高纪
嗨，我是高纪。本周《十字路口》请到的嘉宾是 Paperboy 的创始人蒋阳，以及他们的一位 Founding Engineer，杰特陈。你好，二位，欢迎来到《十字路口》。
酱阳
Hello，高纪，感谢。
高纪
Hello。这里也给大家说明一下，因为酱用英文表达会比较自然，所以我们这一期的内容会比较中英夹杂。我们同时有音频和视频，如果大家需要的话可以去看视频，视频里会配字幕和翻译，帮助大家理解。
我们还是从《十字路口》的传统快问快答开始。有请二位的年龄。
酱阳
I'm 21.
杰特陈
19。
高纪
你们的毕业院校？
酱阳
Maybe at Pratt Institute. I study architecture.
杰特陈
我是上海新虹桥双语学校高中毕业，然后现在在 CMU 读大一，大一刚结束。
高纪
你们的 MBTI 和星座是什么？
蒋阳
ISTJ，双子座。
杰特陈
INTJ，处女座。
高纪
创业之前你们做了一些什么？
酱阳
Paperboy 是我第二家公司。我的第一家公司叫 Million，我们是在 React ecosystem 里做了很多开源的 dev tools，然后我们做了一个产品叫 same.new。普通人可以用一句话输入一个 URL，然后做出和任何网站一模一样的 UI。And uh, Melanie was in YC Winter 2024.
杰特陈
我在 Paperboy 之前是个高中生。我比较喜欢做开源项目，也喜欢打 CTF。我做了一个东西叫 EarthKit：你拍一张照片，然后我用多模态模型去推测这张照片是在哪里拍的，会比传统的纯神经网络模型更好。
高纪
那是哪一年？
杰特陈
差不多一两年前。
高纪
那我们来介绍一下，Paperboy 是一个什么样的产品？
蒋阳
1. Paperboy Needs A New Interface
Paperboy is a company I just started, roughly with the mission of trying to figure out the way I most like to work with AI.
我在去年做完 same.new 以后，对当时用过的那些产品都有一些不满。So Paperboy just started with me trying to hack around different approaches to figure out how to solve this. There are technical problems here, but also form-factor problems. 我不应该需要把自己的文件、email，还有 all this personal information 丢到一个聊天框里。
如果我需要和别人一起合作，同时跟一个 agent 对话，这应该是一个非常容易的方式，让我们在同一个 context window 里协作。Agents 如果知道了我的很多信息以后，应该可以 proactive 地提前帮我做一些事情，但是现在的 chat window 做不到。
另外，现在所有产品基本都是 session-based 的。Session 一多，你就找不到过去要从哪个 context window 继续聊下去了。所以，我觉得这些问题总体来说，都是 model capabilities 和 deployment 之间的差距。我觉得在产品体验以及这些产品上，仍然有很大的机会去探索和创新。
高纪
我们待会会展开一下，但我们先把快问快答做完。你的融资情况是什么？
酱阳
We raised $4.7 million at 25.
高纪
Cool。收入和利润呢？现在还没有发布，对吧？
酱阳
Zero, and negative gross margins. We lose money every day.
高纪
你们估计什么时候会发布产品，让大家用到？
酱阳
We shipped the prototype to friends, and the idea is that our agent learns from OS activities, but there are cost issues. It's too expensive; it doesn't really work.
The next iteration of this product, we're trying to wrap it up by this month, and we're going to try to go to market again.
高纪
好。我们目前的团队规模呢？
酱阳
现在有 12 个全职员工，其中有 10 位 engineer。
高纪
我记得第一次见到酱的时候，你给我看了一个文档。那个文档应该是你写出来，给内部开会时用的一个会议文档。当时第一句话就是，你说，人类和 AI 配合工作的最佳方式很可能还没有被发明出来。
写这句话的时候你看到了什么？到今天，你觉得你看到的事情发生了一些变化吗？
蒋阳
2. The Thesis Keeps Moving
Yeah, that doc is the one I prepared for our very first all-hands. 那个时候公司就我、杜哥、陈、Jet 四位，时间是 11 月 30 号。
We just started working with one thesis, which is that the best way to work with AI hasn't been invented, and we have a shot at being the team that figures out that way. Cursor started really trying to figure out what's the best way to code with AI, and they have been largely successful. I think they were the first company to really take on that goal, and they figured out how important it was to focus on that goalpost before everybody else.
你问我，从这句话说到现在还有什么新的学习？I think what's cool is，这其实是一个 moving goalpost。你无法真正达到市场的预期，只能不断变得更好。因为每一次你提出新的东西，其他人都会看到。如果其他团队和用户有 taste，他们就会发现，也会看到其中的 friction points，而 friction 总是会存在。所以你要做的就是不断变得更好。
I think it's really a moving goalpost. For instance, OpenCloud came out, and we have Cloud CoWork, which is Anthropic's.
