think that like, we we really closely partner with the Frontier Labs on capabilities.
当这些能力发展时，我们也必须对此有坚定的信念。
And we also have to have strong conviction on as those capabilities move.
Notion 的目标是成为你协作和完成工作的最佳场所。
Notion is about being the best place for you to collaborate and do your work.
如果我们工作方式发生变化，这个叙事会如何改变？
And how does that narrative change if the way that we work changes?
是的。
Yeah.
对。
Yeah.
你告诉我你是 Agent Lab 理论的粉丝。
You told me you're a fan of the Agent Lab thesis.
而这正是
And this is
我向很多候选人展示过这个理论。
I show that thesis to so many candidates.
我现在已经把它设置为Chrome的自动填充内容了。
Like, I have it as my Chrome autofill at this point.
它是我访问最频繁的网站之一。
Like, it's one of my most visited
这是不是在说‘你为什么应该在Notion工作，而不是在OpenAI’？
Is this the here's why you should work in Notion and not OpenAI?
这就是它的不同之处。
Here's what's different about it.
是的。
Yeah.
这就是为什么它不仅仅是一个包装层。
And here's why it's not just a wrapper.
我真的认为越来越多的人意识到它不仅仅是一个包装层。
I actually think more and more people understand it's not just a wrapper.
顺便说一下，一开始我们构建的部分确实是基于一些已经运行良好的功能的包装。
And by the way, like, in the beginning, parts of what we build are wrappers on functionality that works well.
那并不是我所说的最能驱动收入的产品，也不一定是用户真正需要的。
That's not really the most I would say that's not the product that that drives revenue, and that's not necessarily always what users need.
Notion 就像是 AWS 的封装。
Notion is the AWS wrapper.
但这个封装做得非常美观，而且打磨得极其精致。
But like the wrapper is very beautiful and like very well polished.
所以我一直想到的一个类比是 Datadog 和 AWS。
So like the analogy that I've been coming back to is Datadog and AWS.
是的。
Yeah.
Datadog 离不开云存储。
Datadog could not exist without cloud storage.
这是基础且关键的支撑。
That it's fundamental that works.
而 AWS 也有自己的 CloudWatch 产品。
And AWS has a CloudWatch product.
但Datadog擅长理解人们在发布产品时对可观测性的需求。
But Datadog is an expert on understanding how people want observability on the products they launch.
而我们擅长理解人们如何希望进行协作。
And we're experts in understanding how people wanna collaborate.
这正是我们的专长所在。
And that's really where our expertise lies
完全正确。
Totally.
无论我们使用什么工具。
Regardless of the tools that we use.
我想知道你是如何看待隐性专长和显性专长的？
I'm curious how you think about implicit versus explicit expertise.
我觉得Datadog在隐性和显性专长上各占一半。
I feel like Datadog is half and half implicit and explicit.
它们能跨越不同市场和行业，理解工程团队通常关注哪些方面。
It's like they understand across markets and industries where engineering teams usually look for.
对于Notion来说，专业知识几乎都集中在边缘，因为作为平台，你的定位非常横向，最终用户并不相同。
With Notion, it's almost like more of the expertise is at the edge because you as a platform, you're, like, so horizontal that the end user is not really the same.
但最终用户总是对的。
But, like, the end user is always Yeah.
可能是工程主管，或者与SRE相关的人。
An engineering lead, a kinda, like, SRE related person.
但对于Notion来说，它可以是任何角色。
With Notion, it can be anything.
所以我很好奇，你们是如何将这种专业知识融入产品中的，而显然，AWS无法构建Notion。
So I'm curious how you put that expertise into a product versus, obviously, AWS cannot build Notion.
在这种情况下，这并不太适用。
It's that doesn't quite work in this case.
它的形态略有不同。
It's a little bit differently shaped.
我认为经典的垂直SaaS，其数据大致就是这样。
I think a classic vertical SaaS, like, the data is kinda like that.
他们非常深入地了解自己的每个客户。
They understand their individual customer very deeply.
这属于一个狭窄的领域。
It's kind of a narrow slice.
Notion 一直是非常通用的，我们的任务始终是平衡这两种某种程度上相互对立的力量。
Notion has always been super horizontal, and our task has always been to balance these two somewhat opposing forces.
我们在倾听客户的需求，了解他们希望我们打造什么。
We're listening to our customers and what they want us to build.
这是一个广泛的领域。
It's a broad slice.
同时，我们也在思考：如何将他们的需求分解成简洁易用的基础组件，以实现最大的投入产出比，并维护整个系统，使其整体清晰、易用。
And then also, we're thinking about, okay, how do we decompose what they want into a nice primitives that are really nice to use and will get us, like, as much bang for the buck as possible, And then maintain the whole system, make it all like super clean and nice to use.
我们仍然拥有更简单的使用路径。
We still have easier journeys.
我们仍然专注于核心功能，实际上，我们团队的失败在于过于关注那些看似酷炫的工具。
We still focus on like core I actually think the failure of our team is when we focus too much on what are cools that are what are tools that are cool tools.
我认为这正是我们效率最低的时候。
I actually think that's when we make have the least velocity.
因为你仍然需要对用户旅程保持一定的专注。
Because you still need some sort of focus on a user journey.
比如，我们每周五都会坐下来，查看最耗令牌的自定义代理对话记录的第99百分位，分析它为什么表现不佳，然后剔除大量任务。
So, like, for instance, we'll all sit down every Friday and look at the p 99 of the most token exhaustive custom agent transcript and just look at why it didn't do well and cut a bunch of tasks.
我们仍然专注于这一点：这应该能正常工作。
Like, we still focus on this has like, this should work.
邮件分类应该能正常工作。
Email triaging should work.
对吧？
Right?
同样，当我们讨论在开始录制之前构建聊天功能时，比如，我该如何导出PDF？
And similarly, like, when we're talking about before building chatting before we started filming about, okay, how can I do PDF export?
这种功能就值得我们开发一个能够访问计算机沙箱、文件系统并具备编写代码能力的工具。
That's functionality that then merits maybe we should build a tool that has access to a computer sandbox and a file system and the ability to write code.
但这是因为我们考虑到用户在日常工作中需要导出PDF，而不是因为我们觉得一个计算机工具可能很酷。
But it's because we're thinking about the fact that our users to do their daily work need to export PDFs not because we're like, I think a computer tool could be cool.
我们先看看会发生什么。
Let's just see what happens.
我们必须专注于某些用户使用流程。
Like, we we have to focus on some user journeys.
否则，我们就缺乏足够的策略来确定优先级。
Otherwise, we just don't have enough strategy to to prioritize.
我觉得你有很多非常强烈的观点。
I think there's a lot of, like, really strong opinions that you've had.
你有没有像莎拉·萨克斯那样的毛巾？
Did you have, like, like, a towel of Sarah Sachs?
你是怎么管理你的团队的？
Like, what how do you run your team?
我觉得你只是积累了很多这样的强烈观点。
Like, I feel like you just have accumulated all these strong opinions.
显然，这部分跟你的Token Town项目有关。
Obviously, part of this is your Token Town thing.
我认为，道家与莎拉·萨克斯合作的方式，取决于你问谁。
I think the Taoist working with Sarah Sachs is it have to it depends if you ask.
我认为这取决于你是我的团队成员、合作伙伴还是供应商。
I think it depends if you're on my team or a partner or a vendor.
是的。
Yeah.
还有其他人希望像你那样管理他们的团队。
There are other people who want to run their teams the way that you're you're you're, like, running these things.
但同样地，当你做自定义代理演示时，你说过我们一直在使用自定义代理，还列出了我们所做的一切的超长清单。
But then also similarly, not Simon, when you did the Custom Agents demo, you had we've been using Custom Agents and here's the super long list of everything that we do.
从来没有人读过它。
No humans ever read it.
这是你说的。
That's what you said.
我当时就想
I was like
是的
Yeah.
所以对我来说，我很快学会并适应的一点是，我的职责并不是提出想法或成为技术专家。
So I think for me, something that I learned very quickly and became very comfortable with was that my job was not to be the ideas person or the technical expert.
我的职责是确保每个人都理解目标，有资源帮助他们优先处理该做什么，并有渠道表达他们认为重要的事情。
My job was to make it so that everybody understood the objective, had a resource to help prioritize what they should work on, and had an avenue to prioritize what they thought was important.