高纪
对，所以最近很多 founder 感到有一点绝望，因为在创业的过程中不断有新的王炸出现。正好你们从创立公司到现在这半年，也是变化非常剧烈的半年，从 Cloud Code 到 Open Cloud，再到 Arms。
所以你的感受是什么？你一开始想象中要做的东西，和今天在做的东西，有因为这些不断出现的庞然大物而发生剧烈变化吗？
酱阳
还真没有什么剧烈的变化。I think the problem space comes from people. 从技术和产品的角度来看，我认为这个 problem space 大概有 3 类问题。
第一，你需要让 agent 真正能够从用户的环境中学习。它必须整合进用户已经在使用的东西里，而数据来源包括电脑上的大量文件以及各种 integrations。
第二，你需要让它个性化。Personalization 意味着你不需要频繁地 prompt 它，可以信任它去处理越来越复杂、越来越困难的事情，也可以信任它做更高风险的决策。这也意味着它要更加可靠。也就是说，你可以真正使用这个模型，让它持续运行越来越长的时间。
第三，从设计的角度来说，它需要极其直观。你不应该需要把它当成一个新工具重新学习。如果你的 agents 是 proactive 的，它们可以自己提出新的想法，那么，什么样的形式可以把所有这些 proactive 的输出整合在一起？它必须存在于用户原本的环境之中，并且是个性化的，也必须能够和你现有的团队很好地协作。
高纪
我们要不要非常简单地安利一下大家，为什么要用 Paperboy？因为听我们播客的很多人，可能已经是重度的 Claude Code、Manus 或者其他 agent 用户了。为什么要在这个时候给 Paperboy 一次机会来用？
蒋阳
3. Paperboy Breaks The Session Model
Yeah, so Claude Code and Manus, to me, are kind of the most successful forms of agents out there. They are session-based and one-to-one, prompt-based. So these are 2 very important factors here。
Session-based 意味着，在它们的 sidebar 里有一堆 workspace，也就是 projects，然后每个 project 下面又有一堆 sessions。每次你想让模型做一些新的事情，就要开启一个新的 session，再把它归类到一个 workspace 里。
第二件事是，你和模型的互动方式是：你 prompt 一些东西，然后等待；接着你再发送一条消息，得到另一个 assistant response。就这样，一轮一轮地对话，然后再发送下一条消息。
这个问题在于，第一，agent 是 reactive 的，你必须非常具体。你可以创建一堆 skills，或者像 `.md` 文件这样的东西，去描述你正在做什么，但你必须主动维护这些内容。不管你投入多少精力去维护，都很难把你对品位和判断的全部了解，以及你认为做事的最佳方式，转译成这些原始的文本文件。
第二个问题是，sessions 并不是连续的，这很糟糕。现在很多人可能已经有几百个，甚至几千个 sessions 了。我知道过去有些 sessions，其中的 context window 里可能包含了一些非常有价值的洞见。但如果我没有刻意把它保存下来，比如明确记录“这是我们在这个项目的所有不同 sessions 中最重要的部分，这是我们需要带到未来的东西”，那么这个 session 基本上就会丢失。
高纪
So, to me, that's like—what is Paperboy?
酱阳
Paperboy 正面解决了这 2 个问题。
第一，agents 必须通过观察你如何使用电脑，自己学习。这里面包括你的截图、键盘操作、嘴部动作、你的会议，包括视频和音频、浏览历史、iMessage。当然，你可以选择给 Paperboy 哪些权限，也可以选择你想让它访问什么。
第二，它需要存在于 chats 里，并且拥有比单个 context window 长得多的、持续存在的对话历史。这些 chats 还需要是可搜索的。标准的产品形态其实就像打开 iMessage 或者微信一样，你有一堆 chats，然后在一个 chat 里持续和其中的参与者对话。
杰特陈
我觉得还可以补充一点关于前面 session 和 context window 的事情。其实现在类似 Claude Code、Manus 这种产品，你可以 argue 说它们有无限的 context window，因为它们有 compaction。然后其实也有更新的产品，比如一家叫 Interaction 的公司出的 Poke、ZOO Computer，甚至 OpenAI，它们也都是以一种不存在 session、但一直保持 chat 形式的方式，去做 agent 和 human 的 interaction。
但是对于 Paperboy 来说，我们和这类产品的一个主要区分点是，这些产品的 context 来源，主要要么是用户和 agent 过去的 chat history，纯粹从这里去 learn；要么就是用户把 email 或者 messages 之类的信息提供给这些 agent，它们可以通过读取这些 email，逐渐更加 personal 地学习关于用户的更多信息。
我们一开始做 prototype 的时候，其实也试图导出用户所有的微信聊天数据，或者 iMessage 的所有聊天数据。但我们很快就发现，这其实并不是一个非常 scalable 的方式。最 scalable 的方式，其实是通过观察用户日常使用电脑的行为，也就是从 OS 层面去 collect user data。
我们发现，这样的话，就会对用户每天在做什么事情有一个非常全面的了解。而且按照信息的浓度来说，用户一般的 computer use 信息浓度是非常高的。你只要观察用户用电脑差不多 60 分钟，就可以学习到用户很多的信息。但如果你只观察用户过去 60 分钟的微信聊天数据，其实学不到太多东西。
所以我们很早的一个 bet，就是通过 OS 层面的 context 去做 user adaptation。
高纪
嗯。《十字路口》之前有一期播客的嘉宾，是 LJ 的 founder。他们其实也在做一个 Mac 或者 Windows 的桌面客户端，以此来 capture 用户尽可能多的上下文和电脑使用习惯。
最近 OpenAI 应该也出了一个叫 Chronicle 的产品，对，这也是类似的一个想法，就是用桌面客户端来抓尽可能多的上下文。所以你们在做的时候，和大家是类似的实践吗？还是你们有什么不一样的想法和做法？
蒋阳
4. The Operating System Learns You
从 Mac 以及用户使用的电脑上抓取 raw data，再把它处理成 memories，这是一个普遍趋势。它不只是会发生在像 Air Jelly 这种刻意专注于此的 startup 上，Codex、Clockwork、Clock Code 也都会这么做。
这有点像是下一个最明显的 context frontier。但当然，不同的人会以非常不同的方式处理这股 raw data，也会以非常不同的方式把它处理成 memories。你选择什么信息，以及你选择如何组织这些信息，都和 agent 的应用场景直接相关。
每个 agent，以及你的应用真正决定要帮助用户做什么，可能都不应该采用同一种结构。你不可能用同一种方式去压缩 raw activity stream。比如说，如果一家公司特别专注于研究用户如何回复 email，那么它们可能会有一套不同的 memory structure，也会用不同的方式去压缩 raw emails，从中学习。
The action space is big enough that there is enough room right now to be able to be the first to get to that. 真正重要的是理解用户在所有应用界面以及各种关系中的状态，然后你就可以去建模用户实际上每天在做什么。
酱阳
Yeah, it’s just that capability alone. I think it’s going to be so obvious. Everyone’s jumping into that, but depending on the application, there’s room to customize and change the algorithm.