我认为这在所有领导工作中都是如此。
And I think it's true with all leadership.
但我认为在AI团队中尤其如此，我们绝大多数最好的想法都来自于原型，来自于那些因为看到用户问题而产生好点子的人。
But I think especially on the AI team, almost all of our best ideas come from prototypes, from people that have a cool idea because they saw a user problem.
如果所有这些想法都必须通过我和产品合作伙伴，或者西蒙和伊万所决定的方向的‘嗅觉测试’，那将是极大的损害。
And it's a huge disservice if all of those ideas have to pass like the sniff test of what me and a product partner or Simon and Ivan decided were the direction.
对吧？
Right?
因为我们正在做的很多工作都是在利用现有能力。
Because a lot of what we're doing is leaning into capabilities.
所以我认为，工程领导的角色从来就不是层级化的，现在也依然不是。
So I think that's the first thing is I don't really view like the role of engineering leadership as hierarchical, nor has it ever been.
但尤其现在，我们非常愿意根据实际成果来调整方向。
But especially now, like very willing to change direction based on like proof is in the pudding.
而我们已经重建了我们的框架三四次。
And I think we have rebuilt our harness three or four times.
当你这么做时，工程领导的第二条原则就是：需要打造一支乐于删除自己代码的团队。
And when you do that, then the second rule of engineering leadership is like, need to build a team that's comfortable deleting their own code.
这支团队要低 ego，以公司利益为先，不会写设计文档只是为了当作自己的晋升材料。
And is very low ego and is driven by what's best for the company and doesn't write design docs because they think it's their promotion packet.
这种文化在我不加入之前，Notion 就已经存在很久了。
And that's a culture that Notion had long before I joined.
但像我们这样愿意迅速集结解决不同问题、重做之前构建过的东西，因为情况变了，这样做在很多公司都会遇到很大阻力。
But like our willingness to just swarm on different problems and redo things that we've built before because something has changed, there's a lot of friction that can happen at companies when you do that.
而在Notion，这种情况不会发生。
And it doesn't happen at Notion.
因为当新人加入时，他们也不希望成为那个说我们不该这么做的人。
And because it doesn't happen when new people join, like, don't wanna be the ones that are saying we shouldn't do this.
我写了这段代码。
I wrote that code.
于是你就创造了一种人人都会采纳的文化。
So then it's you create a culture that everyone adopts.
而这种文化直接来源于西蒙和伊文，因为他们非常开明。
And that culture comes directly, I think, from Simon and Iven, though, because they're very open minded.
你还有什么要补充的吗？
Anything you'd add?
我不是像莎拉那样的管理者。
I'm not a manager like Sarah is.
我的很多职责其实是提前思考，确保我们正在构建正确的功能，然后进行原型设计。
A lot of my role is really to try to think a little bit ahead, make sure that we're building on the right capabilities, and then the prototyping stuff.
而且，是的，始终重新开始真的非常关键。
And, yeah, it's really critical to always just be starting again.
没关系。
It's okay.
这是个新事物。
This is a new thing.
这意味什么？
What does this mean?
如果我们彻底重新思考一切，重写一切呢？
What if we just rethought everything, rewrote everything?
因此，显然每六个月就循环做一次这样的事。
Hence, obviously, just doing that in a loop every six months.
是的。
Yeah.
你相信为此类事情举办内部黑客松吗？
Do you believe in internal hackathons for this stuff?
我认为有两种不同的版本。
I think there's two different versions.
一种是我们有一批稳定的资深工程师，他们来来往往，参与我们所谓的西蒙漩涡项目，并将我们构建的东西投入生产。
So one is we just have a solid bench of senior engineers that come and go on what we call the Simon Vortex and productionizing what we built.
对吧？
Right?
因为当你身处西蒙漩涡时，进展速度非常快。
Because when you're in the Simon Vortex, the velocity is super high.
方向每天都在变化。
The direction changes daily.
它的作用相当于一个臭名昭著的创新实验室。
And it's meant to be like the equivalent of a Scuckworks lab.
我们不需要为这个举办黑客松。
We don't need to do hackathons for that.
我们需要的是可以信赖的资深工程师，让他们进出这些项目。
We need to have senior engineers that we trust to come in and out of those projects.
比如，管理边界非常宽松。
For instance, like, management boundaries are really loose.
比如，你向他汇报，但现在却为她工作。
Like, you report to him, but you work for her right now.
当我们招聘经理时，这一点他们不必在意，因为我们通常会形成工作结构。
Like, that's something that when we hire managers, it's important they don't care about because we tend to form work structures.
是的。
Yeah.
别太有地盘意识。
Don't be too territorial.
我们通常是在交付成果后才形成工作结构，而不是之前，这是历史惯例。
We form work structures after we ship things, not before, just historically.
第二点是我们确实有全公司的黑客松活动。
The second thing is we do have company wide hackathons.
实际上，我们今天早上刚举办了上周黑客松的演示日。
Actually, we just had our demos day for the hackathon we had last week this morning.
这更多是给那些不直接参与项目的人，让他们觉得有时间停下来，学习如何提高自己的效率，或者如何使用Notion自定义代理来构建一些东西。
That's more for people that aren't directly working on the project, feeling like they have the time to pause and learn how to make themselves more productive or how they would use Notion custom agents to build something.
或者，这次黑客松的一部分实际上是鼓励公司全员从零开始构建自己的代理工具循环。
Or part of the hackathon was actually encouraging everyone across the company to build their own agentic tool loop calling from scratch.
就像每篇关于如何做的博客文章里所描述的那样，我觉得。
Following like in every blog post on how to do it, I think.
因为我们希望
Because we want
这是那个复合工程的项目吗？
Is that the compound engineering one?
是的。
Yeah.
我们希望公司每个人都使用Cloud Code，或者任何他们喜欢的编码代理，并理解这一基本理念。
We want everyone to use Cloud Code in the company or whatever the coding agent they please and understand that fundamental.
因此，我们安排了一天半的时间，让所有领导层鼓励各自团队的全体员工参与其中。
So we set aside a day and a half where all leadership encourage everyone on their teams across the company to do it.
所以我们会有这样的黑客马拉松。
So we have hackathons like that.
我会说，有点开玩笑的意思。
I would say, like, facetiously.
我们所构建的每一样东西，起初都像一场黑客马拉松，直到它成熟，穿上正装，作为产品运营上线，之后再配备数据科学家之类的人。
Like, everything we build is a little bit like a hackathon until it graduates and puts on big boy pants and as a product ops rollout later and has assigned data scientists and stuff like
安全审查、企业级支持。
Security review, enterprise stuff.
实际上，安全审查是我们最早引入的环节之一，因为如果后期才介入，会严重拖慢进度并引发很多矛盾，而让他们早期参与反而能打造出更好的产品。
Actually, security review is one of the things that we bring in first because it just slows us down way more and causes a lot of tension, and they build better product if they're involved early.
这可能是最早介入某项工作的人员。
That is probably the first person to get involved in something.
公关批准的回答。
PR approved answer.
不是。
No.
不。
No.
但这不仅仅是公关批准的。
But it's not just PR approved.
它就像是
It's like
这是真实的。
It's real.
它确实是真实的。
It's actually real.
我只是想说
It's I'm just saying
暗色组织。
dark tissue.
是的。
Yeah.
因为我的背景也是，我在Robinhood工作了多年。
Because my background's also I worked at Robinhood for a number of years.
所以合规性之类的事情，如果不能自然掌握，就得通过碰壁来学习。
So compliance and things like that are a little bit more you learn the hard way when it doesn't come naturally.
是的。
Yeah.
我认为黑客松对于提升大众非常重要。
I think the the hackathon is really important for uplifting the general population.
但如果这是你构建新事物的唯一方式，那你就完了。
But if that's the only way you can build new things, you're toast.
它必须是构建这些新事物的日常流程。
It has to be, like, the daily processes of building these new things.
而且我认为，在AI时代，更多的杠杆作用会集中在那些最具好奇心和热情的人身上。
And it has to be about I think I think in the AI era, a lot more leverage accumulates to the most curious and excited people.
所以我们的一切都是为了激发这种能量。
And so it's like, we're all about just, like, activating that energy.
如果有人在周末原型化他们感兴趣且重要的东西，那这应该是我们最应该关注的事情。
If someone's prototyping something on the weekend that they're excited about, and it's important, that should be the main thing that we're doing.
是的。
Yeah.
这不是我们每季度安排一次的黑客马拉松。
It's not a hackathon that we schedule once a quarter.
这是一种日常流程。
It's just daily process.
这是文化的一部分。
It's part of the culture.
这就是我们现在在Notion中推动图像生成的方式。
That's how we shift image generation in Notion now.
它一直是个不错的功能，但之前并不清楚它与产品优先级的关联性。
It was always this thing that would be nice to have, but it wasn't really clear where that was necessarily aligned in product priorities.
这会耗费大量工作。
It'd be a lot of work.
我们数据库集合团队有一位叫吉米的同事，他非常想为封面图片和Notion内部实现图像生成功能。
And we had someone on the database collections team, Jimmy, who was like, I really wanna do image generation for cover photos and Inside Notion.
我们对他说：如果你想做，那就去做吧，欢迎你来实现。
And we're like, if you wanna build it, like it's do it, please.
我们非常支持你。
Like, we encourage you.
我们为他提供了所有资源，让他可以直接使用Gemini，并能追踪令牌使用情况，最终通过我们的端点实现了功能。
We gave him all the resources of working directly with Gemini and being able to track the token usage and it got working through our endpoints.
我们还为他提供了评估支持，一切资源都给了他。
We gave him eval support, everything.
然后，这变成了一整个项目。
And then it became a full project.
是的。
Yeah.
所以，作为领导者，你不能有自我中心。
That's why you can't have, like, ego as a leader.
我们就是这样工作的。
That's how we work.
如今团队规模有多大，包括工程团队和整体团队？
What's the size of the team today, both engineering and overall?
我负责管理我们称之为核心AI能力和基础设施的团队。
I manage the team that's what we'll call, like, core AI capabilities and infrastructure.
大约有50人。
That's about 50 people.
但我们还有负责打包的合作伙伴团队。
But then we have partner teams that do packaging.
比如在侧边聊天、自定义代理和会议笔记中的呈现方式，又涉及另外三四十人。
So how it shows up in the corner chat versus Custom Agents versus Meeting Notes, that's another thirty, forty people.
而Notion中每个拥有产品服务的团队，都负责其代理所交互的工具。
And then every team that has a product service at Notion that a user can interface with owns the tool that the agent interfaces with.
编辑团队。
The editor team.
负责CRDT以实现离线模式的团队，也是负责处理两个代理竞争编辑区块的团队。
The team that did CRDT for offline mode is the same team that handles how two agents edit competing blocks.
这是同一个问题。
It's the same problem.
构建底层SQL引擎的团队，也是负责代理如何请求执行SQL查询并高效完成的团队。
The team that built the underlying SQL engine is the same team that owns how the agent asked it to run a SQL query and it does it performantly.
因此，从这个角度来看，所有从事产品工程的人员都肩负着让系统同时为人类用户和代理提供良好体验的任务。
And so from that regard, anyone working on product engineering is tasked with making them work for customers that are humans and agents.
因为随着时间推移，我们大部分流量将来自使用我们界面的代理，而非人类。
Because over time, a majority of our traffic will be coming from agents using our interface, not humans.
我们的目标是让整个产品团队都为代理构建产品。
And our objective is to make it so that the whole product org is building for agents.
是的。
Yeah.
这在内部会带来哪些变化？
How does it change internally?
展开剩余字幕（还有 480 条）
激活门槛降低了很多。
The activation bar is lowered a lot.
比如，任何人都能相对容易地创建原型，尤其是如果你已经在现有的代码库中了。
Like, anybody can create a prototype very somewhat easily, especially if you're, like, in an existing code base.
你们有没有提高对人们提交的原型类型的要求，不再只是随便看看，而是
Have you raised the bar on what type of prototype people need to bring forward to gonna be taking not, like, seriously, but
我认为在很多方面，门槛都降低了。
I think the bar is lowered in many ways.
我们团队做的一件非常酷的事情是，我们的设计团队专门创建了一个名为设计 Playground 的独立 GitHub 仓库，里面包含了很多辅助组件，方便快速搭建其他界面，现在已经变得相当复杂了。
Like, one thing of our our team build that is really cool is our our design team made a whole separate GitHub repo called the design playground, and it's basically just they create a bunch of, like, helper components and you for quickly I threw in the other UIs, and it's become, like, actually quite sophisticated.
里面甚至还有智能体，这挺有趣的。
Like, has agent in there, and that's pretty fun.
所以我们基本上不再做静态原型了。
So we pretty much they don't do mocks.
他们直接就做出来，
They just make,
比如完整的原型。
like, full prototypes.
几天内就能搞定。
Period days, it works.
他们给你一个网址。
They give you, like, a URL.
他们说，好的。
They're like, okay.
所以我们得做出那个真正的生产版本。
So we have to make the, like, the real production version of that.
对工程师来说，一个原型就是实现一个真正能用的功能开关。
And then for engineers, a prototype looks like just making it a feature flag that actually works.
这就是标准。
That's a bar.
要理解的一点是，Notion的独特之处在于，没有人像Notion的员工那样在工作中如此频繁地使用Notion。
So the thing to understand that's really unique about Notion, one of the reasons I joined, we're super lucky, is no one uses Notion in their job as much as people that work at Notion.
当然。
Of course.
所以我认为很少有公司能做到这样。
So I think there's very few companies.
也许只有在Chrome团队工作的人才会这样。
Maybe if you worked on Chrome, I guess.
但我们内部首先发布所有产品，并获得非常快速的反馈。
But like everything that we ship internally first and get a lot of really quick feedback.
而且有时候我们的开发实例完全出问题了。
And also sometimes our dev instance is totally borked.
你必须调整一堆功能开关才能完成工作，每个人都是这样。
You have to change a bunch of flags to get things done, and that's everyone.
所以做IT工单、供应链采购、招聘的人，都在使用同一个Notion实例，上面启用了大量为这些原型设计的功能开关。
So people that do IT ticketing, people that do supply chain procurement recruiting, everyone is using the same instance of Notion with a lot of flags on for these prototypes people build.
我们的团队里有一位叫布莱恩·莱文的设计师，我认为是他推广了‘演示胜过文档’这个理念。
I mean, so we have this Brian Levin, one of the designers on our team, I think evangelized this concept of demos over memos.
哦。
Oh.
这在构建演示方面非常有效。
Which has been very good for building demos.
我认为这给我们带来了很大的压力，要求我们具备强大的产品信念。
And I think it's put a big pressure point on us to have really strong product conviction.
因为如果任何东西都能被演示，你就必须有一个强大的筛选机制，确保当你投入一定量的工作时，你是在聚焦于一个核心点。
Because if anything can be demoed, you really need a strong filter of making sure that if you're doing x amount of work, you're making the you're focusing on one tower.
你不是在建造一个平平无奇的小土坡。
You're not just building a really flat hill.
我认为这正是我们的产品经理和设计师，以及整个公司，必须对我们要走的路径有更强信念的地方。
That's actually where I think there has to be more conviction from our PMs and our designers, the company really to have conviction of what journey we're going on.
但总体而言，感觉
But overall, feel
它运作得相当不错。
like it works pretty well.
比如，几乎所有的工程师都有足够的判断力，能意识到这个原型在产品中其实没有意义，或者有意义。
Like, people almost all the engineers have good enough taste to realize that this prototype doesn't actually make sense in the product, or it does.
所以，我并不常看到这样的原型。
So it's not that common that I would see a prototype.
就是觉得，哦，这完全说不通。
It's like, oh, this makes no sense.
大家做的事情都是合理的，问题只是我们先做哪些，以及通常是如何控制它们的开启和关闭。
It's people are doing reasonable things, and then it's just a matter of which things we build first, and then often just figuring out how to turn it on and off.
在我们的实验性聊天界面里，大概有一百个复选框。
There's, like, part in the in our, like, experimental chat UI, there's this there's probably, like, a 100 check boxes in there.
让我崩溃。
Kills me.
这些功能之所以能开启或关闭，是有原因的。
There's reason you could turn it on and off.