我也认为，现在通过收集用户屏幕数据，或者 computer use，去 build 这个 context，已经在一定程度上成为了一个行业共识。然后后面更重要的就是说，在这个范式下你具体去做什么事情。目前像 Codex，或者说类似 Littlebird 这种，他们会把这个屏幕数据当做一个 context layer。比如说对于 Codex Chronicle，它目前的 use case 是，你可以通过收集用户屏幕数据，去学习用户一般是怎么去 develop 这个 application 的。
如果你的 use case 不一样，最后这个 pipeline 也会不一样。我觉得这是一个非常新的 domain。在收集用户屏幕数据，以及收集 computer use 情况的基础上，你其实还可以做非常多其他的事情，这需要非常多的 engineering 和 research。比如说，你如何去做最好的 proactive agents：你是在预测用户下一个 keystroke，还是预测用户后面 1 个小时要干什么，还是预测其他的事情？这些都是相对来说还没有被很好 explore 的 problem space。
目前我认为，也没有任何人发现一个绝对最好的 recipe 去做这件事情。所以我觉得，对于一个公司来说，现在去探索这个，还是一个非常好的选择。
荣辉
那如果今天用户安装了 Paperboy，他在第 1 个小时，甚至前 5 分钟，能够感受到的最大价值是什么？你们期待用户在第 1 时间感受到什么样的差异和价值，从而能够留下来？这个 highlight 点是什么？
蒋阳
5. Meeting Prep Builds Trust
We’re probably going to start with meeting prep. An important thing for a small program is that you have to demonstrate to the user a framework of what your product can do. You know, it has to set expectations. One thing about Paperboy is that it gets better the longer you use it, so you sort of need that initial period where the user trusts it to learn.
Yeah, and so you can open it up. You can see a chat box—not a chat box like a prompt box, but a literal chat window. Once you give it permissions to, say, your calendar and your emails, it’s going to start reading what you give it. Then it’s going to start asking a little bit about who you are, based on the context it’s reading about you, and it’s going to start giving you a couple of suggestions, like, “Hey, I see this meeting coming up. Do you want me to look into that for you?”
By interaction, they sort of figured out the recipe here, which is: you just connect to the user’s existing context, and you need to show the user that you are an agent that can actually engage with their personality and proactively, in a non-annoying manner, help the user. So you get the user into this expectation that we are the agent that can actually text you in a way that makes sense.
荣辉
Yeah, so you have Vivian here as well?
Vivian
Hello.
蒋阳
6. Agents Become Team Members
Yeah, Vivian is very funny. She joined from Xiaohongshu. She’s on the Hongshan team. We have Vivian’s Paperboy, which is called Community Vivian. My Paperboy is called AutoJohn, and they exist in Slack right now.
All the time, teammates on Paperboy would just ask AutoJohn questions, so it’s able to handle all the incoming questions and help teammates on the product and design teams find the help they need. For MiniVivian, for example, Vivian does a lot of recruiting work, so MiniVivian is the recruiting intern for her on the team. MiniVivian actually understands everything that I’ve ever said to it across meetings and Slack, including my judgment and taste around what kind of candidates we’re trying to hire and where we want to source from.
It’s able to actually help Vivian surface these candidates across GitHub, Xiaohongshu, and Twitter. It’s something that saved her a lot of time. I think since February, Vivian has not used Claude. She can’t use Claude because it just doesn’t know. You can’t use Claude to research candidates for you. There are so many things you would have to tell it about your judgment of the person. Because it has more context, you’re actually giving it prompts.
荣辉
这个时候，你甚至可以不 prompt。
蒋阳
Yeah, I hate prompting. For the same reason, I just don’t want to prompt. I mean, you have to communicate, but people—we don’t think in prompts. We send text, you know, and we expect the other person to know what we’re talking about. We enjoy a relationship when the communication is high-bandwidth.
I think smart people enjoy being told what we don’t know, especially what we don’t know that we don’t know. Today’s models are smarter than we are, so frankly, I look forward to the day when I can just lay back and have AutoJohn be the smarter, higher-IQ existence than I am for the rest.
荣辉
Okay, yeah. Then, Jet, do you have any aha moments from using Paperboy?
杰特陈
首先就是，确实很多时候，和 John 的 AutoJohn 聊天会比和 John 聊天好。但你在和他的 AutoJohn，也就是他的 agent 替身，对话的时候，会担心它目前的意志还不能完全代表 John，从而产生一些 misunderstanding。
酱阳
So the way I look at it—not just because we are creators of this thing, but the user has to be responsible for their agent eventually. The way AutoJohn gets set up, it’s not just like, one day, boom, you have this existence that you can bridge into Slack. There is an onboarding process where the agent actually asks you questions, like, “Okay, how much am I allowed to share with this specific person?”
By default, it mirrors. For example, I share a lot with Jet. Maybe I don’t share as much with a new engineering hire, and the agent actually knows that from observing all my chats.