但我觉得，好吧。
But I think that okay.
没错，西蒙，确实是这样。
So that is true, Simon.
但作为负责评估团队的人，这确实给平台团队带来了额外的压力。
But, like, being the person that manages the evals team, like, there is a level of intensity that it adds to the platform team.
但如果我们要在Notion中实现图像生成，那么我们处理附件的方式、LLM完成请求的方式——比如Cortex的交互和期望返回的令牌——现在却要返回图像了。
But if we're gonna do image generation in Notion, all of sudden, the way that we do attachments and the way that we our LLM completion, like, Cortex talks and expects tokens back, and now it's getting images back.
我们需要巩固很多平台层面的工作。
There's a lot of platform work that we do need to solidify a little bit.
所以有时候功能在开发环境中要待上几周，才能上线，因为我们还得让它更稳定、符合HIPAA和ZDR合规要求，以及与供应商签订合适的合同等等。
So sometimes it'll be in dev for a couple weeks before it makes it to prod just because we still have to, like, make it robust, make it HIPAA compliant, ZDR compliant, figure out the right contracting with the vendor, whatever it is.
而且我们需要评估它，因为我们希望团队能继续维护他们所构建的内容。
And we need to eval it because we want the team to still maintain what they build.
这是最关键的一点。
That's the one thing.
如果我们有大量原型，就不能只让一小群人去维护这些原型里的东西。
If we have a bunch of prototypes, it can't just be like a small group of people that then maintain whatever in prototypes.
所以我们投入了大量人员在评估和模型行为理解团队上，我们称之为代理开发速度。
So we have invested a lot of people in an eval and model behavior understanding teams that we call it agent dev velocity.
因此，如果我们在这个平台上投入资源，你开发代理的速度会更快。
So your dev velocity building agents can be faster if we invest in that platform.
所以我们有一个完整的团队专门负责亚洲平台速度，让你能够构建自己的评估系统，并在发布后自行维护。
And so we have a whole org dedicated to Asian platform velocity so that you can build your own eval and then maintain it once you ship it.
所以如果新模型发布，我们每一个
So if a new model release comes out and we Every
团队各自维护自己的评估？
team maintains their own eval?
我们维护评估框架。
We maintain the eval framework.
每个团队都拥有自己的评估。
Every team owns their own evals.
其中很多我们已经集成到CI中，或者每晚运行。
And a lot of them we've integrated to opt into CI or we run them nightly.
我们有一个团队，一个自定义代理，用于触发团队查看主要故障。
And we have a team, a custom agent that triggers to a team to look at the major failures.
这非常关键，因为现在我们有这么多不同的服务，其中很多都运行在同一个代理框架上。
That's really critical because if we have all these different services now, a lot of it's on the same agent harness.
所以更容易维护。
So it's easier to maintain.
这仅仅是不同代理框架的打包。
It's just packaging of different agent harnesses.
但代理的新功能，比如我们想更新，像他们弃用了Sonnet 4之类的。
But new functionality of the agent let's say that, like, we wanna update, like, they deprecated Sonnet four or whatever it is.
我们需要自动更新了吗？
And we need to auto update already?
这样很好。
That's so okay.
好的。
Yeah.
三级。
Level three.
那不是很长时间吗？
Wasn't that long?
他们只是35。
They were just 35.
35。
35.
37刚刚被弃用了。
37 just got deprecated.
37。
37.
我认为
I think
是52。
it's a 52.
或者吧。
Or yeah.
不。
No.
不是5.2。
It's not 5.2.
不是5.2。
It's not 5.2.
是的。
Yeah.
是的。
Yeah.
5.4比5.2贵40%。
Five four is 40% more expensive than five two.
所以如果他们废弃了5.2，我会跟你说这件事的。
So if they deprecated five two, you would hear you would hear from me about that one.
但我还有另一个话题要谈。
But another conversation I have.
我有个调皮的评估问题想问你。
I have a cheeky evals question for you.
你有没有注意到任何主要模型提供商出现过隐性的性能下降？
Have you noticed any secret degradation from any of the major model providers?
隐性的性能下降？
Secret degradation?
在高峰时段的走廊上，它突然变得
During the walkway when it's high traffic, it suddenly gets
是的。
Yeah.
我们不仅注意到了不稳定性。
Not just between the we definitely noticed flickiness.
我们确实注意到，特别是某些提供商，在工作时间里速度变慢了，并且
We've definitely noticed particularly for some providers that things are slower during working hours and
这是一个延迟问题，而不是质量问题。
There's a latency argument, not a quality argument.
不。
No.
我认为有趣的质量差异在于，即使一些公司声称他们销售的是同一模型，实际上却涉及量化处理。
I think the quality difference that's interesting is even though companies that say they're selling the same it's really into like quantize quantization.
但像那些声称通过不同供应商（无论是第一方、Betterock、Azure等）销售同一模型的公司，我们确实有时会看到不同的质量表现。
But like, companies that say they're selling the same model through different vendors, whether it be through first party or Betterock, Azure, etcetera, we do see different qualities sometimes.
而这并不一定是他们宣传的内容。
And that's not necessarily what's advertised.
是的。
Yeah.
Kidney 提到过，如果我们对所有供应商进行统一的评估，就能非常清楚地看出谁在偷偷进行量化，而且非常明显。
Kidney went to the point of if you we'd be shipped like this, like, eval across all the providers, and it was, like, very obvious who was secretly quantizing, and it was very
那就是是的。
that's Yeah.
令人尴尬。
Embarrassing.
我们雇佣分包商来帮我们弄清楚这一点。
We hire subprocessors to figure that out for us.
所以我们只想了解哪里出现了倒退，或者哪里得到了优化。
So we just wanna understand where it's regressing or where it's optimized.
有时，如果倒退能优化延迟，我们也能接受这些适当的倒退。
And sometimes we're okay with regressions that optimize latency if they're the appropriate regressions.
我们的工作是确保我们有相应的评估手段，以理解对我们重要的变化。
Our job is to make sure we have the evals to understand the changes that are important to us.
甚至当我们与实验室合作测试模型的预发布版本时，他们也会给我们发送多个快照。
And even like when we're partnering with labs on prereleases of models, they'll send us multiple snapshots.
这与量化关系不大，更多是关于倒退问题。
And this is less about quantization, but more just regressions.
他们曾发布过不是我们想要的快照的模型。
Like, they have shipped models that were not the snapshots that we wanted.
他们根据我们的反馈调整了所发布的快照，因为我们的反馈更侧重于企业级应用，而非编码代理。
And they have changed the snapshots that they shipped based on the feedback that we give because our feedback tends to be more enterprise work focused and not coding agent focused.
当然，这些可能会带来意外的后果。
And definitely, those can be boomers.
我们知道这不是你们想要的版本，但我们愿意帮助你们让它正常运行。
We know that this wasn't the version you wanted, but we'll help you make it work.
我们总是能让它运行起来，但这种情况确实会发生。
We always make it work, but that definitely happens.
是的。
Yeah.
你们有没有一些失败的评估，只是希望当一个更好的模型出现时，它们能最终成功？
Do you have failing evals that you're just hoping Oh.
当一个更好的模型出现时，它们能最终成功吗？
That will have success eventually when a good model comes out?
是的。
Yeah.
所以我觉得我能讲这个话题六十分钟，但我得限制自己。
So I think I could talk about this for sixty minutes, so I will limit myself.
当人们提到评估时，我觉得这确实是个问题。
I think it's a real issue when people say evals.
这其实就是质量。
And it's just like, that's quality.
这就像单元测试，就像说测试一样。
That's like unit it's like saying testing.
这不仅仅是单元测试。
It's not just unit tests.
所以我们有相当于单元测试和回归测试的东西。
So we have the equivalent of unit tests, regression tests.
这些测试运行在持续集成中。
Those live in CI.
这些测试必须在一定的随机误差范围内通过一定比例。
Those have to pass a certain percent within some stochastic error rate.
然后，在你开发产品时，这些评估目前并未通过，但这是达到发布标准的质量。
Then we have, as you're building a product, evals of these aren't passing right now and this is launch quality.
所以我们有一份成绩单，要在这几类用户旅程中，达到80%或90%的通过率才能发布。
So we have a report card and we need to, on these categories, be at 80 or 90 of all of these user journeys to launch.
而我们所谓的前沿或余量评估，则是主动希望达到30%的通过率。
And then what we have, what we call frontier or headroom evals, where we actively want to be at 30% pass rate.
这实际上是我们过去两三个月与Anthropic和OpenAI合作推进的一项工作，因为我们当时遇到了评估饱和的问题。
And that's actually been an effort that we took in partnership with Anthropic and OpenAI in the past maybe two or three months because we actually hit a point where our evals were saturated.
我们无法提供有洞察力的反馈，只能说是没有变得更差。
And we weren't able to really give insightful feedback other than it wasn't worse.
这不仅对我们的合作伙伴没有帮助，对我们自己理解趋势走向也没有帮助，回到之前的类比。
And not only is that not helpful for our partners, it's not helpful for us to understand where the stream is going, going back to that analogy.
因此，我们花了大量时间思考Notion的终极考试应该是什么样子。
And so we spent a lot of time thinking about what Notion's Last Exam looks like.
对吧？
Right?
不只是人类的最终考试。
Not just humanity's Last Exam.
Notion的最终考试。
Notion's Last Exam.
关于它会是什么样子，有很多设想。
And there's a lot of dreams about what that would look like.
我知道我们已经谈了很多关于基准测试和SWICS的内容。
I know we've talked a lot about benchmarking, SWICS.
但是的，Notion的最终考试在公司内部是一个非常重要的项目，我们有专人全职负责。
But, yeah, Notion's Last Exam is a big thing inside the company, and we have people full time staff to it exclusively.
我们有一位数据科学家、一位模型行为工程师，以及一位全职的评估工程师，专门负责那些我们只有30%通过率的评估。
We have a data scientist, a model behavior engineer, and a full time evals engineer just dedicated to the evals that we pass 30% of the time.
你们在招聘什么？
What you're hiring for?
模型行为工程师？
MBEs?
我在招聘。
I am hiring.
MBE是什么？
What is an MBE?
模型行为工程师。
A model behavior engineer.
在我不来之前，模型行为工程师的头衔是数据专员，当时他们正在和西蒙一起用Google表格工作。
Model behavior engineers started with a title data specialist before I joined when they were working with Simon on Google Sheets.
西蒙只是需要一个人去看Google表格，然后说：是，不是。
And Simon just needed someone to look through Google Sheets and say, yes, no.
这个看起来不好。
This looks bad.
这个看起来不错。
This looks good.
所以我们招聘了具有不同语言学背景的人。
And so we hired people with diverse linguistics background.
我们招了一位没完成语言学博士学业的人和一位斯坦福大学比较文学专业的应届毕业生。
We had a linguistics PhD dropout and a Stanford comp lit new grad.
他们非常出色。
And they're amazing.
他们基本上建立了一个新的职能。
And they formed a new function, basically.
随着时间推移，我们组建了一个完整的团队，由一位经理领导，他现在正用编码代理重新定义这个角色。
And over time, we've built a whole team with a manager who's now reinventing what that role is with coding agents.
他们过去是手动检查代码。
So they used to be manually inspecting code.
现在他们主要构建能够为自己或大语言模型裁判编写评估的代理。
Now they're primarily building agents that can write evals for themselves or LLM judges.
有天特别有趣。
There's a really funny day.
我可以发你一张照片，是大约一年半前西蒙在教他们如何使用GitHub。
I can send you the picture where Simon, about a year and a half ago, was teaching them how to use GitHub.
他们在白板前。
And they're on the whiteboard.
我当时想，如果我们数据专家能学会使用GitHub并把这些问题提交到代码里，效率肯定会高很多。
And it was like, okay, I think it would be so much faster if our data specialists learned how to use GitHub and learned how to commit these things into code.
那是过去的事了。
And that was then.
而现在，编程已经变得容易多了。
And now I think coding has been a lot more accessible.
但未来，这将是一个融合了数据科学、测试、产品经理和提示工程师的混合角色。
But moving forward, it's this mix of like data science, test, PM, and prompt engineer.
因为要理解模型能做什么、不能做什么，这本身就是一门手艺。
Because there's craft in understanding, like, even what models can and can't do things.
我们该如何定义这种余地呢？
How do we define, like, that headroom?
我们该如何定义什么是好的体验？
How do we define, like, what a good journey is?
这个模型更好吗？
Is this model better or not?
为什么这个失败了？
Why is this failing?
有一些定性工作，但同时也需要很多直觉和品味。
There's some qualitative work, but then there's also like a lot of instinct and taste to it.
软件工程。
Software engineering.
因此我们有着非常坚定的信念。
And so we have like very firm conviction.
多年来我们一直认为，这是一条独立的职业路径。
And we have had for a number of years now that is its own career path.
我们一直欢迎那些特立独行的人，这么说吧。
And we have always welcomed the misfits, so to speak.
所以我们坚信，你并不需要有工程背景才能在这份工作中做到最好。
So we really firmly believe that you don't need an engineering background to be the best at this job.
而这正是这个特定职位的独特之处。
And that's what's quite unique about this particular role.
是的。
Yeah.
最近我一直对此感到非常兴奋，因为我们努力将评估系统视为一个代理框架。
This is something that I've been pretty excited about recently is we made an effort basically to treat the eval system as like an agent harness.
所以如果你
So if you
仔细想想，你应该能够
think about it, you should be able to have
实现一个端到端的代理，下载数据集，运行评估，迭代修复失败，调试，然后实施修复方案。
an agent end to end, download a dataset, run an eval, iterate on a failure, debug, and then implement a fix.
最终，你应该能够在一个有人类观察外部系统的前提下，驱动整个端到端流程。
And ultimately, you should be able to drive the full end to end process with a human sort of observing the the outer system.
所以，我们在这上面确实下了很大功夫。
So, yeah, we went pretty hard on that.
到目前为止，这非常有效。
That that's worked extremely well so far.
这基本上就是把它变成一个编码代理的问题。
It's like, basically, just to turn it into a coding agent problem.
你的编码代理，或者随便什么。
Your coding agent or just whatever
它应该完全通用。
It it harness be totally general.
是的。
Yeah.
我认为把它固定在某个特定的编码代理上是个错误。
I think it'll be a mistake to fix it on any particular coding agent.
归根结底，这就像命令行工具。
At the end of the day, it's just like CLI tools.
这就像让编码代理编写单元测试一样。
It's like the same way that you would have a coding agent write the unit test.
你应该让编码代理来编写评估代码。
You should have a coding agent write the eval.
是的。
Yeah.
但其中仍然有很多监督成分。
But there's a lot of supervision in that still.
我们只是不认为这种监督必须来自软件工程师，因为其中很多是用户体验研究和其他内容。
We just don't believe that supervision has to come from software engineers because a lot of it is UXRE and whatever.
这些也是负责排查故障并告诉我们下一步应该投入资源的人。
And these are the people that also triage failures and tell us where we should be investing next.
对。
Yeah.
我打算提一个有点尖锐的问题。
I'm gonna go ahead and ask a spicy question.
在Notion，有没有哪一天是没有软件工程师的？
Is there a day there are no software engineers at Notion?
成为软件工程师意味着什么？
What does it mean to be a software engineer?
没错。
Exactly.
我认为事情的发展趋势是，如果我们回看三年前，那时人类还在手动编写所有代码，后来有了自动补全功能。
I think the way things are going is, like, we're on some continuum where if you look back three years ago, humans were typing all the code, and then we had autocomplete.
你仍然在编写所有代码。
You're typing all code.
接着出现了能自动填充行的辅助工具，而现在我们正进入代理执行更长期任务的阶段——你可以调试并修复问题，验证修复是否有效，甚至完成代码审查、合并和部署。
Then we had, like, filling agents filling lines, and now we're getting into, like, agents doing longer range tasks where you can debug and implement a fix and then verify it works and get your PR even, like, merge and deployed.
我认为我们只是在向上提升抽象层级，而人类的角色逐渐转变为观察和维护整个外部系统。
I think we're just moving up the abstraction ladder, and then the human role becomes more about observing and maintaining the outer system.
有一连串的代理在处理合并拉取请求。
There's a string of agents flowing through merging PRs.
哪里出了问题？
What's going off the rails?
需要批准什么？
What do need to approve?