杰特陈
而且我觉得，一般和一个人 interact 其实分很多种。有一种，你可能就纯粹是和 John 聊天，或者说我需要去做一件事情，但是我没有权限，需要 John 去 approve 一个东西。这些东西我觉得目前可能就是我不会去和 AutoJohn 这样说的。
但是实际上，在工作场景中，你很多时候和对方的一些 communication，是因为 John 有整个 Paperboy 公司的 context。然后我作为一个工程师，或者我作为一个 IC，我每天需要想的是：我如何能对这个公司产生最大价值？我去做什么 research，或者做什么 engineering，是最有用的？
因为这个 agent，第一，事实上它有 John 的大部分 context；第二，它通过观察 John 工作的时候，它的 heuristic 其实和 John 的 heuristic 是比较类似的。所以说，我在做决策的时候，无论是 context 方面，还是 heuristic 方面，AutoJohn 都是比较有用的。
7. Text Completion Goes Everywhere
对于我来说，第 1 个 aha moment 其实比较早，就是我们做出 text completion 的时候。对于我平常编程来说，它已经挺有用了。现在有很多 AI terminal 很火，但是实际上那些命令行工具没有很多传统命令行那么流畅，而传统命令行的那些 app 很多时候又没有 AI integration。很多时候你去写脚本会很烦，但是我们做出 text completion 之后，我完全可以写完一堆代码，然后我要去发一个 git commit 的时候，直接在命令行里面输入 `@pb commit`，它就会自动把整个 commit message 写好，然后我按 Enter 就可以发一个 commit。
荣辉
这要不要稍微展开讲一讲？这是一个什么样的 feature？因为刚开始我们没有介绍到。
杰特陈
哦，OK，OK。我们 Paperboy 的开发流程大概是，先把收集操作系统用户数据，以及建立一个比较好的 memory 系统这一套做出来。然后做出这套系统之后，我们在工程上就有一个类似 framework 的东西。你会有一个比较好的 Markdown document，这个 Markdown document 会实时更新，包括用户的职业、用户过去几天做了什么，也包括用户过去几秒钟、几分钟做了些什么事情。
距离用户现在做的事情越近，它的 granularity，也就是颗粒度，就会越好、更细。所以我们的 Paperboy agent 一直会有这个 context。有了这套 context 之后，我们其实就在寻找应用场景。
我们找到的第 1 个比较好的应用场景，就是可以在操作系统的任何一个地方做类似 autocomplete 的东西。比如说你在发微信的时候，可以在文字框里面输入 `@pb`，这是一个激活词，然后后面输入一个比较简短的 instruction 或者 prompt。你也可以不输入，它会猜你在这个时候 @ 它，是在找它来干嘛。
荣辉
对，你不输入，它也会猜一猜你此刻找我是要我来干嘛。
杰特陈
对，因为它有这个 context。比如说有一个同事过来，可能我不需要对他说什么，就一个眼神他就懂了为什么我在找他。
高纪
That’s great, actually.
杰特陈
对。比如说很多时候，你和你的工程师沟通，并不需要描述，你只要点一下屏幕，指一下屏幕，他就知道：看这儿。对，看这儿，他一看就明白，哦，就是这个问题。
当时 Paperboy 已经差不多达到这个效果了。所以不管你是在回微信，还是在做任何其他事情，它都会适时地拥有这些 context。对我来说，在命令行里面，甚至平常在 GitHub 里面发 PR 的时候，它都可以直接把整个 PR description 做出来。
很多时候我发现，它生成的 PR 实际上会比 Cursor 或者 Claude Code 这类基于传统单应用 context 的 agent 更好。因为很多时候我开发一个比较大的 feature，一会儿在和 Claude Code 打交道，一会儿在微信里面和 John 发消息，一会儿又在浏览器里面做 research。其实我可以把所有这些 context 聚合在一起，让它以一个 PR 的形式给我做一个 draft。
我会发现，它对我在做什么事情的理解程度，会比任何基于单应用 context 的 agent 或者 AI 都更好。
我自我怀疑我有 ADHD，我在工作的时候很容易分心。我经常在 talk 的时候，旁边会看 Hacker News。如果 Hacker News 上有什么比较有意思的 article，我可能就会点进去，结果一看就是 30 分钟，回来时 talk 可能已经结束 10 多分钟了。
对于这种情况，我记得我当时每天晚上都会让 Paperboy 给我一个 report，说我今天到底有多 efficient。我会让它去骂我，告诉我什么时候不够高效、什么时候分神了。对我来说，平常工作的时候，当我知道 Paperboy 在盯着我工作，我会效率更高。
荣辉
哦，就是 Big Brother is watching you？
蒋阳
A large portion of my job now is to use different products, do research, and talk to people. I unfortunately can’t code as much as I used to anymore. But Paperboy is great because it actually keeps all those different information-granularity sources in its head, right?
So when I want to go learn about something, or when I have things happening in my head and need something to help me thread them through—
荣辉
Yeah, it’s like connecting the dots.
蒋阳
Yeah, and then it can tell me, “Wait, have you thought about this? You tried this a couple of months ago, and this is the product form, and these are the problems that you recognized in the product,” right?
I know about these things, but when you’re doing that sort of thinking across different planes, you can do it manually. I should just sit myself in front of a piece of paper and sketch it out, but it’s a lot faster to have a very smart model that knows me help me through that process.
荣辉
所以这里有一个例子吗？就是最近一次给你这样的 inspiration？
蒋阳
Oh yeah. Paperboy—we started working on this latest interface 2 weeks ago. I can talk a little bit about the problems we had.
For example, if you’re shipping Paperboy to VCs, and VCs are definitely part of our ICPs, they want a personal CRM module, a meetings-coming-up module, and a deals-of-the-quarter module. They want to keep track of everything that’s happening across all their portfolio companies, and they want to keep track of their relationships with LPs.
Maybe you can ship a sidebar where you have these existing modules in place. Then the question is: okay, we know this thing works for operators, founders, and real estate salesmen. They all need a different sidebar. So what do we do? Do we ship these things as skills, as plugins, or as recipes, and allow the user to pick which one they want to add into their product?
So you’re thinking about this problem, and Paperboy helped you. That didn’t make a lot of sense to me, right? Paperboy pointed out to me that it was sort of like: what application do we use that is—well, you obviously need a list, but it’s an infinite list and it’s actually not annoying?
Then it came back to the IM idea, which is: when you think about what’s going on in our iMessage and WeChat today—WeChat especially in China—there are contacts and a bunch of group chats. And then there are group chats where it’s sort of like, maybe you have 3 people. Okay, it’s just 3 people, but your 3 people can have 4 group chats for different topics.
蒋阳
Yeah, and it's actually a very intuitive way for people to organize their topics and the things that are going on.