有没有什么学习或记忆机制是有效的？
Is there like a learning or memory mechanism that works?
这是一个棘手的工程问题。
So it's a hard engineering problem.
那里还有很多工作要做。
There's a there's a lot to do there.
我觉得我们只是在向上层迁移。
I think we're just we're living up the stack.
这和机器学习工程师经历的转变一样。
The same transition machine learning engineers have made.
我好一阵子没看过PR曲线了。
Like, I haven't looked at a PR curve in a while.
是的。
Yeah.
你以前做这些事，现在自动研究就能做到了。
You used to do this stuff, and now auto research can do it.
对。
Right.
我觉得这取决于你怎么定义软件工程师。
Like, I think it depends on what you define as a software engineer.
是的。
Yes.
这肯定在变化。
That's changing for sure.
我认为诺提恩今年夏天每一位软件工程师都经历了一种身份危机，我们公司的一位工程主管称之为：每个软件工程师都正在经历每个管理者曾经经历过的转变——突然意识到，自己写代码的能力不如委派任务和上下文切换的能力重要。
I think every software engineer at Notion this summer went through like this sheer one of our engineering leads at the company called it like, every software engineer is going through the identity crisis that every manager goes through, where all of a sudden they realize their ability to write code is less important than their ability to delegate and context switch.
我觉得这是一种从软件工程师身份中转变出来的过程。
And I think that is a transition out of being a software engineer.
但是是的。
But Yeah.
成为一名管理者有一个关键区别，那就是它实际上非常技术性。
There's a critical difference to being a manager, which is that it is actually very deeply technical.
问题是人类非常模糊，你不能把一群人当作一个严格的系统来处理，其中代码审查请求可以流入并处于阻塞状态。
The problem humans are very, like, fuzzy, and you can't treat a team of humans like a rigorous system where PRs flow through and can be in, like, a blocked status.
当他们被阻塞时会发生什么？
And then what happens when they're blocked?
对于一组代理，你实际上可以做到这一点。
With a set of agents, you actually can do that.
我认为这其中蕴含着许多有趣的技术严谨性。
Think it's actually there's a lot of interesting technical rigor that goes into that.
这归根结底是一个技术设计问题。
It's a it's a technical design problem, ultimately.
你正在构建的软件工厂的设计是什么？
What is the design of the software factory that you're building?
是的。
Yeah.
我觉得我们在尝试很多不同的东西。
I think we're trying a lot of different things.
最终，你希望设计一个尽可能减少人工干预的系统，但仍然保持你关心的不变性。
Ultimately, you want to design a system that requires as little human intervention as possible, but, like, still maintaining the invariance that you care about.
所以，是的，我们在那里探索很多不同的想法。
So, yeah, we're exploring a lot of different ideas there.
我觉得我谈到了一些我认为重要的事情。
I think I talk about a few things I think are important there.
我认为非常重要的一点是拥有某种规范层。
One thing I think is really important is having some kind of, like, specification layer.
你可以
You can
直接提交 Markdown 文件。
just commit Markdown files.
这效果还不错，但是
That works pretty well, but
能用Notion真好，老兄。
It's nice to be Notion, man.
我只是说，规范的天然归属地就是Notion。
I'm just saying, like, the spec like, yeah, the natural home for specs is Notion.
是的。
Yeah.
它可以是一个页面的数据库。
It can be a database of pages.
是的。
Yeah.
它需要是一种人类可读且易于理解的东西。
It needs to be something that is human readable and imbuable.
我认为这一点非常关键。
And I think that's pretty key.
另一个非常关键的组成部分是自我验证循环。
Another really key component is, like, the self verification loop.
是的。
Yes.
你需要一个非常优秀的测试层。
You need a really good testing layers, basically.
而要实现这一点是一个非常深层的问题。
And that's a really deep problem about getting that.
然后就是出现bug时的处理流程。
And then there's kinda like the workflow of what happens when there's a bug.
它如何流入系统？
How does it flow into the system?
是不是有一个子代理在处理它？
Is it like a sub agent working on it?
它如何创建拉取请求？
How does it make a PR?
这个拉取请求如何被审查和合并？
And how does that get reviewed and merged?
然后就是整个流程的走向。
And then so there's, like, the flow of process.
是的。
Yeah.
很棒。
Cool.
在你们来之前，我们已经确定了这个演示或这些智能体。
One thing we did work out before you guys came in was this demo or this Agents.
智能体演示。
Agent demo.
所以。
So
每次我们做一期节目，都会试用这个产品。
every time we do an episode, we try the product.
对吧？
Right?
我不记得有任何一集我没试过。
I don't think there's ever been an episode that I haven't tried.
‘试’这个词太重了。
Try is a big word.
从第一天起，Lanespace 就一直在用 Notion，但这个是新东西。
Since day one, Lanespace has been on Notion, but this is the this is the new thing.
是的。
Yes.
这是给 Kernel Labs 的，也就是我们所在的这个空间。
So this is for Kernel Labs, which is the space we're in.
下周，我们将开放租户申请。
So next week, we're opening applications for tenants.
所以有一个网页表单。
So there's a web form.
让我们来看看这个表单已经做成了什么样。
Let me we have this form done here.
以前，工作流程是我收到一封邮件，然后查看对方的信息。
Before, the workflow would be I get an email, then I look at the person.
我会想，我该花时间跟这个人聊聊吗？
Was like, should I spend time talking to this person?
然后我回复。
Then I respond.
他们再回我。
They respond back.
所以我做了这个东西。
So I build this.
所以这个名字是自己冒出来的。
So the name, it came up for on its own.
你能说说怎么
Can you maybe how do
它怎么自己想出名字的？
how does it come up with its own name?
是的
Yeah.
这个名字挺贴切的。
That's a pretty apt name.
它就是一个随机的名字生成器。
It's it's just a random it's a random name generator.
哦，原来如此。
Oh, okay.
这很有趣。
That's funny.
它只是突然冒出来的。
It just came
它选了这个名字，真是太好笑了。
The fact that it picked that is is hilarious.
我敢肯定它只是
I'm pretty sure it's just
坚韧收集者。
Resilient Collector.
我想我从未看过那部分代码。
I think I've never looked at the code for that.
我从未质疑过它。
I've never second guessed it.
我觉得这就像一个填字游戏。
I think it's like a Mad Libs situation.
我觉得没错。
Think it's Right.
是的。
Yeah.
这完全是确定性的。
It's totally a a deterministic.
我觉得这很棒。
I thought it was great.
是的。
Yeah.
尽管当你使用AI来自我配置时，它可以更新自己的名称。
It's Although when the if you use the AI to set itself up, it can update its own name.
好的。
Okay.
你是怎么创建它的？
How did you create it?
你只是用了石膏吗？
Did you just do Plaster?
你 okay 了。
Did you okay.
我确实做了。
I did yeah.
我会说，检查一下我的收件箱里关于共享办公空间的申请，保持它对人们可用。
I'll say just check my inbox for applications for a coworking space, keep it available to people.
所以它为我创建了一个数据库，我就在这里。
So it created a database for me, which I have here.
我想数据库就像一个Notion表格，因为所有东西都是用Notion做的。
And I guess database is like a Notion table because everything is Notion.
每当有邮件进来，比如这里，它就会为这个人创建一个新的条目，然后通过网络搜索来丰富他的资料。
And then whenever an email comes in, like here, it just creates a new role for the person, and then it uses web search to enrich the the profile.
所以它会搜索网络，确认这个人是谁。
So it, like, searches the web and is this is who this person is.
它会确认他们想什么时候搬进来，并更新其他所有信息。
This is when does they wanna move in and updates everything else.
这并不是AGI，但对我来说，我不想做这些工作。
This is it's not AGI, but to me, I don't wanna do this work.
所以我觉得我花大约十五分钟就设置了整个系统，我真的很喜欢这一点。
So it feels it took me maybe, like, fifteen minutes to set up the whole thing, and I really like that.
大部分信息都应该存放在这里。
Most of the information should live here.
它不像其他一些工具那样问我‘是的’。
It's not like some other tool asking me Yeah.
把我的东西搬到那里去。
To bring my stuff there.
我可能早就已经创建了一个Notion的东西了。
It's like, I would have probably already created an Notion thing.
所以
So
我们最大的使用场景和收益大多来自于流程中的人工介入环节，其中一个最重要的使用场景是缺陷分类。
Most of our biggest use cases and gains are from that extra layer of human involvement in the process to make it and, like, one of our biggest use cases is bug triaging.
比如，有人在Slack里发了一条消息，你就可以设置一个专属代理，它能自行判断这条消息属于哪个团队，自动在任务数据库中创建任务，并在该Slack频道中回复。
So if someone posts something in Slack, and you just have a custom agent that lives there that has its own routing constitution of what team this belongs to, creates a task in your task database, and then posts in that Slack channel?
这应该是我们最早内部开发的功能之一。
That's one of the first things that we built internally, I think.
它彻底改变了Notion作为一家公司的运作方式。
And it's completely changed the way that Notion functions as a company.
没有任何事情会被遗漏。
Nothing falls through.
大多数事情都不会出错。
Most things don't fall through the crack.
我们不知道自己不知道什么。
We don't know what we don't know.
但它并没有取代人。
But it's not replacing people.
它取代的是流程。
It's replacing processes.
是的。
Yeah.
我很想知道你是如何看待这些系统的可组合性的。
And I'm curious how you think about composability of these things.
我最近在做的另一个是类似补丁填充的东西。
So the other one I was working on is like a piece filler.
所以每当有人注册成为租户并填写租赁合同时，应该有一个类似办公室管理员的代理来处理请求、生成租赁合同，并为他们提供Verkada办公室访问权限等。
So whenever somebody signs up as a tenant, fills up the lease for them, there should probably be some agent that is, like, office manager agent that can handle the request, make the lease, and then give them a verkada access to the office and all of that.
你如何看待这个未来？
How do you think about that future?
是的。
Yeah.
你可以通过两种方式来组合。
So there's two ways you can compose.
一种方式是使用数据原语。
One way is by using, like, the data primitives.
你可以让一个代理向数据库写入数据，另一个代理则监控数据库。
So you can you you can give you have one agent be writing to the database, there's another agent that's watching the database.
这是一种它们可以协调的方式。
So that's one way that they can coordinate that.
这种方式稍微松散一些，效果非常好。
It's, like, a little bit more decoupled and works really well.
或者你可以将它们耦合起来。
Or you can couple them.
所以我认为它还没有发布。
So I think it's actually not released yet.
发布后，下周在代理的设置中，你可以获得调用任何其他代理的权限。
Releasing it, like, next week is in the settings for an agent, you can get access to invoke any other agent.
所以你可以让它们直接通信。
So you can have them just just talk directly.
这就是你所说的。
That's what you Was
递归次数有上限吗？还是可能没有？
there a limit on number of recursions or just Probably.
比如，你可能会陷入无限循环。
Like, you can just get an infinite loop
那样的话。
that way.
有一些
There's some
有点是的。
kind of yeah.
我觉得我只是想
Think it's I'm just like
实际上 somewhere 有一个数字。
There is actually a number somewhere.
这是一个线索。
It's a lead.
我只是觉得，有人会搞砸的。
I'm just like, you're someone's gonna screw it up.
你应该
You should
直接试试看。
just try it and see.
是的。
Yeah.
所有东西都会变成回形针。
Everything's gonna be paperclips.
哦，是的。
Oh, yeah.
是的。
Yeah.
但那确实很有用。
But but that's really useful.
是的。
Yeah.
所以我们就像，前几天我刚帮内部一个人忙了。
So we like, I just I helped someone internally the other day.
他们为我们的市场推广团队构建了三十多个自定义代理，做各种不同的事情。
They had built, like, over 30 custom agents for for our go to market team doing all kinds of different things.
比如，研究客户信息、填写客户资料，或者筛选客户反馈之类的。
For example, like, researching like, filling information about about a customer or, like, triaging customer feedback or something like that.
整整三十多个。
Literally over 30 of them.
然后他给了我一个包含所有代理的数据库。
And then he gave me, like, a database of all the agents.
然后他说没问题。
And then he was okay.
现在我每天收到七十多条通知，都是因为这些代理在各种事情上被卡住了。
And now I'm getting 70 over 70 notifications per day with just the agents are blocked on various things.
然后我就想，哦，明白了。
And then I was like, oh, okay.
不错。
Cool.
很显然，应该做一个管理型代理。
The obvious thing to do there is to make a manager agent
对。
Right.
那会形成阻塞。
That's gonna Blocks.
在你那30个代理之间增加另一层抽象。
Be another abstraction layer in between your 30 agents.
所以，是的，我们设置了一个管理代理，它能调用所有其他代理，并且一直在观察它们。
So, yeah, we set up with a manager agent, and then has access to invoke all the other agents, and it's, like, watching and observing them.
然后它就建立了一层抽象。
And then it's it just creates a layer of abstraction.
所以每天的提醒从70次减少到了大约5次。
So instead of 70 notifications per day, it's, like, five.
而管理代理可以帮助调试和解决这些代理的任何问题。
And then the manager agent can help debug and fix any any problems with the
这就像一个收件箱的概念，因为你实际上是在说它们可以互相发消息。
This is a concept of an inbox or something like because you're you're basically saying that they can message each other.
是的。
Yeah.
使用一个记录系统，比如 Notion。
Use a system of record, which is Notion.
是的。
Yeah.
所以我们实际上确实如此。
So we actually yeah.
我们根本没有创建任何特殊概念。
We didn't make any special concepts at all.
他们感兴趣的是
They're interested in
我本会收到的那些通知。
the motion notifications that I would have got.
他们可以直接向数据库写入一条记录，由其他代理监听；或者直接向代理发送 Webhook。
They can just write a test to a database that the other agent's tasked to listening to, or they can actually call a webhook to the agent.
比如，他们可以直接添加这个代理。
Like, they can just add the agent.
好的。
Okay.
是的。
Yeah.
这正是我们还在处理的事情。
This is something that that we're still working on.
我觉得是的。
I think Yeah.
我们通常的做法是，先以一种可能有点粗糙的方式让它能跑起来。
We like, generally, the way we do these things is you first make it possible and maybe, like, sort of janky way.
所以我认为我设置的方式是，我们创建了一个新的数据库，用来记录自定义代理遇到的问题，然后给他们所有人权限提交问题。
So I think the way I set them up is we created, a new database that was, like, issues that the Custom Agents were were experiencing, and then gave them all access to file an issue.
然后经理可以访问并查看这些问题。
And then the manager has access to to read the issues.
这效果还不错。
And that worked pretty well.
本质上，为这些代理专门建立了一个内部问题追踪系统。
Essentially, gave it its own, like, internal issue tracker just for the agents.
如果这个概念后来显得普遍有用，我们可能会考虑如何将其整合进来。
And then if that becomes a concept that seems useful, generally, maybe we'll think about how to package it in.
但通常来说，只要可能，我们都尽量只使用基础组件来构建。
But generally, we try to just keep it to composing the primitives if we can.
另一个例子是我们没有内置的记忆功能。
Another example of this is we have no built in memory concept.
记忆就是页面和数据库。
Memory is just pages and databases.
所以如果你想给它记忆功能，就给它一个页面，并赋予它编辑该页面的权限。
And so if you wanna give it memory, just give it a page and give it edit access to that page.
然后人类也可以编辑它。
And then human can edit it.
代理可以编辑它。
Agent can edit it.
是的。
Yeah.
所以
And so
这种模式在它上面效果非常好。
that works that pattern works extremely well on it.
根据使用场景，你可以让它只是一个页面，或者是一个完整的数据库，也可以有子页面。
And depending on the use case, you can have it be just a page, or it could be an entire database with or I can have sub pages.
这很直接，所以你可以这么做。
It's pretty honest, so you can do that.
所以当我设置这个时，我连接了我的收件箱，它问我：你想用 Gmail 还是 Notion Mail？
So when I was setting this up, I connected my inbox, and it was like, do you wanna use Gmail or Notion Mail?
我说：我不想用任何一个。
And I'm like, I don't wanna use either.
我只是希望你来处理它。
I just want you to do it.
我想知道你是如何看待Notion Mail、Notion日历以及Notion的这些UI/UX设计的。
I'm curious how you think about Notion Mail, Notion Calendar, all of these kind of UI UX info Notion.