高纪
Yeah, actually, we have a new group for each podcast. Each group is fixed with a few people because we have several podcasts being processed at the same time. At the beginning, it's super chaotic in one group. So every time we start a new podcast, we create a new group, and everyone joins the new group.
蒋阳
Yeah, and the great thing about IM is that if you don't chat in that group anymore, it gets pushed down. You can also choose to right-click or swipe to hide it.
高纪
Right, I do that all the time, so my WeChat is sort of like this inbox.
蒋阳
So that was kind of the inspiration. We know we also need an inbox feature, so what if we don't design an inbox feature? We just turn—what is the thing that itself is an inbox? It's IM. So that's kind of how the latest interface came about.
高纪
所以其实在我们讲的时候，如果大家没有用过 Paperboy，可能就需要一些脑补。我们已经安利了非常多 Paperboy 与众不同的功能特性，以及很独特的一些交互理念和设计。正好也在这里我们来讨论一下：其实像我们刚才讲到的，capture 所有的 context，这是一个共识，大家都会做。
然后我们刚才提到的一些新的交互范式或者一些思考，我觉得其实如果它真的 make sense，我觉得大家也会很快地追上来、扑上来。所以想问问二位怎么看待竞争：你们觉得最主要的竞争对手是谁，以及你们如何在竞争里面保持持续的优势和差异？
蒋阳
8. OpenAI And Anthropic Set The Bar
Only 2 that matter today are OpenAI and Anthropic. Everybody else is behind them, right? And they have all the advantages. They actually own the models. Anthropic has new compute now, and they have great distribution.
All we can pray for is that we're like Cursor, where we have some edge in taste, where we beat them to this new interface, and all of that is for individuals. Product experience isn't going to go away. But then also, the second thing is enterprise. Everybody definitely wants to go into enterprise. I mean, it's where the money is.
People were actually trying to build the first product as AI Slack, and that was our answer initially to what to build if the best way to work with agents hadn't been invented yet.
高纪
所以这也是一个非常热门的话题。现在有很多产品都在尝试解决这个问题：当 agents 进入团队之后，我们应该构建什么样的产品，来和它们一起工作？
So how do you see some of the products that are being discussed in the market now, such as Slack, Multi, Mocks, etc.? 等等，有没有你想讲的？
酱阳
Well, first of all, I love Slack. I think Richard is brilliant. He has a great background. His co-founder is great, and now they're building up a product team as well.
The hard part about AI Slack—and this is from my personal POV—is how do you get an enterprise to switch? AI Slack's business model that makes sense is that you have to sell it to a team. It's the team that's paying for it.
For an enterprise, if there is a way to export all the data from Slack and import it into AI Slack, and it's a one-to-one experience, then maybe it can work. But you still have to build a bunch of connectors, like the integrators, for all the other data that the enterprise has. So that integration work isn't going away.
The second thing is that Slack is honestly not a great product. If you just try to build an interface that's similar to Slack, with channels and threads, to me, I don't enjoy using it. At scale, you kind of have to switch to it because nothing better exists right now, but it's not an enjoyable product to use.
So that's all on the user side and the adoption side, and then there are the agents. As Jet said, we don't believe in the learning rates of the agents if the only sources they get for understanding how to work with people come from the messages you send to them.
People with jobs are busy. They do not have time to care about training agents. They expect things to work for them out of the box, and they expect the product to keep getting better. The reason we moved away from Slack is that the agents in Slack still need more context to learn naturally.
高纪
Cool. Hmm, I think Slack actually still has a lot of network effects, just like what Zhang said, to prevent it from being replaced by new players. For example, Slack actually has a feature called External Connections.
Ronghui
Currently, all the serious companies are using Slack, and communication between companies is now largely done through Slack Connect rather than email. This is hard to replace. I think Slack still has a lot of integration-related network effects, just like WeChat, making it irreplaceable for them.
Compared to that, if you take agents as a layer above, more at the OS level, and make it something that complements Slack, I think it would actually have a better adoption story. So is this what you’re doing—trying to complement Slack in some way?
Jiang
Yeah, we assume that initially, if you try to go into enterprise, they will be communicating on Slack. If you’re building out Slack right now, you can’t really try to sell something to a team larger than maybe 50 people. It would take them some work to switch off Slack.
It makes more sense, again, to bring the agents to where people already work instead of trying to replace Slack and build everything from scratch.
Koji
我其实有注意到 Paperboy 目前的官网，产品还没有发布，但是你们放了一篇 blog post，标题叫作 The Last Interface。那篇 blog 里其实提到有 5 kinds of pace，有 5 种不同的速度。可不可以稍微展开一下？我理解这背后好像是在讲你们如何理解今天的 context layer，还有整个 memory system 等等。
9. Pace Layers Shape The Interface
酱阳
Yeah, I don’t want to go into too much detail because, actually, the choice of saying 5 isn’t because we only have 5 layers. It’s just because that was a sensible categorization.
The important thing is that if you think about it, in San Francisco there’s a nonprofit called the Long Now Foundation, and its founder, Stuart Brand, has published something called “Pace Layers.” It’s sort of like, at the top, there’s fashion, then commerce, politics or governance, infrastructure, civilization, and nature. These are 6 paces that change at different velocities.
Fashion obviously changes the fastest. Nature, as we understand it, doesn’t really change, right? Physics. The world is made up of these different paces, and they all have a relationship with each other. There are reflexive relationships between every layer, and largely they make up our society.
I’ve been fascinated by that since high school, and it’s definitely influenced my worldview of the world that we live in, on both the natural and social sides. When I think about what’s actually efficient and what can scale for representing this world that we live in, in the agents’ world, you need these different pace layers too.
Koji
Yeah, so it’s quite interesting. What would correspond to, for example, nature, which is constant, and what would correspond to fashion, which is always changing?
酱阳
I think for products, if you classify the length of these tasks, there are actually many different categories. For example, when a user replies to a message on WeChat, it only takes 1 second or 10 seconds. We think this is one end of the spectrum.
The other end of the spectrum could be when a user reads 10 very long reports and makes a business decision, and this could take several hours. That’s probably closer to the left side, too. The long side is more like over a month.
Koji
Yeah, like over a month. Yeah, yeah.
酱阳
基本上有一个 spectrum，然后我们认为这个 spectrum 的每一端都可以被某种 AI agent augment，或者说自动化掉。但是每一端它们需要去做的事情其实不一定是一样的。
如果你要去自动化用户回复微信的这个 task，最好的 form factor 有可能是：当你点到那个微信的回复界面，你会自动弹出一个 autocomplete。这个是我们已经在做的。
但是随着这个 time horizon 越来越长，其实你的产品形态就会越来越不确定。你去自动化用户一个几个小时的 task，最好的 form factor 是什么？这其实目前还是一个非常值得被探索的区域，我们也很感兴趣探索这个。
Koji
其实在录这一期播客之前，我和 Paperboy 的团队也有一些交流，发现团队里面好几个成员都说，他们每次和姜沟通的时候，感觉不到他是 2004 年的，就觉得他有超越年龄的成熟。我其实也有这样的感觉。
我想问一下姜，你觉得自己的这样的状态是怎么来的？是因为你做了什么、看到了什么，或者是受到了什么启发、激励？
Jiang
I think I’m still a pretty childish person, especially with Jet around. We have a lot of fun together. I think you’re old.
For the team, at least, I definitely just have a passion for business. We’re trying to do something very hard in a very short amount of time, and so clarity, efficiency, and focus are important.
Koji
这要怎么做到呢？这个，知易行难。
Jiang
For me, that starts with my number-one job: to bring to the team what defines what success looks like, and to make sure that the context every person has about their goals and how they’re actually operating is complete and accurate.
Actually, even for me, the best part of being able to start a company is that I get to learn so much about what I’ve always wanted to explore and understand about the world. That curiosity has probably been the biggest driver for me since high school.
Between high school and college, I took a couple of internships, and all of them were at different companies. I just did different types of problem-solving activities.
At Million, obviously, we started in the B2B business, selling web performance RPA bots to large enterprises. We learned that it was a difficult business to scale. We were 20-year-olds, and we weren’t good at scaling up enterprise relationships.
Since then, I think I’ve been on this trajectory where I’m just trying to learn what markets actually make sense for different players. The way I like to learn is not just through startups, but across different business scales too. It’s like founder-market fit, team-market fit.
I think market picking is one of the most important, if not the most important, traits of a founder early on. It’s making sure that you understand what you’re going into, where all the leverage points are, and what challenges are coming down on you. Then you can actually go build a team, and then there’s the product experience, finding customers, and technology.
Koji
所以这里有什么可以分享的吗？就如何去 choose a market, the right market.
Jiang
Yeah. I think about a bunch of common sense again. It has to be big enough, and “big” isn’t just a dollar sign, but how sustainable is the market? That scale has to keep increasing so your team can continuously develop your product and your market position across a decade.
So when I look at the most successful products in history, it’s never a single product. It’s actually sort of like a product lineup, where the expectation is, when you build this thing, it’s going to take maybe a decade for it to get to a form where it’s really mass consumer, or everybody’s using it. That’s the first thing: you’ve got to start with a problem close to you. It’s all just, to me, a method of understanding the market, but then you need a lot of history to be able to understand if the business dynamics—the economics—are actually favorable, and whether there have been cases in the past or not. There are so many details, and I think I’m just barely scratching the surface. Paperboy is just my chance to kind of prove out what I know.
Koji
Paperboy 应该是你第一次做 CEO，因为上一家公司你是 co-founder。所以，做 CEO 之后，你觉得自己最大的变化是什么？
10. The CEO Owns Every Decision
酱阳
Man, it’s fucking hard. I got to own everything. I mean, as a co-founder, you definitely feel that too. I get frustrated when an engineer, for example, wastes an afternoon, but as a CEO, I’m actually the person who can cost the company the most amount of money because I have so much leverage, right? If I say we go do this and that is actually not a good idea, it’s a position where literally all of your ideas can have such a large cost and reward.
As for team skills, obviously, who do I choose to hire? Where do they come from? And then, once they’re hired, how do you continuously improve together, for the individual and for the team? Making sure every person is obviously in the role that they deserve. I feel like these are basic management skills. This is basically my first time as a manager too, so a lot of that I have to figure out on the spot. How do you give performance reviews? How do you have one-on-ones? How do you make sure that sometimes I have to just talk to the team leader, right, and not to everyone, and micromanage?
Koji
那你是从哪里来学习这些管理技巧的？刚才提到怎么开会、怎么做 one-on-one、怎么做 performance review，你是用什么方式在学呢？
酱阳
Before Paperboy, the one stint I had where I actually stayed at a larger company was at Manus. Pan Pan is an amazing CTO. I mean, I haven’t worked with any other real engineering managers, but he is a great engineering manager. When I was at Manus, we had the opportunity to have a couple of one-on-ones where I just asked him all the questions I wanted to ask about how to organize everything.
And then I also spent time reading. I think the best management book ever written is still High Output Management by Andy Grove. Then Ben Horowitz has The Hard Thing About Hard Things. And then Bill Campbell’s Trillion Dollar Coach is a great biography.
I got myself a CEO coach. We meet weekly for about an hour. That’s helped me a lot, especially the first few calls, just giving me a space to talk about all of the specific problems that I have.
Koji
你怎么找到她的？
酱阳
Yeah, it’s actually the same coach for the investor too.
Koji
可以给我们介绍一下她吗？
酱阳
Yeah. She was an ex-VC executive at a bunch of large corporations for many years, and she’s an expert in coaching first-time CEOs. I never tried getting a therapist before because I never believed in therapy. You know, I learned about how therapy works, right? You just kind of go in and talk about your emotions, but a coach is so much better because you can obviously talk about your emotions, but then you can actually also talk about business. You can talk about everything that’s happening in the business, and it’s confidential.