是的。
Yeah.
当同时有代理以某种方式将它们从你身边推开时，你会怎么想？
When, like, at the same time, you have the agents obstructing them away from you in a way.
你是如何分配产品资源的，
How do you spend, like, the product calories,
这么说吧？
so to speak?
对。
Yeah.
我认为，不必非得使用Notion Mail才能连接邮件功能，我们可以直接连接Gmail或你希望使用的任何邮件服务。
I think it's pretty important that you don't have to use Notion Mail to connect to the mail capability, so we can just connect to Gmail or whatever you want to use.
我们把邮件服务视为非常出色，因为它完全是基于智能代理构建的。
And we're thinking of the mail service as being really great to the extent that it's really agent built.
所以，也许邮件应用只是一个预装的智能代理，帮助你自动化收件箱。
So maybe the mail app is just prepackaged agent that helps you automate your inbox.
是的。
Yeah.
自动分类标签非常棒。
The auto labeling is great.
我认为
I think the
当我们与Gmail集成时，例如，我们可以使用通过MCP或API提供的系列工具。
When we integrate with Gmail, for instance, we have a series of tools available that are available via MCP or API to Gmail.
当我们集成Notion Mail时，我们可以依靠Notion Mail的工程团队为我们打造恰好合适的工具，以优化延迟、性能和质量。
When we integrate with Notion Mail, we have the Notion Mail engineering team to build us the exact right tools that optimize latency, optimize performance and quality.
他们对这一质量负责。
They own that quality.
那里有一些产品负责人，他们直接关注邮件中用户遇到的问题。
There's product leads there that are directly thinking about the user problems that happen in mail.
因此，当我们构建集成和连接时，通常会先原生开发，然后再考虑将其通用化，因为原生开发也更容易实现。
So it tends to be when we build integrations and connections, we build natively first and then think about extending them generally just because it's also easier to build natively first.
所以这通常就是我们逐步推进的方式。
So that tends to be how we phase things out.
说到集成，你提醒了我，我得问问MCP和CLI现在怎么样？
Talking about integrations, you prompted me, so I gotta ask MCP, CLI, what's going on?
你怎么看Yeah。
What's the Yeah.
看法？
Opinion?
我觉得我 definitely 对CLI持乐观和兴奋的态度。
I think I'm definitely bullish and excited about CLIs.
我觉得CLI有几个非常酷的地方。
I think there's a few really cool things about CLIs.
其中一个很酷的地方是它运行在终端环境中，因此获得了额外的强大功能。
So one really cool thing is is that it's in the terminal environment, so it gets a bunch of extra power.
例如，我可以分页并滚动浏览很长的输出内容。
So for example, I can, like, like, paginate and cursor through, like, long outputs.
而且它天然具备渐进式披露的特性。
And it has progressive disclosure inherently.
你不会一次性看到所有工具。
You don't see all the tools at once.
你只会看到 CLI 封装层，然后可以通过帮助命令和读取文件来操作。
It's just you see the CLI wrapper, and you can use the help commands and read files.
我认为最棒的一点是，它本身是自举的。
And then I think the most important thing that's super cool is that there it's also inherently bootstrapped.
所以如果出现问题，代理可以在使用该工具的同一环境中调试并修复自己。
So if there's an issue, the agent can debug and fix itself within the same environment that it uses the tool.
对吧？
Right?
我想今天早上看到一条推文。
I think I saw a tweet this morning.
有人说，我的代理没有浏览器，所以我让它变成一个浏览器工具。
Someone said, my agent didn't have a browser, so I asked it to make it a browser tool.
在不到一百行代码内，它给我们提供了一个小浏览器，封装了Chromium API。
And within a 100 lines of code, it gave us a little browser, like, wrapping the Chromium API.
这简直太不可思议了。
That's pretty incredible.
如果出现bug，它会立即尝试修复。
And then if there was a bug, it would just immediately try to fix it.
另一方面，如果你使用Chrome DevTools MCP，我就遇到过这种情况：有时候传输会出问题。
On the other hand, if you use an if you use, like, the Chrome DevTools MCP, I've had this issue where, like, sometimes the transport gets, like, messed up.
如果出问题了，代理就无法自我修复。
If it gets messed up, the agent has no way to fix itself.
它不再拥有浏览器了。
It no longer has a browser.
它没坏。
It's not broken.
我觉得这相当根本。
I think that's pretty fundamental.
但我会说，它很多缺点是可以修复的。
But I would say, like, a lot of the bad things about it can be fixed.
所以我觉得，渐进式披露可以通过合适的框架来解决。
So I think, like, the progressive disclosure, that can be fixed with the right harness.
显然，没必要一直展示所有工具。
Like, it it obviously doesn't make sense to show it all the tools all the time.
这并不是MCP协议固有的问题。
That's not really inherent to the MCP protocol.
这只是你如何封装和使用它的问题。
It's just, like, how you wrap it and use it.
由于我们当时不了解，很多FCP实现得都很差。
There's many poorly implemented FCPs because we didn't know.
是的。
Yeah.
我的意思是，那时候还太早了。
I I mean, it was just early.
显而易见的做法是，一开始就把所有工具都展示出来。
Like, the obvious thing is, to start with, is to just show it all the tools.
这没问题。
And it's okay.
现在我们有一百个工具了。
Now we have a 100 tools.
是的。
Yeah.
而且工具调用确实能正常工作。
And the tool calling actually works.
这是你成功的代价。
Let's The victim of your success.
给它一个方式，比如过滤或搜索工具。
Give it a way to, like, filter to search the tools.
是的。
Yeah.
我认为， broadly speaking，我对命令行界面非常看好。
I would say, like, broadly speaking, I'm really bullish on CLIs.
我仍然对在环境中使用 MCPs 很看好。
I'm still bullish on MCPs in in a environment.
我认为，特别是 MCPs 非常适合当你需要一个狭窄且轻量的代理时。
I think, in particular, MCPs is really great for when you want, like, a narrow lightweight agent.
我认为确实有很多
I think there's definitely a lot of
适用场景是，你并不需要一个带有计算运行时的完整编码代理。
use cases where you don't want, like, a full coding agent with the compute runtime.
而且你还希望它的权限更加严格。
And also, you want it to be, like, more tightly permissioned.
MCP 本身具有非常强大的权限模型。
MCP inherently has a really strong permission model.
你只能调用这些工具。
Like, all you can do is call the tools.
CLI 则要模糊一些。
A CLI is a little bit murkier.
我能访问 API 密钥吗？
It's can I access the API token?
你有没有正确地重新加密密钥，防止它被窃取？
Are you, like, properly, like, re encrypting the token so it can exfiltrate it?
它带来了很多真实且难以解决的新问题。
It introduces a lot of, like, new issues, which are real and hard to solve.
MCP 就是个简单直接又管用的东西，效果相当不错。
And MCP is just, like, the dumb simple thing that works, and it it is pretty good.
我再补充两个视角，不是关于它在 Notion 上表现如何，而是 Notion 如何同时支持这两个平台。
I'll add two more perspectives, not from it working well for Notion, but how Notion, like, commits to both platforms.
Notion 致力于成为人们开展企业工作时最优秀的记录系统。
Notion is dedicated to being the best system of record for where people do their enterprise work.
因此，只要其他人还在使用 MCP，我们就一定会继续支持我们的 MCP。
So we will always support our MCP insofar as other people are using MCPs.
对吧？
Right?
所以，无论我们的观点如何，我们都为 MCP 投入了大量精力，并正在组建一支出色的团队来进一步拓展这方面的工作。
So regardless of our perspective, we've put a lot of effort into our MCP, and we have a fantastic team that we're building to do more there.
我想说的第二点是，我们大家可能都经常思考这个问题，但最近我一直在认真思考如何确保价值、定价与能力之间保持一致。
And the second thing I'll say, I think we all think a lot, but lately, I've been thinking a lot about making sure there's a value alignment and pricing with capability.
就这个问题本身而言。
Literally on this question.
为执行确定性任务而依赖语言模型显得很浪费。
And needing language to execute deterministic tasks feels wasteful.
对于不需要语言模型的任务，却要求其与第三方服务商交互，也是一种浪费。
And requiring on a language model to interface with third party providers seems wasteful for tasks that don't require it.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。