It’s also been really fun to use Paperboy with the coaching calls, because Paperboy listens to the calls and can actually help me follow up on things that I committed to doing on calls and organize everything.
Koji
你们其实这个 team size 并不小，所以我也比较好奇，在你们看来，在今天有那么多 AI agent 工具的情况下，去 build 一个团队，和过去我们做一个产业或产品团队相比，又有哪些特别大的变化在发生？
酱阳
If you try to make a serious infrastructure build-out, you need people who understand infrastructure. I think we hire 2 kinds of people on the engineering side. One is like Jet, who is young, high-IQ, creative, and prototypes the shit out of every problem he sees.
Then the second kind is that you’ve got to be really solid at what you do. You’ve got to understand the systems; you’ve got to understand the OS. So we hired this guy out of AWS who worked on Windows kernels for the AWS infrastructure for running Windows. I don’t think that’s going to go away anytime soon: the human experts.
Koji
其实 Jet 现在大一，对吧？我们看到硅谷有一些公司，比如 Palantir，直接给高中生 offer，让他们不要读大学了，直接工作，因为认为在 AI 时代不需要大学了。那很显然，Jet 也是优秀的大一学生，但你也在 Paperboy 做 founding engineer，那你为什么现在还要读大学啊？
Jet
It’s got Chinese parents.
上大学的意义，对于大部分人来说，大学还是能提供非常大的意义的。因为至少我在 CMU 这边看见，大部分和我同龄的人，他们其实并不是很确定未来想要做什么事情。然后在不确定的情况下，你可以做的是提升自己的 technical 能力；上大学是个比较好的选项。学业只会占你相对比较少的时间，你大部分其他时间是自己决定的，相当于有 4 年的时间去做这个 exploration。这个 4 年的时间，其实对于大部分人来说会是比较有用的。
但是，如果你在上大学的时候已经很确定你想要做什么事情，并且你在做事情这方面确实有机会，我觉得辍学也是一个理性的选择。
Koji
嗯，对。这个 Jet，你作为一个非常年轻的 founder，也是 first-time founder，你怎么吸引到同学们愿意加入你？你要怎么说服大家的？
Jet
对，是什么方面吸引到我？你观察到的，对大家来说，职业选择是非常重要的人生投资嘛。很多时候你选一个 founder，也是有 intercept、有 slope，然后很多时候 slope 比 intercept 重要很多。
我觉得张是一个 high-slope 的 founder。怎么说呢，如果你看他的 track record，他现在就差不多比我大，却在一个很小的时间内实现了特别多比较厉害的事情。在一定程度上，他是比较有 agency 的，可以很快地学习，然后去做非常多的新事情。对于一个创业者来说，这 arguably 是最重要的一个 quality：你得要头脑清晰，得要能很快地适应各种情况。
首先，你得要认为他是一个 good founder。对我来说，我和张已经工作比较久了，所以我和他合作的时候，有比较高的 confidence，相信我们可以比较顺畅地合作。他也有不错的 research、engineering，甚至 product 的 taste。我觉得在和他工作的过程中，我可以吸取很多经验，然后让我个人也有更好的成长。我个人觉得他还是一个比较 charismatic 的人。
Koji
大型表白现场。谁谁的谁？哎呀，所以这样，你自己怎么看？你有没有一些方法被你抽象总结出来？比如你现在去 approach 一个新的 candidate，你会有一些什么样的方法去说服他？
酱阳
I treat every candidate individually, and there’s not a script I stick to. It has to be mutual. I never hard-sell the company to a candidate.
Recruiting is one thing. What’s more important is my reliability, and that’s something I actually pay a lot of attention to. After I recruit somebody, can I actually deliver what I tell them? Can I carry that and make sure that I deliver on what the team needs, time and time again?
You’ve got to pick people who are great. Man, you can’t talk to mediocre people, and you can tell that from conversations about their past work. What’s especially important is their past life decisions. I really like to learn about whether, if they’re older, they have a decade of decisions behind them, and you can really learn a lot. You can’t fake that.
Koji
除了 Paperboy？
酱阳
Yeah, still love Cursor, man. Still don’t use Codex.
Koji
我特别喜欢 Codex。
酱阳
Yeah, Cursor has always been a great inspiration to me. I am really happy for them, but I hope whatever acquisition happens goes through.
Koji
第一次创业、做 co-founder 的时候，你做的是 billing。那个时候其实有一个收购 offer，来自 David 的那家公司 Cognition。
Jiang
Yeah, there were other offers too, like from Brazil and Century.
Koji
所以当时你们都没有接这些 offer。可以讲讲在面临这样的收购 offer 时，你们是怎么思考、怎么决策的吗？你现在回头看，是更 appreciate 当时的决定，还是会有一些遗憾？
酱阳
First of all, these are acquirers. Whatever offers they make for the talent, it’s not going to be a lot of money, okay? And you’re joining the companies and becoming just another employee, and you have to work on somebody else’s ideas.
I have never been interested in dev tools that much, so if you asked me to join Brazil, Century, or Cognition as an individual, I think I wouldn’t do it. Right, so it doesn’t really make sense. If you can stay independent, and if you can achieve your dreams by staying independent, you stay independent, right? And that’s what every founder wants.
Koji
好，我们再说回来。刚才你提到了你最喜欢的产品是 Cursor，然后是 Codex。可不可以讲讲为什么是 Codex，而不是 Claude Code？
酱阳
Claude Code 的理念不一样。它相当于以最 ambitious 的形式，展示未来 software engineer 会怎么 work。我觉得 Claude Code 最重要的东西，是体现 Opus 有多好，不是它的命令行有多好。我 appreciate Opus 这个 model，但不是很 appreciate Claude Code 的 CLI 本身。
Codex 我喜欢它几个点。一个是，它核心的 agent 是开源的，它的命令行还有整体的 agent loop，实际上都放在 GitHub 上。然后如果你用他们的桌面端 app，把它和比如说 Claude Code 的桌面端 app 比，它整个 app 的 craft 水平会非常非常好。
他们前面买了一个团队叫 Sky。Sky 是做 Apple Shortcuts 的那个团队，OpenAI 会 hire 这种人去做一个 very polished 的产品。你去看 Codex，会发现他们最近出了一个叫 Codex Pets 的东西，非常 polished，而且也很好玩。包括 Codex 的 computer use、browser use，其实都做得非常 polished。
我觉得 Codex 也是第一个以高并行的形式，把不同的 agent 同时工作的方式当作主要 UI 推出来的产品。还有就是，OpenAI 在 infra 层面其实比 Anthropic 的优势大非常多。所以如果你用 Codex 的话，模型大差不差，但是稳定很多。
不好意思，OpenAI 在 infra 上面其实可以 subsidize 更多的 compute，而 Anthropic 他们现在投资得不够，所以没法 subsidize，导致了一系列风波。比如说，如果你的 repo 里面有什么 OpenClaw 或者 Hermes Agent，Claude Code 可能会直接收你 10 倍的价格。反正 Anthropic 现在因为 compute 受限，其实做了很多奇怪的操作。
Koji
如果可以买他们的 secondary，也就是可以买他们的股票，你现在会买 OpenAI 还是买 Anthropic？还是会 30% 和 70% 这样买？
酱阳
我感觉我会买 OpenAI，all in OpenAI，可能就是 OpenAI 80%，Anthropic 20%。
Koji
OK，yeah，什么原因呢？
Jiang
他们两家公司都会 do very well，但是我更喜欢 OpenAI 在做的事情。我觉得 Anthropic 做得最好的事情，是他们的 interpretability 还有 societal impacts，那两个部门做得非常非常好。我特别喜欢那两个东西。
但是 Anthropic 在很多模型上，我觉得他们过于 opinionated，就是说用户该怎么用这个 language model，会非常——不管是道德上还是具体使用习惯上——非常 opinionated。相比之下，OpenAI 的理念是，在最小的局限性之下，用户想要怎么用这个模型就怎么用。我个人会比较 libertarian，所以会喜欢 OpenAI 这种更加 unopinionated 的 values。
还有就是，我觉得 OpenAI 在 compute 上面会占优势。If I’m making investment decisions, I’m not gonna buy either of them at the current valuation. But I appreciate Anthropic’s safety commitment.
Koji
那你们怎么看“模型吞噬一切”这个说法？
酱阳
我觉得你完全可以 argue，很多公司会提供差不多的模型，然后产品的差异性还是在产品公司里面。这不一定会是真的，但是我觉得这是一个 possibility。
还有 Cursor，我觉得是一个不错的例子。他们一开始是产品公司，但是现在他们既是产品公司，又是模型公司。他们现在可以用 xAI 的 compute 去做 frontier coding model。所以我觉得公司不能把自己局限为模型公司还是产品公司。
Koji
模型公司还是产品公司？
酱阳
It’s usually both.
Koji
如果给你们一笔虚拟的资金去做投资，比如 300 万美金，可以投 3 个团队，身边认识的 3 个人，你会怎么投？
酱阳
Slack is good. Just bet on Richard. You know, robotics companies—I think that’s definitely going to be big—and then infra, people who are focused on infra. And then I believe the biggest company similar to Cursor is going to come out of consumer.
酱阳
It’s not going to come out of enterprise AI. It’s going to come out of consumer. Obviously, Paperboy is in that market because I followed that, but consumer is just such a big market. There are so many opportunities.
我觉得 robotics 我也会特别想投，因为我觉得在不断用这一类产品去 automate knowledge work 之后，很有可能最大的市场是用这个 model capability 去进化物理世界的这种 opportunity。然后还有一个就是安全，因为我觉得首先 AI 加 security 是个非常好的领域，这个 demand 基本上是 infinite 的。只要模型越强，attacker 和 defender 的能力都会变强，很多时候这种公司的 TAM 非常大。
The problem with that is, how do you compete with Anthropic getting into that business?
酱阳
Anthropic and OpenAI 肯定会有最好的模型。问题是，你有最好模型之后，你下一个 optimization layer 是什么？你是做比较好的 harness 吗？你是把人类 security research 的各种 heuristic back 到 agent 里面，还是说你去做 infra，比如说非常庞大的一个 multi-agent 系统去不断地挖这些漏洞？
我觉得这个探索空间会非常大。比如说，你这个产品挖的漏洞会比另外一个产品多 10%，那么你这个产品就是非常有用的。这是一个非常 meritocratic 的 market，所以我觉得它非常好。
Koji
We haven’t talked about Long Horizon Law yet, but, yeah, we’ve hardly shipped this thing, right?
Jiang
And then I think the other John Yon[?] shipped its new SWE-bench today or recently, reproducing complete codebases from a set of specs. That’s not going to go away. I think security capability is very, very important for national security.
But I think for entrepreneurs, this is a very good opportunity because the model 最强的人，他有这个 incentive 去把这个模型说只有我们自己国家可以用这个东西。那么相当于你其他国家怎么搞呢？其实很多很多不同的国家或者说其他实体都需要这种模型，所以说这个 market 是非常大的。
Koji
好，我们最后问个问题。你会期待 1 年后的此刻，Paperboy 是一个什么样的状态？你有什么样的期待，或者有什么害怕做错的事？
酱阳
Well, I hope we’re not losing money every day anymore. I want to keep hiring the best people I can find, so the talent density needs to be higher than where we’re at today. There’s so much specifics we can talk about, but I think from a team and business point of view, we actually need cash flow, and we need to keep building a better team.
Okay, 好，谢谢二位，谢谢，非常开心。
Koji
好，嗯，拜拜。